Merge remote-tracking branch 'origin/master' into codex/project-instruction-files

# Conflicts:
#	docs/architecture.md
#	docs/config-catalog.md
#	docs/event-producer-consumer.md
#	docs/module-graph.md
#	packages/core/agent-core/src/index.ts
#	packages/ui/acp-agent/src/index.ts

packages/README.md

@@ -1,6 +1,6 @@
 # Packages
 
-Harness packages, all under the `@deepseek-ai/dsh-*` scope. Each package is a Cordis plugin (microkernel-style): it exports either a default `Service` subclass or a functional plugin, declares its ctx key/events through declaration merging, and exposes extension points through `ctx.effect()`, `ctx.on()`, and `ctx.waterfall()`. Authoring conventions: [AGENTS.md](AGENTS.md) (subtree) and the root [AGENTS.md](../AGENTS.md) § Conventions.
+Harness packages use the `@deepseek-ai/dsh-*` scope. Each is a Cordis `Service` or functional plugin that declares context keys/events through declaration merging and registers extension points with `ctx.effect()`, `ctx.on()`, or `ctx.waterfall()`. Authoring conventions: [package](AGENTS.md) and [root](../AGENTS.md) `AGENTS.md`.
 
 ## Hierarchy
 
@@ -9,13 +9,14 @@ Packages are grouped by modular role at `packages/<group>/<pkg>/`. The group dir
 | Group | Role | Release expectation |
 |---|---|---|
 | [`core/`](core/README.md) | Product API spine: session, system-prompt, tools, agent, and the concrete loop | Product — stable surface |
-| [`prompt/`](prompt/README.md) | Prompt and request-context extensions that shape model inputs | Product — stable surface |
+| [`prompt/`](prompt/README.md) | Prompt and request-context extensions | Product — stable surface |
 | [`llm/`](llm/README.md) | LLM capability family: the abstract service + provider adapters | Product — stable surface |
 | [`bash/`](bash/README.md) | Bash capability family: the executor seam, a local impl, and the model-facing tool | Product — stable surface |
 | [`code-runtime/`](code-runtime/README.md) | Code-execution capability family: the abstract runtime seam for model-written programs + a worker-thread backend | Product — stable surface |
 | [`fs/`](fs/README.md) | Filesystem capability family: the abstract seam, a local impl, and the model-facing file tools | Product — stable surface |
 | [`compact/`](compact/README.md) | Compaction capability family: the abstract seam + a basic backend (tool deferred) | Product — stable surface |
 | [`subagent/`](subagent/README.md) | Subagent capability family: the provider-registry seam and the model-facing delegation tool | Product — stable surface |
+| [`workflow/`](workflow/README.md) | Workflow capability family: the script-engine seam, the worker-thread engine, and the model-facing `workflow` tool | Product — stable surface |
 | [`web/`](web/README.md) | Web capability family: the abstract seam, search/fetch provider impls, and the model-facing web tools | Product — stable surface |
 | [`timeout/`](timeout/README.md) | Tool-call timeout policy: the `tools/execute` deadline enforcer | Product — stable surface |
 | [`todo/`](todo/README.md) | Todo/planning family: the model-facing `todo_write` tool | Product — stable surface |

packages/code-runtime/README.md

@@ -1,6 +1,6 @@
 # code-runtime/ — code-execution capability family
 
-The code-execution capability seam (see [capability seams](../../docs/rfc/implemented/architecture/2026-06-13-capability-seams.md)): an abstract runtime interface for executing one model-written program against host-provided async bindings, capturing what it printed and returned. The consumer is the tool registry's Code Mode, specified alongside the seam in the [Code Mode RFC](../../docs/rfc/proposed/feature/2026-06-15-code-mode.md). **Product** packages.
+The code-execution capability seam (see [capability seams](../../docs/rfc/implemented/architecture/2026-06-13-capability-seams.md)): an abstract runtime interface for executing one model-written program against host-provided async bindings, capturing what it printed and returned. The consumer is the tool registry's [Code Mode](../core/tools/README.md) (`tools: { mode: code }` — the `run_code` tool and the generated TypeScript SDK); design in the [Code Mode RFC](../../docs/rfc/implemented/feature/2026-06-15-code-mode.md). **Product** packages.
 
 | Package | Role | ctx key |
 |---|---|---|

packages/code-runtime/code-runtime-worker/README.md

@@ -1,6 +1,6 @@
 # @deepseek-ai/dsh-code-runtime-worker
 
-Worker-thread implementation of the [`@deepseek-ai/dsh-code-runtime`](../code-runtime/README.md) seam: `WorkerCodeRuntime` runs each program in ONE fresh Node `worker_threads.Worker` — TypeScript in, type-stripped host-side, bindings bridged over the message port, `{ value, logs, error? }` out. **Containment, not a security boundary**: trust posture is bash-equivalent by design (the [Code Mode RFC](../../../docs/rfc/proposed/feature/2026-06-15-code-mode.md) § Trust posture), with containment bash does not have — separate isolate, empty environment, heap cap, hard termination.
+Worker-thread implementation of the [`@deepseek-ai/dsh-code-runtime`](../code-runtime/README.md) seam: `WorkerCodeRuntime` runs each program in ONE fresh Node `worker_threads.Worker` — TypeScript in, type-stripped host-side, bindings bridged over the message port, `{ value, logs, error? }` out. **Containment, not a security boundary**: trust posture is bash-equivalent by design (the [Code Mode RFC](../../../docs/rfc/implemented/feature/2026-06-15-code-mode.md) § Trust posture), with containment bash does not have — separate isolate, empty environment, heap cap, hard termination.
 
 ## Config
 

packages/code-runtime/code-runtime-worker/package.json

@@ -11,6 +11,10 @@
       "types": "./lib/types/index.d.ts",
       "default": "./lib/index.js"
     },
+    "./worker": {
+      "types": "./lib/types/worker.d.ts",
+      "default": "./lib/worker.js"
+    },
     "./src/*": "./src/*",
     "./package.json": "./package.json"
   },

packages/code-runtime/code-runtime/README.md

@@ -2,7 +2,7 @@
 
 The **code-execution seam**: an abstract `CodeRuntime` service (`ctx.codeRuntime`) defining WHAT a code runtime does — run one model-written program against a set of host-provided async bindings and report `{ value, logs, error? }` — without saying HOW.
 
-This package is the interface third of the capability (the bash trio is the template — see [capability seams](../../../docs/rfc/implemented/architecture/2026-06-13-capability-seams.md)): implementations subclass `CodeRuntime` and register the service; the consumer is the tool registry's Code Mode, which generates the model-facing SDK and bridges tool dispatch — both specified in the [Code Mode RFC](../../../docs/rfc/proposed/feature/2026-06-15-code-mode.md), whose first implementation is a Node worker-thread backend. The runtime knows nothing about tools or sessions: it is handed named async functions and a program string, and everything tool-shaped stays with the consumer.
+This package is the interface third of the capability (the bash trio is the template — see [capability seams](../../../docs/rfc/implemented/architecture/2026-06-13-capability-seams.md)): implementations subclass `CodeRuntime` and register the service; the consumer is the tool registry's Code Mode, which generates the model-facing SDK and bridges tool dispatch — both specified in the [Code Mode RFC](../../../docs/rfc/implemented/feature/2026-06-15-code-mode.md), whose first implementation is a Node worker-thread backend. The runtime knows nothing about tools or sessions: it is handed named async functions and a program string, and everything tool-shaped stays with the consumer.
 
 ## Service API (`ctx.codeRuntime`)
 

packages/code-runtime/code-runtime/src/index.ts

@@ -7,7 +7,7 @@
  * substrate (worker thread, separate process, container) and by source
  * language, both declared as readonly descriptors. The design and its
  * consumer (the tool registry's Code Mode) are specified in the Code Mode RFC
- * (docs/rfc/proposed/feature/2026-06-15-code-mode.md).
+ * (docs/rfc/implemented/feature/2026-06-15-code-mode.md).
  *
  * The split mirrors the bash seam (`BashExecutor`): the runtime knows nothing
  * about tools or sessions — it is handed named async functions and a program,

packages/cordis/tool-cordis/src/api-catalog.ts

@@ -197,6 +197,13 @@ export const SERVICE_API: readonly ServiceApiEntry[] = [
       'async fetch(request: WebFetchRequest, exec?: WebExecContext): Promise<WebFetchResult>',
     ],
   },
+  {
+    key: 'workflows',
+    summary: 'Abstract workflow execution service.',
+    methods: [
+      'abstract start(request: WorkflowStartRequest): WorkflowRun',
+    ],
+  },
 ]
 
 /** Every harness event, sorted by name. */
@@ -375,6 +382,42 @@ export const EVENT_API: readonly EventApiEntry[] = [
     signature: '\'tools/pre-execute\'(this: ToolRegistry, exec: ToolExecution, next: () => Promise<PreToolDecision>): Promise<PreToolDecision>',
     summary: 'Waterfall BEFORE a tool runs — the gate where sandbox, permission, and hook plugins allow or deny a call (Claude Code\'s `PreToolUse`).',
   },
+  {
+    name: 'workflow/agent-end',
+    mode: 'emit',
+    signature: '\'workflow/agent-end\'(info: WorkflowRunInfo, agent: WorkflowAgentEndInfo): void',
+    summary: 'One `agent()` call settled (clean result, child failure, or run cancellation).',
+  },
+  {
+    name: 'workflow/agent-start',
+    mode: 'emit',
+    signature: '\'workflow/agent-start\'(info: WorkflowRunInfo, agent: WorkflowAgentInfo): void',
+    summary: 'One `agent()` call started a child run.',
+  },
+  {
+    name: 'workflow/end',
+    mode: 'emit',
+    signature: '\'workflow/end\'(info: WorkflowRunInfo, result: WorkflowResultInfo): void',
+    summary: 'A workflow run settled (any stop reason).',
+  },
+  {
+    name: 'workflow/log',
+    mode: 'emit',
+    signature: '\'workflow/log\'(info: WorkflowRunInfo, message: string): void',
+    summary: 'The script emitted a narration line (a `log(message)` call).',
+  },
+  {
+    name: 'workflow/phase',
+    mode: 'emit',
+    signature: '\'workflow/phase\'(info: WorkflowRunInfo, title: string): void',
+    summary: 'The script entered a phase (a `phase(title)` call) — progress grouping for observers; no execution semantics.',
+  },
+  {
+    name: 'workflow/start',
+    mode: 'emit',
+    signature: '\'workflow/start\'(info: WorkflowRunInfo): void',
+    summary: 'A workflow run started — the script\'s meta block validated, the body about to execute.',
+  },
 ]
 
 /** Shapes of every exported type the SERVICE_API signatures reference (transitively), sorted by name. */
@@ -851,6 +894,34 @@ export const TYPE_API: readonly TypeApiEntry[] = [
     name: 'WebSearchSource',
     declaration: 'export interface WebSearchSource {\n    readonly url: string;\n    readonly title?: string;\n    readonly snippet?: string;\n    readonly publishedAt?: string;\n}',
   },
+  {
+    name: 'WorkflowMeta',
+    declaration: 'export interface WorkflowMeta {\n    name: string;\n    description: string;\n    whenToUse?: string;\n    phases?: WorkflowPhase[];\n}',
+  },
+  {
+    name: 'WorkflowPhase',
+    declaration: 'export interface WorkflowPhase {\n    title: string;\n    detail?: string;\n    model?: string;\n}',
+  },
+  {
+    name: 'WorkflowResult',
+    declaration: 'export interface WorkflowResult {\n    value: unknown;\n    stopReason: WorkflowStopReason;\n    error?: string;\n    agentsStarted: number;\n}',
+  },
+  {
+    name: 'WorkflowRun',
+    declaration: 'export interface WorkflowRun {\n    readonly id: WorkflowRunId;\n    readonly meta: WorkflowMeta;\n    readonly result: Promise<WorkflowResult>;\n    cancel(reason?: string): void;\n    dispose(): Promise<void>;\n}',
+  },
+  {
+    name: 'WorkflowRunId',
+    declaration: 'export type WorkflowRunId = Branded<\'WorkflowRunId\'>;',
+  },
+  {
+    name: 'WorkflowStartRequest',
+    declaration: 'export interface WorkflowStartRequest {\n    script: string;\n    meta: WorkflowMeta;\n    args?: unknown;\n    parent: Agent;\n    signal?: AbortSignal;\n}',
+  },
+  {
+    name: 'WorkflowStopReason',
+    declaration: 'export type WorkflowStopReason = \'completed\' | \'cancelled\' | \'error\';',
+  },
 ]
 
 /** The inherited `ctx` surface (cordis core + loader/hmr/timer), in curated order. */

packages/core/agent-core/src/index.ts

@@ -47,7 +47,7 @@ import z from 'schemastery'
 import LlmService from '@deepseek-ai/dsh-llm'
 import SessionStore from '@deepseek-ai/dsh-session'
 import SystemPrompt, { type Config as SystemPromptConfig } from '@deepseek-ai/dsh-system-prompt'
-import ToolRegistry from '@deepseek-ai/dsh-tools'
+import ToolRegistry, { type Config as ToolsConfig } from '@deepseek-ai/dsh-tools'
 import AgentRegistry from '@deepseek-ai/dsh-agent'
 import * as invariants from '@deepseek-ai/dsh-invariants'
 import * as toolBash from '@deepseek-ai/dsh-tool-bash'
@@ -61,11 +61,13 @@ export const name = 'agent-core'
  * `agents` to the agent loop (an app that pre-creates no agents, like the ACP
  * bridge, simply omits it), `persona` and `toolOrder` to the system-prompt
  * plugin (the deployment's persona section and the explicit model-facing tool
- * order), and `workspaceContext` to the workspace-context plugin. Every
- * field is optional INPUT here because each owner's schema supplies the
- * default (`[]` / `''` / absent — lexicographic / loader defaults); the schema
- * is the INTERSECTION of the owners' own schemas, so validation and defaulting
- * can never drift from them.
+ * order), the `tools` object to the tool registry (its presentation `mode`),
+ * and `workspaceContext` to the workspace-context plugin. Every field is
+ * optional INPUT here because each owner's schema supplies the default (`[]` /
+ * `''` / absent — lexicographic / `native` / loader defaults); the schema is
+ * the INTERSECTION of the owners' own schemas (the registry and workspace
+ * loader nested under their bundle keys), so validation and defaulting cannot
+ * drift from them.
  */
 export interface Config {
   /** The agent-loop `agents` list (see dsh-agent-loop's `Config`). */
@@ -74,6 +76,8 @@ export interface Config {
   persona?: SystemPromptConfig['persona']
   /** The explicit model-facing tool order (see dsh-system-prompt's `Config`). */
   toolOrder?: SystemPromptConfig['toolOrder']
+  /** The tool registry's config — its presentation `mode` (see dsh-tools' `Config`). */
+  tools?: ToolsConfig
   /** Workspace-context loader controls; set `false` for hermetic prompts. */
   workspaceContext?: workspaceContext.Config | false
 }
@@ -83,8 +87,9 @@ export const Config = z.intersect([
   AgentLoop.Config,
   SystemPrompt.Config,
   z.object({
+    tools: ToolRegistry.Config,
     workspaceContext: z.union([z.const(false), workspaceContext.Config]),
-  }) as unknown as z<Pick<Config, 'workspaceContext'>>,
+  }) as unknown as z<Pick<Config, 'tools' | 'workspaceContext'>>,
 ]) as unknown as z<Config>
 
 /**
@@ -111,7 +116,7 @@ export function apply(ctx: Context, config: Config): void {
     persona: config.persona ?? '',
     ...config.toolOrder !== undefined ? { toolOrder: config.toolOrder } : {},
   })
-  ctx.plugin(ToolRegistry)
+  ctx.plugin(ToolRegistry, config.tools ?? {})
   ctx.plugin(AgentRegistry)
   ctx.plugin(invariants)
   ctx.plugin(toolBash)

packages/core/tools/README.md

@@ -1,9 +1,18 @@
 # dsh-tools
 
-Tool registry and execution pipeline. Tool plugins register their schemas and executors; the agent loop executes each call through `tools/pre-execute` (the allow/deny gate) → `tools/execute` (an around-dispatch wrapper for timeout/retry/metrics plugins) → `tools/post-execute` (inspect/replace the result, attach context).
+Tool registry and execution pipeline. Tool plugins register their schemas and executors; the agent loop executes each call through `tools/pre-execute` (the allow/deny gate) → `tools/execute` (an around-dispatch wrapper for timeout/retry/metrics plugins) → `tools/post-execute` (inspect/replace the result, attach context). The registry also owns HOW its tools are presented to the model — its `mode` config selects native function calling, [Code Mode](#code-mode), or both.
 
 ## Service: `ToolRegistry` (ctx key: `tools`)
 
+### Config
+
+```yaml
+tools:
+  mode: native   # native (default) | code | both
+```
+
+`native` contributes every registered tool as a wire function definition — the default, byte-for-byte the pre-config behavior. `code` contributes exactly ONE wire tool, `run_code`, plus the generated `tools:sdk` prompt section (see [Code Mode](#code-mode)). `both` contributes every native definition AND `run_code` + the SDK section. Non-native modes require a loaded `ctx.codeRuntime` with `language: 'typescript'`; a missing or mismatched runtime rejects every prompt assembly with an actionable error, and a `systemPrompt.toolOrder` naming tools the mode no longer contributes rejects the assembly the same way.
+
 ### Public API
 
 - `ctx.tools.register(definition: ToolDefinition): () => void` Register a tool. Disposed with the calling fiber.
@@ -119,6 +128,16 @@ const bash = defineTool({
 })
 ```
 
+### Code Mode
+
+Under `mode: code` (or `both`) the registry turns the tool surface into a programming API, per the [Code Mode RFC](../../../docs/rfc/implemented/feature/2026-06-15-code-mode.md): the model writes a TypeScript program (the body of an async function) and passes it to the ONE wire tool `run_code`; the program runs in `ctx.codeRuntime` (the [code-execution seam](../../code-runtime/README.md) — the shipped backend is a worker thread) with one async binding per registered tool (`await tools.bash({...})`), and ONLY what it prints or returns re-enters the model's context.
+
+- **The SDK section** (`tools:sdk`, order 150): a lazy prompt section regenerating, at each assembly, a `declare const tools: {...}` TypeScript declaration of every registered tool except `run_code` (exotic names via quoted keys), plus fixed usage instructions. Deterministic — lexicographic tool order, byte-identical text for an unchanged tool set (prefix-cache-friendly). The codegen (`jsonSchemaToTs`, exported) is TOTAL: constructs outside the `defineTool` subset degrade to `unknown`, never throw.
+- **The dispatch bridge** (`run_code`'s execute): every binding call is JSON-normalized BEFORE dispatch (a value that does not survive — `BigInt`, circulars — rejects that one call, so the dispatched form and the logged form are the same JSON value by construction), serialized through a per-run queue (even `Promise.all` executes the underlying `ctx.tools.execute()` calls one at a time in submission order — the tool contract carries no concurrency-safety metadata yet), gated by `tools/pre-execute`/`tools/post-execute` like any native call (a deny reaches the program as a binding rejection), and logged as one `tool/code-dispatch` session event (log-only: `deriveMessages()` never surfaces it) with the deterministic sub-id `<parent>:code:<n>`. A failed sub-call REJECTS the program-side promise with the tool's error text — real code error handling, no bespoke envelope. A sub-call's `additionalContext` is deliberately DROPPED (no safe outlet mid-run without breaking tool-call/result adjacency; deferred until a real hook needs it through Code Mode).
+- **Settlement discipline**: the bridge owns a run-scoped abort that follows the outer signal in and fires when the run settles for any reason, so a budget expiry aborts an in-flight sub-tool instead of orphaning it; the bridge then drains its queue BEFORE returning, so every `tool/code-dispatch` lands inside the open turn. A failed run throws `CodeRunFailedError` (`code: 'CODE_RUN_FAILED'`, message = the failure kind + captured logs), which the pipeline converts to a structured `isError` the model self-corrects from.
+
+The wire collapse is the registry's own contribution (`systemPrompt.tools()` is mode-aware), so the logged `request/header` records it for free — under `code`, the assembled tool list is exactly `[run_code]`, pinned by tests and the snapshot goldens. Try it: `pnpm run demo:code-mode` ([the coding-agent example's Code Mode overlay](../../../examples/coding-agent/README.md#code-mode)); `pnpm run demo:code-mode acp` serves the same mode over ACP instead of the REPL.
+
 ### What is NOT here (TODO)
 
 - **Tool shapes review** — when real tools land (e.g. a concurrency-safety hint for parallel execution); phase 1 executes tool calls sequentially.

packages/core/tools/package.json

@@ -23,13 +23,20 @@
   "license": "BSD-3-Clause",
   "peerDependencies": {
     "@deepseek-ai/dsh-agent": "^0.0.1",
+    "@deepseek-ai/dsh-code-runtime": "^0.0.1",
     "@deepseek-ai/dsh-llm": "^0.0.1",
+    "@deepseek-ai/dsh-session": "^0.0.1",
     "@deepseek-ai/dsh-system-prompt": "^0.0.1",
     "cordis": "^4.0.0-rc.6"
   },
+  "dependencies": {
+    "schemastery": "^3.18.0"
+  },
   "devDependencies": {
     "@deepseek-ai/dsh-agent": "workspace:^",
+    "@deepseek-ai/dsh-code-runtime": "workspace:^",
     "@deepseek-ai/dsh-llm": "workspace:^",
+    "@deepseek-ai/dsh-session": "workspace:^",
     "@deepseek-ai/dsh-system-prompt": "workspace:^",
     "cordis": "^4.0.0-rc.6"
   }

packages/core/tools/src/code-mode.ts

@@ -0,0 +1,318 @@
+/**
+ * Code Mode: the `run_code` tool and its dispatch bridge. The model writes a
+ * TypeScript program; the bridge hands it to `ctx.codeRuntime` with one
+ * async binding per registered tool, serializes every binding call through a
+ * per-run queue onto `ToolRegistry.execute()` (so `tools/pre-execute` /
+ * `tools/post-execute` gate sub-calls exactly like native ones), logs each
+ * sub-dispatch as a `tool/code-dispatch` session event, and returns only the
+ * program's curated output. The registry itself decides WHEN this tool
+ * exists (its `mode` config); this module owns only the tool and the bridge.
+ *
+ * @module @deepseek-ai/dsh-tools/src/code-mode
+ */
+
+import { inspect } from 'node:util'
+import { CallId, HarnessError } from '@deepseek-ai/dsh-llm'
+import type { ContentBlock } from '@deepseek-ai/dsh-llm'
+import type { CodeBindingFunction, CodeRunResult, CodeRuntime } from '@deepseek-ai/dsh-code-runtime'
+import type {} from '@deepseek-ai/dsh-session'
+import { defineTool } from './schema.ts'
+import type { ToolDefinition, ToolRegistry } from './index.ts'
+
+declare module '@deepseek-ai/dsh-session' {
+  interface SessionEventMap {
+    /**
+     * One bridged sub-dispatch from a `run_code` program: the parent
+     * `run_code` call id, the deterministic sub-call id
+     * (`<parent>:code:<n>`), the tool `name` with its JSON-normalized
+     * `arguments` — the exact value dispatched, normalized BEFORE dispatch,
+     * so this append can never fail on payload shape — whether the sub-call
+     * errored, and a bounded `resultSummary` of its model-facing text.
+     * Log-only: `deriveMessages()` ignores it, so sub-calls never re-enter
+     * model context; persistence and UIs get every call. Appended inside the
+     * parent `run_code`'s execution (the bridge drains its queue before
+     * returning), so the turn-enclosure invariant holds by construction.
+     */
+    'tool/code-dispatch': { parentCallId: CallId; subCallId: CallId; name: string; arguments: unknown; isError: boolean; resultSummary: string }
+  }
+}
+
+/** The model-facing name of the Code Mode tool. */
+export const RUN_CODE_NAME = 'run_code'
+
+/** The `tools:sdk` section order: inside the 100–199 tool-guidance band, after per-tool guidance sections. */
+export const SDK_SECTION_ORDER = 150
+
+/**
+ * Thrown by `run_code` when the program run itself failed — a program
+ * exception, a budget expiry, an abort, or substrate death. Extends
+ * {@link HarnessError} (`code: 'CODE_RUN_FAILED'`); the registry's execution
+ * pipeline converts it into a structured `isError` result whose text carries
+ * the failure kind plus the captured logs, so the model can self-correct.
+ */
+export class CodeRunFailedError extends HarnessError {
+  constructor(message: string) {
+    super(message, 'CODE_RUN_FAILED')
+    this.name = 'CodeRunFailedError'
+  }
+}
+
+/**
+ * Cap for a `tool/code-dispatch` event's `resultSummary`. A log-ergonomics
+ * constant, not config: the full result already flows to the program; the
+ * summary exists so log readers see what a sub-call returned at a glance.
+ */
+const SUMMARY_MAX_CHARS = 200
+
+/** Bounded inspect for rendering a program's completion value into the model-facing text. */
+const INSPECT_OPTIONS = { depth: 4, maxArrayLength: 100, maxStringLength: 10_000 } as const
+
+/** Join a result's text blocks; a non-text block becomes a placeholder (an MVP limitation, stated in the SDK instructions). */
+function textOf(content: ContentBlock[]): string {
+  return content
+    .map((block) => {
+      switch (block.type) {
+        case 'text': return block.text
+        // ContentBlockMap is merge-extensible — future block kinds land here
+        // deliberately (no assertNever on merge-extensible unions).
+        default: return `[${block.type} content]`
+      }
+    })
+    .join('\n')
+}
+
+/** Bound a sub-call's model-facing text for the log event's `resultSummary`. */
+function summarize(text: string): string {
+  return text.length > SUMMARY_MAX_CHARS ? `${text.slice(0, SUMMARY_MAX_CHARS)}…` : text
+}
+
+/**
+ * JSON-normalize one binding call's argument into TWO independent parses of
+ * the same canonical text: `dispatched` goes to the tool, `logged` to the
+ * `tool/code-dispatch` event — identical by construction (the runtime's
+ * structured-clone boundary is wider than JSON; the session log accepts only
+ * JSON), and separate objects, so a tool mutating its args can neither
+ * desync the log from what was dispatched nor re-poison the append. A value
+ * that does not survive the round-trip (`undefined` — the log rejects it as
+ * event data — `BigInt`, a circular structure, a bare function) rejects that
+ * one call BEFORE dispatch with a model-correctable error: nothing ever
+ * executes unlogged.
+ */
+function jsonNormalizeArgs(value: unknown): { dispatched: unknown; logged: unknown } {
+  if (value === undefined) {
+    throw new Error('tool arguments must be JSON-serializable (call the tool with an arguments object, e.g. `{}`)')
+  }
+  let text: string | undefined
+  try {
+    text = JSON.stringify(value)
+  } catch (error: unknown) {
+    throw new Error(`tool arguments must be JSON-serializable: ${error instanceof Error ? error.message : String(error)}`)
+  }
+  // JSON.stringify's lib type claims `string`, but a bare function or symbol
+  // root really yields `undefined` at runtime — the guard is live.
+  // eslint-disable-next-line @typescript-eslint/no-unnecessary-condition
+  if (text === undefined) throw new Error('tool arguments must be JSON-serializable (got a value JSON cannot represent)')
+  return { dispatched: JSON.parse(text) as unknown, logged: JSON.parse(text) as unknown }
+}
+
+/** Render the program's completion value for the model-facing result text (`''` when the program returned nothing). */
+function renderValue(value: unknown): string {
+  if (value === undefined) return ''
+  return typeof value === 'string' ? value : inspect(value, INSPECT_OPTIONS)
+}
+
+/** The run_code result's `meta` payload (JSON-serializable; `presentResult` narrows it back). */
+interface RunCodeMeta {
+  logs: CodeRunResult['logs']
+  dispatches: number
+}
+
+/** Soft-narrow a result `meta` back to {@link RunCodeMeta} (replay may carry older shapes; presentation must not throw). */
+function asRunCodeMeta(meta: unknown): RunCodeMeta | undefined {
+  if (typeof meta !== 'object' || meta === null) return undefined
+  const m = meta as Record<string, unknown>
+  if (!Array.isArray(m.logs) || typeof m.dispatches !== 'number') return undefined
+  return m as unknown as RunCodeMeta
+}
+
+/**
+ * Build the `run_code` {@link ToolDefinition}: one required `code` parameter,
+ * executed through the dispatch bridge described in the module doc. The
+ * registry registers it under non-native modes.
+ * @param registry - the owning registry (sub-calls go through its `execute`,
+ *   bindings cover its registered tools).
+ * @param requireRuntime - resolves `ctx.codeRuntime` or throws the loud
+ *   misconfiguration error (shared with the registry's assembly-time checks).
+ * @returns the registry-ready definition.
+ */
+export function createRunCodeTool(registry: ToolRegistry, requireRuntime: () => CodeRuntime): ToolDefinition {
+  return defineTool({
+    name: RUN_CODE_NAME,
+    description:
+      'Execute a TypeScript program against the available tools. Write the BODY of an '
+      + 'async function (erasable syntax only; top-level `await` and `return` work) and '
+      + 'call tools as `await tools.name(args)` per the declarations in the system prompt. '
+      + 'Only what you print or return comes back — curate it.',
+    parameters: {
+      code: { type: 'string', required: true, description: 'The program: the body of an async TypeScript function.' },
+    },
+    async execute(args, exec) {
+      const runtime = requireRuntime()
+
+      // The run-scoped abort: follows the outer signal in, and fires when the
+      // run settles for ANY reason, so an in-flight sub-dispatch is aborted
+      // (its executor kills on this signal) instead of orphaned, and
+      // queued-unstarted dispatches are abandoned.
+      const runController = new AbortController()
+      const onOuterAbort = (): void => { runController.abort(exec.signal?.reason) }
+      if (exec.signal?.aborted) onOuterAbort()
+      exec.signal?.addEventListener('abort', onOuterAbort, { once: true })
+
+      let dispatches = 0
+      // The per-run serialization queue: every binding call chains onto the
+      // tail, so even `Promise.all` executes the underlying tool calls one at
+      // a time in submission order (the tool contract carries no
+      // concurrency-safety metadata yet). The fold keeps the tail non-rejecting
+      // so one failed dispatch never poisons the chain.
+      let queue: Promise<void> = Promise.resolve()
+      const enqueue = <T>(task: () => Promise<T>): Promise<T> => {
+        const turn = queue.then(() => {
+          if (runController.signal.aborted) {
+            throw new Error(`run_code run is over (${String(runController.signal.reason)}); tool call abandoned`)
+          }
+          return task()
+        })
+        queue = turn.then(() => undefined, () => undefined)
+        return turn
+      }
+
+      // Read through a call, not a bare property: the abort state genuinely
+      // changes across awaits, and a direct `.aborted` re-check after one
+      // would be narrowed away by control flow analysis.
+      const runOver = (): boolean => runController.signal.aborted
+
+      const binding = (name: string): CodeBindingFunction => async (rawArgs: unknown): Promise<unknown> => {
+        if (runOver()) {
+          throw new Error(`run_code run is over (${String(runController.signal.reason)}); ${name} not dispatched`)
+        }
+        const normalized = jsonNormalizeArgs(rawArgs)
+        const outcome = await enqueue(async () => {
+          const n = ++dispatches
+          const subCallId = CallId(`${String(exec.callId)}:code:${n}`)
+          const result = await registry.execute({
+            callId: subCallId,
+            name,
+            arguments: normalized.dispatched,
+            ...exec.agent ? { agent: exec.agent } : {},
+            signal: runController.signal,
+          })
+          const text = textOf(result.content)
+          // Sub-call `additionalContext` is deliberately DROPPED here: the
+          // loop's buffering (append after the step's tool/results) has no
+          // safe analogue from inside a running run_code — injecting now
+          // would break tool-call/result adjacency. Deferred until a real
+          // hook needs it through Code Mode.
+          exec.agent?.session.append('tool/code-dispatch', {
+            parentCallId: exec.callId,
+            subCallId,
+            name,
+            // The SIBLING parse of the dispatched value: byte-identical JSON,
+            // but a separate object — a tool mutating its args cannot desync
+            // this record from what it actually received.
+            arguments: normalized.logged,
+            isError: result.isError,
+            resultSummary: summarize(text),
+          })
+          return { text, isError: result.isError }
+        })
+        // A budget expiry or outer cancel that lands while this call was in
+        // flight already aborted the dispatch; stop the program now rather
+        // than hand it a result from a run that is over.
+        if (runOver()) {
+          throw new Error(`run_code run is over (${String(runController.signal.reason)}); ${name} result discarded`)
+        }
+        // A failed tool call REJECTS — real code signals failure by throwing,
+        // so try/catch and Promise.all short-circuiting behave as models
+        // expect (the error text is the tool's model-facing result text).
+        if (outcome.isError) throw new Error(outcome.text)
+        return outcome.text
+      }
+
+      // Null-prototype + defineProperty, mirroring the worker-side namespace
+      // build: a registered tool named `__proto__` must become an ordinary
+      // own key (a plain-object assignment would hit the prototype setter,
+      // silently dropping the binding), and the runtime host resolves
+      // binding names as own properties only.
+      const functions: Record<string, CodeBindingFunction> = Object.create(null) as Record<string, CodeBindingFunction>
+      for (const schema of registry.schemas()) {
+        if (schema.name === RUN_CODE_NAME) continue
+        Object.defineProperty(functions, schema.name, { enumerable: true, value: binding(schema.name) })
+      }
+
+      try {
+        let result: CodeRunResult
+        try {
+          result = await runtime.run({
+            program: args.code,
+            bindings: [{ global: 'tools', functions }],
+            signal: runController.signal,
+          })
+        } finally {
+          // Quiescence before returning, whether the runtime fulfilled or
+          // REJECTED (a backend that starts a binding call and then throws
+          // must not leak a live sub-dispatch past this settlement): fire
+          // the run-scoped abort (cancelling an in-flight sub-dispatch,
+          // abandoning queued ones), then await the queue's drain — an
+          // aborted sub-call still settles and logs its event INSIDE the
+          // open turn; nothing can append after we return. `queue` is the
+          // FOLDED tail (every link swallows its rejection into undefined),
+          // so this await cannot itself reject — an abandoned queued call
+          // can never mask the runtime's own failure, returned or thrown;
+          // rejections surface only on the per-call promises the program
+          // holds.
+          runController.abort('run_code settled')
+          await queue
+        }
+
+        if (result.error) {
+          const logsText = result.logs.length > 0 ? `\nCaptured output:\n${result.logs.map(entry => entry.text).join('\n')}` : ''
+          throw new CodeRunFailedError(`code run failed (${result.error.kind}): ${result.error.message}${logsText}`)
+        }
+        const rendered = renderValue(result.value)
+        const parts = [result.logs.map(entry => entry.text).join('\n'), rendered].filter(part => part.length > 0)
+        const meta: RunCodeMeta = { logs: result.logs, dispatches }
+        return {
+          content: [{ type: 'text', text: parts.length > 0 ? parts.join('\n') : '(run_code completed with no output)' }],
+          meta,
+        }
+      } finally {
+        exec.signal?.removeEventListener('abort', onOuterAbort)
+      }
+    },
+    // The program IS the title, the way command tools title their cards with
+    // the command: an execute-card's title is the one slot an ACP client
+    // always shows (Zed's execute cards render no body content and no raw
+    // input without a real terminal attached), so anywhere else the code
+    // would be invisible. Multi-line titles are the execute-card idiom —
+    // capable clients render them whole; others truncate to the first line
+    // and still hold the full program in rawInput.
+    presentCall: args => ({
+      card: 'generic',
+      title: args.code,
+      kind: 'execute',
+      rawInput: args.code,
+    }),
+    // Title omitted on the result: an update replaces only the fields it
+    // carries, so the pending card's program title persists through
+    // completion; the captured output rides as body content.
+    presentResult: (_args, result) => {
+      const meta = asRunCodeMeta(result.meta)
+      if (!meta) return undefined
+      const output = meta.logs.map(entry => entry.text).join('\n')
+      return {
+        card: 'generic',
+        ...output.length > 0 ? { content: [{ type: 'text' as const, text: output }] } : {},
+      }
+    },
+  })
+}

packages/core/tools/src/index.ts

@@ -6,15 +6,26 @@
  * (inspect/replace the result, attach context) for sandbox, permission, and hook
  * plugins to gate or transform a call.
  *
+ * The registry also owns HOW its tools are presented to the model — its
+ * `mode` config: `'native'` (every tool as a wire function definition,
+ * today's behavior and the default), `'code'` (the wire carries exactly one
+ * tool, `run_code`, plus a generated TypeScript SDK prompt section), or
+ * `'both'`. See `code-mode.ts` (the tool + dispatch bridge) and
+ * `ts-types.ts` (the SDK codegen); design in the Code Mode RFC.
+ *
  * @module @deepseek-ai/dsh-tools
  */
 
 import { Context, Service } from 'cordis'
+import z from 'schemastery'
 import type { CallId, ContentBlock, ToolSchema } from '@deepseek-ai/dsh-llm'
 import { HarnessError } from '@deepseek-ai/dsh-llm'
 import type { Agent, HookContext } from '@deepseek-ai/dsh-agent'
 import type {} from '@deepseek-ai/dsh-system-prompt'
+import type { CodeRuntime } from '@deepseek-ai/dsh-code-runtime'
 import type { ToolCallView, ToolResultView } from './presentation.ts'
+import { createRunCodeTool, RUN_CODE_NAME, SDK_SECTION_ORDER } from './code-mode.ts'
+import { renderToolsSdk } from './ts-types.ts'
 
 export {
   defineTool,
@@ -39,6 +50,9 @@ export {
   type StructuredScalar,
 } from './json-schema.ts'
 
+export { CodeRunFailedError, RUN_CODE_NAME } from './code-mode.ts'
+export { jsonSchemaToTs, renderToolsSdk } from './ts-types.ts'
+
 // The render-intent vocabulary a tool declares via `presentCall`/`presentResult`
 // lives in its own UI-facing module; re-export it so `@deepseek-ai/dsh-tools`
 // stays the single public surface for consumers (producers + the ACP bridge).
@@ -298,20 +312,100 @@ function errorInfo(error: unknown): ToolErrorInfo | undefined {
   return error instanceof HarnessError ? { name: error.name, code: error.code } : undefined
 }
 
+/** How the registry presents its tools to the model (see {@link Config.mode}). */
+export type ToolPresentationMode = 'native' | 'code' | 'both'
+
+/** Plugin config: how the registered tools are presented to the model. */
+export interface Config {
+  /**
+   * The presentation mode. `'native'` (the default) contributes every
+   * registered tool as a wire function definition — byte-for-byte today's
+   * behavior. `'code'` contributes exactly ONE wire tool, `run_code`, plus
+   * the generated `tools:sdk` prompt section declaring every other tool as a
+   * TypeScript API the program calls. `'both'` contributes every native
+   * definition AND `run_code` + the SDK section. Non-native modes require a
+   * loaded `ctx.codeRuntime` whose `language` is `'typescript'` — a missing
+   * or mismatched runtime rejects every prompt assembly with an actionable
+   * error (misconfiguration fails loud, before any model request). A
+   * configured `systemPrompt.toolOrder` naming native tools likewise rejects
+   * every assembly under `'code'` (those names are no longer contributed) —
+   * a deployment switching modes updates its order config or drops it.
+   */
+  mode?: ToolPresentationMode
+}
+
 /**
  * Tool registry (`ctx.tools`): tool plugins register definitions; the agent
  * loop executes calls through the `tools/pre-execute` → `tools/execute` →
  * `tools/post-execute` pipeline. The registry contributes its schemas into the
- * system-prompt assembly.
+ * system-prompt assembly — WHICH schemas is governed by its `mode` config
+ * (see {@link Config.mode}); under a non-native mode it also registers the
+ * `run_code` tool and the `tools:sdk` prompt section itself.
  */
 export class ToolRegistry extends Service {
   static inject = ['systemPrompt']
 
-  private store = new Map<string, ToolDefinition>()
+  static Config: z<Config> = z.object({
+    mode: z.union(['native', 'code', 'both'] as const).default('native'),
+  })
 
-  constructor(ctx: Context) {
+  private store = new Map<string, ToolDefinition>()
+  private readonly mode: ToolPresentationMode
+
+  constructor(ctx: Context, config: Config = {}) {
     super(ctx, 'tools')
-    ctx.systemPrompt.tools(() => this.schemas())
+    // The schema already defaulted an omitted mode; the ?? narrows the
+    // optional-input type for direct (non-Loader) construction in tests.
+    this.mode = config.mode ?? 'native'
+    ctx.systemPrompt.tools(() => this.wireSchemas())
+    if (this.mode !== 'native') {
+      this.register(createRunCodeTool(this, () => this.requireCodeRuntime()))
+      ctx.systemPrompt.section({
+        name: 'tools:sdk',
+        order: SDK_SECTION_ORDER,
+        // A lazy thunk over the live store: regenerated at each assembly, in
+        // lexicographic tool order, so an unchanged tool set renders
+        // byte-identical text (prefix-cache-friendly) and a mid-session
+        // registration surfaces exactly like a native-mode tool change.
+        text: () => {
+          this.requireCodeRuntime()
+          return renderToolsSdk(this.schemas().filter(schema => schema.name !== RUN_CODE_NAME))
+        },
+      })
+    }
+  }
+
+  /**
+   * The registry's contribution to the wire tool list, per {@link Config.mode}.
+   * Because `PromptAssembly.tools` is what the loop's request header
+   * snapshots, the mode's collapse is logged and reconstructable for free.
+   * Under a non-native mode this is also the loud misconfiguration gate: no
+   * usable code runtime → every assembly rejects before any model request.
+   */
+  private wireSchemas(): ToolSchema[] {
+    if (this.mode === 'native') return this.schemas()
+    this.requireCodeRuntime()
+    const all = this.schemas()
+    return this.mode === 'code' ? all.filter(schema => schema.name === RUN_CODE_NAME) : all
+  }
+
+  /**
+   * Resolve the code runtime or throw the actionable misconfiguration error.
+   * Read at use time (assembly / run_code execution), NOT via static
+   * `inject`: an inject entry would hold `ctx.tools` — and every tool plugin
+   * behind it — hostage to a code runtime existing even under `mode:
+   * 'native'` (the loop's optional-backend idiom, same as
+   * `sessionPersistence`).
+   */
+  private requireCodeRuntime(): CodeRuntime {
+    const runtime = this.ctx.get('codeRuntime')
+    if (!runtime) {
+      throw new Error(`dsh-tools: mode "${this.mode}" requires a code runtime — load a ctx.codeRuntime implementation (e.g. @deepseek-ai/dsh-code-runtime-worker) or set tools mode to "native"`)
+    }
+    if (runtime.language !== 'typescript') {
+      throw new Error(`dsh-tools: mode "${this.mode}" generates a TypeScript SDK, but the loaded code runtime's language is "${runtime.language}"`)
+    }
+    return runtime
   }
 
   /**

packages/core/tools/src/ts-types.ts

@@ -0,0 +1,121 @@
+/**
+ * Code Mode codegen: the pure projection from registered tool schemas to the
+ * TypeScript SDK text the model programs against (the `tools:sdk` prompt
+ * section). Sibling of `json-schema.ts` — `schemas()` (native function
+ * calling) and this module (the generated `declare const tools` surface) are
+ * two projections of the same store.
+ *
+ * TOTAL by design: {@link jsonSchemaToTs} maps the JSON-Schema subset the
+ * `defineTool` DSL emits and degrades every construct outside it (`$ref`,
+ * `oneOf`, `integer`, future MCP shapes, …) to `unknown` without ever
+ * throwing — codegen must never be the thing that fails an assembly.
+ * Deterministic: a fixed tool set renders byte-identical text (tools in
+ * lexicographic name order), so the section is prefix-cache-friendly.
+ *
+ * @module @deepseek-ai/dsh-tools/src/ts-types
+ */
+
+import type { ToolSchema } from '@deepseek-ai/dsh-llm'
+
+/** Property names that are valid bare TS identifiers; anything else is quoted. */
+const IDENTIFIER = /^[A-Za-z_$][A-Za-z0-9_$]*$/
+
+/** Render an object key: bare when it is a valid identifier, quoted otherwise (every name stays reachable, no aliasing). */
+function renderKey(name: string): string {
+  return IDENTIFIER.test(name) ? name : JSON.stringify(name)
+}
+
+/** One `indent`-deep line prefix (two spaces per level). */
+function pad(indent: number): string {
+  return '  '.repeat(indent)
+}
+
+/** A one-line JSDoc block for a schema `description`, or no lines when there is none. */
+function docLines(description: unknown, indent: number): string[] {
+  if (typeof description !== 'string' || description.length === 0) return []
+  // Keep the doc a single-line comment per property: descriptions are prose
+  // (possibly with newlines); collapse whitespace so the rendered SDK stays
+  // stable and compact. A comment-closer inside the description is escaped so
+  // it cannot terminate the generated JSDoc early.
+  const collapsed = description.replace(/\s+/g, ' ').trim()
+  return [`${pad(indent)}/** ${collapsed.replaceAll('*/', String.raw`*\/`)} */`]
+}
+
+/**
+ * Map one JSON-Schema node to a TypeScript type literal. Handles exactly the
+ * subset the `defineTool` DSL emits — `object` (`properties` + `required`),
+ * `string` (with `enum` → a literal union), `number`, `boolean`, `array`
+ * (`items`) — and returns `unknown` for anything else, without throwing.
+ * @param schema - the JSON-Schema node (any shape; hostile inputs degrade).
+ * @param indent - the indentation level for nested object members.
+ * @returns the TS type text (multi-line for objects with properties).
+ */
+export function jsonSchemaToTs(schema: unknown, indent = 0): string {
+  if (typeof schema !== 'object' || schema === null) return 'unknown'
+  const node = schema as Record<string, unknown>
+  switch (node.type) {
+    case 'string': {
+      if (Array.isArray(node.enum) && node.enum.length > 0 && node.enum.every(value => typeof value === 'string')) {
+        return node.enum.map(value => JSON.stringify(value)).join(' | ')
+      }
+      return 'string'
+    }
+    case 'number': return 'number'
+    case 'boolean': return 'boolean'
+    case 'array': {
+      const item = jsonSchemaToTs(node.items, indent)
+      // Parenthesize a union item type so `('a' | 'b')[]` parses as intended.
+      return item.includes('|') ? `(${item})[]` : `${item}[]`
+    }
+    case 'object': {
+      const properties = node.properties
+      if (typeof properties !== 'object' || properties === null) return 'Record<string, unknown>'
+      const entries = Object.entries(properties as Record<string, unknown>)
+      if (entries.length === 0) return 'Record<string, unknown>'
+      const required = new Set(Array.isArray(node.required) ? node.required.filter(name => typeof name === 'string') : [])
+      const lines: string[] = ['{']
+      for (const [name, prop] of entries) {
+        const description = typeof prop === 'object' && prop !== null ? (prop as Record<string, unknown>).description : undefined
+        lines.push(...docLines(description, indent + 1))
+        lines.push(`${pad(indent + 1)}${renderKey(name)}${required.has(name) ? '' : '?'}: ${jsonSchemaToTs(prop, indent + 1)};`)
+      }
+      lines.push(`${pad(indent)}}`)
+      return lines.join('\n')
+    }
+    default: return 'unknown'
+  }
+}
+
+/** The fixed model-facing usage contract rendered above the declarations (see the Code Mode RFC's "What the model sees"). */
+const SDK_INSTRUCTIONS = `## Writing code for run_code
+
+Pass \`run_code\` the body of an async TypeScript function (erasable syntax only — no \`enum\` or namespaces; type annotations are advisory, the code runs type-stripped). Inside the program:
+
+- Call tools as \`await tools.name(args)\` — quoted access for exotic names: \`tools["my-tool"](args)\`. Every call resolves to the tool's text output as a string. Tool arguments must be JSON-serializable.
+- A FAILED tool call rejects with an \`Error\` carrying the tool's error text — \`try/catch\` it to handle and continue.
+- Calls execute sequentially, even under \`Promise.all\`.
+- Emit results with \`return\` and/or \`console.log(...)\`. ONLY what you print or return comes back to you — intermediate tool results never enter the conversation, so extract just what you need.
+
+The available tools:`
+
+/**
+ * Render the full `tools:sdk` prompt section: the fixed usage instructions
+ * plus one `declare const tools` interface covering every given tool.
+ * Deterministic — tools are emitted in lexicographic name order, so an
+ * unchanged tool set produces byte-identical text across assemblies.
+ * @param schemas - the tool schemas to declare (the caller excludes
+ *   `run_code` itself).
+ * @returns the complete section text.
+ */
+export function renderToolsSdk(schemas: ToolSchema[]): string {
+  const sorted = [...schemas].sort((a, b) => a.name < b.name ? -1 : a.name > b.name ? 1 : 0)
+  const members: string[] = []
+  for (const schema of sorted) {
+    members.push(...docLines(schema.description, 1))
+    members.push(`${pad(1)}${renderKey(schema.name)}(args: ${jsonSchemaToTs(schema.parameters, 1)}): Promise<string>;`)
+  }
+  const declaration = members.length > 0
+    ? `declare const tools: {\n${members.join('\n')}\n}`
+    : 'declare const tools: {}'
+  return `${SDK_INSTRUCTIONS}\n\n\`\`\`ts\n${declaration}\n\`\`\``
+}

packages/core/tools/tests/code-mode.spec.ts

@@ -0,0 +1,640 @@
+import { describe, expect, it } from 'vitest'
+import { Context } from 'cordis'
+import { CallId } from '@deepseek-ai/dsh-llm'
+import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
+import { CodeRuntime } from '@deepseek-ai/dsh-code-runtime'
+import type { CodeRunRequest, CodeRunResult } from '@deepseek-ai/dsh-code-runtime'
+import ToolRegistry, { CodeRunFailedError, RUN_CODE_NAME, defineTool } from '@deepseek-ai/dsh-tools'
+import type { Config, PostToolDecision, ToolExecutionResult } from '@deepseek-ai/dsh-tools'
+import type { Agent } from '@deepseek-ai/dsh-agent'
+import { Session, SessionId } from '@deepseek-ai/dsh-session'
+import type { SessionEventMap } from '@deepseek-ai/dsh-session'
+
+/**
+ * Code Mode unit tier (per the RFC's plan): provider contribution per mode,
+ * misconfiguration rejections, the run_code dispatch bridge (serialization,
+ * abort, JSON normalization, error mapping, events, quiescence), and HMR
+ * safety — all against an in-repo fake runtime, exactly the
+ * interface/implementation/consumer shape the seam promises.
+ */
+
+/** A scriptable in-repo CodeRuntime: each test sets `behavior` to drive the bindings however it needs. */
+class FakeRuntime extends CodeRuntime {
+  readonly language: string
+  readonly isolation = 'fake'
+  behavior: (request: CodeRunRequest) => Promise<CodeRunResult> = () => Promise.resolve({ logs: [] })
+  lastRequest?: CodeRunRequest
+
+  constructor(ctx: Context, config: { language?: string } = {}) {
+    super(ctx)
+    this.language = config.language ?? 'typescript'
+  }
+
+  run(request: CodeRunRequest): Promise<CodeRunResult> {
+    this.lastRequest = request
+    return this.behavior(request)
+  }
+}
+
+interface SetupOptions {
+  mode?: Config['mode']
+  runtime?: false | { language?: string }
+  toolOrder?: string[]
+}
+
+async function setup(options: SetupOptions = {}) {
+  const ctx = new Context()
+  await ctx.plugin(SystemPrompt, { ...options.toolOrder ? { toolOrder: options.toolOrder } : {} })
+  await ctx.plugin(ToolRegistry, { mode: options.mode ?? 'code' })
+  let runtime: FakeRuntime | undefined
+  if (options.runtime !== false) {
+    await ctx.plugin(FakeRuntime, options.runtime ?? {})
+    runtime = ctx.codeRuntime as FakeRuntime
+  }
+  return { ctx, tools: ctx.tools, systemPrompt: ctx.systemPrompt, runtime: runtime! }
+}
+
+/** Register a trivial echo tool; returns the calls it received. */
+function registerEcho(ctx: Context, name = 'echo'): unknown[] {
+  const calls: unknown[] = []
+  ctx.tools.register(defineTool({
+    name,
+    description: `Echo tool ${name}.`,
+    parameters: { value: { type: 'string', required: true } },
+    execute(args) {
+      calls.push(args)
+      return Promise.resolve([{ type: 'text' as const, text: `${name}:${args.value}` }])
+    },
+  }))
+  return calls
+}
+
+/** A structural fake of the owning agent: captures session appends. */
+function fakeAgent(): { agent: Agent; events: { type: string; data: unknown }[] } {
+  const events: { type: string; data: unknown }[] = []
+  const agent = {
+    session: {
+      append: (type: string, data: unknown) => { events.push({ type, data }) },
+    },
+  } as unknown as Agent
+  return { agent, events }
+}
+
+/** Dispatch run_code through the registry pipeline, as the loop would. */
+async function runCode(ctx: Context, code: string, extras: { agent?: Agent; signal?: AbortSignal } = {}): Promise<ToolExecutionResult> {
+  return ctx.tools.execute({
+    callId: CallId('call-1'),
+    name: RUN_CODE_NAME,
+    arguments: { code },
+    ...extras.agent ? { agent: extras.agent } : {},
+    ...extras.signal ? { signal: extras.signal } : {},
+  })
+}
+
+describe('mode-aware wire contribution', () => {
+  it("mode 'native' contributes every schema, no run_code, no SDK section — and needs no runtime", async () => {
+    const { ctx, systemPrompt } = await setup({ mode: 'native', runtime: false })
+    registerEcho(ctx)
+    const assembly = await systemPrompt.assemble()
+    expect(assembly.tools.map(tool => tool.name)).toEqual(['echo'])
+    expect(assembly.sections.some(section => section.name === 'tools:sdk')).toBe(false)
+  })
+
+  it("mode 'code' contributes exactly [run_code] plus the SDK section declaring the other tools", async () => {
+    const { ctx, systemPrompt } = await setup({ mode: 'code' })
+    registerEcho(ctx)
+    const assembly = await systemPrompt.assemble()
+    expect(assembly.tools.map(tool => tool.name)).toEqual([RUN_CODE_NAME])
+    const sdk = assembly.sections.find(section => section.name === 'tools:sdk')
+    expect(sdk?.text).toContain('declare const tools: {')
+    expect(sdk?.text).toContain('echo(args:')
+    expect(sdk?.text).not.toContain('run_code(args:')
+  })
+
+  it("mode 'both' contributes every native schema plus run_code, and the SDK section", async () => {
+    const { ctx, systemPrompt } = await setup({ mode: 'both' })
+    registerEcho(ctx)
+    const assembly = await systemPrompt.assemble()
+    expect(assembly.tools.map(tool => tool.name)).toEqual(['echo', RUN_CODE_NAME])
+    expect(assembly.sections.some(section => section.name === 'tools:sdk')).toBe(true)
+  })
+
+  it("never exposes run_code to programs, even under mode 'both' (no recursive dispatch path)", async () => {
+    const { ctx, runtime } = await setup({ mode: 'both' })
+    registerEcho(ctx)
+    runtime.behavior = (request) => {
+      const functions = request.bindings[0]!.functions
+      return Promise.resolve({
+        logs: [],
+        value: JSON.stringify({
+          names: Object.keys(functions).sort(),
+          // Own-property AND prototype-chain reads both come back empty —
+          // there is no handle a program could re-enter run_code through.
+          runCode: String(functions[RUN_CODE_NAME]),
+        }),
+      })
+    }
+    const result = await runCode(ctx, 'program')
+    expect(result.isError).toBe(false)
+    expect(JSON.parse((result.content[0] as { text: string }).text)).toEqual({ names: ['echo'], runCode: 'undefined' })
+  })
+
+  it('renders byte-identical SDK text across consecutive assemblies of an unchanged tool set', async () => {
+    const { ctx, systemPrompt } = await setup({ mode: 'code' })
+    registerEcho(ctx)
+    const first = await systemPrompt.assemble()
+    const second = await systemPrompt.assemble()
+    const text = (assembly: typeof first) => assembly.sections.find(section => section.name === 'tools:sdk')?.text
+    expect(text(first)).toBe(text(second))
+  })
+
+  it('rejects every assembly when a non-native mode has no code runtime', async () => {
+    const { systemPrompt } = await setup({ mode: 'code', runtime: false })
+    await expect(systemPrompt.assemble()).rejects.toThrow(/requires a code runtime/)
+  })
+
+  it("rejects every assembly when the runtime's language is not typescript", async () => {
+    const { systemPrompt } = await setup({ mode: 'code', runtime: { language: 'python' } })
+    await expect(systemPrompt.assemble()).rejects.toThrow(/language is "python"/)
+  })
+
+  it("rejects the assembly when toolOrder names a native tool that mode 'code' no longer contributes", async () => {
+    const { ctx, systemPrompt } = await setup({ mode: 'code', toolOrder: ['echo', '<unlisted-tools>'] })
+    registerEcho(ctx)
+    await expect(systemPrompt.assemble()).rejects.toThrow(/toolOrder lists unregistered tool "echo"/)
+  })
+
+  it('removes run_code and the SDK section when the registry fiber disposes (HMR safety)', async () => {
+    const ctx = new Context()
+    await ctx.plugin(SystemPrompt, {})
+    await ctx.plugin(FakeRuntime, {})
+    const fiber = await ctx.plugin(ToolRegistry, { mode: 'code' })
+    expect(ctx.tools.get(RUN_CODE_NAME)).toBeDefined()
+    await fiber.dispose()
+    const assembly = await ctx.systemPrompt.assemble()
+    expect(assembly.tools).toEqual([])
+    expect(assembly.sections.some(section => section.name === 'tools:sdk')).toBe(false)
+  })
+})
+
+describe('the run_code dispatch bridge', () => {
+  it('bridges tool calls, returns only the curated output, and logs one event per dispatch', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const calls = registerEcho(ctx)
+    const { agent, events } = fakeAgent()
+    runtime.behavior = async (request) => {
+      const tools = request.bindings[0]!.functions
+      const first = await tools.echo!({ value: 'one' })
+      const second = await tools.echo!({ value: 'two' })
+      return { logs: [{ source: 'console', level: 'log', text: `saw ${String(first)}` }], value: second }
+    }
+    const result = await runCode(ctx, 'const …: string = …', { agent })
+    expect(result.isError).toBe(false)
+    expect(result.content).toEqual([{ type: 'text', text: 'saw echo:one\necho:two' }])
+    expect(calls).toEqual([{ value: 'one' }, { value: 'two' }])
+    const dispatches = events.filter(event => event.type === 'tool/code-dispatch')
+    expect(dispatches.map(event => event.data)).toEqual([
+      { parentCallId: 'call-1', subCallId: 'call-1:code:1', name: 'echo', arguments: { value: 'one' }, isError: false, resultSummary: 'echo:one' },
+      { parentCallId: 'call-1', subCallId: 'call-1:code:2', name: 'echo', arguments: { value: 'two' }, isError: false, resultSummary: 'echo:two' },
+    ])
+    expect(result.meta).toEqual({ logs: [{ source: 'console', level: 'log', text: 'saw echo:one' }], dispatches: 2 })
+  })
+
+  it('serializes Promise.all dispatches: tool executions never overlap, in submission order', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const intervals: [string, string][] = []
+    let active = 0
+    ctx.tools.register(defineTool({
+      name: 'probe',
+      description: 'Records execution overlap.',
+      parameters: { id: { type: 'string', required: true } },
+      async execute(args) {
+        active++
+        expect(active, 'probe executions overlapped').toBe(1)
+        intervals.push(['enter', args.id])
+        await new Promise(resolve => setTimeout(resolve, 20))
+        intervals.push(['exit', args.id])
+        active--
+        return [{ type: 'text' as const, text: args.id }]
+      },
+    }))
+    runtime.behavior = async (request) => {
+      const tools = request.bindings[0]!.functions
+      const values = await Promise.all([tools.probe!({ id: 'a' }), tools.probe!({ id: 'b' }), tools.probe!({ id: 'c' })])
+      return { logs: [], value: values.join(',') }
+    }
+    const result = await runCode(ctx, 'program')
+    expect(result.isError).toBe(false)
+    expect(intervals).toEqual([
+      ['enter', 'a'], ['exit', 'a'],
+      ['enter', 'b'], ['exit', 'b'],
+      ['enter', 'c'], ['exit', 'c'],
+    ])
+    expect(result.content[0]).toEqual({ type: 'text', text: 'a,b,c' })
+  })
+
+  it('rejects the program-side call when the tool errors, with the tool error text', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    ctx.tools.register(defineTool({
+      name: 'fail',
+      description: 'Always fails.',
+      parameters: {},
+      execute(): Promise<never> { return Promise.reject(new Error('deliberate failure')) },
+    }))
+    runtime.behavior = async (request) => {
+      try {
+        await request.bindings[0]!.functions.fail!({})
+        return { logs: [], value: 'unreachable' }
+      } catch (error: unknown) {
+        return { logs: [], value: `caught: ${error instanceof Error ? error.message : String(error)}` }
+      }
+    }
+    const result = await runCode(ctx, 'program')
+    expect(result.content[0]).toEqual({ type: 'text', text: 'caught: Error: deliberate failure' })
+  })
+
+  it('a tools/pre-execute deny reaches the program as a binding rejection', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    registerEcho(ctx)
+    ctx.on('tools/pre-execute', (exec, next) => {
+      if (exec.name === 'echo') return Promise.resolve({ kind: 'deny' as const, reason: 'not on my watch' })
+      return next()
+    })
+    runtime.behavior = async (request) => {
+      try {
+        await request.bindings[0]!.functions.echo!({ value: 'x' })
+        return { logs: [], value: 'unreachable' }
+      } catch (error: unknown) {
+        return { logs: [], value: `denied: ${error instanceof Error ? error.message : String(error)}` }
+      }
+    }
+    const result = await runCode(ctx, 'program')
+    expect(result.content[0]?.type).toBe('text')
+    expect((result.content[0] as { text: string }).text).toContain('not on my watch')
+  })
+
+  it('rejects a binding argument that does not survive JSON normalization, dispatching nothing', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const calls = registerEcho(ctx)
+    const { agent, events } = fakeAgent()
+    runtime.behavior = async (request) => {
+      try {
+        await request.bindings[0]!.functions.echo!({ value: 'x', big: 1n })
+        return { logs: [], value: 'unreachable' }
+      } catch (error: unknown) {
+        return { logs: [], value: error instanceof Error ? error.message : String(error) }
+      }
+    }
+    const result = await runCode(ctx, 'program', { agent })
+    expect((result.content[0] as { text: string }).text).toContain('JSON-serializable')
+    expect(calls).toEqual([])
+    expect(events.filter(event => event.type === 'tool/code-dispatch')).toEqual([])
+  })
+
+  it('dispatches the JSON-normalized value: what the tool sees is what the event logs', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const calls = registerEcho(ctx)
+    const { agent, events } = fakeAgent()
+    runtime.behavior = async (request) => {
+      // A Date survives structured clone but is not JSON; the bridge
+      // normalizes it to its JSON form (an ISO string) BEFORE dispatch.
+      await request.bindings[0]!.functions.echo!({ value: 'x', when: new Date(0) }).catch(() => undefined)
+      return { logs: [] }
+    }
+    await runCode(ctx, 'program', { agent })
+    expect(calls).toEqual([{ value: 'x', when: '1970-01-01T00:00:00.000Z' }])
+    const dispatch = events.find(event => event.type === 'tool/code-dispatch')?.data as SessionEventMap['tool/code-dispatch']
+    expect(dispatch.arguments).toEqual({ value: 'x', when: '1970-01-01T00:00:00.000Z' })
+  })
+
+  it('suppresses sub-call additionalContext (deliberately; pinned)', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    registerEcho(ctx)
+    ctx.on('tools/post-execute', (exec, _result, next): Promise<PostToolDecision> => {
+      if (exec.name === 'echo') {
+        return Promise.resolve({
+          kind: 'accept' as const,
+          additionalContext: { content: [{ type: 'text' as const, text: 'context for the next request' }], source: { kind: 'plugin' as const, plugin: 'test' } },
+        })
+      }
+      return next()
+    })
+    runtime.behavior = async (request) => {
+      await request.bindings[0]!.functions.echo!({ value: 'x' })
+      return { logs: [], value: 'done' }
+    }
+    const result = await runCode(ctx, 'program')
+    expect(result.isError).toBe(false)
+    // The sub-call's context has no safe outlet mid-run; the parent result
+    // must not carry it either.
+    expect(result.additionalContext).toBeUndefined()
+  })
+
+  it('converts a failed run into a structured isError result carrying kind, message, and captured logs', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    runtime.behavior = () => Promise.resolve({
+      logs: [{ source: 'console', level: 'log', text: 'got this far' }],
+      error: { kind: 'timeout', message: 'compute budget exhausted (300ms busy)' },
+    })
+    const result = await runCode(ctx, 'program')
+    expect(result.isError).toBe(true)
+    expect(result.error).toEqual({ name: 'CodeRunFailedError', code: 'CODE_RUN_FAILED' })
+    const text = (result.content[0] as { text: string }).text
+    expect(text).toContain('code run failed (timeout)')
+    expect(text).toContain('compute budget exhausted')
+    expect(text).toContain('got this far')
+  })
+
+  it('CodeRunFailedError is a HarnessError with the CODE_RUN_FAILED code', () => {
+    const error = new CodeRunFailedError('boom')
+    expect(error.code).toBe('CODE_RUN_FAILED')
+    expect(error.name).toBe('CodeRunFailedError')
+  })
+
+  it('aborting the outer signal aborts the in-flight sub-dispatch and abandons queued ones', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const seen: string[] = []
+    let sawAbort = false
+    ctx.tools.register(defineTool({
+      name: 'slow',
+      description: 'Slow tool observing its signal.',
+      parameters: { id: { type: 'string', required: true } },
+      async execute(args, exec) {
+        seen.push(args.id)
+        await new Promise<void>((resolve) => {
+          const timer = setTimeout(resolve, 500)
+          exec.signal?.addEventListener('abort', () => { sawAbort = true; clearTimeout(timer); resolve() }, { once: true })
+        })
+        return [{ type: 'text' as const, text: args.id }]
+      },
+    }))
+    const controller = new AbortController()
+    runtime.behavior = async (request) => {
+      const tools = request.bindings[0]!.functions
+      const calls = [tools.slow!({ id: 'first' }).catch(() => 'rejected'), tools.slow!({ id: 'second' }).catch(() => 'rejected')]
+      setTimeout(() => { controller.abort('user-cancel') }, 50)
+      await Promise.all(calls)
+      // A real runtime would be terminated by the abort; the fake honors the
+      // contract by reporting the abort as the run failure.
+      return { logs: [], error: { kind: 'abort', message: 'user-cancel' } }
+    }
+    const result = await runCode(ctx, 'program', { signal: controller.signal })
+    expect(result.isError).toBe(true)
+    expect((result.content[0] as { text: string }).text).toContain('code run failed (abort)')
+    expect(seen).toEqual(['first'])
+    expect(sawAbort).toBe(true)
+  })
+
+  it('a runtime that starts a binding call and then REJECTS still reaches quiescence before returning', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const { agent, events } = fakeAgent()
+    let sawAbort = false
+    let started!: () => void
+    const inFlight = new Promise<void>((resolve) => { started = resolve })
+    ctx.tools.register(defineTool({
+      name: 'slow',
+      description: 'Slow tool observing its signal.',
+      parameters: { id: { type: 'string', required: true } },
+      async execute(args, exec) {
+        started()
+        await new Promise<void>((resolve) => {
+          const timer = setTimeout(resolve, 500)
+          exec.signal?.addEventListener('abort', () => { sawAbort = true; clearTimeout(timer); resolve() }, { once: true })
+        })
+        return [{ type: 'text' as const, text: args.id }]
+      },
+    }))
+    runtime.behavior = async (request) => {
+      // Start a sub-dispatch, keep its rejection held, and fail the run once
+      // the tool is genuinely in flight — a seam error AFTER work has begun.
+      // The bridge's settlement still owes quiescence: without the finally,
+      // run_code would return now and the slow tool would finish (and log)
+      // afterwards.
+      request.bindings[0]!.functions.slow!({ id: 'orphan' }).catch(() => 'held')
+      await inFlight
+      throw new Error('backend exploded')
+    }
+    const result = await runCode(ctx, 'program', { agent })
+    expect(result.isError).toBe(true)
+    expect((result.content[0] as { text: string }).text).toContain('backend exploded')
+    // Quiescence held: the in-flight sub-dispatch was aborted and its event
+    // logged INSIDE the run_code execution, not after it returned.
+    expect(sawAbort).toBe(true)
+    expect(events.filter(event => event.type === 'tool/code-dispatch').map(event => (event.data as { name: string }).name)).toEqual(['slow'])
+  })
+
+  it('runs without an owning agent: dispatches work, event logging is skipped', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const calls = registerEcho(ctx)
+    runtime.behavior = async (request) => {
+      await request.bindings[0]!.functions.echo!({ value: 'x' })
+      return { logs: [], value: 'ok' }
+    }
+    const result = await runCode(ctx, 'program')
+    expect(result.isError).toBe(false)
+    expect(calls).toEqual([{ value: 'x' }])
+  })
+
+  it('executing run_code under a missing runtime is a structured isError, not a crash', async () => {
+    const ctx = new Context()
+    await ctx.plugin(SystemPrompt, {})
+    await ctx.plugin(ToolRegistry, { mode: 'code' })
+    const result = await runCode(ctx, 'program')
+    expect(result.isError).toBe(true)
+    expect((result.content[0] as { text: string }).text).toContain('requires a code runtime')
+  })
+
+  it('presents the PROGRAM as the execute-card title on both call and result (the one slot execute cards always show)', async () => {
+    const { ctx } = await setup({ mode: 'code' })
+    const tool = ctx.tools.get(RUN_CODE_NAME)!
+    // The program IS the title, mirroring how command tools title their cards
+    // with the command: an ACP client's execute-card header is the only
+    // always-visible slot (Zed renders no body content and no raw input for
+    // execute-kind cards without a real terminal).
+    expect(tool.presentCall?.({ code: 'return 1' })).toEqual({
+      card: 'generic',
+      title: 'return 1',
+      kind: 'execute',
+      rawInput: 'return 1',
+    })
+    const view = tool.presentResult?.({ code: 'return 1' }, {
+      content: [{ type: 'text', text: 'model-facing' }],
+      isError: false,
+      meta: { logs: [{ source: 'console', level: 'log', text: 'printed' }], dispatches: 1 },
+    })
+    // The result omits the title — an update replaces only provided fields,
+    // so the pending card's program title persists through completion.
+    expect(view).toEqual({
+      card: 'generic',
+      content: [{ type: 'text', text: 'printed' }],
+    })
+    // No captured output → no content either; everything pending persists.
+    expect(tool.presentResult?.({ code: 'x' }, { content: [], isError: false, meta: { logs: [], dispatches: 2 } }))
+      .toEqual({ card: 'generic' })
+    // Replay with an unrecognizable meta falls back to the generic rendering.
+    expect(tool.presentResult?.({ code: 'x' }, { content: [], isError: false, meta: { other: true } })).toBeUndefined()
+    expect(tool.presentResult?.({ code: 'x' }, { content: [], isError: false })).toBeUndefined()
+  })
+
+  it('renders non-text sub-result blocks as placeholders and truncates long event summaries', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const { agent, events } = fakeAgent()
+    const long = 'x'.repeat(300)
+    ctx.tools.register(defineTool({
+      name: 'mixed',
+      description: 'Returns mixed content.',
+      parameters: {},
+      execute() {
+        return Promise.resolve([
+          { type: 'text' as const, text: long },
+          { type: 'reasoning' as const, text: 'hidden' },
+        ])
+      },
+    }))
+    runtime.behavior = async (request) => {
+      const value = await request.bindings[0]!.functions.mixed!({})
+      return { logs: [], value }
+    }
+    const result = await runCode(ctx, 'program', { agent })
+    expect(result.isError).toBe(false)
+    expect((result.content[0] as { text: string }).text).toBe(`${long}\n[reasoning content]`)
+    const dispatch = events.find(event => event.type === 'tool/code-dispatch')?.data as SessionEventMap['tool/code-dispatch']
+    expect(dispatch.resultSummary.length).toBe(201)
+    expect(dispatch.resultSummary.endsWith('…')).toBe(true)
+  })
+
+  it('rejects undefined, JSON-throwing, and JSON-unrepresentable binding arguments BEFORE dispatch', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const calls = registerEcho(ctx)
+    const { agent, events } = fakeAgent()
+    runtime.behavior = async (request) => {
+      const echo = request.bindings[0]!.functions.echo!
+      const catchMessage = (promise: Promise<unknown>) => promise.then(() => 'resolved', (error: unknown) => error instanceof Error ? error.message : String(error))
+      return {
+        logs: [],
+        value: [
+          // Root undefined must reject up front: the event log rejects it as
+          // data, and nothing may execute unlogged.
+          await catchMessage(echo(undefined)),
+          // A toJSON that throws a NON-Error propagates out of JSON.stringify.
+          await catchMessage(echo({ toJSON() { throw 'raw-throw' } })),
+          // A bare function is a value JSON cannot represent at all.
+          await catchMessage(echo(() => 1)),
+        ].join(' | '),
+      }
+    }
+    const result = await runCode(ctx, 'program', { agent })
+    const text = (result.content[0] as { text: string }).text
+    expect(text).toContain('call the tool with an arguments object')
+    expect(text).toContain('JSON-serializable: raw-throw')
+    expect(text).toContain('a value JSON cannot represent')
+    // None of the three dispatched, none logged.
+    expect(calls).toEqual([])
+    expect(events.filter(event => event.type === 'tool/code-dispatch')).toEqual([])
+  })
+
+  it('logs the value the tool RECEIVED even when the tool mutates its arguments', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const { agent, events } = fakeAgent()
+    ctx.tools.register(defineTool({
+      name: 'mutator',
+      description: 'Mutates its own args object.',
+      parameters: { list: { type: 'array', required: true } },
+      execute(args) {
+        args.list.push('injected-by-tool')
+        return Promise.resolve([{ type: 'text' as const, text: 'mutated' }])
+      },
+    }))
+    runtime.behavior = async (request) => {
+      await request.bindings[0]!.functions.mutator!({ list: ['original'] })
+      return { logs: [] }
+    }
+    const result = await runCode(ctx, 'program', { agent })
+    expect(result.isError).toBe(false)
+    const dispatch = events.find(event => event.type === 'tool/code-dispatch')?.data as SessionEventMap['tool/code-dispatch']
+    expect(dispatch.arguments).toEqual({ list: ['original'] })
+  })
+
+  it('exposes a tool named __proto__ as an ordinary own binding', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    ctx.tools.register(defineTool({
+      name: '__proto__',
+      description: 'A prototype-colliding tool name.',
+      parameters: {},
+      execute() { return Promise.resolve([{ type: 'text' as const, text: 'proto-tool-ok' }]) },
+    }))
+    runtime.behavior = async (request) => {
+      const functions = request.bindings[0]!.functions
+      expect(Object.getPrototypeOf(functions)).toBeNull()
+      const value = await functions['__proto__']!({})
+      return { logs: [], value }
+    }
+    const result = await runCode(ctx, 'program')
+    expect(result.isError).toBe(false)
+    expect(result.content[0]).toEqual({ type: 'text', text: 'proto-tool-ok' })
+  })
+
+  it('renders a non-string completion value inspect-style', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    runtime.behavior = () => Promise.resolve({ logs: [], value: { n: 42 } })
+    const result = await runCode(ctx, 'program')
+    expect((result.content[0] as { text: string }).text).toBe('{ n: 42 }')
+  })
+
+  it('reports a pre-aborted outer signal as the run failure without dispatching anything', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const calls = registerEcho(ctx)
+    runtime.behavior = (request) => {
+      // The fake honors the seam contract for an already-aborted signal.
+      if (request.signal?.aborted) return Promise.resolve({ logs: [], error: { kind: 'abort' as const, message: String(request.signal.reason) } })
+      return Promise.resolve({ logs: [], value: 'unreachable' })
+    }
+    const controller = new AbortController()
+    controller.abort('too-late')
+    const result = await runCode(ctx, 'program', { signal: controller.signal })
+    expect(result.isError).toBe(true)
+    expect((result.content[0] as { text: string }).text).toContain('code run failed (abort)')
+    expect(calls).toEqual([])
+  })
+
+  it('rejects a binding invoked after the run is over without dispatching it', async () => {
+    const { ctx, runtime } = await setup({ mode: 'code' })
+    const calls = registerEcho(ctx)
+    const controller = new AbortController()
+    runtime.behavior = async (request) => {
+      controller.abort('cancelled-mid-run')
+      const message = await request.bindings[0]!.functions.echo!({ value: 'x' })
+        .then(() => 'resolved', (error: unknown) => error instanceof Error ? error.message : String(error))
+      return { logs: [], value: message }
+    }
+    const result = await runCode(ctx, 'program', { signal: controller.signal })
+    expect(result.isError).toBe(false)
+    expect((result.content[0] as { text: string }).text).toContain('not dispatched')
+    expect(calls).toEqual([])
+  })
+
+  it('a tool/code-dispatch event never derives a model message', () => {
+    const session = new Session(SessionId('code-mode-derive'))
+    session.append('user/message', { content: [{ type: 'text', text: 'hi' }], source: { kind: 'user' } }, { surfaceOp: 'append' })
+    session.append('tool/code-dispatch', {
+      parentCallId: CallId('p1'),
+      subCallId: CallId('p1:code:1'),
+      name: 'echo',
+      arguments: { value: 'x' },
+      isError: false,
+      resultSummary: 'echo:x',
+    })
+    const derived = session.deriveMessages()
+    expect(derived).toHaveLength(1)
+    expect(derived[0]?.role).toBe('user')
+  })
+
+  it('defaults to native mode under direct construction with no config', async () => {
+    const ctx = new Context()
+    await ctx.plugin(SystemPrompt, {})
+    const registry = new ToolRegistry(ctx)
+    expect(registry.get(RUN_CODE_NAME)).toBeUndefined()
+    const assembly = await ctx.systemPrompt.assemble()
+    expect(assembly.sections.some(section => section.name === 'tools:sdk')).toBe(false)
+  })
+})

packages/core/tools/tests/gen-tool-catalog.spec.ts

@@ -35,7 +35,7 @@ describe('gen-tool-catalog collectToolCatalog', () => {
   it('boots every shipped tool package and harvests its model-facing schemas', async () => {
     const catalog = await collectToolCatalog()
     const names = catalog.flatMap(entry => entry.schemas.map(s => s.name)).sort()
-    expect(names).toEqual(['ask_user_question', 'bash', 'bash_kill', 'bash_output', 'cordis_inspect', 'cordis_mount', 'cordis_unmount', 'edit', 'read', 'subagent', 'todo_write', 'web_fetch', 'web_search', 'write'])
+    expect(names).toEqual(['ask_user_question', 'bash', 'bash_kill', 'bash_output', 'cordis_inspect', 'cordis_mount', 'cordis_unmount', 'edit', 'read', 'run_code', 'subagent', 'todo_write', 'web_fetch', 'web_search', 'workflow', 'write'])
     // Every tool carries a JSON-Schema `parameters` object (what the model sees).
     for (const entry of catalog) {
       for (const schema of entry.schemas) {

packages/core/tools/tests/ts-types.spec.ts

@@ -0,0 +1,124 @@
+import { describe, expect, it } from 'vitest'
+import { jsonSchemaToTs, renderToolsSdk } from '@deepseek-ai/dsh-tools/src/ts-types.ts'
+import { schemaSpecToJsonSchema } from '@deepseek-ai/dsh-tools'
+import type { ToolSchema } from '@deepseek-ai/dsh-llm'
+
+describe('jsonSchemaToTs', () => {
+  it('maps the defineTool DSL subset', () => {
+    const cases: [unknown, string][] = [
+      [{ type: 'string' }, 'string'],
+      [{ type: 'number' }, 'number'],
+      [{ type: 'boolean' }, 'boolean'],
+      [{ type: 'string', enum: ['a', 'b'] }, '"a" | "b"'],
+      [{ type: 'array', items: { type: 'number' } }, 'number[]'],
+      [{ type: 'array', items: { type: 'string', enum: ['x', 'y'] } }, '("x" | "y")[]'],
+      [{ type: 'array' }, 'unknown[]'],
+      [{ type: 'object' }, 'Record<string, unknown>'],
+      [{ type: 'object', properties: {} }, 'Record<string, unknown>'],
+    ]
+    for (const [schema, expected] of cases) {
+      expect(jsonSchemaToTs(schema), JSON.stringify(schema)).toBe(expected)
+    }
+  })
+
+  it('renders objects with required/optional keys, nested shapes, and per-property docs', () => {
+    const schema = schemaSpecToJsonSchema({
+      path: { type: 'string', required: true, description: 'Absolute file path' },
+      limit: { type: 'number' },
+      opts: {
+        type: 'object',
+        properties: { deep: { type: 'boolean', required: true } },
+      },
+    })
+    expect(jsonSchemaToTs(schema)).toBe([
+      '{',
+      '  /** Absolute file path */',
+      '  path: string;',
+      '  limit?: number;',
+      '  opts?: {',
+      '    deep: boolean;',
+      '  };',
+      '}',
+    ].join('\n'))
+  })
+
+  it('is total: unsupported or hostile constructs degrade to unknown, never throw', () => {
+    const cases: unknown[] = [
+      undefined,
+      null,
+      42,
+      'string-schema',
+      {},
+      { type: 'integer' },
+      { type: 'null' },
+      { oneOf: [{ type: 'string' }] },
+      { $ref: '#/defs/x' },
+      { type: 'object', properties: 7 },
+      { type: 'object', properties: { bad: { $ref: 'x' } } },
+      { type: 'string', enum: [1, 2] },
+      { type: 'string', enum: [] },
+    ]
+    for (const schema of cases) {
+      expect(() => jsonSchemaToTs(schema), JSON.stringify(schema)).not.toThrow()
+    }
+    expect(jsonSchemaToTs({ type: 'integer' })).toBe('unknown')
+    expect(jsonSchemaToTs({ oneOf: [] })).toBe('unknown')
+    expect(jsonSchemaToTs({ type: 'object', properties: 7 })).toBe('Record<string, unknown>')
+    expect(jsonSchemaToTs({ type: 'object', properties: { bad: { $ref: 'x' } }, required: ['bad'] })).toContain('bad: unknown;')
+    // A non-string-only enum degrades to plain string; an empty one too.
+    expect(jsonSchemaToTs({ type: 'string', enum: [1, 2] })).toBe('string')
+    expect(jsonSchemaToTs({ type: 'string', enum: [] })).toBe('string')
+    // A hostile required list only accepts string members.
+    expect(jsonSchemaToTs({ type: 'object', properties: { a: { type: 'string' } }, required: [7] })).toContain('a?: string;')
+    // A property VALUE that is not an object degrades to unknown (and can
+    // carry no description).
+    expect(jsonSchemaToTs({ type: 'object', properties: { weird: 42 } })).toContain('weird?: unknown;')
+  })
+
+  it('escapes a comment-closer inside a description so the generated JSDoc cannot end early', () => {
+    const rendered = jsonSchemaToTs({
+      type: 'object',
+      properties: { glob: { type: 'string', description: 'a pattern like packages/*/tool-*/ over here' } },
+    })
+    expect(rendered).not.toContain('tool-*/ over')
+    expect(rendered).toContain(String.raw`tool-*\/ over`)
+  })
+})
+
+describe('renderToolsSdk', () => {
+  const bash: ToolSchema = {
+    name: 'bash',
+    description: 'Run a shell command.',
+    parameters: schemaSpecToJsonSchema({ command: { type: 'string', required: true } }) as unknown as Record<string, unknown>,
+  }
+  const exotic: ToolSchema = {
+    name: 'my-mcp.tool',
+    description: 'Exotic name.',
+    parameters: schemaSpecToJsonSchema({}) as unknown as Record<string, unknown>,
+  }
+
+  it('declares every tool in lexicographic order with quoted keys for exotic names', () => {
+    const text = renderToolsSdk([exotic, bash])
+    expect(text).toContain('declare const tools: {')
+    expect(text.indexOf('bash(args:')).toBeGreaterThan(0)
+    expect(text).toContain('"my-mcp.tool"(args:')
+    expect(text.indexOf('bash(args:')).toBeLessThan(text.indexOf('"my-mcp.tool"(args:'))
+    expect(text).toContain('): Promise<string>;')
+    expect(text).toContain('/** Run a shell command. */')
+    // The fixed instruction lines the model relies on.
+    expect(text).toContain('erasable syntax only')
+    expect(text).toContain('rejects with an `Error`')
+    expect(text).toContain('sequentially, even under `Promise.all`')
+    expect(text).toContain('JSON-serializable')
+  })
+
+  it('is deterministic: same tool set, byte-identical text regardless of input order', () => {
+    expect(renderToolsSdk([bash, exotic])).toBe(renderToolsSdk([exotic, bash]))
+    // Equal names sort stably (the comparator's equal arm).
+    expect(renderToolsSdk([bash, bash])).toBe(renderToolsSdk([bash, bash]))
+  })
+
+  it('renders an empty declaration for an empty tool set', () => {
+    expect(renderToolsSdk([])).toContain('declare const tools: {}')
+  })
+})

packages/core/tools/tsconfig.json

@@ -8,6 +8,12 @@
     "src"
   ],
   "references": [
+    {
+      "path": "../../core/session"
+    },
+    {
+      "path": "../../code-runtime/code-runtime"
+    },
     {
       "path": "../../../vendor/cosmokit"
     },

packages/hooks/hooks-claude/README.md

@@ -37,7 +37,7 @@ The hooks **themselves** run in the agent's session workspace: for the agent-sco
 | `SessionStart` | `agent/session-start` (emit) | additionalContext → `agent.inject()` into the new session (cannot block) |
 | `UserPromptSubmit` | `agent/prompt-submit` (waterfall) | `deny` → `PromptDecision.block`; additionalContext-only → delegate via `next()` then fold context onto the downstream decision (a later listener can still block/rewrite) |
 | `PreToolUse` | `tools/pre-execute` (waterfall) | `deny` → `PreToolDecision.deny`; `ask` → `PreToolDecision.ask` |
-| `PostToolUse` | `tools/post-execute` (waterfall) | `deny` → `block` with feedback; additionalContext-only → delegate via `next()` then fold context onto the downstream decision |
+| `PostToolUse` | `tools/post-execute` (waterfall) | `deny` → `block` with feedback; additionalContext-only → delegate via `next()` then fold context onto the downstream decision (a Code Mode sub-call’s context is dropped by the run_code bridge — see [the pipeline doc](../../../docs/tool-execution-pipeline.md)) |
 | `Stop` | `agent/turn-continuation` (waterfall) | a blocking Stop hook forces `continue`, feeding its reason as next-step steering |
 | `SubagentStart` | `subagent/start` (emit) | additionalContext → `agent.inject()` into the live child |
 | `SubagentStop` | `subagent/end` (emit) | observe-only |

packages/hooks/hooks-codex/README.md

@@ -43,7 +43,7 @@ The hooks themselves run in the agent's session workspace: for the agent-scoped
 | `SessionStart` | `agent/session-start` (emit) | a plain-stdout hook's output → additionalContext → `agent.inject()` |
 | `UserPromptSubmit` | `agent/prompt-submit` (waterfall) | `block` (exit 2) → `PromptDecision.block`; additionalContext-only → delegate via `next()` then fold context onto the downstream decision |
 | `PreToolUse` | `tools/pre-execute` (waterfall) | `block` → `PreToolDecision.deny` (no `allow`/`ask`) |
-| `PostToolUse` | `tools/post-execute` (waterfall) | `block` → `block` with feedback; additionalContext-only → delegate via `next()` then fold context onto the downstream decision |
+| `PostToolUse` | `tools/post-execute` (waterfall) | `block` → `block` with feedback; additionalContext-only → delegate via `next()` then fold context onto the downstream decision (a Code Mode sub-call’s context is dropped by the run_code bridge — see [the pipeline doc](../../../docs/tool-execution-pipeline.md)) |
 | `Stop` | `agent/turn-continuation` (waterfall) | a blocking Stop hook forces `continue` with the reason as next-step steering |
 
 A tool call's payload carries the real `tool_name` (the same value the matcher tests) and Codex's `tool_input: { command }` shape (the `command` arg when present, else `''`). The matcher subject is the tool name (`PreToolUse`/`PostToolUse`) or the session source (`SessionStart`); `UserPromptSubmit`/`Stop` ignore matchers.

packages/subagent/README.md

@@ -8,9 +8,10 @@ The subagent seam: an agent delegating work to a child agent. Like the [bash](..
 | `subagent-inprocess/` | Shared in-process run driver (pure lib; registers nothing) | — |
 | `subagent-spawn/` | In-process backend: a fresh child agent | (registers on `ctx.subagents`) |
 | `subagent-fork/` | In-process backend: a child seeded with the parent's completed-turn prefix | (registers on `ctx.subagents`) |
+| `subagent-subprocess/` | Shared out-of-process machinery: env scrub, dispose ladder, isolated config dirs (pure lib; registers nothing) | — |
 | `subagent-acp/` | Out-of-process backend: a child agent in a spawned subprocess, driven over ACP | (registers on `ctx.subagents`) |
 | `tool-subagent/` | Model-facing `subagent` delegation tool over `ctx.subagents` | (registers on `ctx.tools`) |
 
-The interface lives at `subagent/subagent/`. The in-process `subagent-spawn` / `subagent-fork` backends share the `subagent-inprocess` driver (a pure library — both depend on it, neither on the other), and the out-of-process `subagent-acp` backend ships alongside them here; the test-only `dsh-subagent-mock` (in [support](../support/README.md)) is separate. All **product** packages except the mock.
+The interface lives at `subagent/subagent/`. The in-process `subagent-spawn` / `subagent-fork` backends share the `subagent-inprocess` driver (a pure library — both depend on it, neither on the other), the out-of-process `subagent-acp` backend builds on the `subagent-subprocess` library (the credential env scrub, the dispose ladder, isolated config dirs) and ships alongside them here; the test-only `dsh-subagent-mock` (in [support](../support/README.md)) is separate. All **product** packages except the mock.
 
 The proposal and design rationale: [docs/rfc/implemented/feature/2026-06-21-subagent-capability-seam.md](../../docs/rfc/implemented/feature/2026-06-21-subagent-capability-seam.md).

packages/subagent/subagent-acp/README.md

@@ -57,7 +57,7 @@ A spawn/transport/RPC failure resolves `error` (or `aborted` if a cancel was req
 
 ## Environment scrub
 
-Credential-shaped ambient vars (`/KEY|SECRET|TOKEN/i`) are NOT forwarded to the child by default — the parent harness's own secrets must not leak into a spawned process implicitly. The child's OWN credentials are supplied explicitly via `config.env`, layered AFTER the scrub, so an intended `DEEPSEEK_API_KEY` survives while an incidental `AWS_SECRET_ACCESS_KEY` does not.
+The child env is built by [`buildChildEnv` from `@deepseek-ai/dsh-subagent-subprocess`](../subagent-subprocess/README.md) — the ambient env minus credential-shaped vars, with `config.env` layered on top after the scrub; the pattern and full semantics live there. For this backend that means the parent harness's own secrets never leak into the spawned agent implicitly, while the child's OWN `DEEPSEEK_API_KEY` is supplied deliberately via `config.env` and survives.
 
 ## Testing
 

packages/subagent/subagent-acp/package.json

@@ -25,6 +25,7 @@
     "@deepseek-ai/dsh-agent": "^0.0.1",
     "@deepseek-ai/dsh-llm": "^0.0.1",
     "@deepseek-ai/dsh-subagent": "^0.0.1",
+    "@deepseek-ai/dsh-subagent-subprocess": "^0.0.1",
     "cordis": "^4.0.0-rc.6"
   },
   "dependencies": {
@@ -35,6 +36,7 @@
     "@deepseek-ai/dsh-agent": "workspace:^",
     "@deepseek-ai/dsh-llm": "workspace:^",
     "@deepseek-ai/dsh-subagent": "workspace:^",
+    "@deepseek-ai/dsh-subagent-subprocess": "workspace:^",
     "@cordisjs/plugin-loader": "^1.0.0-rc.4",
     "cordis": "^4.0.0-rc.6"
   }

packages/subagent/subagent-acp/src/run.ts

@@ -22,7 +22,7 @@
  * @module @deepseek-ai/dsh-subagent-acp/run
  */
 
-import { spawn, type ChildProcess } from 'node:child_process'
+import { spawn } from 'node:child_process'
 import { randomUUID } from 'node:crypto'
 import { Readable, Writable } from 'node:stream'
 import {
@@ -40,6 +40,7 @@ import {
 import { AgentId } from '@deepseek-ai/dsh-agent'
 import type { ContentBlock } from '@deepseek-ai/dsh-llm'
 import type { SubagentResult, SubagentRun, SubagentStartRequest, SubagentStopReason } from '@deepseek-ai/dsh-subagent'
+import { buildChildEnv, disposeChildProcess, spawnFailure } from '@deepseek-ai/dsh-subagent-subprocess'
 
 /**
  * How the client answers a child's `session/request_permission`. The first cut
@@ -110,31 +111,6 @@ export const DEFAULT_DISPOSE_EOF_GRACE_MS = 6_000
 /** Default grace between SIGTERM and SIGKILL on dispose (the `disposeGraceMs` config; mirrors the bash executor). */
 export const DEFAULT_DISPOSE_GRACE_MS = 3_000
 
-/**
- * Credential-shaped ambient env vars are NOT forwarded to the child by default
- * (the parent harness's own `DEEPSEEK_API_KEY`/secrets must not leak into a
- * spawned process implicitly). Same pattern as the bash executor. The child
- * agent needs its OWN credentials to reach a model — those are supplied
- * explicitly via {@link AcpRunSpec.env}, which is layered on top AFTER the
- * scrub, so an intended `DEEPSEEK_API_KEY` survives while an incidental
- * `AWS_SECRET_ACCESS_KEY` does not.
- */
-export const SENSITIVE_ENV_PATTERN = /KEY|SECRET|TOKEN/i
-
-/**
- * The ambient env minus credential-shaped vars, plus the spec's explicit env.
- * @param extra - explicit vars layered on top AFTER the scrub, so a
- * credential-shaped name supplied deliberately still reaches the child.
- * @returns the environment to spawn the child with.
- */
-export function buildChildEnv(extra: Record<string, string>): NodeJS.ProcessEnv {
-  const env: NodeJS.ProcessEnv = {}
-  for (const [key, value] of Object.entries(process.env)) {
-    if (!SENSITIVE_ENV_PATTERN.test(key)) env[key] = value
-  }
-  return { ...env, ...extra }
-}
-
 /**
  * Map an ACP {@link StopReason} to a harness {@link SubagentStopReason}.
  * @param reason - the terminal reason from the child's `session/prompt` response.
@@ -196,24 +172,6 @@ function toError(value: unknown): Error {
   return value instanceof Error ? value : new Error(String(value))
 }
 
-/** Resolve once the child process exits (any code/signal); immediate if gone. */
-function waitForExit(child: ChildProcess): Promise<void> {
-  // Already-exited fast path: dispose guards on exitCode before calling, so in
-  // tests the child is always still alive here.
-  /* v8 ignore next */
-  if (child.exitCode !== null || child.signalCode !== null) return Promise.resolve()
-  return new Promise<void>(resolve => child.once('exit', () => { resolve() }))
-}
-
-/** Resolve `true` if the child exits within `ms`, `false` on timeout. */
-function exitsWithin(child: ChildProcess, ms: number): Promise<boolean> {
-  return Promise.race([
-    waitForExit(child).then(() => true),
-    // `.unref()` so a pending grace timer never keeps the parent's loop alive.
-    new Promise<boolean>(resolve => setTimeout(() => { resolve(false) }, ms).unref()),
-  ])
-}
-
 /**
  * Start an out-of-process ACP child for `request` and return a {@link SubagentRun}.
  *
@@ -254,13 +212,11 @@ export function startAcpRun(request: SubagentStartRequest, spec: AcpRunSpec): Su
     env: buildChildEnv(spec.env),
     stdio: ['pipe', 'pipe', 'inherit'],
   })
-  // A spawn-level failure (e.g. ENOENT for a bad command) is emitted as an
-  // `error` event, NOT a thrown exception — without a listener Node treats it as
-  // an unhandled error and crashes the parent. Capture it into a promise the
-  // result path races, so a bad command settles `error` like any child failure.
-  const spawnFailed = new Promise<Error>((resolve) => {
-    child.once('error', (err) => { resolve(err) })
-  })
+  // Same-tick capture (the library's contract): a spawn-level failure (e.g.
+  // ENOENT for a bad command) is an `error` EVENT that would crash the parent
+  // unheard; the result path races this promise, so a bad command settles
+  // `error` like any child failure.
+  const spawnFailed = spawnFailure(child)
 
   // Accumulate the child's streamed assistant text — the SubagentResult output.
   const output: string[] = []
@@ -393,33 +349,19 @@ export function startAcpRun(request: SubagentStartRequest, spec: AcpRunSpec): Su
     },
     async dispose(): Promise<void> {
       request.signal?.removeEventListener('abort', onAbort)
-      // Reach quiescence, not merely request it (dispose must AWAIT the child
-      // actually stopping). If the child is already gone, nothing to do.
-      if (child.exitCode !== null || child.signalCode !== null) return
-      const eofGraceMs = spec.disposeEofGraceMs
-      const graceMs = spec.disposeGraceMs
-      // 1. Graceful: end the ACP request stream (stdin EOF) and let the child
-      //    quiesce ON ITS OWN. Our acp-agent has NO SIGTERM handler in a normal
-      //    session — it tears down via the server bridge's connection-close path
-      //    (conn.closed → per-agent dispose → final session/flush), driven by the
-      //    stdin EOF, NOT by a signal. A prompt response can resolve from a
-      //    turn/end BEFORE that post-turn flush lands, so the child still has
-      //    durable work owed when dispose runs. Give the EOF-driven quiesce a real
-      //    window — wider than a single signal-grace, since the child's own
-      //    teardown may itself be awaiting a signal-trapping grandchild (a bash
-      //    subprocess in its own SIGTERM→SIGKILL grace) plus a flush — and only
-      //    escalate if it overruns. Sending SIGTERM in the same tick (or too soon)
-      //    would default-terminate the child mid-flush, orphaning its nested work.
-      child.stdin.end()
-      if (await exitsWithin(child, eofGraceMs)) return
-      // 2. SIGTERM, then escalate to SIGKILL if it still does not exit within the
-      //    grace period — a child that ignores EOF and traps SIGTERM must not
-      //    wedge dispose forever (the seam requires bounded quiescence).
-      child.kill('SIGTERM')
-      if (await exitsWithin(child, graceMs)) return
-      // 3. Force-kill and await the (now-certain) exit.
-      child.kill('SIGKILL')
-      await waitForExit(child)
+      // Quiescent teardown via the shared ladder (stdin EOF → SIGTERM →
+      // SIGKILL, awaiting the actual exit). For THIS child the EOF tier is the
+      // one that matters: our acp-agent has NO SIGTERM handler in a normal
+      // session — it tears down via the server bridge's connection-close path
+      // (conn.closed → per-agent dispose → final session/flush), driven by the
+      // stdin EOF, NOT by a signal — and a prompt response can resolve from a
+      // turn/end BEFORE that post-turn flush lands, so the child still has
+      // durable work owed when dispose runs (hence the wide EOF grace; see
+      // DEFAULT_DISPOSE_EOF_GRACE_MS).
+      await disposeChildProcess(child, {
+        disposeEofGraceMs: spec.disposeEofGraceMs,
+        disposeGraceMs: spec.disposeGraceMs,
+      })
     },
   }
 }

packages/subagent/subagent-acp/tests/subagent-acp.spec.ts

@@ -6,9 +6,10 @@ import { tmpdir } from 'node:os'
 import { join } from 'node:path'
 import { fileURLToPath } from 'node:url'
 import SubagentService from '@deepseek-ai/dsh-subagent'
+import { buildChildEnv, SENSITIVE_ENV_PATTERN } from '@deepseek-ai/dsh-subagent-subprocess'
 import type { Agent } from '@deepseek-ai/dsh-agent'
 import * as acp from '../src/index.ts'
-import { acpStopReason, acpContentText, buildChildEnv, DEFAULT_DISPOSE_EOF_GRACE_MS, DEFAULT_DISPOSE_GRACE_MS, SENSITIVE_ENV_PATTERN, startAcpRun, toAcpPrompt, type AcpRunSpec } from '../src/run.ts'
+import { acpStopReason, acpContentText, DEFAULT_DISPOSE_EOF_GRACE_MS, DEFAULT_DISPOSE_GRACE_MS, startAcpRun, toAcpPrompt, type AcpRunSpec } from '../src/run.ts'
 
 /**
  * Keyless integration tests for the ACP subagent backend. Each spawns a REAL

packages/subagent/subagent-acp/tsconfig.json

@@ -25,6 +25,9 @@
     },
     {
       "path": "../subagent"
+    },
+    {
+      "path": "../subagent-subprocess"
     }
   ]
 }

packages/subagent/subagent-subprocess/README.md

@@ -0,0 +1,40 @@
+# @deepseek-ai/dsh-subagent-subprocess
+
+Shared machinery for **out-of-process subagent backends** — providers that spawn an external agent as a child process, such as the [ACP backend](../subagent-acp/README.md). A pure library (no provider, no registration, no Config): what every spawn-a-CLI-child backend needs to keep the parent deployment's credentials out of the child, tear the child down to quiescence, and isolate it from the host user's on-disk CLI state. Design rationale: [the Claude Code / Codex subagent backends RFC](../../../docs/rfc/proposed/feature/2026-07-07-claude-code-and-codex-subagent-backends.md).
+
+Every tunable is a **parameter**: the dispose ladder takes its grace periods per call, the config-dir helper takes an optional pinned path. Defaults live in each consuming plugin's Config (defaulted, validated fields changeable from `cordis.yml`), never in this library.
+
+## What it exports
+
+### `SENSITIVE_ENV_PATTERN` / `buildChildEnv(extra)`
+
+The credential env scrub (same pattern as the [bash executor](../../bash/bash-local/README.md)): the child env is the ambient env minus credential-shaped vars (`/KEY|SECRET|TOKEN/i`), with `extra` layered on top AFTER the scrub. `PATH`, `HOME`, `TMPDIR`, locale, and proxy vars survive, so the child CLI runs normally; the parent's own secrets never leak implicitly, while an explicitly supplied credential (the child's OWN key in a backend's `env` config) still reaches the child.
+
+### `spawnFailure(child)`
+
+Spawn-failure capture: a promise that resolves (never rejects) with the child's first `error` event. A spawn failure such as `ENOENT` is an event, not a thrown exception — without a listener Node crashes the parent process — so call this in the same tick as `spawn()` and race it in the run's result path; a bad command then settles as an ordinary child-level failure. For a child that spawns cleanly the promise never settles.
+
+### `waitForExit(child)` / `exitsWithin(child, ms)`
+
+Exit waits over a `ChildProcess`: resolve once the child exits by any code or signal (immediately if it is already gone), or race that against a timer (`true` = exited in time). The race cleans up after itself on both outcomes — the pending timer is `unref()`ed and cleared on exit, the exit listener removed on timeout — so repeated calls (the dispose ladder's tiers, a poll loop) never accumulate listeners on the child.
+
+### `disposeChildProcess(child, graces)`
+
+The three-tier dispose ladder. Resolves only once the child has ACTUALLY exited — quiescence reached, not merely requested (see [defensive patterns](../../../docs/defensive-patterns.md)):
+
+1. stdin EOF (when stdin is piped), then wait `graces.disposeEofGraceMs` — a cooperative child quiesces on its own, its flushes and nested-subprocess teardown intact;
+2. `SIGTERM`, then wait `graces.disposeGraceMs`;
+3. `SIGKILL`, then await the now-certain exit — a child that ignores EOF and traps `SIGTERM` cannot wedge dispose forever.
+
+The two graces (`DisposeLadderGraces`) come from the consuming plugin's `disposeEofGraceMs`/`disposeGraceMs` Config fields; the EOF window is deliberately a separate — usually wider — grace than the signal tier, since a cooperative child's EOF teardown may itself await a signal-trapping grandchild plus a final flush.
+
+### `createIsolatedConfigDir(prefix, pinnedPath?)`
+
+A per-run isolated config directory for an external CLI child (the target of `CLAUDE_CONFIG_DIR` / `CODEX_HOME`-style redirection), so child behavior is a function of deployment config alone — never of whatever `~/.claude` / `~/.codex`-style state exists on the host. Returns an `IsolatedConfigDir` handle: `path` goes into the child env, `remove()` runs on dispose.
+
+- **Fresh (default)**: a private (0700) `mkdtemp` dir under the OS temp root; `remove()` deletes it best-effort (never rejects — a leftover temp dir beats a failed dispose) and is idempotent.
+- **Pinned** (`pinnedPath` set): the path is returned as-is — never created, never removed. A deployment that pins a directory to share child state across runs owns that directory's lifecycle.
+
+## Testing
+
+`tests/subagent-subprocess.spec.ts`: the env scrub and config-dir helpers run against the real process env and real filesystem (the rm-failure path injects its rejection at the fs boundary — a real recursive-rm failure is not portably provokable, and root ignores permission bits); the exit waits and the dispose ladder run against a scriptable fake child, driving each escalation tier deterministically. The [ACP backend suite](../subagent-acp/README.md) exercises the same ladder against real subprocesses (EOF-cooperative, EOF-ignoring, and SIGTERM-trapping children) end to end.

packages/subagent/subagent-subprocess/package.json

@@ -0,0 +1,30 @@
+{
+  "name": "@deepseek-ai/dsh-subagent-subprocess",
+  "description": "Shared out-of-process subagent machinery: credential env scrub, spawn-failure capture, child-exit waits, the EOF-to-SIGTERM-to-SIGKILL dispose ladder, and isolated config dirs (pure lib; registers nothing)",
+  "version": "0.0.1",
+  "private": true,
+  "type": "module",
+  "main": "lib/index.js",
+  "types": "lib/types/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./lib/types/index.d.ts",
+      "default": "./lib/index.js"
+    },
+    "./src/*": "./src/*",
+    "./package.json": "./package.json"
+  },
+  "files": [
+    "lib/index.js",
+    "lib/types/**/*.d.ts",
+    "lib/types/**/*.d.ts.map",
+    "src"
+  ],
+  "license": "BSD-3-Clause",
+  "peerDependencies": {
+    "cordis": "^4.0.0-rc.6"
+  },
+  "devDependencies": {
+    "cordis": "^4.0.0-rc.6"
+  }
+}

packages/subagent/subagent-subprocess/src/index.ts

@@ -0,0 +1,219 @@
+/**
+ * Shared machinery for OUT-OF-PROCESS subagent backends — providers that spawn
+ * an external agent as a child process and must keep the parent deployment's
+ * credentials out of it, tear it down to quiescence, and isolate it from the
+ * host user's on-disk CLI state. The pieces: the credential env scrub
+ * ({@link SENSITIVE_ENV_PATTERN} / {@link buildChildEnv}), the spawn-failure
+ * capture ({@link spawnFailure}), the child-exit waits ({@link waitForExit} /
+ * {@link exitsWithin}), the stdin-EOF → SIGTERM → SIGKILL dispose ladder
+ * ({@link disposeChildProcess}), and the per-run isolated config dir
+ * ({@link createIsolatedConfigDir}).
+ *
+ * This package owns no provider and registers nothing; it is a pure library
+ * the out-of-process backend packages depend on (the `subagent-inprocess`
+ * shape, for the process boundary). Every tunable — the ladder's grace
+ * periods, a pinned config dir — is a PARAMETER here: defaults belong in each
+ * consuming plugin's Config, per the no-hardcoded-tunables rule.
+ *
+ * @module @deepseek-ai/dsh-subagent-subprocess
+ */
+
+import type { ChildProcess } from 'node:child_process'
+import { mkdtemp, rm } from 'node:fs/promises'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+
+/**
+ * Credential-shaped ambient env vars are NOT forwarded to a child by default
+ * (the parent harness's own `DEEPSEEK_API_KEY`/secrets must not leak into a
+ * spawned process implicitly). Same pattern as the bash executor. The child
+ * agent needs its OWN credentials to reach a model — those are supplied
+ * explicitly via the `extra` layer of {@link buildChildEnv}, which lands AFTER
+ * the scrub, so an intended `DEEPSEEK_API_KEY` survives while an incidental
+ * `AWS_SECRET_ACCESS_KEY` does not.
+ */
+export const SENSITIVE_ENV_PATTERN = /KEY|SECRET|TOKEN/i
+
+/**
+ * The ambient env minus credential-shaped vars, plus the caller's explicit
+ * env. `PATH`, `HOME`, `TMPDIR`, locale, and proxy vars survive the scrub, so
+ * a child CLI runs normally; only {@link SENSITIVE_ENV_PATTERN}-shaped names
+ * are dropped.
+ * @param extra - explicit vars layered on top AFTER the scrub, so a
+ * credential-shaped name supplied deliberately still reaches the child.
+ * @returns the environment to spawn the child with.
+ */
+export function buildChildEnv(extra: Record<string, string>): NodeJS.ProcessEnv {
+  const env: NodeJS.ProcessEnv = {}
+  for (const [key, value] of Object.entries(process.env)) {
+    if (!SENSITIVE_ENV_PATTERN.test(key)) env[key] = value
+  }
+  return { ...env, ...extra }
+}
+
+/**
+ * Capture the child's spawn-level failure as a promise the run's result path
+ * can race. A spawn failure (e.g. `ENOENT` for a bad command) is emitted as an
+ * `error` EVENT, not a thrown exception — and without a listener Node treats
+ * it as an unhandled error and crashes the parent process. Call this in the
+ * SAME TICK as `spawn()`, so no window exists for the event to fire unheard.
+ * @param child - the just-spawned child process.
+ * @returns a promise that RESOLVES (never rejects) with the child's first
+ * `error` event; for a child that spawns cleanly it never settles.
+ */
+export function spawnFailure(child: ChildProcess): Promise<Error> {
+  return new Promise<Error>((resolve) => {
+    child.once('error', (err) => { resolve(err) })
+  })
+}
+
+/**
+ * Resolve once the child process exits (any code/signal); immediate if it is
+ * already gone.
+ * @param child - the child process to await.
+ */
+export function waitForExit(child: ChildProcess): Promise<void> {
+  if (child.exitCode !== null || child.signalCode !== null) return Promise.resolve()
+  return new Promise<void>(resolve => child.once('exit', () => { resolve() }))
+}
+
+/**
+ * Race the child's exit against a timer. Neither outcome leaves anything
+ * behind on the child: the exit listener is removed on timeout and the timer
+ * is cleared on exit, so repeated calls (the dispose ladder's tiers, a poll
+ * loop) never accumulate listeners.
+ * @param child - the child process to watch.
+ * @param ms - the wait window in milliseconds.
+ * @returns `true` if the child exits within `ms` (immediately if it is
+ * already gone), `false` on timeout.
+ */
+export function exitsWithin(child: ChildProcess, ms: number): Promise<boolean> {
+  if (child.exitCode !== null || child.signalCode !== null) return Promise.resolve(true)
+  return new Promise<boolean>((resolve) => {
+    const onExit = (): void => {
+      clearTimeout(timer)
+      resolve(true)
+    }
+    // `.unref()` so a pending grace timer never keeps the parent's loop alive.
+    const timer = setTimeout(() => {
+      child.removeListener('exit', onExit)
+      resolve(false)
+    }, ms).unref()
+    child.once('exit', onExit)
+  })
+}
+
+/**
+ * The two grace periods of the dispose ladder, supplied per call by the
+ * consuming backend — each plugin carries them as defaulted, validated
+ * `disposeEofGraceMs`/`disposeGraceMs` Config fields, so teardown timing is
+ * deployment-tunable and this library hardcodes nothing.
+ */
+export interface DisposeLadderGraces {
+  /**
+   * Tier-1 window (ms): after stdin EOF, how long the child gets to quiesce
+   * ON ITS OWN — flush durable state, tear down its own nested subprocesses —
+   * before the parent escalates to `SIGTERM`. A separate (usually WIDER)
+   * grace than {@link DisposeLadderGraces.disposeGraceMs}: a cooperative
+   * child's EOF-driven teardown may itself be waiting on a signal-trapping
+   * grandchild plus a final flush, needing more than one signal-grace of
+   * headroom.
+   */
+  disposeEofGraceMs: number
+  /** Tier-2 window (ms): between `SIGTERM` and the `SIGKILL` escalation. */
+  disposeGraceMs: number
+}
+
+/**
+ * Tear a child process down to QUIESCENCE: resolves only once the child has
+ * actually exited (or was already gone), never merely after requesting it.
+ * Three-tier escalation —
+ *
+ * 1. stdin EOF (when stdin is piped), then wait `disposeEofGraceMs`: a
+ *    cooperative child quiesces on its own, its teardown and flushes intact;
+ * 2. `SIGTERM`, then wait `disposeGraceMs`;
+ * 3. `SIGKILL`, then await the (now-certain) exit — a child that ignores EOF
+ *    and traps `SIGTERM` must not wedge dispose forever.
+ *
+ * @param child - the child process to tear down.
+ * @param graces - the two grace periods, from the consuming plugin's Config.
+ */
+export async function disposeChildProcess(child: ChildProcess, graces: DisposeLadderGraces): Promise<void> {
+  // Already gone: nothing to reap.
+  if (child.exitCode !== null || child.signalCode !== null) return
+  // 1. Graceful: end the request stream (stdin EOF) and let the child quiesce
+  //    on its own. Sending SIGTERM in the same tick (or too soon) would
+  //    default-terminate a cooperative child mid-flush, orphaning its nested
+  //    work. A child spawned without a stdin pipe skips straight to the wait.
+  child.stdin?.end()
+  if (await exitsWithin(child, graces.disposeEofGraceMs)) return
+  // 2. SIGTERM, escalating if the child still does not exit within the grace.
+  child.kill('SIGTERM')
+  if (await exitsWithin(child, graces.disposeGraceMs)) return
+  // 3. Force-kill and await the (now-certain) exit.
+  child.kill('SIGKILL')
+  await waitForExit(child)
+}
+
+/**
+ * A per-run config directory handle for an external CLI child — the target of
+ * `CLAUDE_CONFIG_DIR` / `CODEX_HOME`-style redirection. Hand {@link path} to
+ * the child's environment; call {@link remove} on dispose.
+ */
+export interface IsolatedConfigDir {
+  /** The directory to point the child at. */
+  path: string
+  /**
+   * Best-effort cleanup: removes the directory (recursively) iff this handle
+   * CREATED it — a pinned directory is never removed. Idempotent; never
+   * rejects (a leftover dir under the OS temp root is preferable to a failed
+   * dispose).
+   */
+  remove(): Promise<void>
+}
+
+/**
+ * An isolated config dir for one child run, so the child's behavior is a
+ * function of deployment config alone — never of whatever `~/.claude` /
+ * `~/.codex`-style state happens to exist on the host machine. Two modes:
+ *
+ * - no `pinnedPath` (the default): creates a FRESH private (0700) `mkdtemp`
+ *   dir under the OS temp root; {@link IsolatedConfigDir.remove} deletes it
+ *   best-effort;
+ * - `pinnedPath` set (a deployment deliberately sharing state across runs):
+ *   the pinned path is returned as-is — never created, never removed — the
+ *   deployment owns that directory's lifecycle.
+ *
+ * @param prefix - the `mkdtemp` name prefix for a fresh dir (e.g.
+ * `dsh-subagent-codex-`); ignored when `pinnedPath` is set.
+ * @param pinnedPath - a deployment-pinned directory to use instead of a
+ * fresh one.
+ * @returns the directory handle: `path` for the child env, `remove()` for
+ * dispose.
+ */
+export async function createIsolatedConfigDir(prefix: string, pinnedPath?: string): Promise<IsolatedConfigDir> {
+  if (pinnedPath !== undefined) {
+    return {
+      path: pinnedPath,
+      remove(): Promise<void> {
+        // A pinned dir is deployment-owned state (config the user asked to
+        // persist across runs); removing it here would destroy it. No-op.
+        return Promise.resolve()
+      },
+    }
+  }
+  const path = await mkdtemp(join(tmpdir(), prefix))
+  return {
+    path,
+    async remove(): Promise<void> {
+      try {
+        await rm(path, { recursive: true, force: true })
+      } catch {
+        // Best-effort by contract: swallows rm failures (EACCES/EBUSY-style —
+        // e.g. the dead child left an unreadable entry behind). The dir lives
+        // under the OS temp root, which reclaims it; failing dispose over
+        // cleanup would be worse than a leftover temp dir.
+      }
+    },
+  }
+}

packages/subagent/subagent-subprocess/tests/subagent-subprocess.spec.ts

@@ -0,0 +1,326 @@
+import { describe, expect, it, vi } from 'vitest'
+import { EventEmitter } from 'node:events'
+import { existsSync } from 'node:fs'
+import { mkdtemp, rm, stat, writeFile } from 'node:fs/promises'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import type { ChildProcess } from 'node:child_process'
+import {
+  buildChildEnv,
+  createIsolatedConfigDir,
+  disposeChildProcess,
+  exitsWithin,
+  SENSITIVE_ENV_PATTERN,
+  spawnFailure,
+  waitForExit,
+} from '../src/index.ts'
+
+// `rm` is wrapped (real-passthrough by default) so ONE test can inject a
+// rejection deterministically. A real recursive-rm failure is not portably
+// provokable — permission tricks (a chmod-000 subtree) fail only for
+// unprivileged users and are ignored by root — so this is the fs boundary
+// the testing policy sanctions mocking; everything else stays the real fs.
+vi.mock('node:fs/promises', async (importOriginal) => {
+  const actual = await importOriginal<typeof import('node:fs/promises')>()
+  return { ...actual, rm: vi.fn(actual.rm) }
+})
+
+/**
+ * Unit tests for the shared out-of-process machinery. The env scrub and the
+ * isolated-config-dir helpers run against the REAL process env and REAL
+ * filesystem (one exception: the rm-failure path injects its rejection at the
+ * mocked fs boundary, see above); the exit waits and the dispose ladder run
+ * against a scriptable fake child so each escalation tier's timing is driven
+ * deterministically (the ACP backend's suite exercises the same ladder
+ * against real subprocesses end to end).
+ */
+
+/** What fells a scripted {@link FakeChild}. */
+type LethalTrigger = 'eof' | NodeJS.Signals
+
+/** Per-scenario script for a {@link FakeChild}. */
+interface FakeChildScript {
+  /**
+   * The one trigger that makes the child exit (SIGKILL always does,
+   * uncatchable, like a real process). Omitted: only SIGKILL fells it.
+   */
+  diesOn?: LethalTrigger
+  /** Delay (ms) between the lethal trigger and the exit event. */
+  delayMs?: number
+  /** `false` models a child spawned without a stdin pipe. */
+  stdin?: boolean
+}
+
+/**
+ * A scriptable stand-in for a ChildProcess carrying exactly the surface the
+ * helpers read: `exitCode`/`signalCode`, `stdin.end()`, `kill()`, and the
+ * `exit` event.
+ */
+class FakeChild extends EventEmitter {
+  exitCode: number | null = null
+  signalCode: NodeJS.Signals | null = null
+  readonly kills: NodeJS.Signals[] = []
+  stdinEnded = false
+  readonly stdin: { end: () => void } | null
+
+  constructor(private readonly script: FakeChildScript = {}) {
+    super()
+    this.stdin = script.stdin === false
+      ? null
+      : { end: () => { this.stdinEnded = true; this.maybeDie('eof') } }
+  }
+
+  kill(signal: NodeJS.Signals): boolean {
+    this.kills.push(signal)
+    this.maybeDie(signal)
+    return true
+  }
+
+  private maybeDie(trigger: LethalTrigger): void {
+    // SIGKILL is uncatchable — it always fells the child; any other trigger
+    // only when the scenario scripts it as the lethal one.
+    if (trigger !== 'SIGKILL' && this.script.diesOn !== trigger) return
+    setTimeout(() => {
+      if (trigger === 'eof') this.exitCode = 0
+      else this.signalCode = trigger
+      this.emit('exit', this.exitCode, this.signalCode)
+    }, this.script.delayMs ?? 0)
+  }
+}
+
+/** The helpers take a real ChildProcess; the fake carries the read surface. */
+function asChild(fake: FakeChild): ChildProcess {
+  return fake as unknown as ChildProcess
+}
+
+describe('buildChildEnv / SENSITIVE_ENV_PATTERN', () => {
+  it('drops credential-shaped ambient vars (KEY/SECRET/TOKEN, case-insensitive)', () => {
+    process.env.DSH_PROC_TEST_API_KEY = 'leak'
+    process.env.dsh_proc_test_secret = 'leak'
+    process.env.DSH_PROC_TEST_TOKEN = 'leak'
+    try {
+      const env = buildChildEnv({})
+      expect(env.DSH_PROC_TEST_API_KEY).toBeUndefined()
+      expect(env.dsh_proc_test_secret).toBeUndefined()
+      expect(env.DSH_PROC_TEST_TOKEN).toBeUndefined()
+    } finally {
+      delete process.env.DSH_PROC_TEST_API_KEY
+      delete process.env.dsh_proc_test_secret
+      delete process.env.DSH_PROC_TEST_TOKEN
+    }
+  })
+
+  it('forwards normal ambient vars', () => {
+    expect(SENSITIVE_ENV_PATTERN.test('PATH')).toBe(false)
+    expect(buildChildEnv({}).PATH).toBe(process.env.PATH)
+  })
+
+  it('layers extras AFTER the scrub, so a deliberate credential-shaped name survives', () => {
+    process.env.DSH_PROC_TEST_EXTRA_TOKEN = 'ambient-leak'
+    try {
+      const env = buildChildEnv({ DSH_PROC_TEST_EXTRA_TOKEN: 'explicit' })
+      // The ambient value was scrubbed; ONLY the explicit opt-in reaches the child.
+      expect(env.DSH_PROC_TEST_EXTRA_TOKEN).toBe('explicit')
+    } finally {
+      delete process.env.DSH_PROC_TEST_EXTRA_TOKEN
+    }
+  })
+
+  it('an extra overrides the ambient value of a non-credential var', () => {
+    process.env.DSH_PROC_TEST_PLAIN = 'ambient'
+    try {
+      expect(buildChildEnv({ DSH_PROC_TEST_PLAIN: 'override' }).DSH_PROC_TEST_PLAIN).toBe('override')
+    } finally {
+      delete process.env.DSH_PROC_TEST_PLAIN
+    }
+  })
+})
+
+describe('spawnFailure', () => {
+  it('resolves (never rejects) with the first error event', async () => {
+    const fake = new FakeChild()
+    const failure = spawnFailure(asChild(fake))
+    const err = new Error('spawn ENOENT')
+    fake.emit('error', err)
+    await expect(failure).resolves.toBe(err)
+  })
+
+  it('never settles for a child that spawns cleanly and exits', async () => {
+    const fake = new FakeChild({ diesOn: 'SIGTERM' })
+    const failure = spawnFailure(asChild(fake))
+    fake.kill('SIGTERM')
+    await waitForExit(asChild(fake))
+    // A clean lifecycle emits `exit`, never `error` — the capture stays
+    // pending forever, so a race against it is decided by the other arms.
+    const settled = await Promise.race([
+      failure.then(() => 'settled'),
+      new Promise<string>(resolve => setTimeout(() => { resolve('pending') }, 30)),
+    ])
+    expect(settled).toBe('pending')
+  })
+})
+
+describe('waitForExit / exitsWithin', () => {
+  it('resolves immediately for a child that already exited by code', async () => {
+    const fake = new FakeChild()
+    fake.exitCode = 0
+    await expect(waitForExit(asChild(fake))).resolves.toBeUndefined()
+  })
+
+  it('resolves immediately for a child that already died by signal', async () => {
+    const fake = new FakeChild()
+    fake.signalCode = 'SIGTERM'
+    await expect(waitForExit(asChild(fake))).resolves.toBeUndefined()
+  })
+
+  it('resolves on the exit event of a live child', async () => {
+    const fake = new FakeChild({ diesOn: 'SIGTERM', delayMs: 5 })
+    const exited = waitForExit(asChild(fake))
+    fake.kill('SIGTERM')
+    await expect(exited).resolves.toBeUndefined()
+    expect(fake.signalCode).toBe('SIGTERM')
+  })
+
+  it('exitsWithin resolves true immediately for an already-exited child (no listener attached)', async () => {
+    const fake = new FakeChild()
+    fake.exitCode = 0
+    await expect(exitsWithin(asChild(fake), 1000)).resolves.toBe(true)
+    expect(fake.listenerCount('exit')).toBe(0)
+  })
+
+  it('exitsWithin resolves true when the child exits inside the window', async () => {
+    const fake = new FakeChild({ diesOn: 'SIGTERM', delayMs: 5 })
+    fake.kill('SIGTERM')
+    await expect(exitsWithin(asChild(fake), 1000)).resolves.toBe(true)
+    // The once-listener fired and the grace timer was cleared — nothing lingers.
+    expect(fake.listenerCount('exit')).toBe(0)
+  })
+
+  it('exitsWithin resolves false on timeout for a child that never exits', async () => {
+    const fake = new FakeChild() // nothing short of SIGKILL fells it; no signal sent
+    await expect(exitsWithin(asChild(fake), 20)).resolves.toBe(false)
+    // The timeout arm removed its exit listener: repeated waits (a poll loop,
+    // the ladder's tiers) never accumulate listeners on the same child.
+    expect(fake.listenerCount('exit')).toBe(0)
+  })
+})
+
+describe('disposeChildProcess', () => {
+  it('returns immediately for an already-exited child (no EOF, no signals)', async () => {
+    const fake = new FakeChild()
+    fake.exitCode = 0
+    await disposeChildProcess(asChild(fake), { disposeEofGraceMs: 1000, disposeGraceMs: 1000 })
+    expect(fake.stdinEnded).toBe(false)
+    expect(fake.kills).toEqual([])
+  })
+
+  it('returns immediately for a child already dead by signal', async () => {
+    const fake = new FakeChild()
+    fake.signalCode = 'SIGKILL'
+    await disposeChildProcess(asChild(fake), { disposeEofGraceMs: 1000, disposeGraceMs: 1000 })
+    expect(fake.stdinEnded).toBe(false)
+    expect(fake.kills).toEqual([])
+  })
+
+  it('tier 1: a cooperative child quiesces on stdin EOF — no signal is ever sent', async () => {
+    const fake = new FakeChild({ diesOn: 'eof', delayMs: 5 })
+    await disposeChildProcess(asChild(fake), { disposeEofGraceMs: 1000, disposeGraceMs: 1000 })
+    expect(fake.stdinEnded).toBe(true)
+    expect(fake.kills).toEqual([])
+    expect(fake.exitCode).toBe(0)
+  })
+
+  it('tier 2: a child that ignores EOF but honors SIGTERM dies on the middle rung', async () => {
+    const fake = new FakeChild({ diesOn: 'SIGTERM', delayMs: 5 })
+    await disposeChildProcess(asChild(fake), { disposeEofGraceMs: 20, disposeGraceMs: 1000 })
+    expect(fake.stdinEnded).toBe(true)
+    expect(fake.kills).toEqual(['SIGTERM'])
+    expect(fake.signalCode).toBe('SIGTERM')
+  })
+
+  it('tier 3: a SIGTERM-trapping child is SIGKILLed, and dispose resolves only after the exit', async () => {
+    const fake = new FakeChild({ delayMs: 5 }) // only SIGKILL fells it
+    await disposeChildProcess(asChild(fake), { disposeEofGraceMs: 20, disposeGraceMs: 20 })
+    expect(fake.kills).toEqual(['SIGTERM', 'SIGKILL'])
+    // Quiescence, not a request: at resolution the child has ACTUALLY exited
+    // (the exit event landed, despite the scripted post-SIGKILL delay).
+    expect(fake.signalCode).toBe('SIGKILL')
+  })
+
+  it('walks the ladder for a child spawned without a stdin pipe', async () => {
+    const fake = new FakeChild({ stdin: false, diesOn: 'SIGTERM', delayMs: 5 })
+    await disposeChildProcess(asChild(fake), { disposeEofGraceMs: 20, disposeGraceMs: 1000 })
+    expect(fake.kills).toEqual(['SIGTERM'])
+  })
+})
+
+describe('createIsolatedConfigDir', () => {
+  it('creates a fresh private mkdtemp dir under the OS temp root', async () => {
+    const dir = await createIsolatedConfigDir('dsh-subagent-subprocess-test-')
+    try {
+      expect(dir.path.startsWith(join(tmpdir(), 'dsh-subagent-subprocess-test-'))).toBe(true)
+      const st = await stat(dir.path)
+      expect(st.isDirectory()).toBe(true)
+      // Private (0700) per the defensive-patterns temp-dir rule.
+      expect(st.mode & 0o777).toBe(0o700)
+    } finally {
+      await dir.remove()
+    }
+  })
+
+  it('creates a distinct dir per call (per-run isolation)', async () => {
+    const a = await createIsolatedConfigDir('dsh-subagent-subprocess-test-')
+    const b = await createIsolatedConfigDir('dsh-subagent-subprocess-test-')
+    try {
+      expect(a.path).not.toBe(b.path)
+    } finally {
+      await a.remove()
+      await b.remove()
+    }
+  })
+
+  it('remove() deletes a fresh dir recursively and is idempotent', async () => {
+    const dir = await createIsolatedConfigDir('dsh-subagent-subprocess-test-')
+    await writeFile(join(dir.path, 'settings.json'), '{}')
+    await dir.remove()
+    expect(existsSync(dir.path)).toBe(false)
+    // Second remove: nothing left to delete, still resolves.
+    await expect(dir.remove()).resolves.toBeUndefined()
+  })
+
+  it('returns a pinned dir verbatim and NEVER removes it', async () => {
+    const pinned = await mkdtemp(join(tmpdir(), 'dsh-subagent-subprocess-pinned-'))
+    try {
+      const dir = await createIsolatedConfigDir('ignored-prefix-', pinned)
+      expect(dir.path).toBe(pinned)
+      await dir.remove()
+      // The deployment owns a pinned dir's lifecycle — remove() must not touch it.
+      expect(existsSync(pinned)).toBe(true)
+    } finally {
+      await rm(pinned, { recursive: true, force: true })
+    }
+  })
+
+  it('does not create a missing pinned path (the deployment owns its lifecycle)', async () => {
+    const missing = join(tmpdir(), `dsh-subagent-subprocess-missing-${process.pid}`)
+    const dir = await createIsolatedConfigDir('ignored-prefix-', missing)
+    expect(dir.path).toBe(missing)
+    expect(existsSync(missing)).toBe(false)
+    await dir.remove()
+    expect(existsSync(missing)).toBe(false)
+  })
+
+  it('remove() is best-effort: an rm rejection resolves instead of rejecting', async () => {
+    const dir = await createIsolatedConfigDir('dsh-subagent-subprocess-locked-')
+    try {
+      // The swallow contract is error-kind agnostic; EACCES stands in for the
+      // family (EBUSY, a vanished mount, …) that best-effort must absorb.
+      vi.mocked(rm).mockRejectedValueOnce(Object.assign(new Error('EACCES: permission denied'), { code: 'EACCES' }))
+      await expect(dir.remove()).resolves.toBeUndefined()
+      // The injected rejection consumed the only rm call — nothing was deleted.
+      expect(existsSync(dir.path)).toBe(true)
+    } finally {
+      await rm(dir.path, { recursive: true, force: true })
+    }
+  })
+})

packages/subagent/subagent-subprocess/tsconfig.json

@@ -0,0 +1,11 @@
+{
+  "extends": "../../../tsconfig.base.json",
+  "compilerOptions": {
+    "rootDir": "src",
+    "outDir": "lib/types"
+  },
+  "include": [
+    "src"
+  ],
+  "references": []
+}

packages/support/acp-snapshot/README.md

@@ -6,7 +6,7 @@ Three layers, importable separately:
 
 - **`runScenario` (harness)** — boots the real agent bin as a subprocess via tsx (unbuilt, Loader path), drives it over ACP JSON-RPC stdio from a deterministic `input.json` script, tees raw stdout for the golden + purity check, and harvests every persisted session JSONL (parent + subagent children, primary-first) after a graceful stdin-EOF shutdown. Parameterized by `AgentUnderTest` (`binScript`, `configPath`, `tsconfigPath` — absolute paths; the subprocess cwd is a temp dir outside the repo).
 - **Normalizers** — pure functions turning the two captured surfaces into stable text: `normalizeStdout` (JSON-RPC ids → first-seen sequence; UUIDs/cwd → tokens; doubles as the stdout-purity check), `normalizeSessionLog` (times zeroed, `seq` kept), and the composable `scrubRequestHeaders` (header bulk → `{{system}}`/`{{tools}}`, structure kept — [pinned-header RFC](../../../docs/rfc/implemented/testing/2026-07-06-pin-request-header-content-in-one-scenario.md)).
-- **`defineAcpSnapshotSuite` (factory)** — registers the whole describe/it tree for a scenario table: per-scenario golden + re-persisted-log compares, record-mode fixture write-back, the per-suite header pin with its live uniformity guard, and the fixture guard block (no orphan scenario dirs, required files present, exactly one pin, non-pinning fixtures header-scrubbed). Must be called at vitest collection time.
+- **`defineAcpSnapshotSuite` (factory)** — registers the whole describe/it tree for a scenario table: per-scenario golden + re-persisted-log compares, record-mode fixture write-back, the per-header-class pin with its live uniformity guard, and the fixture guard block (no orphan scenario dirs, required files present, exactly one pin per class, pinning fixtures well-formed, non-pinning fixtures header-scrubbed). Must be called at vitest collection time.
 
 A consuming `*.snapshot.ts` is the scenario table plus one factory call:
 
@@ -26,11 +26,13 @@ defineAcpSnapshotSuite({
     tsconfigPath: fileURLToPath(new URL('../../../tsconfig.json', import.meta.url)),
   },
   snapshotsDir: join(dirname(fileURLToPath(import.meta.url)), 'snapshots'),
-  scenarios: SCENARIOS, // exactly one entry sets pinsHeader
+  scenarios: SCENARIOS, // exactly one entry per header class sets pinsHeader
   mode: process.env.DSH_SNAPSHOT === 'record' ? 'record' : 'replay',
 })
 ```
 
+A scenario booting a differently-composed tree sets its own `configPath` (an overlay whose basename still ends in `cordis.yml`, so the bin's replay swap finds the sibling `*cordis.snapshot.yml`) and, when that composition changes the request header, its own `headerClass` with its own pinning scenario — the acp-agent example's Code Mode scenarios are the template.
+
 The example also ships a `cordis.snapshot.yml` replay overlay next to its `cordis.yml` (the bin swaps them under `DSH_SNAPSHOT=replay` — [single-source replay config RFC](../../../docs/rfc/implemented/testing/2026-07-04-single-source-acp-replay-config.md)); replay fixtures are served by [`dsh-llm-replay`](../llm-replay/README.md), which this package points at via the `DSH_SNAPSHOT_*` env vars it sets on the child. Fixture roles, record/replay semantics, and scenario-table fields are documented on `Scenario` and in the [snapshot RFC](../../../docs/rfc/implemented/testing/2026-06-19-acp-snapshot-tests.md).
 
 Constraints: `suite.ts` imports vitest, so the package is importable only inside a vitest run (the harness and normalizers have no such dependency but ship from the same entry). ACP-specific by design — the harness speaks the SDK's `ClientSideConnection`. Permission round-trips are scriptable: `InputScript.permissionAnswers` is a FIFO queue of option-kind selections (`allow_once`, `reject_once`, …) the client maps to the agent-issued `optionId` at answer time; an absent or exhausted queue answers `cancelled`, and a kind the request never offered rejects the run (the agent is answered `cancelled`, so a tolerant agent cannot absorb the scenario bug).

packages/support/acp-snapshot/src/harness.ts

@@ -166,6 +166,15 @@ export interface RunOptions {
    * start from an empty workspace.
    */
   workspaceDir?: string
+  /**
+   * Alternate LIVE config path for the boot (absolute), overriding
+   * {@link AgentUnderTest.configPath} for this run. A scenario needing a
+   * differently-composed tree (the Code Mode scenarios) ships an overlay
+   * whose basename still ends in `cordis.yml`, so the bin's replay swap
+   * resolves the sibling `*cordis.snapshot.yml` the same way it does for
+   * the default.
+   */
+  configPath?: string
 }
 
 /**
@@ -209,7 +218,7 @@ export async function runScenario(input: InputScript, opts: RunOptions): Promise
 
     child = spawn(
       process.execPath,
-      ['--import', tsxLoader, opts.agent.binScript, opts.agent.configPath],
+      ['--import', tsxLoader, opts.agent.binScript, opts.configPath ?? opts.agent.configPath],
       { cwd, env, stdio: ['pipe', 'pipe', 'pipe'] },
     )
 

packages/support/acp-snapshot/src/suite.ts

@@ -11,13 +11,14 @@
  * before comparing).
  *
  * Request-header content (the composed system prompt + tool schemas riding on
- * `request/header` events) is pinned by exactly ONE scenario per suite — the
- * one with `pinsHeader` — and scrubbed to `{{system}}`/`{{tools}}` tokens in
- * every other fixture and compare, so a prompt or tool-schema edit churns one
- * committed line instead of every fixture. A per-run uniformity guard keeps
- * the single pin sound: every live header must equal the pinned one, and no
- * header-delta may appear outside the pinning scenario (see the
- * pinned-header RFC,
+ * `request/header` events) is pinned by exactly ONE scenario per HEADER CLASS
+ * — scenarios that boot the same config compose the same header; each class's
+ * `pinsHeader` scenario commits it verbatim — and scrubbed to
+ * `{{system}}`/`{{tools}}` tokens in every other fixture and compare, so a
+ * prompt or tool-schema edit churns one committed line per class instead of
+ * every fixture. A per-run uniformity guard keeps each pin sound: every live
+ * header must equal its class's pinned one, and no header-delta may appear
+ * outside a pinning scenario (see the pinned-header RFC,
  * docs/rfc/implemented/testing/2026-07-06-pin-request-header-content-in-one-scenario.md).
  *
  * `pnpm run test:snapshot:record` (DSH_SNAPSHOT=record + -u) re-records the
@@ -80,18 +81,38 @@ export interface Scenario {
    * Whether THIS scenario's fixtures keep the full request-header content (the
    * composed system prompt and tool schema list on `request/header` /
    * `request/header-delta` events) and compare it verbatim. Exactly one
-   * scenario per suite pins it; every other scenario stores and compares that
-   * content as `{{system}}`/`{{tools}}` tokens ({@link scrubRequestHeaders}),
+   * scenario per HEADER CLASS ({@link headerClass}) pins it; every other
+   * scenario of that class stores and compares that content as
+   * `{{system}}`/`{{tools}}` tokens ({@link scrubRequestHeaders}),
    * so a system prompt or tool-schema change shows up as ONE committed-fixture
-   * diff, not one per scenario. One pin suffices because header composition is
-   * suite-uniform (parent, spawn child, and fork child all compose the same
-   * prompt-modulo-cwd and the same tools) — and that premise is ASSERTED, not
-   * assumed: every non-pinning run's live headers must equal the pinned
-   * fixture's (normalized), so a session-dependent header (say, a restricted
-   * subagent toolset) fails loud until it gets its own pinning scenario.
+   * diff per class, not one per scenario. One pin per class suffices because
+   * header composition is class-uniform (parent, spawn child, and fork child
+   * all compose the same prompt-modulo-cwd and the same tools) — and that
+   * premise is ASSERTED, not assumed: every non-pinning run's live headers
+   * must equal its class's pinned fixture's (normalized), so a
+   * session-dependent header (say, a restricted subagent toolset) fails loud
+   * until it gets its own pinning scenario.
    * Defaults to false.
    */
   pinsHeader?: boolean
+  /**
+   * Which header-composition class this scenario belongs to. Scenarios that
+   * boot the same config compose the same header; each class has exactly one
+   * {@link pinsHeader} scenario, and the uniformity guard compares every
+   * other member against ITS class's pin. Defaults to `'default'`; a
+   * scenario booting an alternate config ({@link configPath}) whose tool
+   * list or prompt sections differ by construction carries its own class.
+   */
+  headerClass?: string
+  /**
+   * Alternate LIVE config path (absolute) this scenario boots instead of
+   * {@link AgentUnderTest.configPath} — an overlay composing a different
+   * tree (its basename must still end in `cordis.yml` so the bin's replay
+   * swap finds the sibling `*cordis.snapshot.yml`). A scenario whose
+   * overlay changes the composed header also needs its own
+   * {@link headerClass}.
+   */
+  configPath?: string
 }
 
 /** One suite's inputs: the agent to boot, where its fixtures live, and its scenario table. */
@@ -183,10 +204,11 @@ export function headerDeltaCount(rawLog: string): number {
 /**
  * Register the suite: one `describe` per scenario (the golden/log compares and
  * the header-uniformity guard) plus the fixture guard block (no orphan
- * scenario dirs, required files present, exactly one pin, non-pinning fixtures
- * header-scrubbed). Must run at vitest collection time — it calls
- * `describe`/`it`. Throws immediately if no scenario pins the header (the
- * uniformity guard would have nothing to compare against).
+ * scenario dirs, required files present, exactly one pin per header class,
+ * pinning fixtures well-formed, non-pinning fixtures header-scrubbed). Must
+ * run at vitest collection time — it calls `describe`/`it`. Throws
+ * immediately if any header class lacks a pinning scenario or carries two
+ * (the uniformity guard needs exactly one comparison anchor per class).
  *
  * @param options The agent, snapshots directory, scenario table, and mode.
  */
@@ -194,9 +216,23 @@ export function defineAcpSnapshotSuite(options: SnapshotSuiteOptions): void {
   const { agent, snapshotsDir, scenarios, mode } = options
   const RECORDING = mode === 'record'
 
-  /** The suite's single header-pinning scenario. Guarded here (and by a meta-test) so the pin cannot silently vanish. */
-  const pinningScenario = scenarios.find(s => s.pinsHeader === true)
-  if (pinningScenario === undefined) throw new Error('acp-snapshot: no scenario pins the request-header content')
+  /** The class a scenario's header composition belongs to (see {@link Scenario.headerClass}). */
+  const classOf = (scenario: Scenario): string => scenario.headerClass ?? 'default'
+
+  /** Each header class's single pinning scenario. Guarded here (and by meta-tests) so a pin cannot silently vanish or split. */
+  const pinningByClass = new Map<string, Scenario>()
+  for (const scenario of scenarios) {
+    if (scenario.pinsHeader !== true) continue
+    const cls = classOf(scenario)
+    const existing = pinningByClass.get(cls)
+    if (existing) throw new Error(`acp-snapshot: header class "${cls}" pinned by both ${existing.name} and ${scenario.name}`)
+    pinningByClass.set(cls, scenario)
+  }
+  for (const scenario of scenarios) {
+    if (!pinningByClass.has(classOf(scenario))) {
+      throw new Error(`acp-snapshot: no scenario pins the request-header content of class "${classOf(scenario)}" (needed by ${scenario.name})`)
+    }
+  }
 
   for (const scenario of scenarios) {
     describe(`snapshot: ${scenario.name}`, () => {
@@ -217,6 +253,9 @@ export function defineAcpSnapshotSuite(options: SnapshotSuiteOptions): void {
           // replays from its own script. In RECORD they are harvested, not read.
           ...!RECORDING && childSessions > 0 ? { childFiles: childFixturePaths(dir, childSessions) } : {},
           ...existsSync(workspaceDir) ? { workspaceDir } : {},
+          // A scenario booting an overlay tree passes its own live config; the
+          // bin's replay swap derives the sibling `*cordis.snapshot.yml` from it.
+          ...scenario.configPath !== undefined ? { configPath: scenario.configPath } : {},
         })
 
         // Scrub every volatile id the run produced: the ACP server-issued session
@@ -277,19 +316,22 @@ export function defineAcpSnapshotSuite(options: SnapshotSuiteOptions): void {
           }
         }
 
-        // Header-uniformity guard: the single pin is sound only while every
-        // session in the suite composes the SAME header and keeps it for the
-        // whole run. Assert both halves live. (1) Every request/header the run
-        // produced (parent, spawn child, fork child, initial or resume) must
-        // equal the pinned fixture's header after each side is normalized
-        // against its own volatile values. (2) No request/header-delta may
-        // appear at all — a mid-run header change diverges from the pin by
-        // construction, and its content would be invisible under the scrub. If
-        // either fails, either the header changed (update the pin: re-record or
-        // hand-edit the pinning scenario's fixture) or composition became
-        // session-dependent by design (give the divergent shape its own
-        // pinning scenario).
+        // Header-uniformity guard: a class's single pin is sound only while
+        // every session in that class composes the SAME header and keeps it
+        // for the whole run. Assert both halves live. (1) Every
+        // request/header the run produced (parent, spawn child, fork child,
+        // initial or resume) must equal the CLASS's pinned fixture's header
+        // after each side is normalized against its own volatile values.
+        // (2) No request/header-delta may appear at all — a mid-run header
+        // change diverges from the pin by construction, and its content
+        // would be invisible under the scrub. If either fails, either the
+        // header changed (update the pin: re-record or hand-edit the pinning
+        // scenario's fixture) or composition became session-dependent by
+        // design (give the divergent shape its own pinning scenario and
+        // class).
         if (scenario.pinsHeader !== true) {
+          /* v8 ignore next -- construction guarantees the pin exists; a miss would fail the one-header assertion loudly. */
+          const pinningScenario = pinningByClass.get(classOf(scenario)) ?? scenario
           const pinnedFixture = await readFile(join(snapshotsDir, pinningScenario.name, 'session.jsonl'), 'utf8')
           const pinned = normalizedHeaders(pinnedFixture, fixtureContext(pinnedFixture))
           expect(pinned.length, `the pinning fixture (${pinningScenario.name}) must carry exactly one request/header`)
@@ -347,11 +389,35 @@ export function defineAcpSnapshotSuite(options: SnapshotSuiteOptions): void {
       }
     })
 
-    it('exactly one scenario pins the request-header content', () => {
-      // Zero pins would drop the prompt/schema surface from the suite entirely;
-      // two would split it. One pin per suite is the design (pinned-header RFC);
-      // WHICH scenario pins is the scenario table's reviewable choice.
-      expect(scenarios.filter(s => s.pinsHeader === true).map(s => s.name)).toEqual([pinningScenario.name])
+    it('exactly one scenario pins the request-header content of each header class', () => {
+      // Zero pins would drop a class's prompt/schema surface from the suite
+      // entirely; two would split it. One pin per class is the design
+      // (pinned-header RFC); WHICH scenario pins is the scenario table's
+      // reviewable choice.
+      const pins = new Map<string, string[]>()
+      for (const scenario of scenarios.filter(s => s.pinsHeader === true)) {
+        const cls = classOf(scenario)
+        pins.set(cls, [...pins.get(cls) ?? [], scenario.name])
+      }
+      expect(Object.fromEntries([...pins].map(([cls, names]) => [cls, names.length]))).toEqual(
+        Object.fromEntries([...pinningByClass.keys()].map(cls => [cls, 1])))
+      for (const scenario of scenarios) {
+        expect(pinningByClass.has(classOf(scenario)), `class "${classOf(scenario)}" (scenario ${scenario.name}) has a pin`).toBe(true)
+      }
+    })
+
+    it('every pinning fixture carries exactly one request/header and no deltas', async () => {
+      // The live uniformity guard runs only in NON-pinning scenarios, so a
+      // class made of just its pinning scenario would otherwise accept a
+      // re-recorded pin with several headers or a mid-run header-delta —
+      // shapes the pin design cannot represent. Assert the committed pins
+      // directly.
+      for (const scenario of pinningByClass.values()) {
+        const fixture = await readFile(join(snapshotsDir, scenario.name, 'session.jsonl'), 'utf8')
+        const headers = normalizedHeaders(fixture, fixtureContext(fixture))
+        expect(headers.length, `${scenario.name}: a pinning fixture must carry exactly one request/header`).toBe(1)
+        expect(headerDeltaCount(fixture), `${scenario.name}: a pinning fixture must carry no request/header-delta`).toBe(0)
+      }
     })
 
     it('committed fixtures carry request-header content ONLY in the pinning scenario', async () => {

packages/support/acp-snapshot/tests/suite.spec.ts

@@ -34,12 +34,18 @@ const AGENT = {
 const REPLAY_DIR = fileURLToPath(new URL('./fixtures/suite', import.meta.url))
 const RECORD_SRC = fileURLToPath(new URL('./fixtures/record-suite', import.meta.url))
 
+// The replay suite doubles as the header-CLASS coverage: every scenario names
+// the same explicit class (the record suite exercises the 'default' fallback),
+// and plain-turn boots through a per-scenario configPath override (the same
+// dummy path the agent default carries — the plumbing, not the composition,
+// is what this suite can exercise; the real overlay boot is the acp-agent
+// example's code-mode scenarios).
 const REPLAY_SCENARIOS: Scenario[] = [
-  { name: 'pin-turn', hasModelTurn: true, recorded: true, pinsHeader: true },
-  { name: 'plain-turn', hasModelTurn: true, recorded: true, childSessions: 1 },
-  { name: 'no-model', hasModelTurn: false, recorded: false },
-  { name: 'blocked-log', hasModelTurn: false, comparesLog: true, recorded: false },
-  { name: 'authored-error', hasModelTurn: true, recorded: false, overridden: true },
+  { name: 'pin-turn', hasModelTurn: true, recorded: true, pinsHeader: true, headerClass: 'main' },
+  { name: 'plain-turn', hasModelTurn: true, recorded: true, childSessions: 1, headerClass: 'main', configPath: AGENT.configPath },
+  { name: 'no-model', hasModelTurn: false, recorded: false, headerClass: 'main' },
+  { name: 'blocked-log', hasModelTurn: false, comparesLog: true, recorded: false, headerClass: 'main' },
+  { name: 'authored-error', hasModelTurn: true, recorded: false, overridden: true, headerClass: 'main' },
 ]
 
 const RECORD_SCENARIOS: Scenario[] = [
@@ -70,7 +76,7 @@ describe('defineAcpSnapshotSuite: record mode', () => {
 })
 
 describe('defineAcpSnapshotSuite: registration contract', () => {
-  it('throws when no scenario pins the request-header content', () => {
+  it("throws when a scenario's header class has no pinning scenario", () => {
     expect(() => {
       defineAcpSnapshotSuite({
         agent: AGENT,
@@ -78,7 +84,33 @@ describe('defineAcpSnapshotSuite: registration contract', () => {
         scenarios: [{ name: 'pinless', hasModelTurn: true, recorded: true }],
         mode: 'replay',
       })
-    }).toThrow(/no scenario pins/)
+    }).toThrow(/no scenario pins the request-header content of class "default"/)
+    // A pinned class does not cover a DIFFERENT class's members.
+    expect(() => {
+      defineAcpSnapshotSuite({
+        agent: AGENT,
+        snapshotsDir: REPLAY_DIR,
+        scenarios: [
+          { name: 'pinned', hasModelTurn: true, recorded: true, pinsHeader: true },
+          { name: 'classless-orphan', hasModelTurn: true, recorded: true, headerClass: 'other' },
+        ],
+        mode: 'replay',
+      })
+    }).toThrow(/class "other" \(needed by classless-orphan\)/)
+  })
+
+  it('throws when two scenarios pin the same header class', () => {
+    expect(() => {
+      defineAcpSnapshotSuite({
+        agent: AGENT,
+        snapshotsDir: REPLAY_DIR,
+        scenarios: [
+          { name: 'first-pin', hasModelTurn: true, recorded: true, pinsHeader: true },
+          { name: 'second-pin', hasModelTurn: true, recorded: true, pinsHeader: true },
+        ],
+        mode: 'replay',
+      })
+    }).toThrow(/header class "default" pinned by both first-pin and second-pin/)
   })
 })
 

packages/ui/acp-agent/package.json

@@ -37,6 +37,7 @@
     "@deepseek-ai/dsh-agent-core": "^0.0.1",
     "@deepseek-ai/dsh-workspace-context": "^0.0.1",
     "@deepseek-ai/dsh-session-persistence-jsonl": "^0.0.1",
+    "@deepseek-ai/dsh-tools": "^0.0.1",
     "@deepseek-ai/dsh-user-interaction": "^0.0.1",
     "cordis": "^4.0.0-rc.6",
     "schemastery": "^3.17.0"
@@ -47,6 +48,7 @@
     "@deepseek-ai/dsh-app-boot": "workspace:^",
     "@deepseek-ai/dsh-acp": "workspace:^",
     "@deepseek-ai/dsh-agent-core": "workspace:^",
+    "@deepseek-ai/dsh-tools": "workspace:^",
     "@deepseek-ai/dsh-system-prompt": "workspace:^",
     "@deepseek-ai/dsh-workspace-context": "workspace:^",
     "@deepseek-ai/dsh-session-persistence-jsonl": "workspace:^",

packages/ui/acp-agent/src/index.ts

@@ -35,6 +35,7 @@ import z from 'schemastery'
 import * as acp from '@deepseek-ai/dsh-acp'
 import * as agentCore from '@deepseek-ai/dsh-agent-core'
 import * as workspaceContext from '@deepseek-ai/dsh-workspace-context'
+import ToolRegistry, { type Config as ToolsConfig } from '@deepseek-ai/dsh-tools'
 import SessionPersistenceJsonl from '@deepseek-ai/dsh-session-persistence-jsonl'
 import UserInteractionService from '@deepseek-ai/dsh-user-interaction'
 
@@ -46,7 +47,8 @@ export const name = 'acp-agent'
  * pre-created agent — ACP creates agents at `session/new`); `persona` is the
  * deployment persona (forwarded to the system-prompt plugin); `toolOrder` is
  * the explicit model-facing tool order (forwarded to the system-prompt plugin);
- * `persistenceRoot` is the JSONL backend's directory.
+ * `tools` is the tool registry's config (its presentation `mode`, forwarded
+ * through agent-core); `persistenceRoot` is the JSONL backend's directory.
  */
 export interface Config {
   /** Model name for ACP-created agents (must have a registered adapter). */
@@ -55,6 +57,8 @@ export interface Config {
   persona?: string
   /** Explicit model-facing tool order (the system-prompt plugin's `toolOrder` config; see dsh-system-prompt). */
   toolOrder?: string[]
+  /** Tool-registry config — its presentation `mode` (forwarded through agent-core; see dsh-tools). */
+  tools?: ToolsConfig
   /** Directory the JSONL session backend writes under. Defaults to `./.sessions`. */
   persistenceRoot?: string
   /** Controls automatic AGENTS.md/CLAUDE.md loading; set `false` for hermetic prompts. */
@@ -68,6 +72,7 @@ export const Config: z<Config> = z.object({
   // order" (the owning dsh-system-prompt schema does the same), while
   // schemastery's native [] default would read as an invalid configured list.
   toolOrder: z.array(z.string()).default(undefined as unknown as string[]),
+  tools: ToolRegistry.Config,
   persistenceRoot: z.string().default('./.sessions'),
   workspaceContext: z.union([z.const(false), workspaceContext.Config]),
 }) as unknown as z<Config>
@@ -83,6 +88,7 @@ export function apply(ctx: Context, config: Config): void {
   ctx.plugin(agentCore, {
     ...config.persona !== undefined ? { persona: config.persona } : {},
     ...config.toolOrder !== undefined ? { toolOrder: config.toolOrder } : {},
+    ...config.tools !== undefined ? { tools: config.tools } : {},
     ...config.workspaceContext !== undefined ? { workspaceContext: config.workspaceContext } : {},
   })
   ctx.plugin(UserInteractionService)

packages/ui/stdio-agent/package.json

@@ -39,6 +39,7 @@
     "@deepseek-ai/dsh-agent-core": "^0.0.1",
     "@deepseek-ai/dsh-workspace-context": "^0.0.1",
     "@deepseek-ai/dsh-session": "^0.0.1",
+    "@deepseek-ai/dsh-tools": "^0.0.1",
     "@deepseek-ai/dsh-session-persistence-jsonl": "^0.0.1",
     "@deepseek-ai/dsh-tool-ask-user": "^0.0.1",
     "@deepseek-ai/dsh-user-interaction": "^0.0.1",
@@ -56,6 +57,7 @@
     "@deepseek-ai/dsh-system-prompt": "workspace:^",
     "@deepseek-ai/dsh-workspace-context": "workspace:^",
     "@deepseek-ai/dsh-session": "workspace:^",
+    "@deepseek-ai/dsh-tools": "workspace:^",
     "@deepseek-ai/dsh-session-persistence-jsonl": "workspace:^",
     "@deepseek-ai/dsh-tool-ask-user": "workspace:^",
     "@deepseek-ai/dsh-user-interaction": "workspace:^",

packages/ui/stdio-agent/src/index.ts

@@ -43,6 +43,7 @@ import ConsoleExporter from '@cordisjs/plugin-logger-console'
 import z from 'schemastery'
 import { AgentId } from '@deepseek-ai/dsh-agent'
 import { SessionId } from '@deepseek-ai/dsh-session'
+import ToolRegistry, { type Config as ToolsConfig } from '@deepseek-ai/dsh-tools'
 import * as agentCore from '@deepseek-ai/dsh-agent-core'
 import * as workspaceContext from '@deepseek-ai/dsh-workspace-context'
 import SessionPersistenceJsonl from '@deepseek-ai/dsh-session-persistence-jsonl'
@@ -67,6 +68,8 @@ export interface Config {
   persona?: string
   /** Explicit model-facing tool order (the system-prompt plugin's `toolOrder` config; see dsh-system-prompt). */
   toolOrder?: string[]
+  /** Tool-registry config — its presentation `mode` (forwarded through agent-core; see dsh-tools). */
+  tools?: ToolsConfig
   /** Directory the JSONL session backend writes under. Defaults to `./.sessions`. */
   persistenceRoot?: string
   /** stdin-chat banner printed once on start. Defaults to `'ready.'`. */
@@ -88,6 +91,7 @@ export const Config: z<Config> = z.object({
   // order" (the owning dsh-system-prompt schema does the same), while
   // schemastery's native [] default would read as an invalid configured list.
   toolOrder: z.array(z.string()).default(undefined as unknown as string[]),
+  tools: ToolRegistry.Config,
   persistenceRoot: z.string().default('./.sessions'),
   welcome: z.string().default('ready.'),
   resumeSessionId: z.string(),
@@ -106,6 +110,7 @@ export function apply(ctx: Context, config: Config): void {
   ctx.plugin(agentCore, {
     ...config.persona !== undefined ? { persona: config.persona } : {},
     ...config.toolOrder !== undefined ? { toolOrder: config.toolOrder } : {},
+    ...config.tools !== undefined ? { tools: config.tools } : {},
     agents: [{
       id: AgentId('main'),
       model: config.model,

packages/workflow/README.md

@@ -0,0 +1,13 @@
+# workflow/ — dynamic-workflow capability family
+
+The workflow seam: a model-written JavaScript orchestration script that fans out subagents at scale (phases, structured per-agent results, concurrency caps), modeled on Claude Code's dynamic workflows. A capability seam (see [capability seams](../../docs/rfc/implemented/architecture/2026-06-13-capability-seams.md)) in the bash shape: ONE engine implementation per context registers as `ctx.workflows`; the model-facing tool consumes it.
+
+| Package | Role | ctx key |
+|---|---|---|
+| `workflow/` | Abstract workflow seam: service base class + run vocabulary + `workflow/*` events | `ctx.workflows` |
+| `workflow-workerthread/` | `node:worker_threads` engine: one worker per run; the script's vm context lives inside the worker, `agent()` bridges to `ctx.subagents` over the message port | (provides `ctx.workflows`) |
+| `tool-workflow/` | Model-facing `workflow` tool over `ctx.workflows` | (registers on `ctx.tools`) |
+
+The interface lives at `workflow/workflow/`. The engine's `agent()` hook rides the [subagent seam](../subagent/README.md) (any registered provider; the shipped examples use `spawn`), and `agent({ schema })` rides the structured-output support the in-process backends implement. The worker thread isolates the SCRIPT — the host never blocks on it, and a cancelled run's post-grace termination is real — but it is NOT a security boundary; an isolated-vm/separate-process engine (actual sandboxing) swaps in behind the same interface if that ever matters.
+
+The proposal, decisions, and deferred work: [docs/rfc/implemented/feature/2026-07-05-dynamic-workflows.md](../../docs/rfc/implemented/feature/2026-07-05-dynamic-workflows.md).

packages/workflow/tool-workflow/README.md

@@ -0,0 +1,22 @@
+# @deepseek-ai/dsh-tool-workflow
+
+The model-facing **`workflow` tool**: run a JavaScript orchestration script that fans out subagents, and return the script's final value. Pure schema + lifecycle shaping over [`ctx.workflows`](../workflow/README.md) — script parsing, execution, caps, and cancellation live behind the seam, so a hardened engine swaps in without touching what the model sees.
+
+## What the model sees
+
+Two parameters: `script` (required — the full `export const meta = {...}` + body text; the tool DESCRIPTION carries the complete authoring contract: hooks, semantics, the supported schema subset) and `args` (optional JSON object exposed to the script as the `args` global; a bare list is wrapped as a field, a deliberate deviation from Claude Code's any-JSON `args` so the wire schema stays honest). The plugin also contributes a `tool:<toolName>` system-prompt section carrying the usage policy — use the tool only on an explicit user ask for a workflow / large orchestration; prefer plain subagent calls for one or two delegations — per the convention that tool guidance ships with the tool plugin, never in the deployment persona.
+
+## Lifecycle
+
+Collection is SYNCHRONOUS this cut (like [`dsh-tool-subagent`](../../subagent/tool-subagent/README.md)): `execute` starts a run and awaits `run.result` inside a `try/finally` that always disposes the run, so the script and its children reach quiescence on every path. `exec.signal` is bridged to `run.cancel()` (including the already-aborted-before-start case). A non-`completed` stop reason maps to an `isError` result reporting the reason — never partial output as success; a parse/meta failure thrown synchronously by `start()` becomes an `isError` the model can correct from. The completed result renders the meta name, the agent count, and the return value as JSON, truncated at `maxResultChars` with an explicit notice.
+
+## Render intent
+
+Decided up front (per the [render-intent RFC](../../../docs/rfc/implemented/architecture/2026-07-02-tool-render-intent-union.md)): a `generic` card titled `workflow: <meta.name>`, the name sniffed TEXTUALLY from `args.script` (presentation must be a pure function of args, so it cannot ask the engine to parse); the script text rides as `rawInput`. The result keeps the generic card.
+
+## Config
+
+| Key | Default | Meaning |
+|---|---|---|
+| `toolName` | `workflow` | The model-facing tool name to register. |
+| `maxResultChars` | `50000` | Rendered-result ceiling; longer JSON is truncated with a notice. |

packages/workflow/tool-workflow/package.json

@@ -0,0 +1,46 @@
+{
+  "name": "@deepseek-ai/dsh-tool-workflow",
+  "description": "Model-facing workflow tool: run a JavaScript orchestration script over ctx.workflows",
+  "version": "0.0.1",
+  "private": true,
+  "type": "module",
+  "main": "lib/index.js",
+  "types": "lib/types/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./lib/types/index.d.ts",
+      "default": "./lib/index.js"
+    },
+    "./src/*": "./src/*",
+    "./package.json": "./package.json"
+  },
+  "files": [
+    "lib/index.js",
+    "lib/types/**/*.d.ts",
+    "lib/types/**/*.d.ts.map",
+    "src"
+  ],
+  "license": "BSD-3-Clause",
+  "peerDependencies": {
+    "@deepseek-ai/dsh-agent": "^0.0.1",
+    "@deepseek-ai/dsh-llm": "^0.0.1",
+    "@deepseek-ai/dsh-system-prompt": "^0.0.1",
+    "@deepseek-ai/dsh-tools": "^0.0.1",
+    "@deepseek-ai/dsh-workflow": "^0.0.1",
+    "cordis": "^4.0.0-rc.6"
+  },
+  "dependencies": {
+    "schemastery": "^3.18.0"
+  },
+  "devDependencies": {
+    "@deepseek-ai/dsh-agent": "workspace:^",
+    "@deepseek-ai/dsh-llm": "workspace:^",
+    "@deepseek-ai/dsh-session": "workspace:^",
+    "@deepseek-ai/dsh-subagent": "workspace:^",
+    "@deepseek-ai/dsh-system-prompt": "workspace:^",
+    "@deepseek-ai/dsh-tools": "workspace:^",
+    "@deepseek-ai/dsh-workflow": "workspace:^",
+    "@deepseek-ai/dsh-workflow-workerthread": "workspace:^",
+    "cordis": "^4.0.0-rc.6"
+  }
+}

packages/workflow/tool-workflow/src/index.ts

@@ -0,0 +1,215 @@
+/**
+ * The model-facing `workflow` tool: run a JavaScript orchestration script that
+ * fans out subagents, and return the script's final value. Pure schema +
+ * lifecycle shaping — script parsing, execution, caps, and cancellation live
+ * behind `ctx.workflows` (`@deepseek-ai/dsh-workflow`), so a hardened engine
+ * swaps in without touching what the model sees.
+ *
+ * Collection is SYNCHRONOUS this cut (like `dsh-tool-subagent`): `execute`
+ * starts a run and awaits `run.result` inside a `try/finally` that always
+ * disposes the run, so the script and its children are torn down on every
+ * path. A non-`completed` stop reason maps to an `isError` tool result (by
+ * throwing) rather than returning partial output as success. Background
+ * collection is deferred to the cross-tool background redesign.
+ *
+ * Render intent (decided up front, per the render-intent RFC): a `generic`
+ * card whose title carries the workflow's `meta.name`, read directly from the
+ * call's `meta` parameter — presentation is a pure function of `args`.
+ *
+ * Usage policy ships with the tool as a `tool:<toolName>` system-prompt
+ * section (explicit-ask-only guidance) — tool guidance lives in tool plugins,
+ * never in the deployment persona.
+ *
+ * @module @deepseek-ai/dsh-tool-workflow
+ */
+
+import type { Context } from 'cordis'
+import z from 'schemastery'
+import { defineTool } from '@deepseek-ai/dsh-tools'
+import type { ToolCallView, ToolResultView } from '@deepseek-ai/dsh-tools'
+import type { ContentBlock } from '@deepseek-ai/dsh-llm'
+import type { WorkflowResult, WorkflowRun } from '@deepseek-ai/dsh-workflow'
+// Declaration merge only: makes ctx.systemPrompt visible for the section registration.
+import type {} from '@deepseek-ai/dsh-system-prompt'
+
+export const name = 'tool-workflow'
+export const inject = ['tools', 'workflows', 'systemPrompt']
+
+/** Config: the model-facing tool name plus result rendering caps. */
+export interface Config {
+  /** The model-facing tool name to register (default `workflow`). */
+  toolName?: string
+  /** Rendered-result ceiling, in characters: a longer JSON value is truncated with a notice (default 50000). */
+  maxResultChars?: number
+}
+
+export const Config: z<Config> = z.object({
+  toolName: z.string().default('workflow'),
+  maxResultChars: z.natural().min(1).default(50_000),
+})
+
+type ResolvedConfig = Required<Config>
+
+/**
+ * The script-authoring contract, embedded in the tool description. This IS the
+ * model-facing spec: the meta block, the hooks and their exact semantics, and
+ * the supported schema subset.
+ */
+const DESCRIPTION = `Run a JavaScript workflow script that orchestrates subagents at scale. Use this for work that fans out across many independent pieces — an audit over many files, a migration, multi-angle research, adversarial verification of findings — where you write the orchestration as a script instead of delegating turn by turn.
+
+The workflow's identity rides the \`meta\` parameter as JSON: required \`name\` (short kebab-case) and \`description\` strings, optional \`whenToUse\` string and \`phases\` array (\`{title, detail?, model?}\`). The \`script\` parameter is the plain JavaScript body ONLY (NOT TypeScript, and NO \`export const meta\` statement — meta is a parameter, not code), running with top-level await; end with \`return <value>\` — the value must be JSON-serializable and is this tool's result.
+
+Script-body hooks:
+- \`agent(prompt, opts?): Promise<any>\` — run one subagent to completion. Without \`opts.schema\` it resolves to the child's final text; with \`opts.schema\` (an object-rooted JSON Schema using ONLY type/properties/required/additionalProperties/items/enum/const — no oneOf/pattern/format/numeric bounds) it resolves to the validated object. Resolves \`null\` when the child fails (filter with \`.filter(Boolean)\`). Other opts: \`label\` (display), \`phase\` (progress group), \`model\` (override). Anything else (\`effort\`/\`isolation\`/\`agentType\`) is rejected loudly.
+- \`pipeline(items, ...stages): Promise<any[]>\` — run each item through the stages independently with NO barrier between stages (prefer this for multi-stage work). Each stage receives \`(prev, item, index)\`. An ordinary stage throw drops that ITEM to \`null\` and skips its remaining stages.
+- \`parallel(thunks): Promise<any[]>\` — run zero-argument functions concurrently and await ALL of them (a barrier; use only when a stage genuinely needs every prior result together). A throwing thunk resolves to \`null\`.
+- \`phase(title)\` — start a progress phase; \`log(message)\` — narrate progress; \`args\` — the tool call's \`args\` input, verbatim.
+
+Misused hooks (bad arguments, unknown options, unsupported schemas, tripped caps) throw errors that ALWAYS kill the script — they never dissolve into a per-item \`null\`.
+
+Constraints: concurrency and total-agent caps apply; no filesystem, network, timers, or Node.js APIs are provided — the agents do the work, the script only coordinates them. The run executes in the foreground: this call returns when the whole script finishes.`
+
+type WorkflowCallArgs = {
+  script: string
+  meta: { name: string; description: string; whenToUse?: string; phases?: { title: string; detail?: string; model?: string }[] }
+  args?: Record<string, unknown>
+}
+
+/** The pending-state card: a generic card titled by the workflow's meta name. */
+function presentWorkflowCall(args: WorkflowCallArgs): ToolCallView {
+  return {
+    card: 'generic',
+    title: `workflow: ${args.meta.name}`,
+    rawInput: args.script,
+  }
+}
+
+/** The completed-state card: keep the pending title; render the result content as-is. */
+function presentWorkflowResult(args: WorkflowCallArgs, result: { content: ContentBlock[]; isError: boolean }): ToolResultView {
+  void args
+  void result
+  return { card: 'generic' }
+}
+
+/** A non-`completed` stop reason means the script did not finish cleanly. */
+function stopReasonError(result: WorkflowResult): string | undefined {
+  switch (result.stopReason) {
+    case 'completed':
+      return undefined
+    case 'cancelled':
+      return `workflow run was cancelled${result.error !== undefined ? ` (${result.error})` : ''}`
+    case 'error':
+      return `workflow run failed: ${result.error ?? 'unknown error'}`
+    /* v8 ignore start -- defensive: WorkflowStopReason is a closed union, exhaustive by construction; a future variant fails here loudly */
+    default:
+      return `workflow run ended abnormally (${String(result.stopReason satisfies never)})`
+    /* v8 ignore stop */
+  }
+}
+
+/** Render the run's outcome text: the meta name, agent count, and the JSON value (capped). */
+function renderResult(run: WorkflowRun, result: WorkflowResult, maxChars: number): string {
+  // The engine returns JSON data (null for a valueless script), so stringify never yields undefined.
+  const rendered = JSON.stringify(result.value, null, 2)
+  const clipped = rendered.length > maxChars
+    ? `${rendered.slice(0, maxChars)}\n… [truncated: ${rendered.length - maxChars} more characters]`
+    : rendered
+  return `workflow "${run.meta.name}" completed (${result.agentsStarted} agent${result.agentsStarted === 1 ? '' : 's'}).\nReturn value:\n${clipped}`
+}
+
+export function apply(ctx: Context, config: Config): void {
+  // schemastery (the exported Config schema) has already filled the defaulted
+  // fields; the assertion records that resolution, not a hidden fallback.
+  const { toolName, maxResultChars } = config as ResolvedConfig
+  // Usage policy ships with the tool (the master convention: tool guidance
+  // lives in tool plugins as prompt sections, not in the deployment persona).
+  ctx.systemPrompt.section({
+    name: `tool:${toolName}`,
+    order: 115,
+    text: `Use the ${toolName} tool ONLY when the user explicitly asks for a workflow or for large multi-agent orchestration: you write a JavaScript script (the tool description documents the exact format) that fans work out across many subagents with phases and structured results. For one or two delegations, prefer plain subagent calls.`,
+  })
+  ctx.tools.register(defineTool({
+    name: toolName,
+    description: DESCRIPTION,
+    parameters: {
+      script: {
+        type: 'string',
+        required: true,
+        description: 'The plain-JS workflow script body (top-level await allowed; NO `export const meta` statement; end with `return <json-value>`).',
+      },
+      meta: {
+        type: 'object',
+        required: true,
+        description: 'The workflow identity block (plain JSON — never code).',
+        properties: {
+          name: { type: 'string', required: true, description: 'Short kebab-case workflow name.' },
+          description: { type: 'string', required: true, description: 'One-line description of what the workflow does.' },
+          whenToUse: { type: 'string', description: 'Optional guidance on when this workflow applies.' },
+          phases: {
+            type: 'array',
+            description: 'Optional phase declarations matched by phase() calls.',
+            items: {
+              type: 'object',
+              properties: {
+                title: { type: 'string', required: true, description: 'The phase title phase() calls match by exact string.' },
+                detail: { type: 'string', description: 'Optional one-line description of the phase.' },
+                model: { type: 'string', description: 'Optional model override this phase is expected to use.' },
+              },
+            },
+          },
+        },
+      },
+      args: {
+        type: 'object',
+        description: 'Optional JSON input exposed to the script as the `args` global (wrap a bare list as a field, e.g. {"files": [...]}).',
+      },
+    },
+    async execute(args, exec): Promise<ContentBlock[]> {
+      const parent = exec.agent
+      if (!parent) {
+        // The loop sets `exec.agent` for every model-driven call; its absence
+        // means a non-agent caller invoked the tool directly, which has no
+        // parent to attribute the children to. Fail loud rather than guess.
+        throw new Error('workflow tool requires a calling agent (exec.agent was undefined)')
+      }
+
+      // Meta/body validation failures (META_INVALID/SCRIPT_PARSE) throw
+      // synchronously here and become isError results via the registry — the
+      // model sees the violation list and can correct the call.
+      const run: WorkflowRun = ctx.workflows.start({
+        script: args.script,
+        meta: args.meta,
+        ...args.args !== undefined ? { args: args.args } : {},
+        parent,
+        ...exec.signal ? { signal: exec.signal } : {},
+      })
+
+      // Bridge the tool's abort signal to the run: if the parent step is
+      // aborted while the script is in flight, cancel the whole run. The
+      // engine also receives `signal` directly, but an explicit bridge keeps
+      // the tool's contract local (and covers an engine that ignores it).
+      const onAbort = (): void => { run.cancel('parent step aborted') }
+      exec.signal?.addEventListener('abort', onAbort, { once: true })
+      // `addEventListener` does NOT fire for a signal already aborted before
+      // this line — cancel explicitly in that case.
+      if (exec.signal?.aborted) run.cancel('parent step aborted')
+
+      try {
+        const result = await run.result
+        const error = stopReasonError(result)
+        if (error !== undefined) {
+          // Map a non-clean finish to an isError result (the registry turns a
+          // throw into an isError). Report the reason, not partial output.
+          throw new Error(error)
+        }
+        return [{ type: 'text', text: renderResult(run, result, maxResultChars) }]
+      } finally {
+        exec.signal?.removeEventListener('abort', onAbort)
+        // Always reach run quiescence — never leak a live script or children.
+        await run.dispose()
+      }
+    },
+    presentCall: args => presentWorkflowCall(args),
+    presentResult: (args, result) => presentWorkflowResult(args, result),
+  }))
+}

packages/workflow/tool-workflow/tests/tool-workflow.spec.ts

@@ -0,0 +1,254 @@
+import { describe, expect, it, vi } from 'vitest'
+import { Context } from 'cordis'
+import Loader from '@cordisjs/plugin-loader'
+import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
+import ToolRegistry from '@deepseek-ai/dsh-tools'
+import type { ToolExecutionResult } from '@deepseek-ai/dsh-tools'
+import { AgentId } from '@deepseek-ai/dsh-agent'
+import type { Agent } from '@deepseek-ai/dsh-agent'
+import { WorkflowRunId, WorkflowService } from '@deepseek-ai/dsh-workflow'
+import type { WorkflowResult, WorkflowRun, WorkflowStartRequest } from '@deepseek-ai/dsh-workflow'
+import { CallId } from '@deepseek-ai/dsh-llm'
+import SubagentService from '@deepseek-ai/dsh-subagent'
+import WorkerWorkflowEngine from '@deepseek-ai/dsh-workflow-workerthread'
+import * as toolWorkflow from '../src/index.ts'
+
+/** A controllable engine standing in behind ctx.workflows (the tool's only seam). */
+class StubEngine extends WorkflowService {
+  requests: WorkflowStartRequest[] = []
+  cancels: string[] = []
+  disposed = 0
+  settle!: (result: WorkflowResult) => void
+  startError: Error | undefined
+
+  start(request: WorkflowStartRequest): WorkflowRun {
+    if (this.startError) throw this.startError
+    this.requests.push(request)
+    const result = new Promise<WorkflowResult>((resolve) => { this.settle = resolve })
+    request.signal?.addEventListener('abort', () => {
+      this.settle({ value: null, stopReason: 'cancelled', error: 'signal', agentsStarted: 0 })
+    }, { once: true })
+    return {
+      id: WorkflowRunId('run-1'),
+      meta: { name: 'stub-flow', description: 'd' },
+      result,
+      cancel: (reason?: string) => {
+        this.cancels.push(reason ?? 'cancelled')
+        this.settle({ value: null, stopReason: 'cancelled', ...reason !== undefined ? { error: reason } : {}, agentsStarted: 0 })
+      },
+      dispose: () => {
+        this.disposed += 1
+        return Promise.resolve()
+      },
+    }
+  }
+}
+
+async function setup(config?: { toolName?: string; maxResultChars?: number }) {
+  const ctx = new Context()
+  await ctx.plugin(SystemPrompt)
+  await ctx.plugin(ToolRegistry)
+  await ctx.plugin(StubEngine)
+  await ctx.plugin(toolWorkflow, config ?? {})
+  const engine = ctx.workflows as StubEngine
+  const parent = { id: AgentId('caller'), options: {} } as unknown as Agent
+  return { ctx, engine, parent }
+}
+
+const SCRIPT = 'return 1'
+const META = { name: 'audit', description: 'd' }
+
+function execute(ctx: Context, args: unknown, extra?: { agent?: Agent; signal?: AbortSignal }): Promise<ToolExecutionResult> {
+  return ctx.tools.execute({
+    callId: CallId('call-1'),
+    name: 'workflow',
+    arguments: args,
+    ...extra?.agent ? { agent: extra.agent } : {},
+    ...extra?.signal ? { signal: extra.signal } : {},
+  })
+}
+
+describe('dsh-tool-workflow', () => {
+  it('starts a run with the script/args/parent/signal and renders the completed value', async () => {
+    const { ctx, engine, parent } = await setup()
+    const controller = new AbortController()
+    const pending = execute(ctx, { script: SCRIPT, meta: META, args: { files: ['a.ts'] } }, { agent: parent, signal: controller.signal })
+    await vi.waitFor(() => { expect(engine.requests.length).toBe(1) })
+    expect(engine.requests[0]).toMatchObject({ script: SCRIPT, meta: META, args: { files: ['a.ts'] }, parent })
+    expect(engine.requests[0]!.signal).toBe(controller.signal)
+    engine.settle({ value: { findings: [1, 2] }, stopReason: 'completed', agentsStarted: 7 })
+    const result = await pending
+    expect(result.isError).toBe(false)
+    const rendered = (result.content[0] as { text: string }).text
+    expect(rendered).toContain('workflow "stub-flow" completed (7 agents)')
+    expect(rendered).toContain('"findings"')
+    expect(engine.disposed).toBe(1)
+  })
+
+  it('maps a non-completed stop reason to an isError result (and still disposes)', async () => {
+    const { ctx, engine, parent } = await setup()
+    const pending = execute(ctx, { script: SCRIPT, meta: META }, { agent: parent })
+    await vi.waitFor(() => { expect(engine.requests.length).toBe(1) })
+    engine.settle({ value: null, stopReason: 'error', error: 'script threw: boom', agentsStarted: 2 })
+    const result = await pending
+    expect(result.isError).toBe(true)
+    expect((result.content[0] as { text: string }).text).toContain('workflow run failed: script threw: boom')
+    expect(engine.disposed).toBe(1)
+  })
+
+  it('reports a cancelled run distinctly (with and without a reason)', async () => {
+    const { ctx, engine, parent } = await setup()
+    const pending = execute(ctx, { script: SCRIPT, meta: META }, { agent: parent })
+    await vi.waitFor(() => { expect(engine.requests.length).toBe(1) })
+    engine.settle({ value: null, stopReason: 'cancelled', error: 'user', agentsStarted: 0 })
+    const result = await pending
+    expect(result.isError).toBe(true)
+    expect((result.content[0] as { text: string }).text).toContain('workflow run was cancelled (user)')
+
+    const bare = execute(ctx, { script: SCRIPT, meta: META }, { agent: parent })
+    await vi.waitFor(() => { expect(engine.requests.length).toBe(2) })
+    engine.settle({ value: null, stopReason: 'cancelled', agentsStarted: 0 })
+    expect(((await bare).content[0] as { text: string }).text.trim().endsWith('cancelled')).toBe(true)
+  })
+
+  it('an error result without a message renders the unknown-error fallback', async () => {
+    const { ctx, engine, parent } = await setup()
+    const pending = execute(ctx, { script: SCRIPT, meta: META }, { agent: parent })
+    await vi.waitFor(() => { expect(engine.requests.length).toBe(1) })
+    engine.settle({ value: null, stopReason: 'error', agentsStarted: 0 })
+    expect(((await pending).content[0] as { text: string }).text).toContain('unknown error')
+  })
+
+  it('cancels the run when exec.signal aborts MID-FLIGHT (the abort bridge)', async () => {
+    const { ctx, engine, parent } = await setup()
+    const controller = new AbortController()
+    const pending = execute(ctx, { script: SCRIPT, meta: META }, { agent: parent, signal: controller.signal })
+    await vi.waitFor(() => { expect(engine.requests.length).toBe(1) })
+    controller.abort()
+    const result = await pending
+    expect(result.isError).toBe(true)
+    expect(engine.cancels).toContain('parent step aborted')
+    expect(engine.disposed).toBe(1)
+  })
+
+  it('a synchronous engine start throw (meta/parse failure) becomes an isError result', async () => {
+    const { ctx, engine, parent } = await setup()
+    engine.startError = new Error('invalid meta: meta.name must be a non-empty string')
+    const result = await execute(ctx, { script: 'nope', meta: { name: '', description: 'd' } }, { agent: parent })
+    expect(result.isError).toBe(true)
+    expect((result.content[0] as { text: string }).text).toContain('meta.name must be a non-empty string')
+  })
+
+  it('requires a calling agent (fails loud without exec.agent)', async () => {
+    const { ctx, engine } = await setup()
+    const result = await execute(ctx, { script: SCRIPT, meta: META })
+    expect(result.isError).toBe(true)
+    expect((result.content[0] as { text: string }).text).toContain('requires a calling agent')
+    expect(engine.requests.length).toBe(0)
+  })
+
+  it('validates its own arguments via the schema DSL (missing script)', async () => {
+    const { ctx, parent } = await setup()
+    const result = await execute(ctx, {}, { agent: parent })
+    expect(result.isError).toBe(true)
+    expect(result.error?.code).toBe('INVALID_ARGS')
+  })
+
+  it('cancels the run when exec.signal is ALREADY aborted at call time', async () => {
+    const { ctx, engine, parent } = await setup()
+    const controller = new AbortController()
+    controller.abort()
+    const result = await execute(ctx, { script: SCRIPT, meta: META }, { agent: parent, signal: controller.signal })
+    expect(result.isError).toBe(true)
+    expect(engine.cancels).toContain('parent step aborted')
+    expect(engine.disposed).toBe(1)
+  })
+
+  it('truncates an oversized rendered value with a notice (maxResultChars)', async () => {
+    const { ctx, engine, parent } = await setup({ maxResultChars: 40 })
+    const pending = execute(ctx, { script: SCRIPT, meta: META }, { agent: parent })
+    await vi.waitFor(() => { expect(engine.requests.length).toBe(1) })
+    engine.settle({ value: { blob: 'x'.repeat(500) }, stopReason: 'completed', agentsStarted: 1 })
+    const rendered = ((await pending).content[0] as { text: string }).text
+    expect(rendered).toContain('[truncated:')
+    expect(rendered.length).toBeLessThan(400)
+  })
+
+  it('registers under a configured toolName and unregisters on fiber dispose (HMR safety)', async () => {
+    const ctx = new Context()
+    await ctx.plugin(SystemPrompt)
+    await ctx.plugin(ToolRegistry)
+    await ctx.plugin(StubEngine)
+    const fiber = await ctx.plugin(toolWorkflow, { toolName: 'orchestrate' })
+    expect(ctx.tools.get('orchestrate')).toBeDefined()
+    expect(ctx.tools.get('workflow')).toBeUndefined()
+    // The usage-policy prompt section rides the same registration: present
+    // under the CONFIGURED name (its guidance names the tool it describes)…
+    const sections = (await ctx.systemPrompt.assemble()).sections
+    const section = sections.find(s => s.name === 'tool:orchestrate')
+    expect(section?.text).toContain('orchestrate')
+    expect(sections.some(s => s.name === 'tool:workflow')).toBe(false)
+    await fiber.dispose()
+    expect(ctx.tools.get('orchestrate')).toBeUndefined()
+    // …and gone with the fiber — a reload must not leak a stale section.
+    expect((await ctx.systemPrompt.assemble()).sections.some(s => s.name === 'tool:orchestrate')).toBe(false)
+  })
+
+  it('presents a generic pending card titled by the meta name, with the script as rawInput', async () => {
+    const { ctx } = await setup()
+    const tool = ctx.tools.get('workflow')!
+    const view = tool.presentCall!({ script: SCRIPT, meta: META })
+    expect(view).toMatchObject({ card: 'generic', title: 'workflow: audit', rawInput: SCRIPT })
+  })
+
+  it('presentResult keeps the generic card; presentation is pure and replay-safe on malformed args', async () => {
+    const { ctx } = await setup()
+    const tool = ctx.tools.get('workflow')!
+    expect(tool.presentResult!({ script: SCRIPT, meta: META }, { content: [], isError: false })).toEqual({ card: 'generic' })
+    // defineTool soft-validates presentation args: a malformed logged shape
+    // (wrong fields entirely, or a call missing its meta) falls back to
+    // undefined instead of throwing mid-replay.
+    expect(tool.presentCall!({ not: 'the schema' })).toBeUndefined()
+    expect(tool.presentCall!({ script: SCRIPT })).toBeUndefined()
+  })
+
+  it('has the namespace-plugin export shape (no stray default)', () => {
+    expect('default' in toolWorkflow).toBe(false)
+    expect(toolWorkflow.name).toBe('tool-workflow')
+    expect(toolWorkflow.inject).toEqual(['tools', 'workflows', 'systemPrompt'])
+    const loader = Object.create(Loader.prototype) as Loader
+    const unwrapped = loader.unwrapExports(toolWorkflow) as Record<string, unknown>
+    expect(unwrapped).toBe(toolWorkflow)
+    expect(typeof unwrapped.apply).toBe('function')
+  })
+
+  describe('composition with the REAL worker-thread engine (the mock above must stay honest)', () => {
+    it('an abort releases the tool even when the script parks on a promise no hook owns', async () => {
+      // Regression for the review-found turn wedge: the tool awaits
+      // run.result BEFORE its disposing finally, the registry and the loop
+      // await the tool — so if cancellation could not settle result (a script
+      // parked on `await new Promise(() => {})`), an aborted turn stayed
+      // wedged forever. The seam now guarantees result settles within the
+      // grace of cancel(); this drives that guarantee through the real
+      // registry + real tool + real engine.
+      const ctx = new Context()
+      await ctx.plugin(SystemPrompt)
+      await ctx.plugin(ToolRegistry)
+      await ctx.plugin(SubagentService)
+      await ctx.plugin(WorkerWorkflowEngine, { disposeGraceMs: 30 })
+      await ctx.plugin(toolWorkflow, {})
+      const parent = { id: AgentId('caller'), options: {} } as unknown as Agent
+      const controller = new AbortController()
+      const pending = execute(ctx, {
+        script: 'await new Promise(() => {})\nreturn 1',
+        meta: { name: 'stuck', description: 'parks forever' },
+      }, { agent: parent, signal: controller.signal })
+      // Give the run a beat to start (past its synchronous slice), then abort.
+      await new Promise(resolve => setTimeout(resolve, 20))
+      controller.abort('user abort')
+      const result = await pending
+      expect(result.isError).toBe(true)
+      expect((result.content[0] as { text: string }).text).toContain('cancelled')
+    })
+  })
+})

packages/workflow/tool-workflow/tsconfig.json

@@ -0,0 +1,36 @@
+{
+  "extends": "../../../tsconfig.base.json",
+  "compilerOptions": {
+    "rootDir": "src",
+    "outDir": "lib/types"
+  },
+  "include": [
+    "src"
+  ],
+  "references": [
+    {
+      "path": "../../../vendor/cosmokit"
+    },
+    {
+      "path": "../../../vendor/cordis"
+    },
+    {
+      "path": "../../../vendor/schemastery"
+    },
+    {
+      "path": "../../core/agent"
+    },
+    {
+      "path": "../../llm/llm"
+    },
+    {
+      "path": "../../core/tools"
+    },
+    {
+      "path": "../../core/system-prompt"
+    },
+    {
+      "path": "../workflow"
+    }
+  ]
+}

packages/workflow/workflow-workerthread/README.md

@@ -0,0 +1,49 @@
+# @deepseek-ai/dsh-workflow-workerthread
+
+The [`WorkflowService`](../workflow/README.md) implementation, on **`node:worker_threads`**: each run gets its OWN worker thread (one run = one worker, no pooling — a run is heavyweight, so the ~tens-of-ms thread spin-up is noise), the script executes in a vm context INSIDE that worker with the workflow hooks injected, and every `agent()` call bridges back over the message port to [`ctx.subagents`](../../subagent/README.md) on the host. Child agents are I/O-bound LLM loops and stay on the host event loop; the thread isolates the SCRIPT, the only part that can spin synchronously.
+
+## Trust premise: what the thread buys (and what it does not)
+
+Workflow scripts are **model-written** — the same trust level as the model's existing bash access — so this engine defends against **buggy** scripts, never hostile ones. A worker thread is NOT a security boundary: the vm context inside it is escapable by construction (`node:vm` shares object machinery with its surrounding realm, so a script can reach the `Function` constructor via `globalThis.constructor.constructor` and from it `process` and every Node builtin), and an escapee holds the same process privileges as the host — Node's permission model is process-wide. The absent globals are API surface that keeps honest scripts portable, not walls. What the thread concretely buys:
+
+- **The host never blocks**: `start()` returns without running any script code on the host; a synchronous spin anywhere in the script occupies the worker's loop, not the harness's.
+- **Termination is real**: a script that outlives its post-cancel grace is `worker.terminate()`d — nothing of it survives `dispose()`, where an in-process engine could only abandon the spin on its own loop.
+- **No ambient credentials**: the worker spawns with an EMPTY environment (`env: {}` plus hermetic `execArgv`, the same stance as `dsh-code-runtime-worker`; the unbuilt dev shape forwards exactly one loader variable, `TSX_TSCONFIG_PATH` — path plumbing, not a secret), so an escapee reading `process.env` finds no harness secrets — ambient-channel hardening only; the process-wide privileges above (fs and the rest) remain, so a genuine sandbox is still the engine swap.
+- **Serialization by construction**: everything crossing the thread is structured-clone data, and plain JSON before that — the `materializeFromRealm` walk rejects loud what JSON cannot carry, which is also what makes every postMessage hop total.
+
+What the seam guarantees regardless, because benign scripts hit these constantly: `result` never rejects, a dropped hook promise never becomes an unhandled rejection, values JSON cannot carry are rejected **loud** instead of silently mangled, and hook misuse is fatal instead of dissolving into a per-item `null`. Genuine sandboxing (containing what an escaped script may touch) remains an isolated-vm/separate-process engine swap behind the seam, still deferred.
+
+## The script contract it executes
+
+- **Meta as DATA** (`validateMeta`, host-side): the workflow's identity arrives on the start request as plain JSON (the tool carries it as its schema-validated `meta` parameter — never as script text) and is shape-validated loud, every violation named (`name`/`description` required; unknown fields rejected). The engine deliberately evaluates NO script text to obtain meta: an evaluated meta literal could smuggle getters that run on the host outside any vm timeout — the exact spin the worker thread exists to isolate. A body that still opens with a Claude Code-style `export const meta` statement is rejected with a pointed `SCRIPT_PARSE` message.
+- **Hooks**: `agent(prompt, {label, phase, schema, model})` (schema = the [structured-output subset](../../core/tools/README.md), forwarded as `outputSchema`; result = validated object, or final text without a schema; a failed child resolves `null`), `parallel(thunks)`, `pipeline(items, ...stages)` with NO cross-stage barrier and `(prev, item, index)` stage callbacks, `phase(title)`, `log(message)`, and the `args` global. Anything else — `effort`/`isolation`/`agentType`, unknown options, malformed arguments, schemas outside the subset — throws a FATAL `WorkflowError` that `parallel`/`pipeline` re-throw rather than nulling (see the seam README's failure discipline).
+- **No ambient APIs**: no timers, filesystem, or Node APIs are injected into the context (absence is API surface, not containment — see the trust premise).
+
+## How a run executes
+
+`start()` shape-validates the meta DATA host-side and parse-checks the body with the identical wrapper the worker compiles (`new vm.Script`, discarded), preserving the seam's synchronous `META_INVALID`/`SCRIPT_PARSE` throws; one redundant parse per run is the deliberate price. It then spawns the worker (`src/worker.ts` unbuilt via an explicit tsx `execArgv`; the sibling `lib/worker.js` bundle when built) with the meta, body, `args`, and worker-side limits as `workerData`.
+
+Inside the worker, `runWorkerSession` builds the execution core (hooks, combinators, concurrency semaphore, caps, fatal-error discipline) over a **child port**: `agent()` sends `child-start` and the host starts the child on `ctx.subagents` (parent attribution, the shared per-run abort signal, `outputSchema`/`model` pass-through), replying with the child id, its settlement (a JSON projection; an infrastructure REJECTION crosses as `child-failed` and stays the fatal `AGENT_RESULT`), and dispose acks. Observer narration (`phase`/`log`/`agent-start`/`agent-end`) crosses as messages and re-emits as the seam's `workflow/*` events. A **ready→go handshake** gates the body: a cancellation racing worker boot arrives before `go`, so a run cancelled before start never executes the body at all.
+
+## The value boundary
+
+Values LEAVING the script (hook options/schemas, the script's return) are materialized by `materializeFromRealm`: a plain recursive walk that rejects loud everything JSON cannot carry (exotic prototypes, functions, symbols, cycles, sparse arrays, non-finite numbers, nested `undefined`), copying into plain containers via `defineProperty` so a `"__proto__"` key becomes a data property, never a prototype mutation. Getters are read ordinarily — the RESULT is what crosses; a read that throws fails loud; both run in the WORKER, never on the host. Values ENTERING the realm (`args`, `agent()` results, hook promises and their failures, combinator arrays) are handed over directly as worker-realm values — the script is trusted, so outer prototypes are not a leak; `args` is cloned once at start so a script scribbling on it cannot mutate the caller's object. One script-visible consequence: an error thrown by a hook is built OUTSIDE the script's vm context, so `e instanceof Error` inside the script is `false` — branch on `e.name`/`e.code` instead (the combinators recognize fatality by `instanceof` against their own realm's class, which a script-built object can never pass, so fatal-vs-null cannot be forged or dissolved).
+
+## Cancellation, death, disposal
+
+Per-run limits: a concurrency semaphore (`maxConcurrentAgents`), a total-`agent()` cap (`maxTotalAgents`), and a per-call item cap (`maxItemsPerCall`), all config. `cancel()` posts the cancel to the worker (its hooks start throwing `CANCELLED`; the script dies at its next await) and cancels every host-side child NOW on **both seam channels** — the shared request signal aborts AND each registered child's explicit `cancel()` is called host-side, because the seam leaves a provider free to honor either channel and a worker wedged in a synchronous spin could not relay its own per-child cancel RPCs (those later land as idempotent no-ops). The grace then arms: a run still unsettled `disposeGraceMs` later force-settles `cancelled` and the worker is **terminated**. A cancellation that lands before the body runs (the ready→go handshake) reports `cancelled` without executing anything; a worker `result` racing an in-flight host cancellation reports `cancelled` too (first-wins settlement — the seam-visible result had not settled when cancellation was requested); post-cancel `phase`/`log` narration is suppressed host-side, while cancelled children still deliver their paired `agent-end`.
+
+A worker that dies unexpectedly (an OOM, a script reaching `process.exit` through the documented vm escape) settles the run `stopReason: 'error'` with the exit diagnostics — or `'cancelled'` when a cancel was in flight — and the host-side child registry is what winds every surviving child down. `dispose()` = cancel + immediate host-driven disposal of every registered child (a wedged worker can relay no dispose RPC, so child teardown overlaps the grace instead of starting after it; the worker's own dispose RPCs join the same per-child disposal) + bounded wait (result, then child-registry quiescence, capped by the grace) + unconditional `worker.terminate()`: the thread never outlives its run. Once a run settles, stray children a script fired without awaiting are cancelled too, and `dispose()` waits for their disposal (bounded by the grace) before returning. `agent-start`/`agent-end` pairing is host-guaranteed the same way: forwarded starts live in a ledger, worker-reported ends pair them on the graceful paths, and the termination paths (grace force-settle, worker death) synthesize the missing ends (outcome `cancelled`) before the run settles — a start still in flight across the force-settle can surface after `workflow/end`, immediately paired the same way.
+
+**Engine-specific limitations**: worker startup is paid per run; on a termination path `agentsStarted` reports the HOST-observed count (accepted `child-start`s — calls still queued worker-side for a concurrency slot are unknowable then); and a returned promise or thenable resolves per JavaScript semantics BEFORE materialization — that is what makes an un-awaited `return agent('x')` work — with the value-boundary guard applying to the resolution.
+
+## Config
+
+| Key | Default | Meaning |
+|---|---|---|
+| `provider` | `spawn` | The `ctx.subagents` provider children run on (host-side). |
+| `maxConcurrentAgents` | `0` (auto) | Concurrent `agent()` ceiling; `0` resolves to `min(16, max(1, cores - 2))`. |
+| `maxTotalAgents` | `1000` | Total `agent()` calls one run may start (runaway-loop backstop). |
+| `maxItemsPerCall` | `4096` | Items accepted by one `parallel()`/`pipeline()` call. |
+| `syncTimeoutMs` | `5000` | vm timeout for the script's initial synchronous slice (in the worker). |
+| `disposeGraceMs` | `5000` | How long a cancelled run may stay unsettled before force-settle + terminate; also bounds `dispose()`. |

packages/workflow/workflow-workerthread/package.json

@@ -0,0 +1,56 @@
+{
+  "name": "@deepseek-ai/dsh-workflow-workerthread",
+  "description": "worker-thread workflow engine: executes model-written orchestration scripts off the host event loop, bridging agent() calls back to ctx.subagents",
+  "version": "0.0.1",
+  "private": true,
+  "type": "module",
+  "main": "lib/index.js",
+  "types": "lib/types/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./lib/types/index.d.ts",
+      "default": "./lib/index.js"
+    },
+    "./worker": {
+      "types": "./lib/types/worker.d.ts",
+      "default": "./lib/worker.js"
+    },
+    "./src/*": "./src/*",
+    "./package.json": "./package.json"
+  },
+  "files": [
+    "lib/index.js",
+    "lib/worker.js",
+    "lib/types/**/*.d.ts",
+    "lib/types/**/*.d.ts.map",
+    "src"
+  ],
+  "license": "BSD-3-Clause",
+  "peerDependencies": {
+    "@deepseek-ai/dsh-agent": "^0.0.1",
+    "@deepseek-ai/dsh-brand": "^0.0.1",
+    "@deepseek-ai/dsh-llm": "^0.0.1",
+    "@deepseek-ai/dsh-subagent": "^0.0.1",
+    "@deepseek-ai/dsh-tools": "^0.0.1",
+    "@deepseek-ai/dsh-workflow": "^0.0.1",
+    "cordis": "^4.0.0-rc.6"
+  },
+  "dependencies": {
+    "schemastery": "^3.18.0"
+  },
+  "devDependencies": {
+    "@deepseek-ai/dsh-agent": "workspace:^",
+    "@deepseek-ai/dsh-agent-loop": "workspace:^",
+    "@deepseek-ai/dsh-brand": "workspace:^",
+    "@deepseek-ai/dsh-invariants": "workspace:^",
+    "@deepseek-ai/dsh-llm": "workspace:^",
+    "@deepseek-ai/dsh-session": "workspace:^",
+    "@deepseek-ai/dsh-subagent": "workspace:^",
+    "@deepseek-ai/dsh-subagent-spawn": "workspace:^",
+    "@deepseek-ai/dsh-system-prompt": "workspace:^",
+    "@deepseek-ai/dsh-tools": "workspace:^",
+    "@deepseek-ai/dsh-workflow": "workspace:^",
+    "cordis": "^4.0.0-rc.6",
+    "tsx": "^4.19.2"
+  }
+}

packages/workflow/workflow-workerthread/src/host.ts

@@ -0,0 +1,473 @@
+/**
+ * The host half of one worker-engine run: spawn the Worker, bridge its child
+ * RPC onto `ctx.subagents`, fan its observer messages into the engine's
+ * events, and own cancellation, the settle-within-grace guarantee, and child
+ * cleanup. The worker's lifetime IS the run's lifetime: `dispose()` always
+ * ends with `worker.terminate()`, so no thread outlives its run.
+ *
+ * The run's `result` promise settles exactly once, from whichever of these
+ * lands first: the worker's `result` message (a host-side cancellation in
+ * flight overrides a non-cancelled report — the seam-visible result had not
+ * settled when cancellation was requested), an unexpected worker death
+ * (`error`/`messageerror`/premature `exit` → `stopReason: 'error'`, or
+ * `'cancelled'` when a cancel was in flight), or the post-cancel grace timer
+ * (a script that never settles is force-settled `cancelled` and its worker
+ * terminated — the real kill an in-process engine could not perform).
+ *
+ * Children live in a host-side registry (callId → run): the worker drives
+ * their disposal by RPC on the graceful path, `dispose()` host-drives every
+ * registered child's disposal immediately (a wedged worker can relay no
+ * dispose RPC, and child teardown must overlap the grace, not start after
+ * it), and the registry is what lets the host abort and dispose every
+ * survivor when the worker dies or is terminated mid-flight. The three
+ * paths share ONE disposal per child (memoized by callId; the seam's
+ * dispose() is idempotent anyway, the memo keeps the bookkeeping and the
+ * containment warn single). Lifecycle pairing is host-guaranteed the same
+ * way: every forwarded `agent-start` lives in a ledger, and a start the
+ * dead or terminated worker never paired is closed by a synthesized
+ * `agent-end` (outcome `'cancelled'`) before the run settles. On a
+ * termination path `agentsStarted` reports the
+ * HOST-observed count (accepted `child-start` messages) — `agent()` calls
+ * still queued worker-side for a concurrency slot are unknowable then; the
+ * worker's own count rides the result message on every graceful path.
+ *
+ * @module @deepseek-ai/dsh-workflow-workerthread/host
+ */
+
+import { fileURLToPath } from 'node:url'
+import { Worker } from 'node:worker_threads'
+import type { WorkerOptions } from 'node:worker_threads'
+import type { Context } from 'cordis'
+import type { Agent } from '@deepseek-ai/dsh-agent'
+import { assertNever } from '@deepseek-ai/dsh-llm'
+import type { SubagentRun } from '@deepseek-ai/dsh-subagent'
+import type { WorkflowAgentEndInfo, WorkflowAgentInfo, WorkflowMeta, WorkflowResult, WorkflowRun, WorkflowRunId } from '@deepseek-ai/dsh-workflow'
+import { renderThrown } from './realm.ts'
+import type { ExecutionObserver } from './runtime.ts'
+import { HostToWorkerType, WorkerToHostType } from './protocol.ts'
+import type { HostToWorkerPayloads, WorkerToHostMessage } from './protocol.ts'
+import type { ChildStartRequest, WorkerInit } from './types.ts'
+
+/**
+ * Resolve the worker entry and spawn options for the current runtime shape.
+ * Unbuilt (tsx demos, vitest — `import.meta.url` points into `src/`), the
+ * entry is the TypeScript sibling and the worker needs the tsx loader
+ * registered explicitly: a worker thread inherits no transform pipeline from
+ * vitest (vite transforms in-process, not via a node loader), and passing
+ * execArgv explicitly also shields the worker from any loader flags the
+ * parent was started with. Built (`lib/index.js`), the entry is the sibling
+ * bundle the package tsdown config emits and no loader is needed (execArgv
+ * pinned empty — hermetic, like the environment).
+ *
+ * Both shapes spawn with an EMPTY environment (`env: {}`): the documented vm
+ * escape reaches `process`, and the harness's ambient credentials
+ * (`DEEPSEEK_API_KEY` et al.) must not ride along — the same stance as
+ * `dsh-code-runtime-worker`, stronger than the scrubbed env the
+ * defensive-patterns rule requires for spawned commands (a shell needs PATH;
+ * this worker needs nothing). Sole exception: the unbuilt shape forwards
+ * `TSX_TSCONFIG_PATH` when the parent carries it (loader plumbing the paths
+ * map depends on outside the repo cwd, not a secret). This closes the
+ * AMBIENT channel only — an escapee still holds process-wide privileges
+ * like fs access (the README's trust premise stands).
+ * @param init - the run payload, passed as `workerData`.
+ * @returns the entry URL and the Worker options to spawn it with.
+ */
+function resolveWorkerSpawn(init: WorkerInit): { entry: URL; options: WorkerOptions } {
+  /* v8 ignore next 3 -- the built-output arm: tests always run unbuilt (src/); the built-worker e2e exercises this shape for real */
+  if (!import.meta.url.endsWith('.ts')) {
+    return { entry: new URL('./worker.js', import.meta.url), options: { workerData: init, env: {}, execArgv: [] } }
+  }
+  // Lazy tsx resolution: only the unbuilt shape needs it, so the built
+  // bundle never requires tsx to be installed. TSX_TSCONFIG_PATH is the one
+  // variable forwarded through the scrub: tsx finds a tsconfig by searching
+  // UP from the worker's cwd, and a parent running with its cwd outside the
+  // repo (the ACP snapshot harness pins the tsconfig through this exact
+  // variable) would otherwise lose the dsh-* paths map and resolve workspace
+  // imports to unbuilt lib/ bundles. Loader plumbing, not a secret.
+  return {
+    entry: new URL('./worker.ts', import.meta.url),
+    options: {
+      workerData: init,
+      env: process.env.TSX_TSCONFIG_PATH === undefined ? {} : { TSX_TSCONFIG_PATH: process.env.TSX_TSCONFIG_PATH },
+      execArgv: ['--import', fileURLToPath(import.meta.resolve('tsx'))],
+    },
+  }
+}
+
+/**
+ * One live worker-engine run — the seam's {@link WorkflowRun}, returned by
+ * `start()` directly. Owns the Worker, the child registry, and the result
+ * settlement; `result` never rejects. `meta` is this handle's OWN clone
+ * (event payloads carry separate clones), so a consumer mutating it corrupts
+ * nothing.
+ */
+export class WorkerRun implements WorkflowRun {
+  /** Settles exactly once with the run's outcome; never rejects. */
+  readonly result: Promise<WorkflowResult>
+  private settleResolve!: (result: WorkflowResult) => void
+  private settled = false
+  private cancelReason: string | undefined
+  private graceTimer: NodeJS.Timeout | undefined
+  private readonly worker: Worker
+  /** Set on `exit`: the thread is gone, so posting has nowhere to go. */
+  private workerGone = false
+  /** Accepted `child-start` messages — the terminate-path `agentsStarted` (see module doc). */
+  private hostStarted = 0
+  /** Live children by callId; an entry leaves ONLY after its dispose settles (quiescence = empty). */
+  private readonly children = new Map<number, SubagentRun>()
+  /** In-flight child disposals by callId — the memo that gives every path (worker RPC, dispose(), reap) ONE shared disposal per child. */
+  private readonly childDisposals = new Map<number, Promise<void>>()
+  /** Started-but-not-ended agents by seq — the pairing ledger the HOST guarantees (see {@link endAgent}). */
+  private readonly liveAgents = new Map<number, WorkflowAgentInfo>()
+  private readonly quiescenceWaiters: (() => void)[] = []
+  /** The per-run abort fanout every child start request carries. */
+  private readonly controller = new AbortController()
+  private disposed: Promise<void> | undefined
+
+  constructor(
+    private readonly ctx: Context,
+    readonly id: WorkflowRunId,
+    readonly meta: WorkflowMeta,
+    private readonly parent: Agent,
+    init: WorkerInit,
+    private readonly provider: string,
+    private readonly disposeGraceMs: number,
+    private readonly observer: ExecutionObserver,
+    signal: AbortSignal | undefined,
+  ) {
+    this.result = new Promise<WorkflowResult>((resolve) => { this.settleResolve = resolve })
+    // workerData rides the structured clone: args are plain JSON by the seam
+    // contract, so the clone is total and doubles as the caller-isolation
+    // copy (a clone failure throws loud out of start()).
+    const { entry, options } = resolveWorkerSpawn(init)
+    this.worker = new Worker(entry, options)
+    this.worker.on('message', (message: WorkerToHostMessage) => { this.onMessage(message) })
+    this.worker.on('error', (error) => { this.onWorkerDeath(`workflow worker failed: ${renderThrown(error)}`) })
+    /* v8 ignore next -- messageerror: not constructible from the engine's own protocol (every payload is JSON data) */
+    this.worker.on('messageerror', (error) => { this.onWorkerDeath(`workflow worker message failed to deserialize: ${renderThrown(error)}`) })
+    this.worker.on('exit', (code) => {
+      this.workerGone = true
+      this.onWorkerDeath(`workflow worker exited before the run settled (exit code ${code})`)
+    })
+    if (signal?.aborted) {
+      this.cancel('workflow start signal already aborted')
+    } else {
+      signal?.addEventListener('abort', () => { this.cancel('workflow signal aborted') }, { once: true })
+    }
+  }
+
+  /**
+   * Cancel the run: the worker is told (its hooks start throwing and the
+   * script dies at its next await), every host-side child is cancelled NOW on
+   * BOTH seam channels — the shared request signal aborts and each registered
+   * child's explicit `cancel()` is called (the seam leaves a provider free to
+   * honor either, and a worker wedged in a synchronous spin could not relay
+   * its own per-child cancel RPCs until far too late) — and the grace timer
+   * arms: a run still unsettled `disposeGraceMs` later force-settles
+   * `cancelled` and its worker is TERMINATED. Idempotent; the first reason
+   * wins.
+   * @param reason - human-readable cause (default `'workflow cancelled'`).
+   */
+  cancel(reason?: string): void {
+    // A settled run has nothing left to cancel: without this guard the
+    // ordinary consumer path (await result, then dispose -> cancel) would arm
+    // a grace timer nothing ever clears, pinning the run and its Worker
+    // closure until the grace expires - a bounded leak per completed run.
+    if (this.settled || this.cancelReason !== undefined) return
+    this.cancelReason = reason ?? 'workflow cancelled'
+    this.post(HostToWorkerType.Cancel, { reason: this.cancelReason })
+    this.controller.abort(this.cancelReason)
+    // The explicit channel is driven host-side, not left to the worker: a
+    // provider honoring only run.cancel() must not wait on a wedged worker's
+    // ChildCancel relay (those later RPCs land as idempotent no-ops).
+    for (const run of this.children.values()) run.cancel(this.cancelReason)
+    this.graceTimer = setTimeout(() => {
+      // The worker may no longer speak (it is about to be terminated): pair
+      // every stranded start before the run settles, so ends precede
+      // workflow/end.
+      this.endStrandedAgents()
+      this.settleResult(this.cancelledResult(this.hostStarted))
+      void this.worker.terminate()
+    }, this.disposeGraceMs)
+    // unref'd: an armed grace timer must never hold the process open.
+    this.graceTimer.unref()
+  }
+
+  /**
+   * Cancel + bounded settle + termination. Host-drives every registered
+   * child's disposal IMMEDIATELY — a wedged worker can relay no dispose RPC,
+   * and deferring child teardown to the post-terminate reap would spend the
+   * whole grace waiting for a quiescence that cannot start, then return with
+   * the disposals still in flight — so child disposal overlaps the same
+   * grace the worker gets to settle (the worker's own dispose RPCs join the
+   * shared per-child disposal). Waits (at most the grace) for the result and
+   * child quiescence, then terminates the worker unconditionally — the
+   * thread never outlives its run — and reaps whatever children remain
+   * (their disposal is contained, not awaited past the grace, the same
+   * abandonment the seam documents for a slow-disposing child). Idempotent;
+   * safe on every path.
+   * @returns resolves when the run's resources are released or abandoned.
+   */
+  dispose(): Promise<void> {
+    this.disposed ??= (async () => {
+      this.cancel('workflow disposed')
+      for (const [callId, run] of [...this.children]) void this.disposeChild(callId, run)
+      await Promise.race([
+        (async () => {
+          await this.result
+          await this.childQuiescence()
+        })(),
+        sleep(this.disposeGraceMs),
+      ])
+      await this.worker.terminate()
+      this.reapChildren('workflow disposed')
+    })()
+    return this.disposed
+  }
+
+  /** Post one message to the worker (payload looked up from the tag's map entry), tolerating a thread that is already gone. */
+  private post<T extends HostToWorkerType>(type: T, payload: HostToWorkerPayloads[T]): void {
+    if (this.workerGone) return
+    try {
+      this.worker.postMessage({ type, ...payload })
+    } catch (error: unknown) {
+      // Only a teardown race can land here (every engine message is JSON
+      // data, so serialization cannot fail); there is nothing left to
+      // deliver to — log and move on.
+      /* v8 ignore next -- postMessage teardown race (a throw between exit and its event): not constructible in-process */
+      this.ctx.logger.warn(`workflow-workerthread: postMessage failed: ${renderThrown(error)}`)
+    }
+  }
+
+  private onMessage(message: WorkerToHostMessage): void {
+    switch (message.type) {
+      case WorkerToHostType.Ready:
+        this.post(HostToWorkerType.Go, {})
+        break
+      case WorkerToHostType.Phase:
+        // Post-cancel narration is suppressed host-side: worker-side the
+        // hooks throw once the cancel message is PROCESSED, but narration
+        // already in flight (or emitted while the cancel crossed the
+        // boundary) must not reach observers — nothing is emitted after
+        // cancel() returns.
+        if (this.cancelReason === undefined) this.observer.phase(message.title)
+        break
+      case WorkerToHostType.Log:
+        if (this.cancelReason === undefined) this.observer.log(message.message)
+        break
+      case WorkerToHostType.AgentStart:
+        this.liveAgents.set(message.info.seq, message.info)
+        this.observer.agentStart(message.info)
+        break
+      case WorkerToHostType.AgentEnd:
+        // NOT suppressed on cancel: cancelled children report their paired
+        // agent-end with outcome 'cancelled'. The gate (with the termination
+        // paths' synthesis) is what makes the one-pair-per-started-child
+        // contract hold on every stop path.
+        this.endAgent(message.info)
+        break
+      case WorkerToHostType.ChildStart:
+        this.onChildStart(message.callId, message.request)
+        break
+      case WorkerToHostType.ChildCancel:
+        this.children.get(message.callId)?.cancel(message.reason)
+        break
+      case WorkerToHostType.ChildDispose:
+        this.onChildDispose(message.callId)
+        break
+      case WorkerToHostType.Result:
+        this.onResult(message.result)
+        break
+      /* v8 ignore next 2 -- closed engine-owned union; the arm only makes adding a message type a compile error */
+      default:
+        assertNever(message, 'worker-to-host message')
+    }
+  }
+
+  private onChildStart(callId: number, request: ChildStartRequest): void {
+    if (this.cancelReason !== undefined) {
+      // The worker's start raced our cancel: refuse — a child must never
+      // start on an already-aborted signal (a provider subscribing only to
+      // future abort events would never observe it).
+      this.post(HostToWorkerType.ChildStartError, { callId, rendered: `workflow run cancelled: ${this.cancelReason}` })
+      return
+    }
+    this.hostStarted += 1
+    let run: SubagentRun
+    try {
+      run = this.ctx.subagents.start(this.provider, {
+        prompt: [{ type: 'text', text: request.prompt }],
+        parent: this.parent,
+        signal: this.controller.signal,
+        ...request.schema !== undefined ? { outputSchema: request.schema } : {},
+        ...request.model !== undefined ? { agentOptions: { model: request.model } } : {},
+      })
+    } catch (error: unknown) {
+      this.post(HostToWorkerType.ChildStartError, { callId, rendered: renderThrown(error) })
+      return
+    }
+    this.children.set(callId, run)
+    this.post(HostToWorkerType.ChildStarted, { callId, childId: run.id })
+    run.result.then(
+      (result) => {
+        this.post(HostToWorkerType.ChildSettled, {
+          callId,
+          result: {
+            output: result.output,
+            ...result.structured !== undefined ? { structured: result.structured } : {},
+            stopReason: result.stopReason,
+          },
+        })
+      },
+      (error: unknown) => { this.post(HostToWorkerType.ChildFailed, { callId, rendered: renderThrown(error) }) },
+    )
+  }
+
+  private onChildDispose(callId: number): void {
+    const run = this.children.get(callId)
+    if (run === undefined) {
+      // Already disposed host-side (a dispose() drive or a death reap beat
+      // the RPC) — the ack is still owed (the worker-side wrapper awaits it).
+      this.post(HostToWorkerType.ChildDisposed, { callId })
+      return
+    }
+    // disposeChild never rejects (containment is inside), so the ack always follows.
+    void this.disposeChild(callId, run).then(() => { this.post(HostToWorkerType.ChildDisposed, { callId }) })
+  }
+
+  /**
+   * Start (or join) one registered child's disposal; the registry entry
+   * leaves when it settles. Memoized per callId: the worker's dispose RPC,
+   * the dispose() host drive, and the reap can all land on the same child —
+   * the child's `dispose()` runs once and every caller awaits that one
+   * settlement. A rejection is contained (the subagent seam's dispose() is
+   * not supposed to reject, but a backend that does anyway must not break
+   * quiescence): logged, and the child still leaves the registry.
+   * @param callId - the child's registry key.
+   * @param run - the registered child (the caller looked it up).
+   * @returns resolves when the disposal settled either way; never rejects.
+   */
+  private disposeChild(callId: number, run: SubagentRun): Promise<void> {
+    let disposal = this.childDisposals.get(callId)
+    if (disposal === undefined) {
+      disposal = run.dispose().then(
+        () => { this.finishChild(callId) },
+        (error: unknown) => {
+          this.ctx.logger.warn(`workflow-workerthread: child dispose failed: ${renderThrown(error)}`)
+          this.finishChild(callId)
+        },
+      )
+      this.childDisposals.set(callId, disposal)
+    }
+    return disposal
+  }
+
+  /** Drop a child from the registry (and its disposal memo), releasing quiescence waiters at zero. */
+  private finishChild(callId: number): void {
+    this.children.delete(callId)
+    this.childDisposals.delete(callId)
+    if (this.children.size === 0) {
+      for (const waiter of this.quiescenceWaiters.splice(0)) waiter()
+    }
+  }
+
+  /** Resolves once the child registry is empty (every disposal settled). */
+  private childQuiescence(): Promise<void> {
+    if (this.children.size === 0) return Promise.resolve()
+    return new Promise((resolve) => { this.quiescenceWaiters.push(resolve) })
+  }
+
+  /** Abort + dispose every registered child (worker death / final teardown); disposal is contained, not awaited. */
+  private reapChildren(reason: string): void {
+    this.controller.abort(this.cancelReason ?? reason)
+    for (const [callId, run] of [...this.children]) {
+      run.cancel(this.cancelReason ?? reason)
+      void this.disposeChild(callId, run)
+    }
+  }
+
+  private onResult(result: WorkflowResult): void {
+    // The worker's settle-reap already child-cancel()s every stray; this
+    // abort fires the seam signal too, for providers that only honor the
+    // request signal (both channels, on every path).
+    if (this.cancelReason === undefined) this.controller.abort('workflow settled')
+    if (this.cancelReason !== undefined && result.stopReason !== 'cancelled') {
+      // The script settled while our cancel was crossing the thread boundary
+      // — the seam-visible result had NOT settled when cancellation was
+      // requested, so report cancelled (the vm drive()'s post-settle check,
+      // relocated to the receiving side of the race).
+      this.settleResult(this.cancelledResult(result.agentsStarted))
+      return
+    }
+    this.settleResult(result)
+  }
+
+  /** An unexpected worker death (or the expected exit after termination). */
+  private onWorkerDeath(message: string): void {
+    // Whatever the worker left behind must not leak — abort + dispose it all.
+    if (this.children.size > 0) this.reapChildren('workflow worker gone')
+    // The thread is gone: no more worker-authored agent-ends can arrive —
+    // pair every stranded start (a start that crossed between the grace
+    // force-settle and this exit included) before the run settles.
+    this.endStrandedAgents()
+    // settleResult no-ops on an already-settled run (the expected exit after
+    // a dispose's terminate lands here too).
+    if (this.cancelReason !== undefined) {
+      this.settleResult(this.cancelledResult(this.hostStarted))
+      return
+    }
+    this.settleResult({ value: null, stopReason: 'error', error: message, agentsStarted: this.hostStarted })
+  }
+
+  /**
+   * The single agent-end emission gate: forwards `end` iff its start is still
+   * unpaired in the ledger, so every forwarded `workflow/agent-start` gets
+   * EXACTLY one `workflow/agent-end` — the worker's own report where it can
+   * speak, a host-synthesized one where it cannot ({@link endStrandedAgents}).
+   * @param end - the settlement to emit (worker-reported or synthesized).
+   */
+  private endAgent(end: WorkflowAgentEndInfo): void {
+    /* v8 ignore next -- a real end still in flight across the grace force-settle: not orderable in-process */
+    if (!this.liveAgents.delete(end.seq)) return
+    this.observer.agentEnd(end)
+  }
+
+  /**
+   * Synthesize the missing `agent-end` for every started-but-unpaired agent,
+   * outcome `'cancelled'`: the reap cancels every child, and a real
+   * settlement racing the force-settle loses to the cancellation — the same
+   * first-wins override {@link onResult} applies to the run's own result.
+   * Called where the worker can no longer speak (the grace force-settle,
+   * worker death), BEFORE settleResult, so the paired ends reach observers
+   * before `workflow/end`.
+   */
+  private endStrandedAgents(): void {
+    for (const info of [...this.liveAgents.values()]) {
+      this.endAgent({ ...info, outcome: 'cancelled' })
+    }
+  }
+
+  private cancelledResult(agentsStarted: number): WorkflowResult {
+    // cancel() is the only writer of cancelReason and every caller checks it
+    // first; the fallback guards the type, not a reachable path.
+    /* v8 ignore next */
+    const reason = this.cancelReason ?? 'workflow cancelled'
+    return { value: null, stopReason: 'cancelled', error: `workflow run cancelled: ${reason}`, agentsStarted }
+  }
+
+  /** First settle wins; disarms the grace timer. */
+  private settleResult(result: WorkflowResult): void {
+    if (this.settled) return
+    this.settled = true
+    clearTimeout(this.graceTimer)
+    this.settleResolve(result)
+  }
+}
+
+/** A plain timer sleep (the dispose grace); unref'd so it never holds the process open. */
+function sleep(ms: number): Promise<void> {
+  return new Promise((resolve) => {
+    const timer = setTimeout(resolve, ms)
+    timer.unref()
+  })
+}

packages/workflow/workflow-workerthread/src/index.ts

@@ -0,0 +1,203 @@
+/**
+ * The `node:worker_threads` workflow engine: the {@link WorkflowService}
+ * implementation. Runs each script in its OWN worker thread (one run = one
+ * worker, no pooling — a run is heavyweight, so thread spin-up is noise): the
+ * body executes in a vm context INSIDE the worker with the workflow hooks
+ * injected, and `agent()` calls bridge back to `ctx.subagents` over the
+ * message port — child agents are I/O-bound LLM loops and stay on the host
+ * event loop; the thread isolates the SCRIPT, the only part that can spin
+ * synchronously.
+ *
+ * TRUST PREMISE: scripts are MODEL-WRITTEN — the same trust level as the
+ * model's existing bash access — so this engine defends against BUGGY
+ * scripts, never hostile ones. A worker thread is NOT a security boundary:
+ * the vm context inside it is escapable by construction, and an escapee
+ * holds the same process privileges as the host (Node's permission model is
+ * process-wide); genuine sandboxing (isolated-vm, a separate process) is an
+ * engine swap behind the seam. What the thread buys, concretely:
+ *
+ * - `start()` never blocks the host: the script's initial synchronous slice
+ *   (and any later synchronous spin) occupies the WORKER's event loop, not
+ *   the harness's.
+ * - Termination is REAL: a script that outlives its post-cancel grace is
+ *   `worker.terminate()`d — nothing of the script survives `dispose()`,
+ *   where an in-process engine could only abandon the spin on its own loop.
+ * - The value boundary is serialization by construction: everything crossing
+ *   the thread is structured-clone data (and plain JSON before that, by the
+ *   materialization walk in ./realm.ts).
+ *
+ * Engine-specific limitations: worker startup (~tens of ms) is paid per run;
+ * on a termination path `agentsStarted` reports the host-observed child
+ * count (calls still queued worker-side for a slot are unknowable — see
+ * ./host.ts); and a worker that dies unexpectedly (an OOM, a script reaching
+ * `process.exit` through the documented vm escape) settles the run
+ * `stopReason: 'error'` with the exit diagnostics.
+ *
+ * Plugin export shape: a default-exported {@link WorkflowService} subclass
+ * (the class-based service form, like `dsh-bash-local`).
+ *
+ * @module @deepseek-ai/dsh-workflow-workerthread
+ */
+
+import { randomUUID } from 'node:crypto'
+import { availableParallelism } from 'node:os'
+import * as vm from 'node:vm'
+import type { Context } from 'cordis'
+import z from 'schemastery'
+import WorkflowService, { WorkflowError, WorkflowRunId } from '@deepseek-ai/dsh-workflow'
+import type { WorkflowRun, WorkflowRunInfo, WorkflowStartRequest } from '@deepseek-ai/dsh-workflow'
+import { WorkerRun } from './host.ts'
+import { validateMeta } from './meta.ts'
+import type { WorkerInit, WorkerLimits } from './types.ts'
+
+export { validateMeta } from './meta.ts'
+export { HostToWorkerType, WorkerToHostType } from './protocol.ts'
+export type { HostToWorkerMessage, HostToWorkerPayloads, WorkerToHostMessage, WorkerToHostPayloads } from './protocol.ts'
+export { materializeFromRealm, MaterializeError } from './realm.ts'
+export { WorkflowExecution, type ExecutionObserver } from './runtime.ts'
+export { requireParentPort, runWorkerSession } from './session.ts'
+export type {
+  ChildHandle,
+  ChildPort,
+  ChildResult,
+  ChildStartRequest,
+  WorkerInit,
+  WorkerLimits,
+} from './types.ts'
+
+/** Plugin config (all optional — `static Config` supplies the defaults). */
+export interface Config {
+  /** The `ctx.subagents` provider children run on (default `spawn`). */
+  provider?: string
+  /** Concurrent `agent()` ceiling; `0` (the default) auto-resolves to `min(16, max(1, cores - 2))`. */
+  maxConcurrentAgents?: number
+  /** Total `agent()` calls one run may start — the runaway-loop backstop (default 1000). */
+  maxTotalAgents?: number
+  /** Items accepted by a single `parallel()`/`pipeline()` call (default 4096). */
+  maxItemsPerCall?: number
+  /** vm timeout for the script's initial synchronous slice, inside the worker (default 5000 ms). */
+  syncTimeoutMs?: number
+  /**
+   * How long after a cancellation an unsettled script may keep running before
+   * the run force-settles `cancelled` and its worker is TERMINATED (default
+   * 5000 ms); also bounds `dispose()`.
+   */
+  disposeGraceMs?: number
+}
+
+type ResolvedConfig = Required<Config>
+
+/** A body that still carries the Claude Code-style meta header (meta rides the seam as data here). */
+const META_STATEMENT = /^\s*export\s+const\s+meta\b/
+
+/**
+ * Parse-check the body with the SAME wrapper the worker-side runtime
+ * compiles, so `start()` keeps the seam's synchronous `SCRIPT_PARSE` throw
+ * (the worker's own compile happens a thread away, after `start()` returned).
+ * One redundant parse per run, bought deliberately for the contract. A body
+ * opening with `export const meta` gets a pointed message instead of the
+ * wrapper's bare SyntaxError — the model's likeliest authoring slip.
+ */
+function assertBodyParses(body: string, name: string): void {
+  if (META_STATEMENT.test(body)) {
+    throw new WorkflowError('workflow meta rides the `meta` request field, not the script: remove the `export const meta = {...}` statement from the body', 'SCRIPT_PARSE')
+  }
+  try {
+    // Parse only — the script object is discarded, nothing executes.
+    void new vm.Script(`(async () => {\n${body}\n})()`, { filename: `workflow:${name}`, lineOffset: -1 })
+  } catch (error: unknown) {
+    throw new WorkflowError(`workflow script does not parse: ${String(error)}`, 'SCRIPT_PARSE', { cause: error })
+  }
+}
+
+/**
+ * The worker-thread engine service. `start()` validates the script up front
+ * (meta + a host-side body parse) and returns a {@link WorkflowRun} whose
+ * `result` never rejects; the `workflow/*` events fire around the run per
+ * the seam contract.
+ */
+export class WorkerWorkflowEngine extends WorkflowService {
+  static inject = ['subagents']
+
+  static Config: z<Config> = z.object({
+    provider: z.string().default('spawn'),
+    maxConcurrentAgents: z.natural().default(0),
+    maxTotalAgents: z.natural().min(1).default(1000),
+    maxItemsPerCall: z.natural().min(1).default(4096),
+    syncTimeoutMs: z.natural().min(1).default(5000),
+    disposeGraceMs: z.natural().default(5000),
+  })
+
+  private readonly config: ResolvedConfig
+
+  constructor(ctx: Context, config: Config) {
+    super(ctx)
+    // schemastery (static Config) has already filled the defaulted fields;
+    // the assertion records that resolution, not a hidden fallback.
+    this.config = config as ResolvedConfig
+  }
+
+  /**
+   * Validate and execute a workflow script in a fresh worker thread. Throws
+   * {@link WorkflowError} synchronously (`META_INVALID` for a malformed meta
+   * block, `SCRIPT_PARSE` for a body that does not compile) for a request
+   * that cannot begin; once a run is returned, every failure resolves through
+   * `result.stopReason` instead.
+   * @param request - the script body, its meta data and `args`, the parent
+   *   agent, and an optional cancel signal.
+   * @returns the live run (its `result` resolves when the script settles).
+   */
+  start(request: WorkflowStartRequest): WorkflowRun {
+    const meta = validateMeta(request.meta)
+    assertBodyParses(request.script, meta.name)
+    const id = WorkflowRunId(randomUUID())
+    // The event payloads and the run handle get SEPARATE meta clones: a
+    // listener mutating its snapshot must not corrupt the holder's view.
+    const info: WorkflowRunInfo = { id, meta: structuredClone(meta) }
+    const limits: WorkerLimits = {
+      maxConcurrentAgents: this.config.maxConcurrentAgents === 0
+        ? Math.min(16, Math.max(1, availableParallelism() - 2))
+        : this.config.maxConcurrentAgents,
+      maxTotalAgents: this.config.maxTotalAgents,
+      maxItemsPerCall: this.config.maxItemsPerCall,
+      syncTimeoutMs: this.config.syncTimeoutMs,
+    }
+    const init: WorkerInit = {
+      meta,
+      body: request.script,
+      ...request.args !== undefined ? { args: request.args } : {},
+      limits,
+    }
+    const workerRun = new WorkerRun(
+      this.ctx,
+      id,
+      structuredClone(meta),
+      request.parent,
+      init,
+      this.config.provider,
+      this.config.disposeGraceMs,
+      {
+        phase: (title) => { this.emitWorkflowEvent('workflow/phase', info, title) },
+        log: (message) => { this.emitWorkflowEvent('workflow/log', info, message) },
+        agentStart: (agent) => { this.emitWorkflowEvent('workflow/agent-start', info, agent) },
+        agentEnd: (agent) => { this.emitWorkflowEvent('workflow/agent-end', info, agent) },
+      },
+      request.signal,
+    )
+
+    this.emitWorkflowEvent('workflow/start', info)
+    // `workflow/end` fires as the (never-rejecting) result settles, with the
+    // outcome DATA only — the value stays with the run's holder.
+    void workerRun.result.then((settled) => {
+      this.emitWorkflowEvent('workflow/end', info, {
+        stopReason: settled.stopReason,
+        ...settled.error !== undefined ? { error: settled.error } : {},
+        agentsStarted: settled.agentsStarted,
+      })
+    })
+
+    return workerRun
+  }
+}
+
+export default WorkerWorkflowEngine

packages/workflow/workflow-workerthread/src/meta.ts

@@ -0,0 +1,85 @@
+/**
+ * Meta validation: check the caller-provided {@link WorkflowMeta} DATA against
+ * the shape contract and reject everything else loud, every violation named.
+ * Meta arrives as plain JSON through the seam (the model-facing tool carries
+ * it as a schema-validated object parameter) — the engine never evaluates
+ * script text to obtain it, so no script-controlled code can run on the host
+ * here (an evaluated meta literal could smuggle getters that spin the host
+ * outside any vm timeout, the exact escape the worker thread exists to
+ * prevent).
+ *
+ * @module @deepseek-ai/dsh-workflow-workerthread/meta
+ */
+
+import { WorkflowError } from '@deepseek-ai/dsh-workflow'
+import type { WorkflowMeta, WorkflowPhase } from '@deepseek-ai/dsh-workflow'
+
+/** Collect shape violations for a meta value (plain JSON data by the seam contract). */
+function validateMetaShape(meta: unknown): { meta?: WorkflowMeta; violations: string[] } {
+  const violations: string[] = []
+  if (typeof meta !== 'object' || meta === null || Array.isArray(meta)) {
+    return { violations: ['meta must be an object'] }
+  }
+  const record = meta as Record<string, unknown>
+  const known = new Set(['name', 'description', 'whenToUse', 'phases'])
+  for (const key of Object.keys(record)) {
+    if (!known.has(key)) violations.push(`meta.${key} is not a recognized field (name/description/whenToUse/phases)`)
+  }
+  if (typeof record.name !== 'string' || record.name.length === 0) violations.push('meta.name must be a non-empty string')
+  if (typeof record.description !== 'string' || record.description.length === 0) violations.push('meta.description must be a non-empty string')
+  if (record.whenToUse !== undefined && typeof record.whenToUse !== 'string') violations.push('meta.whenToUse must be a string')
+  const phases: WorkflowPhase[] = []
+  if (record.phases !== undefined) {
+    if (!Array.isArray(record.phases)) {
+      violations.push('meta.phases must be an array')
+    } else {
+      record.phases.forEach((phase, index) => {
+        if (typeof phase !== 'object' || phase === null || Array.isArray(phase)) {
+          violations.push(`meta.phases[${index}] must be an object`)
+          return
+        }
+        const entry = phase as Record<string, unknown>
+        for (const key of Object.keys(entry)) {
+          if (!['title', 'detail', 'model'].includes(key)) violations.push(`meta.phases[${index}].${key} is not a recognized field`)
+        }
+        if (typeof entry.title !== 'string' || entry.title.length === 0) violations.push(`meta.phases[${index}].title must be a non-empty string`)
+        if (entry.detail !== undefined && typeof entry.detail !== 'string') violations.push(`meta.phases[${index}].detail must be a string`)
+        if (entry.model !== undefined && typeof entry.model !== 'string') violations.push(`meta.phases[${index}].model must be a string`)
+        if (violations.length === 0) {
+          phases.push({
+            title: entry.title as string,
+            ...entry.detail !== undefined ? { detail: entry.detail as string } : {},
+            ...entry.model !== undefined ? { model: entry.model as string } : {},
+          })
+        }
+      })
+    }
+  }
+  if (violations.length > 0) return { violations }
+  return {
+    violations,
+    meta: {
+      name: record.name as string,
+      description: record.description as string,
+      ...record.whenToUse !== undefined ? { whenToUse: record.whenToUse as string } : {},
+      ...record.phases !== undefined ? { phases } : {},
+    },
+  }
+}
+
+/**
+ * Validate a caller-provided meta value against the {@link WorkflowMeta}
+ * contract. Throws `META_INVALID` naming every violation (unknown fields,
+ * missing/mistyped `name`/`description`, malformed `phases`); the returned
+ * meta is a NORMALIZED copy built from the validated fields, so the engine
+ * never aliases the caller's object.
+ * @param value - the meta data from the start request (plain JSON by the seam contract).
+ * @returns the validated, normalized meta block.
+ */
+export function validateMeta(value: unknown): WorkflowMeta {
+  const { meta, violations } = validateMetaShape(value)
+  if (meta === undefined) {
+    throw new WorkflowError(`invalid meta: ${violations.join('; ')}`, 'META_INVALID')
+  }
+  return meta
+}

packages/workflow/workflow-workerthread/src/protocol.ts

@@ -0,0 +1,115 @@
+/**
+ * The host⇄worker wire protocol: one string-valued enum of message tags per
+ * direction, a payload map giving each tag its parameters (the single source
+ * of truth), and the message unions derived from them. Everything in a
+ * payload is plain JSON data by construction (the runtime materializes
+ * script values before they reach a message; the host projects seam results
+ * down to their JSON fields), so the structured-clone hop never meets a
+ * value it cannot carry.
+ *
+ * Both directions are CLOSED (engine-owned): each side switches on `type`
+ * and ends with `assertNever` — an unknown message is a protocol bug, never
+ * something to skip silently. Senders go through a generic
+ * `post(type, payload)` whose payload parameter is looked up from the map,
+ * so a tag/payload mismatch is a compile error at the call site.
+ *
+ * @module @deepseek-ai/dsh-workflow-workerthread/protocol
+ */
+
+import type { WorkflowAgentEndInfo, WorkflowAgentInfo, WorkflowResult } from '@deepseek-ai/dsh-workflow'
+import type { ChildResult, ChildStartRequest } from './types.ts'
+
+/** Message tags the worker sends the host (the wire values are the tag strings). */
+export enum WorkerToHostType {
+  /** The startup handshake: the session is listening and awaits {@link HostToWorkerType.Go}. */
+  Ready = 'ready',
+  /** Observer narration: a `phase(title)` call. */
+  Phase = 'phase',
+  /** Observer narration: a `log(message)` call. */
+  Log = 'log',
+  /** Observer lifecycle: one `agent()` call started a child. */
+  AgentStart = 'agent-start',
+  /** Observer lifecycle: one `agent()` call settled. */
+  AgentEnd = 'agent-end',
+  /** Child RPC: start a child on the host (answered by ChildStarted or ChildStartError). */
+  ChildStart = 'child-start',
+  /** Child RPC: cancel a started child (fire-and-forget). */
+  ChildCancel = 'child-cancel',
+  /** Child RPC: dispose a started child (answered by ChildDisposed). */
+  ChildDispose = 'child-dispose',
+  /** The run's single terminal result. */
+  Result = 'result',
+}
+
+/** The payload each worker→host tag carries. */
+export interface WorkerToHostPayloads {
+  /** Ready carries nothing. */
+  [WorkerToHostType.Ready]: Record<never, never>
+  /** The phase title, verbatim. */
+  [WorkerToHostType.Phase]: { title: string }
+  /** The logged message, verbatim. */
+  [WorkerToHostType.Log]: { message: string }
+  /** The call's sequence number, label, phase, and child id. */
+  [WorkerToHostType.AgentStart]: { info: WorkflowAgentInfo }
+  /** The call identity plus its outcome. */
+  [WorkerToHostType.AgentEnd]: { info: WorkflowAgentEndInfo }
+  /** The RPC correlation id and the prompt plus validated options. */
+  [WorkerToHostType.ChildStart]: { callId: number; request: ChildStartRequest }
+  /** The RPC correlation id and the cancel reason (undefined = unspecified). */
+  [WorkerToHostType.ChildCancel]: { callId: number; reason: string | undefined }
+  /** The RPC correlation id of the child to dispose. */
+  [WorkerToHostType.ChildDispose]: { callId: number }
+  /** The run's terminal outcome. */
+  [WorkerToHostType.Result]: { result: WorkflowResult }
+}
+
+/** Message tags the host sends the worker (the wire values are the tag strings). */
+export enum HostToWorkerType {
+  /** Releases the startup gate: run the script body. */
+  Go = 'go',
+  /** Cancel the run: hooks start throwing and the script dies at its next await. */
+  Cancel = 'cancel',
+  /** Child RPC reply: the start succeeded (exactly one of ChildStarted/ChildStartError per ChildStart). */
+  ChildStarted = 'child-started',
+  /** Child RPC reply: the start was refused or threw. */
+  ChildStartError = 'child-start-error',
+  /** Child RPC: a started child's result RESOLVED (its JSON projection). */
+  ChildSettled = 'child-settled',
+  /** Child RPC: a started child's result REJECTED (an infrastructure fault, rendered). */
+  ChildFailed = 'child-failed',
+  /** Child RPC reply: a requested disposal completed. */
+  ChildDisposed = 'child-disposed',
+}
+
+/** The payload each host→worker tag carries. */
+export interface HostToWorkerPayloads {
+  /** Go carries nothing. */
+  [HostToWorkerType.Go]: Record<never, never>
+  /** The cancel reason, canonical for the whole run. */
+  [HostToWorkerType.Cancel]: { reason: string }
+  /** The RPC correlation id and the child agent's id (minted by the subagent seam). */
+  [HostToWorkerType.ChildStarted]: { callId: number; childId: string }
+  /** The RPC correlation id and the rendered start failure. */
+  [HostToWorkerType.ChildStartError]: { callId: number; rendered: string }
+  /** The RPC correlation id and the child's terminal result projection. */
+  [HostToWorkerType.ChildSettled]: { callId: number; result: ChildResult }
+  /** The RPC correlation id and the rendered infrastructure fault. */
+  [HostToWorkerType.ChildFailed]: { callId: number; rendered: string }
+  /** The RPC correlation id of the completed disposal. */
+  [HostToWorkerType.ChildDisposed]: { callId: number }
+}
+
+/**
+ * One worker→host message of tag `T`; unparameterized, the closed union over
+ * every tag (a discriminated union — `switch` on `type` narrows).
+ */
+export type WorkerToHostMessage<T extends WorkerToHostType = WorkerToHostType> =
+  { [K in T]: { type: K } & WorkerToHostPayloads[K] }[T]
+
+/**
+ * One host→worker message of tag `T`; unparameterized, the closed union over
+ * every tag (a discriminated union — `switch` on `type` narrows).
+ */
+export type HostToWorkerMessage<T extends HostToWorkerType = HostToWorkerType> =
+  { [K in T]: { type: K } & HostToWorkerPayloads[K] }[T]
+

packages/workflow/workflow-workerthread/src/realm.ts

@@ -0,0 +1,174 @@
+/**
+ * The engine's value boundary: copy script-realm values into plain JSON data
+ * — loud about everything JSON cannot carry — and render thrown script
+ * values to failure text. The script runs in a vm context INSIDE the worker
+ * thread, so "host" here means the worker-side JavaScript around that
+ * context; everything that later crosses the thread boundary is JSON by this
+ * walk, which is what makes the postMessage hop total.
+ *
+ * TRUST PREMISE (everything in this module hangs on it): workflow scripts are
+ * MODEL-WRITTEN, the same trust level as the model's existing bash access, so
+ * this boundary guards against BUGGY scripts, not hostile ones. It rejects
+ * loud what JSON would silently mangle — functions, symbols, bigints,
+ * non-finite numbers, nested `undefined`, cycles, sparse arrays, exotic
+ * prototypes — because accepted-then-ignored is this repo's banned failure
+ * mode. It does NOT defend against adversarial values: the walk reads
+ * properties ordinarily (a getter runs, and whatever it returns is what
+ * crosses), {@link renderThrown} reads `stack`/`message`/`String()` directly,
+ * and a proxy is walked through its traps. A hostile script gains nothing
+ * worth defending here — the vm context inside the worker is escapable by
+ * construction, so hostile-value containment would be cost without a threat
+ * model (what the worker thread DOES buy is that a spin occupies the
+ * worker's loop, not the host's, and termination is real).
+ *
+ * The host→realm direction needs no machinery at all: hooks hand the script
+ * plain values of the worker realm, prototypes included — the script is
+ * trusted. One consequence is documented in the engine README: an error
+ * thrown by a hook is built OUTSIDE the script's vm context, so an in-script
+ * `instanceof Error` check is false; read `name`/`code`/`message` instead.
+ *
+ * @module @deepseek-ai/dsh-workflow-workerthread/realm
+ */
+
+/** Thrown by {@link materializeFromRealm}; the caller wraps it into the right `WorkflowError` code. */
+export class MaterializeError extends Error {
+  constructor(public readonly path: string, public readonly reason: string) {
+    super(`${path}: ${reason}`)
+    this.name = 'MaterializeError'
+  }
+}
+
+/**
+ * Render a thrown value to failure text without ever throwing: prefer the
+ * `stack` (host or realm — a realm error's `stack` is a plain string read),
+ * fall back to `message`, then `String()`. Reading those properties MAY run
+ * script code (a getter, `toString`) — accepted under the module's trust
+ * premise; if that code itself throws, a fixed label is returned instead.
+ * @param error - the thrown value, of any shape and any realm.
+ * @returns human-readable text for the failure report; prefers the stack.
+ */
+export function renderThrown(error: unknown): string {
+  try {
+    const stack = (error as { stack?: unknown } | null | undefined)?.stack
+    if (typeof stack === 'string' && stack.length > 0) return stack
+    const message = (error as { message?: unknown } | null | undefined)?.message
+    if (typeof message === 'string' && message.length > 0) return message
+    return String(error)
+  } catch {
+    // A throwing accessor/toString on the thrown value — rendering must be
+    // total (drive()'s never-reject contract), so fall back to a fixed label.
+    return '[unrenderable thrown value]'
+  }
+}
+
+/**
+ * Whether an object's prototype chain is data-shaped: `null`, or a prototype
+ * whose own prototype is `null` (the realm's `Object.prototype` — which we
+ * cannot compare by identity across realms). A `Date`/`Map`/class instance
+ * has a longer chain and is rejected.
+ */
+function hasPlainPrototype(value: object): boolean {
+  const proto: unknown = Object.getPrototypeOf(value)
+  if (proto === null) return true
+  return Object.getPrototypeOf(proto) === null
+}
+
+/**
+ * Copy `value` (typically from the vm realm) into plain host JSON data.
+ * Throws {@link MaterializeError} naming the offending path for anything JSON
+ * cannot carry losslessly. Properties are read ordinarily — a getter runs and
+ * its RESULT is materialized; a read that throws surfaces as a
+ * {@link MaterializeError} carrying the rendered failure. `undefined` is
+ * accepted only at the ROOT (a script with no `return` value) — the caller
+ * decides what it means; an `undefined` nested INSIDE a container is a
+ * violation.
+ * @param value - the realm value to materialize.
+ * @param root - the path label for the root value (error messages).
+ * @returns the host-realm copy (plain objects/arrays/scalars only).
+ */
+export function materializeFromRealm(value: unknown, root = 'value'): unknown {
+  if (value === undefined) return undefined
+  try {
+    return materialize(value, root, new Set())
+  } catch (error: unknown) {
+    if (error instanceof MaterializeError) throw error
+    // A property read ran script code that threw; total-ize it so callers can
+    // keep the narrow MaterializeError contract.
+    throw new MaterializeError(root, `reading the value threw: ${renderThrown(error)}`)
+  }
+}
+
+function materialize(value: unknown, path: string, seen: Set<object>): unknown {
+  switch (typeof value) {
+    case 'boolean':
+    case 'string':
+      return value
+    case 'number': {
+      if (!Number.isFinite(value)) throw new MaterializeError(path, 'non-finite numbers are not JSON data')
+      return value
+    }
+    case 'bigint':
+      throw new MaterializeError(path, 'bigints are not JSON data')
+    case 'function':
+      throw new MaterializeError(path, 'functions cannot cross the workflow value boundary')
+    case 'symbol':
+      throw new MaterializeError(path, 'symbols cannot cross the workflow value boundary')
+    case 'undefined':
+      throw new MaterializeError(path, 'undefined is not JSON data')
+    case 'object':
+      break
+  }
+  if (value === null) return null
+  const objectValue: object = value
+  if (seen.has(objectValue)) throw new MaterializeError(path, 'circular references are not JSON data')
+  seen.add(objectValue)
+  try {
+    if (Array.isArray(objectValue)) return materializeArray(objectValue, path, seen)
+    return materializeObject(objectValue, path, seen)
+  } finally {
+    seen.delete(objectValue)
+  }
+}
+
+function materializeArray(value: unknown[], path: string, seen: Set<object>): unknown[] {
+  const out: unknown[] = []
+  for (let index = 0; index < value.length; index++) {
+    if (!(index in value)) throw new MaterializeError(`${path}[${index}]`, 'sparse arrays are not JSON data')
+    out.push(materialize(value[index], `${path}[${index}]`, seen))
+  }
+  // Own enumerable props beyond the indices (e.g. `arr.total = 3`) would be
+  // silently dropped by JSON — reject them instead.
+  for (const key of Object.keys(value)) {
+    const index = Number(key)
+    if (!Number.isInteger(index) || index < 0 || index >= value.length) {
+      throw new MaterializeError(`${path}.${key}`, 'arrays with non-index properties are not JSON data')
+    }
+  }
+  if (Object.getOwnPropertySymbols(value).length > 0) {
+    throw new MaterializeError(path, 'symbol-keyed properties cannot cross the workflow value boundary')
+  }
+  return out
+}
+
+function materializeObject(value: object, path: string, seen: Set<object>): Record<string, unknown> {
+  if (!hasPlainPrototype(value)) {
+    throw new MaterializeError(path, 'only plain objects and arrays are JSON data (exotic prototype)')
+  }
+  if (Object.getOwnPropertySymbols(value).length > 0) {
+    throw new MaterializeError(path, 'symbol-keyed properties cannot cross the workflow value boundary')
+  }
+  const out: Record<string, unknown> = {}
+  // Object.keys = own enumerable string keys, matching JSON.stringify's
+  // property selection exactly (non-enumerable props never reach JSON output).
+  for (const key of Object.keys(value)) {
+    // defineProperty, never assignment: a "__proto__" key must become an OWN
+    // data property of the copy, not a prototype mutation.
+    Object.defineProperty(out, key, {
+      value: materialize((value as Record<string, unknown>)[key], `${path}.${key}`, seen),
+      enumerable: true,
+      writable: true,
+      configurable: true,
+    })
+  }
+  return out
+}

packages/workflow/workflow-workerthread/src/runtime.ts

@@ -0,0 +1,522 @@
+/**
+ * Per-run execution state for the engine's THREAD side: the script's vm
+ * context and its injected hooks (`agent`/`parallel`/`pipeline`/`phase`/
+ * `log`/`args`), the concurrency semaphore and caps, cancellation, and the
+ * drive loop that turns a script settlement into a {@link WorkflowResult}.
+ * Children are started by RPC to the host through a {@link ChildPort}, so
+ * this module never touches a cordis context — it runs inside the worker
+ * thread.
+ *
+ * Value boundary (the trust premise lives in ./realm.ts): values ENTERING the
+ * worker-side host code from the script (hook options, schemas, the return
+ * value) are materialized by `materializeFromRealm` — a plain walk that
+ * rejects loud everything JSON cannot carry, which also makes every value
+ * safe for the later postMessage hop. Values ENTERING the realm (`args`,
+ * `agent()` results, hook promises and their failures, combinator arrays) are
+ * handed over DIRECTLY as worker-realm values: the script is model-written
+ * and trusted, so outer prototypes are not a leak. `args` is cloned once at
+ * start so a script scribbling on it cannot mutate the session's init object
+ * (a benign-bug guard; the postMessage clone already isolated the caller).
+ *
+ * Failure discipline: fatal {@link WorkflowError}s (bad hook arguments,
+ * unsupported options/schemas, tripped caps, host start refusals and child
+ * result rejections, cancellation) ALWAYS propagate through
+ * `parallel`/`pipeline` — recognized by `instanceof` against this realm's
+ * class, which a script inside the vm context cannot forge — and the per-item
+ * `null` is reserved for child-run failures and ordinary in-stage script
+ * errors. Every hook-returned promise gets a no-op rejection consumer, so a
+ * dropped promise cannot surface an unhandled rejection (which would kill the
+ * worker and read as an engine fault).
+ *
+ * There is deliberately NO worker-side abandon channel: a script that never
+ * settles after a cancel simply never posts a result, and the HOST enforces
+ * the settles-within-grace guarantee by force-settling `cancelled` and
+ * terminating the worker — the real kill an in-process engine could not have.
+ *
+ * @module @deepseek-ai/dsh-workflow-workerthread/runtime
+ */
+
+import * as vm from 'node:vm'
+import { AgentId } from '@deepseek-ai/dsh-agent'
+import type { ContentBlock } from '@deepseek-ai/dsh-llm'
+import { assertSupportedOutputSchema, OutputSchemaError } from '@deepseek-ai/dsh-tools'
+import type { StructuredOutputSchema } from '@deepseek-ai/dsh-tools'
+import { isFatalWorkflowError, WorkflowError } from '@deepseek-ai/dsh-workflow'
+import type {
+  WorkflowAgentEndInfo,
+  WorkflowAgentInfo,
+  WorkflowMeta,
+  WorkflowResult,
+} from '@deepseek-ai/dsh-workflow'
+import { materializeFromRealm, MaterializeError, renderThrown } from './realm.ts'
+import type { ChildHandle, ChildPort, WorkerLimits } from './types.ts'
+
+/** The observers the execution reports progress through (the session posts them to the host). */
+export interface ExecutionObserver {
+  phase(title: string): void
+  log(message: string): void
+  agentStart(info: WorkflowAgentInfo): void
+  agentEnd(info: WorkflowAgentEndInfo): void
+}
+
+/** The `agent()` options the script may pass; everything else rejects loud. */
+const SUPPORTED_AGENT_OPTIONS = new Set(['label', 'phase', 'schema', 'model'])
+/** Deferred Claude Code options we name explicitly in the rejection message. */
+const DEFERRED_AGENT_OPTIONS = new Set(['effort', 'isolation', 'agentType'])
+
+/** Flatten a child's final output blocks to text (the non-schema `agent()` result). */
+function outputText(blocks: ContentBlock[]): string {
+  return blocks
+    .filter((block): block is Extract<ContentBlock, { type: 'text' }> => block.type === 'text')
+    .map(block => block.text)
+    .join('')
+}
+
+/** A short display label derived from the prompt when the script passes none. */
+function defaultLabel(prompt: string): string {
+  const newline = prompt.indexOf('\n')
+  const line = newline === -1 ? prompt : prompt.slice(0, newline)
+  return line.length <= 48 ? line : `${line.slice(0, 47)}…`
+}
+
+/**
+ * One live script execution inside the worker. Constructed per run by the
+ * session; `drive()` is called exactly once and NEVER rejects — every failure
+ * becomes a {@link WorkflowResult} with a non-`completed` stop reason.
+ */
+export class WorkflowExecution {
+  /** 1-based count of `agent()` calls started (the `agentsStarted` result field). */
+  private started = 0
+  private activeSlots = 0
+  private readonly slotWaiters: { resolve(): void; reject(error: unknown): void }[] = []
+  private cancelReason: string | undefined
+  private cancelError: WorkflowError | undefined
+  private readonly controller = new AbortController()
+  private currentPhase: string | undefined
+  private readonly context: vm.Context
+  private readonly compiled: vm.Script
+
+  constructor(
+    meta: WorkflowMeta,
+    body: string,
+    args: unknown,
+    private readonly limits: WorkerLimits,
+    private readonly observer: ExecutionObserver,
+    private readonly children: ChildPort,
+  ) {
+    // Compile FIRST: a body syntax error must throw out of the constructor
+    // before any realm state exists. The host pre-parses the identical
+    // wrapper, so under one Node version this throw is unreachable in
+    // production — the session still maps it to an error result defensively.
+    // lineOffset compensates for the wrapper line, so stack traces carry the
+    // script's own line numbers.
+    try {
+      this.compiled = new vm.Script(`(async () => {\n${body}\n})()`, {
+        filename: `workflow:${meta.name}`,
+        lineOffset: -1,
+      })
+    } catch (error: unknown) {
+      throw new WorkflowError(`workflow script does not parse: ${String(error)}`, 'SCRIPT_PARSE', { cause: error })
+    }
+
+    this.context = vm.createContext({}, { name: `workflow:${meta.name}` })
+
+    const globals: Record<string, unknown> = {
+      agent: (prompt: unknown, opts?: unknown) => this.contain(this.agent(prompt, opts)),
+      parallel: (thunks: unknown) => this.contain(this.parallel(thunks)),
+      pipeline: (items: unknown, ...stages: unknown[]) => this.contain(this.pipeline(items, stages)),
+      phase: (title: unknown) => { this.phase(title) },
+      log: (message: unknown) => { this.log(message) },
+      // Cloned once: a script scribbling on args must not mutate the
+      // session's init object (a benign-bug guard; args is plain JSON by the
+      // seam contract and already crossed one structured clone as workerData,
+      // so this clone is total).
+      args: args === undefined ? undefined : structuredClone(args),
+    }
+    for (const [key, value] of Object.entries(globals)) {
+      // Data properties on the contextified global; frozen shape not required —
+      // a script overwriting its own hooks only sabotages itself.
+      ;(this.context as Record<string, unknown>)[key] = typeof value === 'function' ? Object.freeze(value) : value
+    }
+  }
+
+  /**
+   * Whether the run has been cancelled. A METHOD, not an inline property
+   * read: `cancel()` mutates `cancelReason` concurrently (the session's
+   * message handler), and an inline read after an `await` gets narrowed by
+   * control flow into an always-false comparison.
+   */
+  private isCancelled(): boolean {
+    return this.cancelReason !== undefined
+  }
+
+  /**
+   * Shared hook entry guard: after {@link cancel}, EVERY hook throws
+   * `CANCELLED` at its next call — cancellation is the next HOOK boundary,
+   * not just the next `agent()`, so a script that caught one cancelled
+   * rejection cannot keep emitting progress through `phase`/`log` or enter a
+   * combinator.
+   */
+  private throwIfCancelled(): void {
+    if (this.isCancelled()) throw this.cancelledError()
+  }
+
+  /**
+   * Cancel the run: in-flight children get a cancel RPC (the shared abort
+   * fanout), waiting `agent()` slots reject, and every future hook call
+   * throws `CANCELLED` — the script dies at its next await. A script that
+   * never settles anyway (parked on a promise no hook owns) is the HOST's
+   * problem: its grace timer force-settles the run and terminates the
+   * worker. Idempotent; the first reason wins.
+   * @param reason - human-readable cause, carried on the CANCELLED error and
+   * into child cancel RPCs. Required: every caller (the session's cancel
+   * message, drive()'s settle-reap) has a concrete reason.
+   */
+  cancel(reason: string): void {
+    if (this.cancelReason !== undefined) return
+    this.cancelReason = reason
+    this.cancelError = new WorkflowError(`workflow run cancelled: ${this.cancelReason}`, 'CANCELLED')
+    this.controller.abort(this.cancelReason)
+    for (const waiter of this.slotWaiters.splice(0)) waiter.reject(this.cancelledError())
+  }
+
+  /**
+   * Run the script to settlement. Resolves — never rejects — with the run's
+   * {@link WorkflowResult}: the materialized return value on `completed`, the
+   * failure message on `error`, and `cancelled` when the script died of
+   * cancellation. After settlement, any stray children a script fired without
+   * awaiting are cancelled (their `agent()` wrappers dispose them via RPC).
+   * @returns the settled outcome — this promise NEVER rejects (the seam's
+   * `result`-never-rejects contract); every failure maps to a variant.
+   */
+  async drive(): Promise<WorkflowResult> {
+    try {
+      // Cancelled before the body ever ran (an already-aborted start signal,
+      // relayed by the host before its `go`): the script must not execute at
+      // all, let alone report `completed`.
+      if (this.isCancelled()) throw this.cancelledError()
+      const scriptPromise = this.compiled.runInContext(this.context, { timeout: this.limits.syncTimeoutMs }) as Promise<unknown>
+      const raw: unknown = await this.contain(Promise.resolve(scriptPromise))
+      // Cancelled while the body ran: a script that settled without touching
+      // another hook (or without any) must still report `cancelled` — the
+      // holder asked for cancellation and `completed` would be a lie.
+      if (this.isCancelled()) throw this.cancelledError()
+      const value = raw === undefined ? null : this.materializeResult(raw)
+      return { value, stopReason: 'completed', agentsStarted: this.started }
+    } catch (error: unknown) {
+      // Any failure after cancel() reports `cancelled` with the canonical
+      // reason — the reject path mirrors the resolve path's post-settle check.
+      if (this.isCancelled()) {
+        return { value: null, stopReason: 'cancelled', error: this.cancelledError().message, agentsStarted: this.started }
+      }
+      // renderThrown is total (thrown values of any realm), so this arm
+      // cannot throw — drive() resolving is the `result` never-rejects seam
+      // contract.
+      return { value: null, stopReason: 'error', error: renderThrown(error), agentsStarted: this.started }
+    } finally {
+      // Reap strays: a script that fired agent() calls without awaiting them
+      // leaves live children behind after settlement — cancel them all. (The
+      // per-call wrappers dispose each child; the contain() consumer keeps
+      // their rejections from going unhandled.)
+      if (this.cancelReason === undefined) this.cancel('workflow settled')
+    }
+  }
+
+  /**
+   * Attach a no-op rejection consumer WITHOUT changing what the caller
+   * receives: if the script drops the promise (no await), cancellation cannot
+   * become an unhandled rejection (which would kill the worker thread); if
+   * the script does await it, it still observes the rejection.
+   */
+  private contain<T>(promise: Promise<T>): Promise<T> {
+    promise.catch(() => { /* consumed: see method contract — a dropped hook promise must not surface an unhandled rejection */ })
+    return promise
+  }
+
+  private cancelledError(): WorkflowError {
+    // cancel() arms cancelError before any caller can observe isCancelled()
+    // === true; the fallback guards the type, not a reachable path.
+    /* v8 ignore next */
+    return this.cancelError ?? new WorkflowError('workflow run cancelled', 'CANCELLED')
+  }
+
+  /** Materialize the script's return value; violations become RESULT_UNSERIALIZABLE. */
+  private materializeResult(raw: unknown): unknown {
+    try {
+      return materializeFromRealm(raw, 'workflow result')
+    } catch (error: unknown) {
+      /* v8 ignore next -- defensive rethrow arm: materializeFromRealm only throws MaterializeError */
+      if (!(error instanceof MaterializeError)) throw error
+      throw new WorkflowError(
+        `the workflow's return value is not plain JSON data — ${error.message}. Return only JSON-serializable objects/arrays/scalars.`,
+        'RESULT_UNSERIALIZABLE',
+        { cause: error },
+      )
+    }
+  }
+
+  /**
+   * Acquire one concurrency slot (FIFO). Cancellation rejects QUEUED waiters
+   * (see {@link cancel}); the callers guard their own entry and post-acquire
+   * windows, so no cancelled-precheck is duplicated here.
+   */
+  private acquireSlot(): Promise<void> {
+    if (this.activeSlots < this.limits.maxConcurrentAgents) {
+      this.activeSlots += 1
+      return Promise.resolve()
+    }
+    return new Promise<void>((resolve, reject) => {
+      this.slotWaiters.push({
+        resolve: () => {
+          this.activeSlots += 1
+          resolve()
+        },
+        reject,
+      })
+    })
+  }
+
+  private releaseSlot(): void {
+    this.activeSlots -= 1
+    const next = this.slotWaiters.shift()
+    if (next) next.resolve()
+  }
+
+  /** The `agent(prompt, opts)` hook. */
+  private async agent(rawPrompt: unknown, rawOpts: unknown): Promise<unknown> {
+    this.throwIfCancelled()
+    if (typeof rawPrompt !== 'string' || rawPrompt.length === 0) {
+      throw new WorkflowError('agent() requires a non-empty prompt string', 'INVALID_ARGUMENT')
+    }
+    const opts = this.readAgentOptions(rawOpts)
+    if (this.started >= this.limits.maxTotalAgents) {
+      throw new WorkflowError(
+        `this run reached its total agent cap (${this.limits.maxTotalAgents}) — a runaway-loop backstop; raise maxTotalAgents in the engine config if the scale is intentional`,
+        'AGENT_CAP',
+      )
+    }
+    this.started += 1
+    const seq = this.started
+    const label = opts.label ?? defaultLabel(rawPrompt)
+    const phase = opts.phase ?? this.currentPhase
+
+    await this.acquireSlot()
+    try {
+      // Re-check after the acquire: the await yields at least one microtask
+      // tick even when a slot is free, and a queued waiter resumes a tick
+      // after its release — a cancel() landing in either window must not
+      // reach the host (which would refuse anyway, but the refusal reads as
+      // a start failure rather than the cancellation it is).
+      this.throwIfCancelled()
+      let run: ChildHandle
+      try {
+        run = await this.children.startAgent({
+          prompt: rawPrompt,
+          ...opts.schema !== undefined ? { schema: opts.schema } : {},
+          ...opts.model !== undefined ? { model: opts.model } : {},
+        })
+      } catch (error: unknown) {
+        // The host refuses starts once the run is cancelled — a refusal that
+        // races our own cancel state must read as the cancellation it is,
+        // not as a broken seam.
+        if (this.isCancelled()) throw this.cancelledError()
+        throw new WorkflowError(`agent() could not start a child: ${renderThrown(error)}`, 'AGENT_START', { cause: error })
+      }
+      // The start round-trip yields to the event loop, so a cancel CAN land
+      // between the host starting the child and this continuation running —
+      // wind the fresh child down instead of leaving it live behind a dead
+      // script.
+      if (this.isCancelled()) {
+        run.cancel(this.cancelReason)
+        await run.dispose()
+        throw this.cancelledError()
+      }
+      const info: WorkflowAgentInfo = { seq, label, ...phase !== undefined ? { phase } : {}, childId: AgentId(run.id) }
+      this.observer.agentStart(info)
+      // Cancellation reaches the child through an explicit cancel RPC per
+      // child (the host also aborts its own per-run signal, but the seam
+      // leaves a provider free to honor either channel, so both are driven).
+      const onAbort = (): void => { run.cancel(this.cancelReason) }
+      this.controller.signal.addEventListener('abort', onAbort, { once: true })
+      try {
+        let result
+        try {
+          result = await run.result
+        } catch (error: unknown) {
+          // A rejected child result is an INFRASTRUCTURE fault relayed by the
+          // host — distinct from a child that failed and resolved. Pair the
+          // lifecycle before propagating, and propagate FATAL: an ordinary
+          // throw would dissolve to a per-item null inside the combinators,
+          // and a broken provider must not read as a failed child.
+          if (this.isCancelled()) {
+            this.observer.agentEnd({ ...info, outcome: 'cancelled' })
+            throw this.cancelledError()
+          }
+          this.observer.agentEnd({ ...info, outcome: 'failed' })
+          throw new WorkflowError(`child agent run failed: ${renderThrown(error)}`, 'AGENT_RESULT', { cause: error })
+        }
+        if (result.stopReason === 'completed') {
+          if (opts.schema !== undefined) {
+            // The provider honored outputSchema (capability-gated at start), so
+            // a completed run without a structured value is a child failure.
+            if (result.structured === undefined) {
+              this.observer.agentEnd({ ...info, outcome: 'failed' })
+              return null
+            }
+            this.observer.agentEnd({ ...info, outcome: 'completed' })
+            return result.structured
+          }
+          this.observer.agentEnd({ ...info, outcome: 'completed' })
+          return outputText(result.output)
+        }
+        // A cancelled RUN kills the script; a child that failed for its own
+        // reasons resolves null (scripts .filter(Boolean) per the CC contract).
+        if (this.isCancelled()) {
+          this.observer.agentEnd({ ...info, outcome: 'cancelled' })
+          throw this.cancelledError()
+        }
+        this.observer.agentEnd({ ...info, outcome: 'failed' })
+        return null
+      } finally {
+        this.controller.signal.removeEventListener('abort', onAbort)
+        await run.dispose()
+      }
+    } finally {
+      this.releaseSlot()
+    }
+  }
+
+  /** Materialize + validate the `agent()` options bag from the realm. */
+  private readAgentOptions(rawOpts: unknown): { label?: string; phase?: string; model?: string; schema?: StructuredOutputSchema } {
+    if (rawOpts === undefined) return {}
+    let opts: unknown
+    try {
+      opts = materializeFromRealm(rawOpts, 'agent() options')
+    } catch (error: unknown) {
+      /* v8 ignore next -- defensive rethrow arm: materializeFromRealm only throws MaterializeError */
+      if (!(error instanceof MaterializeError)) throw error
+      throw new WorkflowError(`agent() options must be plain JSON data — ${error.message}`, 'INVALID_ARGUMENT', { cause: error })
+    }
+    if (typeof opts !== 'object' || opts === null || Array.isArray(opts)) {
+      throw new WorkflowError('agent() options must be an object', 'INVALID_ARGUMENT')
+    }
+    const record = opts as Record<string, unknown>
+    for (const key of Object.keys(record)) {
+      if (SUPPORTED_AGENT_OPTIONS.has(key)) continue
+      if (DEFERRED_AGENT_OPTIONS.has(key)) {
+        throw new WorkflowError(`agent() option "${key}" is deferred and not supported by this engine (supported: label, phase, schema, model)`, 'UNSUPPORTED_OPTION')
+      }
+      throw new WorkflowError(`agent() option "${key}" is not recognized (supported: label, phase, schema, model)`, 'UNSUPPORTED_OPTION')
+    }
+    for (const key of ['label', 'phase', 'model'] as const) {
+      if (record[key] !== undefined && typeof record[key] !== 'string') {
+        throw new WorkflowError(`agent() option "${key}" must be a string`, 'INVALID_ARGUMENT')
+      }
+    }
+    let schema: StructuredOutputSchema | undefined
+    if (record.schema !== undefined) {
+      try {
+        assertSupportedOutputSchema(record.schema)
+        schema = record.schema
+      } catch (error: unknown) {
+        /* v8 ignore next -- defensive rethrow arm: assertSupportedOutputSchema only throws OutputSchemaError */
+        if (!(error instanceof OutputSchemaError)) throw error
+        throw new WorkflowError(`agent() schema is outside the supported subset — ${error.message}`, 'UNSUPPORTED_SCHEMA', { cause: error })
+      }
+    }
+    return {
+      ...record.label !== undefined ? { label: record.label as string } : {},
+      ...record.phase !== undefined ? { phase: record.phase as string } : {},
+      ...record.model !== undefined ? { model: record.model as string } : {},
+      ...schema !== undefined ? { schema } : {},
+    }
+  }
+
+  /** The `parallel(thunks)` hook: each thunk caught → `null`; fatal errors propagate. */
+  private async parallel(rawThunks: unknown): Promise<unknown[]> {
+    this.throwIfCancelled()
+    if (!Array.isArray(rawThunks)) {
+      throw new WorkflowError('parallel() requires an array of zero-argument functions', 'INVALID_ARGUMENT')
+    }
+    this.assertItemCap(rawThunks.length, 'parallel()')
+    const thunks = rawThunks.map((thunk, index) => {
+      if (typeof thunk !== 'function') {
+        throw new WorkflowError(`parallel() item ${index} is not a function`, 'INVALID_ARGUMENT')
+      }
+      return thunk as () => unknown
+    })
+    return Promise.all(thunks.map(async (thunk) => {
+      try {
+        return await thunk()
+      } catch (error: unknown) {
+        // Hook failures are WorkflowErrors built OUTSIDE the script's realm;
+        // fatality is recognized by `instanceof` against this realm's class —
+        // a script-built object can never pass it, so fatality cannot be
+        // forged (nor accidentally dissolved).
+        if (isFatalWorkflowError(error)) throw error
+        return null
+      }
+    }))
+  }
+
+  /** The `pipeline(items, ...stages)` hook: per-item stage chains, NO cross-stage barrier. */
+  private async pipeline(rawItems: unknown, rawStages: unknown[]): Promise<unknown[]> {
+    this.throwIfCancelled()
+    if (!Array.isArray(rawItems)) {
+      throw new WorkflowError('pipeline() requires an items array', 'INVALID_ARGUMENT')
+    }
+    this.assertItemCap(rawItems.length, 'pipeline()')
+    if (rawStages.length === 0) {
+      throw new WorkflowError('pipeline() requires at least one stage function', 'INVALID_ARGUMENT')
+    }
+    const stages = rawStages.map((stage, index) => {
+      if (typeof stage !== 'function') {
+        throw new WorkflowError(`pipeline() stage ${index} is not a function`, 'INVALID_ARGUMENT')
+      }
+      return stage as (previous: unknown, item: unknown, index: number) => unknown
+    })
+    return Promise.all(rawItems.map(async (item: unknown, index) => {
+      let value: unknown = item
+      try {
+        for (const stage of stages) {
+          value = await stage(value, item, index)
+        }
+        return value
+      } catch (error: unknown) {
+        // An ordinary stage throw drops the ITEM to null and skips its
+        // remaining stages; a fatal WorkflowError (see parallel()) kills the
+        // whole script.
+        if (isFatalWorkflowError(error)) throw error
+        return null
+      }
+    }))
+  }
+
+  private assertItemCap(length: number, hook: string): void {
+    if (length > this.limits.maxItemsPerCall) {
+      throw new WorkflowError(
+        `${hook} received ${length} items — over the per-call cap (${this.limits.maxItemsPerCall}); split the work or raise maxItemsPerCall in the engine config`,
+        'ITEM_CAP',
+      )
+    }
+  }
+
+  /** The `phase(title)` hook: sets the current label for subsequent `agent()` calls and notifies observers. */
+  private phase(title: unknown): void {
+    this.throwIfCancelled()
+    if (typeof title !== 'string' || title.length === 0) {
+      throw new WorkflowError('phase() requires a non-empty title string', 'INVALID_ARGUMENT')
+    }
+    this.currentPhase = title
+    this.observer.phase(title)
+  }
+
+  /** The `log(message)` hook: narration to observers. */
+  private log(message: unknown): void {
+    this.throwIfCancelled()
+    if (typeof message !== 'string') {
+      throw new WorkflowError('log() requires a message string', 'INVALID_ARGUMENT')
+    }
+    this.observer.log(message)
+  }
+}

packages/workflow/workflow-workerthread/src/session.ts

@@ -0,0 +1,210 @@
+/**
+ * The worker-side half of the engine: {@link runWorkerSession} wires one
+ * MessagePort to one {@link WorkflowExecution} — hook progress and child
+ * starts go out as messages, run control and child lifecycle come back in —
+ * and posts the run's terminal result exactly once. Deliberately separated
+ * from the thread bootstrap (./worker.ts): the whole session is drivable
+ * in-process over a `MessageChannel`, which is where its unit coverage lives
+ * (code inside a real Worker is invisible to the main process's coverage).
+ *
+ * Startup handshake: the session posts `ready` and runs the script only
+ * after the host's `go` — without it, a cancellation racing the worker's
+ * boot could arrive AFTER the script's initial synchronous slice already
+ * ran, and a run cancelled before start must not execute the body at all.
+ * A `cancel` arriving instead of `go` still releases the gate: `drive()`
+ * sees the cancelled state and settles without running the body.
+ *
+ * @module @deepseek-ai/dsh-workflow-workerthread/session
+ */
+
+import type { MessagePort } from 'node:worker_threads'
+import { assertNever } from '@deepseek-ai/dsh-llm'
+import { HostToWorkerType, WorkerToHostType } from './protocol.ts'
+import type { HostToWorkerMessage, WorkerToHostPayloads } from './protocol.ts'
+import { renderThrown } from './realm.ts'
+import { WorkflowExecution } from './runtime.ts'
+import type { ExecutionObserver } from './runtime.ts'
+import type {
+  ChildHandle,
+  ChildPort,
+  ChildResult,
+  ChildStartRequest,
+  WorkerInit,
+} from './types.ts'
+
+/** The book-keeping for one in-flight child RPC (keyed by callId). */
+interface PendingChild {
+  started: PromiseWithResolvers<string>
+  settled: PromiseWithResolvers<ChildResult>
+  disposed: PromiseWithResolvers<void>
+}
+
+/** The typed post half of the port: each tag pairs with ITS payload from the map (a mismatch is a compile error at the call site). */
+type Post = <T extends WorkerToHostType>(type: T, payload: WorkerToHostPayloads[T]) => void
+
+/**
+ * The worker-side handle for one started child agent ({@link ChildHandle}):
+ * every member is an RPC to the host keyed by this call's `callId`, resolved
+ * by the session's message handler through the bridge's pending entry.
+ */
+class RpcChildHandle implements ChildHandle {
+  readonly result: Promise<ChildResult>
+
+  constructor(
+    private readonly post: Post,
+    private readonly callId: number,
+    private readonly entry: PendingChild,
+    readonly id: string,
+  ) {
+    this.result = entry.settled.promise
+  }
+
+  cancel(reason?: string): void {
+    this.post(WorkerToHostType.ChildCancel, { callId: this.callId, reason })
+  }
+
+  dispose(): Promise<void> {
+    this.post(WorkerToHostType.ChildDispose, { callId: this.callId })
+    return this.entry.disposed.promise
+  }
+}
+
+/**
+ * The worker-side child-RPC bridge ({@link ChildPort}): allocates callIds,
+ * posts the start/cancel/dispose RPCs, and owns the per-call pending
+ * book-keeping the session's message handler settles via the `onChild*`
+ * entry points.
+ */
+class ChildRpcBridge implements ChildPort {
+  private nextCallId = 0
+  private readonly pending = new Map<number, PendingChild>()
+
+  constructor(private readonly post: Post) {}
+
+  async startAgent(request: ChildStartRequest): Promise<ChildHandle> {
+    this.nextCallId += 1
+    const callId = this.nextCallId
+    const entry: PendingChild = {
+      started: Promise.withResolvers<string>(),
+      settled: Promise.withResolvers<ChildResult>(),
+      disposed: Promise.withResolvers<void>(),
+    }
+    // Containment: when the start is refused (or the run torn down) the
+    // settled promise may never gain a consumer — it must not surface as an
+    // unhandled rejection and kill the worker.
+    entry.settled.promise.catch(() => { /* consumed: unconsumed child settlement after a refused start */ })
+    this.pending.set(callId, entry)
+    this.post(WorkerToHostType.ChildStart, { callId, request })
+    const childId = await entry.started.promise
+    return new RpcChildHandle(this.post, callId, entry, childId)
+  }
+
+  /** The host started the child; releases the `startAgent` await. */
+  onChildStarted(callId: number, childId: string): void {
+    this.pending.get(callId)?.started.resolve(childId)
+  }
+
+  /** The host refused the start; `startAgent` rejects with the rendered cause. */
+  onChildStartError(callId: number, rendered: string): void {
+    this.pending.get(callId)?.started.reject(new Error(rendered))
+  }
+
+  /** The child's terminal result arrived. */
+  onChildSettled(callId: number, result: ChildResult): void {
+    this.pending.get(callId)?.settled.resolve(result)
+  }
+
+  /** The child's `result` rejected host-side (an infrastructure fault, relayed as fatal). */
+  onChildFailed(callId: number, rendered: string): void {
+    this.pending.get(callId)?.settled.reject(new Error(rendered))
+  }
+
+  /** The host acked the dispose; the call's book-keeping is complete. */
+  onChildDisposed(callId: number): void {
+    const entry = this.pending.get(callId)
+    this.pending.delete(callId)
+    entry?.disposed.resolve()
+  }
+}
+
+/**
+ * Narrow the nullable `parentPort` the bootstrap reads from
+ * `node:worker_threads`.
+ * @param port - `parentPort` as imported (null on the main thread).
+ * @returns the port, non-null.
+ */
+export function requireParentPort(port: MessagePort | null): MessagePort {
+  if (port === null) throw new Error('the workflow worker entry must be loaded inside a worker thread (no parentPort)')
+  return port
+}
+
+/**
+ * Run one workflow script to settlement against `port`, posting the terminal
+ * result message exactly once; resolves after that post (stray children may
+ * still be winding down through the port — the host owns their teardown and
+ * ultimately terminates the thread). Never rejects: a constructor failure
+ * (unparseable body — host pre-parse makes this a Node-version-skew signal)
+ * is reported as an `error` result rather than dying without a result.
+ * @param port - the channel to the host (the real `parentPort`, or one side
+ *   of an in-process `MessageChannel` in tests).
+ * @param init - the run payload the host provided as `workerData`.
+ */
+export async function runWorkerSession(port: MessagePort, init: WorkerInit): Promise<void> {
+  const post: Post = (type, payload) => {
+    port.postMessage({ type, ...payload })
+  }
+  const children = new ChildRpcBridge(post)
+
+  const observer: ExecutionObserver = {
+    phase: (title) => { post(WorkerToHostType.Phase, { title }) },
+    log: (message) => { post(WorkerToHostType.Log, { message }) },
+    agentStart: (info) => { post(WorkerToHostType.AgentStart, { info }) },
+    agentEnd: (info) => { post(WorkerToHostType.AgentEnd, { info }) },
+  }
+
+  let execution: WorkflowExecution
+  try {
+    execution = new WorkflowExecution(init.meta, init.body, init.args, init.limits, observer, children)
+  } catch (error: unknown) {
+    post(WorkerToHostType.Result, { result: { value: null, stopReason: 'error', error: renderThrown(error), agentsStarted: 0 } })
+    return
+  }
+
+  const gate = Promise.withResolvers<void>()
+  port.on('message', (message: HostToWorkerMessage) => {
+    switch (message.type) {
+      case HostToWorkerType.Go:
+        gate.resolve()
+        break
+      case HostToWorkerType.Cancel:
+        execution.cancel(message.reason)
+        // A cancel doubles as the gate release: drive() checks the cancelled
+        // state before running the body, so the script never executes.
+        gate.resolve()
+        break
+      case HostToWorkerType.ChildStarted:
+        children.onChildStarted(message.callId, message.childId)
+        break
+      case HostToWorkerType.ChildStartError:
+        children.onChildStartError(message.callId, message.rendered)
+        break
+      case HostToWorkerType.ChildSettled:
+        children.onChildSettled(message.callId, message.result)
+        break
+      case HostToWorkerType.ChildFailed:
+        children.onChildFailed(message.callId, message.rendered)
+        break
+      case HostToWorkerType.ChildDisposed:
+        children.onChildDisposed(message.callId)
+        break
+      /* v8 ignore next 2 -- closed engine-owned union; the arm only makes adding a message type a compile error */
+      default:
+        assertNever(message, 'host-to-worker message')
+    }
+  })
+
+  post(WorkerToHostType.Ready, {})
+  await gate.promise
+  const result = await execution.drive()
+  post(WorkerToHostType.Result, { result })
+}

packages/workflow/workflow-workerthread/src/types.ts

@@ -0,0 +1,97 @@
+/**
+ * Non-protocol wire vocabulary for the worker-thread engine: the `workerData` init
+ * payload and the child-port interfaces the worker-side runtime consumes.
+ * The host⇄worker MESSAGE protocol lives in ./protocol.ts; everything here
+ * that a message transports (`ChildStartRequest`, `ChildResult`) is plain
+ * JSON data by construction, so the structured-clone hop never meets a value
+ * it cannot carry. Types only, per the package convention.
+ *
+ * @module @deepseek-ai/dsh-workflow-workerthread/types
+ */
+
+import type { ContentBlock } from '@deepseek-ai/dsh-llm'
+import type { StructuredOutputSchema } from '@deepseek-ai/dsh-tools'
+import type { WorkflowMeta } from '@deepseek-ai/dsh-workflow'
+
+/**
+ * The per-run limits the worker-side runtime enforces. The host keeps the
+ * knobs only it can act on (`provider`, `disposeGraceMs`).
+ */
+export interface WorkerLimits {
+  /** Concurrent `agent()` ceiling (already auto-resolved; ≥ 1). */
+  maxConcurrentAgents: number
+  /** Total `agent()` calls per run (the runaway-loop backstop). */
+  maxTotalAgents: number
+  /** Items accepted by one `parallel()`/`pipeline()` call. */
+  maxItemsPerCall: number
+  /** vm timeout for the script's initial synchronous slice (inside the worker). */
+  syncTimeoutMs: number
+}
+
+/** The `workerData` payload one run is initialized with (host → worker, once, at spawn). */
+export interface WorkerInit {
+  /** The validated meta block (plain data off the start request, validated host-side). */
+  meta: WorkflowMeta
+  /** The plain-JS script body, exactly as the start request carried it. */
+  body: string
+  /** The run's `args` value; the workerData structured clone is the copy that isolates the caller. */
+  args?: unknown
+  /** The worker-enforced limits. */
+  limits: WorkerLimits
+}
+
+/** What the worker asks the host to start for one `agent()` call (options already validated worker-side). */
+export interface ChildStartRequest {
+  /** The child's prompt text. */
+  prompt: string
+  /** The structured-output schema, if the call passed one (already subset-checked). */
+  schema?: StructuredOutputSchema
+  /** The per-child model override, if the call passed one. */
+  model?: string
+}
+
+/**
+ * The JSON projection of a child's `SubagentResult` crossing the port. The
+ * seam's `stopReason` union is merge-extensible, so it degrades to `string`
+ * on the wire — the runtime only ever branches on `'completed'`.
+ */
+export interface ChildResult {
+  /** The child's final assistant output blocks. */
+  output: ContentBlock[]
+  /** The structured value, present iff the request carried a schema AND the provider honored it. */
+  structured?: unknown
+  /** Why the child run ended (`'completed'` is the only value the runtime branches on). */
+  stopReason: string
+}
+
+/**
+ * The worker-side handle for one started child — the RPC mirror of the
+ * subagent seam's run handle, reduced to what the runtime consumes.
+ */
+export interface ChildHandle {
+  /** The child agent's id (minted host-side by the subagent seam). */
+  readonly id: string
+  /**
+   * Resolves with the child's terminal {@link ChildResult}; REJECTS only when
+   * the host reports an infrastructure fault (`child-failed`) — a child that
+   * failed for its own reasons resolves with a non-`completed` stop reason.
+   */
+  readonly result: Promise<ChildResult>
+  /** Ask the host to cancel the child (fire-and-forget). */
+  cancel(reason?: string): void
+  /** Ask the host to dispose the child; resolves on the host's ack. */
+  dispose(): Promise<void>
+}
+
+/**
+ * The worker-side port the runtime starts child agents through — the seam
+ * that lets the execution core stay ignorant of the thread boundary.
+ */
+export interface ChildPort {
+  /**
+   * Start one child agent on the host (the `agent()` hook's start half).
+   * @param request - the prompt and validated options.
+   * @returns the child handle; rejects when the host refuses the start.
+   */
+  startAgent(request: ChildStartRequest): Promise<ChildHandle>
+}

packages/workflow/workflow-workerthread/src/worker.ts

@@ -0,0 +1,18 @@
+/**
+ * The worker-thread entry the engine spawns: bootstrap ./session.ts on the
+ * real `parentPort`. Deliberately a single statement — every piece of logic
+ * lives in `runWorkerSession`, which the unit suite drives in-process over a
+ * `MessageChannel` (code inside a real Worker is invisible to main-process
+ * coverage); loading this module on the main thread throws via
+ * `requireParentPort`, which is how the suite covers the file itself.
+ *
+ * @module @deepseek-ai/dsh-workflow-workerthread/worker
+ */
+
+import { parentPort, workerData } from 'node:worker_threads'
+import { requireParentPort, runWorkerSession } from './session.ts'
+import type { WorkerInit } from './types.ts'
+
+// workerData is `any` at the node:worker_threads boundary; the engine is the
+// only spawner and always provides a WorkerInit.
+void runWorkerSession(requireParentPort(parentPort), workerData as WorkerInit)

packages/workflow/workflow-workerthread/tests/built-worker.e2e.ts

@@ -0,0 +1,57 @@
+import { existsSync } from 'node:fs'
+import { rm, writeFile } from 'node:fs/promises'
+import { join } from 'node:path'
+import { execFile } from 'node:child_process'
+import { promisify } from 'node:util'
+import { fileURLToPath } from 'node:url'
+import { describe, expect, it } from 'vitest'
+
+const packageRoot = fileURLToPath(new URL('..', import.meta.url))
+const builtIndex = join(packageRoot, 'lib', 'index.js')
+const builtWorker = join(packageRoot, 'lib', 'worker.js')
+const run = promisify(execFile)
+
+/**
+ * The BUILT-output guard for the worker entry: every other suite runs
+ * unbuilt (src/ + tsx), so nothing else proves that `lib/index.js` resolves
+ * its sibling `lib/worker.js` and that the bundle boots a worker under plain
+ * node (no tsx loader). Keyless — a zero-agent script needs no provider —
+ * and self-skips until `pnpm run build` has produced the bundles.
+ */
+describe.skipIf(!existsSync(builtIndex) || !existsSync(builtWorker))('built worker entry (lib/worker.js)', () => {
+  it('the built engine spawns its built worker under plain node and completes a run', async () => {
+    // ESM resolves bare specifiers from the IMPORTING FILE's location, so the
+    // driver must live inside the package for its node_modules to apply — a
+    // temp-named file at the package root, removed on the way out.
+    const driver = join(packageRoot, `.built-worker-driver-${process.pid}.mjs`)
+    try {
+      await writeFile(driver, `
+import { Context } from 'cordis'
+import SubagentService from '@deepseek-ai/dsh-subagent'
+import WorkerWorkflowEngine from '@deepseek-ai/dsh-workflow-workerthread'
+
+const ctx = new Context()
+await ctx.plugin(SubagentService)
+await ctx.plugin(WorkerWorkflowEngine, {})
+const run = ctx.workflows.start({
+  script: 'return 6 * 7',
+  meta: { name: 'built-smoke', description: 'built worker smoke' },
+  // A zero-agent script never touches the provider, so a bare id suffices.
+  parent: { id: 'built-smoke-parent', options: {} },
+})
+const result = await run.result
+await run.dispose()
+if (result.stopReason !== 'completed' || result.value !== 42) {
+  console.error('unexpected result: ' + JSON.stringify(result))
+  process.exit(1)
+}
+console.log('built-worker-smoke-ok')
+`, 'utf8')
+      // Plain node — no tsx loader anywhere; the bundle must stand on its own.
+      const { stdout } = await run(process.execPath, [driver], { cwd: packageRoot, timeout: 60_000 })
+      expect(stdout).toContain('built-worker-smoke-ok')
+    } finally {
+      await rm(driver, { force: true })
+    }
+  }, 120_000)
+})

packages/workflow/workflow-workerthread/tests/integration.spec.ts

@@ -0,0 +1,91 @@
+import { describe, expect, it } from 'vitest'
+import { Context } from 'cordis'
+import LlmService from '@deepseek-ai/dsh-llm'
+import SessionStore from '@deepseek-ai/dsh-session'
+import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
+import ToolRegistry from '@deepseek-ai/dsh-tools'
+import AgentRegistry, { AgentId } from '@deepseek-ai/dsh-agent'
+import AgentLoop from '@deepseek-ai/dsh-agent-loop'
+import * as Invariants from '@deepseek-ai/dsh-invariants'
+import SubagentService from '@deepseek-ai/dsh-subagent'
+import * as spawn from '@deepseek-ai/dsh-subagent-spawn'
+import { STRUCTURED_OUTPUT_TOOL } from '@deepseek-ai/dsh-subagent-inprocess'
+import { MockAdapter, textResponse, toolCallResponse } from '../../../core/agent-loop/tests/mock-adapter.ts'
+import WorkerWorkflowEngine from '../src/index.ts'
+
+type Script = ConstructorParameters<typeof MockAdapter>[0]
+
+/**
+ * The whole in-process stack, keyless, with the script in a REAL worker
+ * thread: the engine drives the REAL spawn backend (with its
+ * structured runtime) on a real agent loop; the scripted mock MODEL is the
+ * only mocked boundary. This is the guard the unit suites structurally
+ * cannot give — the MessageChannel suite fakes the host, and the host suite
+ * stubs the subagent seam.
+ */
+async function setup(script: Script) {
+  const ctx = new Context()
+  const adapter = new MockAdapter(script)
+  await ctx.plugin(LlmService)
+  await ctx.plugin(SessionStore)
+  await ctx.plugin(SystemPrompt)
+  await ctx.plugin(ToolRegistry)
+  await ctx.plugin(AgentRegistry)
+  await ctx.plugin(Invariants)
+  await ctx.plugin(AgentLoop, { agents: [] })
+  await ctx.plugin(SubagentService)
+  await ctx.plugin(spawn, { providerName: 'spawn' })
+  await ctx.plugin(WorkerWorkflowEngine, {})
+  ctx.llm.registerAdapter(['mock'], adapter)
+  const parent = ctx.agentLoop.create(AgentId('parent'), { model: 'mock' })
+  return { ctx, parent, adapter }
+}
+
+describe('dsh-workflow-workerthread over the real in-process stack', () => {
+  it('runs a two-stage workflow: a plain child, then a schema child through the structured runtime', async () => {
+    const { ctx, parent } = await setup([
+      textResponse('the file list is a.ts'),
+      toolCallResponse('c1', STRUCTURED_OUTPUT_TOOL, { verdict: 'real', confidence: 0.9 }),
+    ])
+    const childIds: string[] = []
+    ctx.on('workflow/agent-start', (_info, agent) => { childIds.push(agent.childId) })
+    const run = ctx.workflows.start({
+      meta: { name: 'integration', description: 'plain + structured children' },
+      script: `phase('Read')
+const prose = await agent('read the repo')
+phase('Judge')
+const judged = await agent('judge: ' + prose, {
+  schema: { type: 'object', properties: { verdict: { type: 'string', enum: ['real', 'bogus'] }, confidence: { type: 'number' } }, required: ['verdict'] },
+})
+return { prose, verdict: judged.verdict, confidence: judged.confidence }`,
+      parent,
+    })
+    const result = await run.result
+    expect(result.stopReason).toBe('completed')
+    expect(result.value).toEqual({ prose: 'the file list is a.ts', verdict: 'real', confidence: 0.9 })
+    expect(result.agentsStarted).toBe(2)
+    await run.dispose()
+    // Both children were disposed to quiescence — no live child agents remain.
+    expect(childIds.length).toBe(2)
+    for (const childId of childIds) {
+      expect(ctx.agents.get(AgentId(childId))).toBeUndefined()
+    }
+  })
+
+  it('a child that fails against its schema (nudges exhausted) reaches the script as null', async () => {
+    const { ctx, parent } = await setup([
+      textResponse('prose only'),
+      textResponse('still prose after the nudge'),
+    ])
+    const run = ctx.workflows.start({
+      meta: { name: 'null-path', description: 'schema failure maps to null' },
+      script: `const judged = await agent('judge it', { schema: { type: 'object', properties: { v: { type: 'string' } } } })
+return { got: judged === null ? 'null' : 'value' }`,
+      parent,
+    })
+    const result = await run.result
+    expect(result.stopReason).toBe('completed')
+    expect(result.value).toEqual({ got: 'null' })
+    await run.dispose()
+  })
+})

packages/workflow/workflow-workerthread/tests/meta.spec.ts

@@ -0,0 +1,88 @@
+import { describe, expect, it } from 'vitest'
+import { WorkflowError } from '@deepseek-ai/dsh-workflow'
+import { validateMeta } from '../src/meta.ts'
+
+/** Assert a META_INVALID throw whose message matches every given fragment. */
+function expectInvalid(value: unknown, ...fragments: string[]): void {
+  let thrown: unknown
+  try {
+    validateMeta(value)
+  } catch (error: unknown) {
+    thrown = error
+  }
+  expect(thrown).toBeInstanceOf(WorkflowError)
+  expect((thrown as WorkflowError).code).toBe('META_INVALID')
+  for (const fragment of fragments) {
+    expect((thrown as WorkflowError).message).toContain(fragment)
+  }
+}
+
+describe('validateMeta', () => {
+  it('accepts a minimal meta and returns a normalized copy (no aliasing of the input)', () => {
+    const input = { name: 'audit', description: 'audit the repo' }
+    const meta = validateMeta(input)
+    expect(meta).toEqual({ name: 'audit', description: 'audit the repo' })
+    expect(meta).not.toBe(input)
+    input.name = 'mutated'
+    expect(meta.name).toBe('audit')
+  })
+
+  it('accepts the full shape and rebuilds phases entry by entry', () => {
+    const meta = validateMeta({
+      name: 'migrate',
+      description: 'migrate call sites',
+      whenToUse: 'large mechanical sweeps',
+      phases: [
+        { title: 'Discover' },
+        { title: 'Transform', detail: 'one agent per file', model: 'deepseek-v4-pro' },
+      ],
+    })
+    expect(meta).toEqual({
+      name: 'migrate',
+      description: 'migrate call sites',
+      whenToUse: 'large mechanical sweeps',
+      phases: [
+        { title: 'Discover' },
+        { title: 'Transform', detail: 'one agent per file', model: 'deepseek-v4-pro' },
+      ],
+    })
+  })
+
+  it('rejects non-object values loud', () => {
+    expectInvalid(undefined, 'meta must be an object')
+    expectInvalid('a string', 'meta must be an object')
+    expectInvalid(null, 'meta must be an object')
+    expectInvalid([{ name: 'x', description: 'd' }], 'meta must be an object')
+  })
+
+  it('rejects unknown fields by name (accepted-then-ignored is banned)', () => {
+    expectInvalid({ name: 'x', description: 'd', color: 'red' }, 'meta.color is not a recognized field')
+  })
+
+  it('rejects missing or mistyped name/description/whenToUse', () => {
+    expectInvalid({ description: 'd' }, 'meta.name must be a non-empty string')
+    expectInvalid({ name: '', description: 'd' }, 'meta.name must be a non-empty string')
+    expectInvalid({ name: 'x' }, 'meta.description must be a non-empty string')
+    expectInvalid({ name: 'x', description: 42 }, 'meta.description must be a non-empty string')
+    expectInvalid({ name: 'x', description: 'd', whenToUse: 3 }, 'meta.whenToUse must be a string')
+  })
+
+  it('rejects malformed phases, entry by entry', () => {
+    expectInvalid({ name: 'x', description: 'd', phases: 'Scan' }, 'meta.phases must be an array')
+    expectInvalid({ name: 'x', description: 'd', phases: ['Scan'] }, 'meta.phases[0] must be an object')
+    expectInvalid({ name: 'x', description: 'd', phases: [{ title: '' }] }, 'meta.phases[0].title must be a non-empty string')
+    expectInvalid({ name: 'x', description: 'd', phases: [{ title: 'Scan', order: 1 }] }, 'meta.phases[0].order is not a recognized field')
+    expectInvalid({ name: 'x', description: 'd', phases: [{ title: 'Scan', detail: 9 }] }, 'meta.phases[0].detail must be a string')
+    expectInvalid({ name: 'x', description: 'd', phases: [{ title: 'Scan', model: 9 }] }, 'meta.phases[0].model must be a string')
+  })
+
+  it('names EVERY violation in one throw, not just the first', () => {
+    expectInvalid(
+      { description: 7, extra: true, phases: [{ title: 'Scan' }, 'bad'] },
+      'meta.extra is not a recognized field',
+      'meta.name must be a non-empty string',
+      'meta.description must be a non-empty string',
+      'meta.phases[1] must be an object',
+    )
+  })
+})

packages/workflow/workflow-workerthread/tests/realm.spec.ts

@@ -0,0 +1,145 @@
+import { describe, expect, it } from 'vitest'
+import * as vm from 'node:vm'
+import { materializeFromRealm, MaterializeError, renderThrown } from '../src/realm.ts'
+
+/** Evaluate an expression inside a fresh vm realm and hand back the raw realm value. */
+function inRealm(expression: string): unknown {
+  return vm.runInNewContext(`(${expression})`)
+}
+
+/** The MaterializeError message for a value that must be rejected (throws if accepted). */
+function rejection(value: unknown): string {
+  try {
+    materializeFromRealm(value)
+  } catch (error: unknown) {
+    if (error instanceof MaterializeError) return error.message
+    throw error
+  }
+  throw new Error('expected the value to be rejected')
+}
+
+describe('materializeFromRealm', () => {
+  it('copies realm objects/arrays/scalars into host plain data', () => {
+    const value = inRealm("{ a: 1, b: 'x', c: true, d: null, list: [1, [2, { deep: 'y' }]] }")
+    const out = materializeFromRealm(value) as Record<string, unknown>
+    expect(out).toEqual({ a: 1, b: 'x', c: true, d: null, list: [1, [2, { deep: 'y' }]] })
+    // The copy is HOST data: prototypes are the host intrinsics.
+    expect(Object.getPrototypeOf(out)).toBe(Object.prototype)
+    expect(Array.isArray(out.list)).toBe(true)
+    // And it round-trips through JSON byte-identically (the whole point).
+    expect(JSON.parse(JSON.stringify(out))).toEqual(out)
+  })
+
+  it('accepts undefined ONLY at the root (a valueless script return)', () => {
+    expect(materializeFromRealm(undefined)).toBeUndefined()
+    expect(rejection(inRealm('{ a: undefined }'))).toContain('value.a')
+  })
+
+  it('invokes getters ordinarily — the getter RESULT is what crosses (trust premise)', () => {
+    const counter = inRealm(`
+      (() => {
+        globalThis.reads = 0
+        return { get x() { globalThis.reads += 1; return globalThis.reads } }
+      })()
+    `)
+    expect(materializeFromRealm(counter)).toEqual({ x: 1 })
+  })
+
+  it('a getter that THROWS surfaces as a MaterializeError carrying the rendered failure', () => {
+    const hostile = inRealm("{ get x() { throw new Error('read failed') } }")
+    const message = rejection(hostile)
+    expect(message).toContain('reading the value threw')
+    expect(message).toContain('read failed')
+  })
+
+  it('a "__proto__" key becomes an OWN data property of the copy, never a prototype mutation', () => {
+    const value: unknown = vm.runInNewContext('JSON.parse(\'{"__proto__": {"polluted": 1}, "ok": 2}\')')
+    const out = materializeFromRealm(value) as Record<string, unknown>
+    expect(Object.getPrototypeOf(out)).toBe(Object.prototype)
+    expect(Object.prototype.hasOwnProperty.call(out, '__proto__')).toBe(true)
+    expect(out.ok).toBe(2)
+    // The host Object.prototype was NOT touched.
+    expect(({} as Record<string, unknown>).polluted).toBeUndefined()
+  })
+
+  it('rejects functions, symbols (keys and values), and bigints with path-qualified messages', () => {
+    expect(rejection(inRealm('{ fn: () => 1 }'))).toContain('value.fn')
+    expect(rejection(inRealm("{ [Symbol('k')]: 1 }"))).toContain('symbol-keyed')
+    expect(rejection(inRealm("{ s: Symbol('v') }"))).toContain('value.s')
+    expect(rejection(inRealm('{ big: 1n }'))).toContain('value.big')
+    expect(rejection(inRealm("[Symbol('x')]"))).toContain('value[0]')
+    const taggedArray = inRealm("(() => { const a = [1]; a[Symbol('t')] = 1; return a })()")
+    expect(rejection(taggedArray)).toContain('symbol-keyed')
+  })
+
+  it('rejects non-finite numbers and undefined values inside containers', () => {
+    expect(rejection(inRealm('{ n: NaN }'))).toContain('non-finite')
+    expect(rejection(inRealm('[Infinity]'))).toContain('non-finite')
+  })
+
+  it('rejects exotic prototypes (Date, Map, class instances) but accepts null-prototype data', () => {
+    expect(rejection(inRealm('{ d: new Date(0) }'))).toContain('exotic prototype')
+    expect(rejection(inRealm('new Map()'))).toContain('exotic prototype')
+    expect(rejection(inRealm('(() => { class C { constructor() { this.x = 1 } } return new C() })()')))
+      .toContain('exotic prototype')
+    expect(materializeFromRealm(inRealm('Object.assign(Object.create(null), { a: 1 })'))).toEqual({ a: 1 })
+  })
+
+  it('rejects cycles and accepts the same object reused as a sibling (a DAG)', () => {
+    expect(rejection(inRealm('(() => { const o = {}; o.self = o; return o })()'))).toContain('circular')
+    const dag = inRealm('(() => { const leaf = { v: 1 }; return { a: leaf, b: leaf } })()')
+    expect(materializeFromRealm(dag)).toEqual({ a: { v: 1 }, b: { v: 1 } })
+  })
+
+  it('rejects sparse arrays and non-index array properties; an array getter element materializes its value', () => {
+    expect(rejection(inRealm('[1, , 3]'))).toContain('sparse')
+    expect(rejection(inRealm('(() => { const a = [1]; a.total = 3; return a })()')))
+      .toContain('non-index')
+    expect(materializeFromRealm(inRealm('(() => { const a = [1]; Object.defineProperty(a, 0, { get: () => 7, enumerable: true }); return a })()')))
+      .toEqual([7])
+  })
+
+  it('skips non-enumerable own properties (matching JSON.stringify exactly)', () => {
+    const value = inRealm(`(() => {
+      const o = { visible: 1 }
+      Object.defineProperty(o, 'hidden', { value: () => 1, enumerable: false })
+      return o
+    })()`)
+    expect(materializeFromRealm(value)).toEqual({ visible: 1 })
+  })
+
+  it('works on plain host values too (the boundary is realm-agnostic)', () => {
+    expect(materializeFromRealm({ a: [1, 'x'] })).toEqual({ a: [1, 'x'] })
+    expect(materializeFromRealm('str')).toBe('str')
+    expect(materializeFromRealm(3)).toBe(3)
+    expect(materializeFromRealm(false)).toBe(false)
+    expect(materializeFromRealm(null)).toBeNull()
+  })
+})
+
+describe('renderThrown', () => {
+  it('prefers the stack, for host and realm errors alike', () => {
+    const host = renderThrown(new Error('host failure'))
+    expect(host).toContain('host failure')
+    expect(host).toContain('at ') // a real stack, not just the message
+    const realmError: unknown = vm.runInNewContext('(() => { try { throw new Error("realm failure") } catch (e) { return e } })()')
+    expect(renderThrown(realmError)).toContain('realm failure')
+  })
+
+  it('falls back from stack to message to String()', () => {
+    expect(renderThrown({ stack: 'custom data stack' })).toBe('custom data stack')
+    const stackless = new Error('stackless failure')
+    delete stackless.stack
+    expect(renderThrown(stackless)).toBe('stackless failure')
+    expect(renderThrown({ code: 42 })).toBe('[object Object]')
+    expect(renderThrown('plain')).toBe('plain')
+    expect(renderThrown(42)).toBe('42')
+    expect(renderThrown(undefined)).toBe('undefined')
+    expect(renderThrown(null)).toBe('null')
+  })
+
+  it('is total: a value whose accessors/toString throw renders as a fixed label', () => {
+    expect(renderThrown({ get stack() { throw new Error('nope') } })).toBe('[unrenderable thrown value]')
+    expect(renderThrown({ [Symbol.toPrimitive]() { throw new Error('nope') } })).toBe('[unrenderable thrown value]')
+  })
+})

packages/workflow/workflow-workerthread/tests/session.spec.ts

@@ -0,0 +1,504 @@
+import { describe, expect, it, vi } from 'vitest'
+import { MessageChannel } from 'node:worker_threads'
+import type { MessagePort } from 'node:worker_threads'
+import { HostToWorkerType, WorkerToHostType } from '../src/protocol.ts'
+import type { HostToWorkerMessage, WorkerToHostMessage } from '../src/protocol.ts'
+import { requireParentPort, runWorkerSession } from '../src/session.ts'
+import type { ChildResult, WorkerInit } from '../src/types.ts'
+
+/** Default limits for in-process sessions (concurrency pinned; auto is machine-derived). */
+function limits(overrides?: Partial<WorkerInit['limits']>): WorkerInit['limits'] {
+  return { maxConcurrentAgents: 8, maxTotalAgents: 1000, maxItemsPerCall: 4096, syncTimeoutMs: 5000, ...overrides }
+}
+
+/** Wrap a body in the minimal valid meta header (the session receives it pre-extracted). */
+function init(body: string, args?: unknown, limitOverrides?: Partial<WorkerInit['limits']>): WorkerInit {
+  return {
+    meta: { name: 'test-flow', description: 'a test workflow' },
+    body,
+    ...args !== undefined ? { args } : {},
+    limits: limits(limitOverrides),
+  }
+}
+
+/** One scripted host over the other end of a MessageChannel. */
+interface FakeHost {
+  port: MessagePort
+  messages: WorkerToHostMessage[]
+  /** Messages of one type, as they arrive. */
+  ofType<T extends WorkerToHostMessage['type']>(type: T): Extract<WorkerToHostMessage, { type: T }>[]
+  send(message: HostToWorkerMessage): void
+  /** Resolves with the terminal result message. */
+  result(): Promise<Extract<WorkerToHostMessage, { type: 'result' }>['result']>
+  close(): void
+}
+
+interface FakeHostOptions {
+  /** Auto-respond to child-start: reply started + settled per child index. Omit a reply to leave the child pending. */
+  reply?: (request: { prompt: string; schema?: unknown; model?: string }, index: number) => ChildResult | undefined
+  /** Reject the start instead (child-start-error) when returning a string. */
+  refuse?: (index: number) => string | undefined
+  /** Auto-send `go` on `ready` (default true). */
+  go?: boolean
+  /** Manual mode: do NOT auto-answer child-start at all (the test scripts the replies). */
+  manual?: boolean
+}
+
+/**
+ * Drive runWorkerSession IN-PROCESS over a MessageChannel: this is where the
+ * worker-side files earn their coverage — code inside a real Worker is
+ * invisible to main-process coverage. The fake host mirrors the real host's
+ * protocol discipline (one started/start-error per start; settled/disposed
+ * follow).
+ */
+function fakeHost(options?: FakeHostOptions): FakeHost {
+  const channel = new MessageChannel()
+  const messages: WorkerToHostMessage[] = []
+  const resultGate = Promise.withResolvers<Extract<WorkerToHostMessage, { type: 'result' }>['result']>()
+  let childIndex = 0
+  channel.port1.on('message', (message: WorkerToHostMessage) => {
+    messages.push(message)
+    switch (message.type) {
+      case WorkerToHostType.Ready:
+        if (options?.go !== false) channel.port1.postMessage({ type: HostToWorkerType.Go } satisfies HostToWorkerMessage)
+        break
+      case WorkerToHostType.ChildStart: {
+        if (options?.manual) break
+        const index = childIndex
+        childIndex += 1
+        const refusal = options?.refuse?.(index)
+        if (refusal !== undefined) {
+          channel.port1.postMessage(
+            { type: HostToWorkerType.ChildStartError, callId: message.callId, rendered: refusal } satisfies HostToWorkerMessage,
+          )
+          break
+        }
+        channel.port1.postMessage({ type: HostToWorkerType.ChildStarted, callId: message.callId, childId: `child-${index}` } satisfies HostToWorkerMessage)
+        const reply = options?.reply?.(message.request, index)
+        if (reply !== undefined) {
+          channel.port1.postMessage(
+            { type: HostToWorkerType.ChildSettled, callId: message.callId, result: reply } satisfies HostToWorkerMessage,
+          )
+        }
+        break
+      }
+      case WorkerToHostType.ChildDispose:
+        channel.port1.postMessage({ type: HostToWorkerType.ChildDisposed, callId: message.callId } satisfies HostToWorkerMessage)
+        break
+      case WorkerToHostType.Result:
+        resultGate.resolve(message.result)
+        break
+      default:
+        break
+    }
+  })
+  return {
+    port: channel.port2,
+    messages,
+    ofType: type => messages.filter((message): message is never => message.type === type),
+    send: (message) => { channel.port1.postMessage(message) },
+    result: () => resultGate.promise,
+    close: () => { channel.port1.close() },
+  }
+}
+
+/** A completed text child result. */
+function text(reply: string): ChildResult {
+  return { output: [{ type: 'text', text: reply }], stopReason: 'completed' }
+}
+
+describe('runWorkerSession over an in-process MessageChannel', () => {
+  it('runs a script end to end: ready/go handshake, phases, log, agents, result', async () => {
+    const host = fakeHost({ reply: (_request, index) => text(`answer-${index}`) })
+    const session = runWorkerSession(host.port, init(`
+      phase('Scan')
+      log('starting with ' + args.files.length + ' files')
+      const answers = await pipeline(args.files, (prev, item) => agent('read ' + item))
+      return { answers }
+    `, { files: ['a.ts', 'b.ts'] }))
+    const result = await host.result()
+    await session
+    expect(result.stopReason).toBe('completed')
+    expect(result.agentsStarted).toBe(2)
+    expect(result.value).toEqual({ answers: ['answer-0', 'answer-1'] })
+    expect(host.messages[0]!.type).toBe('ready')
+    expect(host.ofType(WorkerToHostType.Phase).map(m => m.title)).toEqual(['Scan'])
+    expect(host.ofType(WorkerToHostType.Log).map(m => m.message)).toEqual(['starting with 2 files'])
+    expect(host.ofType(WorkerToHostType.AgentStart).map(m => m.info.childId)).toEqual(['child-0', 'child-1'])
+    expect(host.ofType(WorkerToHostType.AgentEnd).every(m => m.info.outcome === 'completed')).toBe(true)
+    host.close()
+  })
+
+  it('agent({schema}) forwards the schema on the start request and returns the structured value', async () => {
+    const host = fakeHost({ reply: () => ({ output: [], structured: { files: ['x.ts'] }, stopReason: 'completed' }) })
+    void runWorkerSession(host.port, init(`
+      const found = await agent('list files', { schema: { type: 'object', properties: { files: { type: 'array', items: { type: 'string' } } } }, model: 'deepseek-v4-pro' })
+      return { first: found.files[0] }
+    `))
+    const result = await host.result()
+    expect(result.value).toEqual({ first: 'x.ts' })
+    const start = host.ofType(WorkerToHostType.ChildStart)[0]!
+    expect(start.request.schema).toEqual({ type: 'object', properties: { files: { type: 'array', items: { type: 'string' } } } })
+    expect(start.request.model).toBe('deepseek-v4-pro')
+    host.close()
+  })
+
+  it('a schema child completing WITHOUT a structured value resolves null with a failed outcome', async () => {
+    const host = fakeHost({ reply: () => text('prose, no structure') })
+    void runWorkerSession(host.port, init("return await agent('p', { schema: { type: 'object' } })"))
+    const result = await host.result()
+    expect(result.value).toBeNull()
+    expect(host.ofType(WorkerToHostType.AgentEnd)[0]!.info.outcome).toBe('failed')
+    host.close()
+  })
+
+  it('a child settling non-completed resolves null (scripts filter), never throwing into the script', async () => {
+    const host = fakeHost({ reply: (_request, index) => index === 0 ? { output: [], stopReason: 'error' } : text('ok') })
+    void runWorkerSession(host.port, init("return await parallel([() => agent('one'), () => agent('two')])"))
+    const result = await host.result()
+    expect(result.value).toEqual([null, 'ok'])
+    expect(host.ofType(WorkerToHostType.AgentEnd).map(m => m.info.outcome)).toEqual(expect.arrayContaining(['failed', 'completed']))
+    host.close()
+  })
+
+  it('a start refusal (child-start-error) is a fatal AGENT_START that kills the script through a combinator', async () => {
+    const host = fakeHost({ refuse: () => 'no provider here' })
+    void runWorkerSession(host.port, init("return await pipeline([1], () => agent('p'))"))
+    const result = await host.result()
+    expect(result.stopReason).toBe('error')
+    expect(result.error).toContain('agent() could not start a child')
+    expect(result.error).toContain('no provider here')
+    host.close()
+  })
+
+  it('a child-failed message (infrastructure rejection) is fatal AGENT_RESULT with the paired failed outcome', async () => {
+    const host = fakeHost()
+    void runWorkerSession(host.port, init(`
+      try { await agent('p'); return 'unreachable' } catch (e) { return { name: e.name, code: e.code, fatal: e.fatal } }
+    `))
+    await vi.waitFor(() => { expect(host.ofType(WorkerToHostType.ChildStart).length).toBe(1) })
+    const callId = host.ofType(WorkerToHostType.ChildStart)[0]!.callId
+    host.send({ type: HostToWorkerType.ChildStarted, callId, childId: 'child-0' })
+    host.send({ type: HostToWorkerType.ChildFailed, callId, rendered: 'backend exploded' })
+    const result = await host.result()
+    expect(result.value).toMatchObject({ name: 'WorkflowError', code: 'AGENT_RESULT', fatal: true })
+    expect(host.ofType(WorkerToHostType.AgentEnd)[0]!.info.outcome).toBe('failed')
+    host.close()
+  })
+
+  it('cancel before go: the body never runs at all and the result is cancelled (a second cancel is a no-op)', async () => {
+    const host = fakeHost({ go: false })
+    const session = runWorkerSession(host.port, init("log('ran')\nreturn 123"))
+    await vi.waitFor(() => { expect(host.messages.some(m => m.type === WorkerToHostType.Ready)).toBe(true) })
+    host.send({ type: HostToWorkerType.Cancel, reason: 'aborted before start' })
+    // Idempotence: the first reason wins; a duplicate cancel changes nothing.
+    host.send({ type: HostToWorkerType.Cancel, reason: 'a later reason that must lose' })
+    const result = await host.result()
+    await session
+    expect(result.stopReason).toBe('cancelled')
+    expect(result.error).toContain('aborted before start')
+    expect(result.error).not.toContain('must lose')
+    expect(result.value).toBeNull()
+    expect(host.ofType(WorkerToHostType.Log)).toEqual([])
+    host.close()
+  })
+
+  it('a script with no return value resolves value: null', async () => {
+    const host = fakeHost({ reply: () => text('ok') })
+    void runWorkerSession(host.port, init("await agent('p')"))
+    const result = await host.result()
+    expect(result.stopReason).toBe('completed')
+    expect(result.value).toBeNull()
+    host.close()
+  })
+
+  it('cancel mid-run: in-flight children get cancel RPCs, hooks throw at entry, the run reports cancelled', async () => {
+    const host = fakeHost()
+    void runWorkerSession(host.port, init(`
+      phase('before')
+      try { await agent('x') } catch (e) {}
+      try { phase('after') } catch (e) {}
+      try { log('after') } catch (e) {}
+      try { await parallel([() => 'ran']) } catch (e) {}
+      try { await pipeline(['item'], p => p) } catch (e) {}
+      return 'survived by catching'
+    `))
+    await vi.waitFor(() => { expect(host.ofType(WorkerToHostType.ChildStart).length).toBe(1) })
+    const callId = host.ofType(WorkerToHostType.ChildStart)[0]!.callId
+    host.send({ type: HostToWorkerType.ChildStarted, callId, childId: 'child-0' })
+    host.send({ type: HostToWorkerType.Cancel, reason: 'stop everything' })
+    // The real host settles the aborted child; mirror it.
+    host.send({ type: HostToWorkerType.ChildSettled, callId, result: { output: [], stopReason: 'aborted' } })
+    const result = await host.result()
+    expect(result.stopReason).toBe('cancelled')
+    expect(result.error).toContain('stop everything')
+    expect(host.ofType(WorkerToHostType.ChildCancel).map(m => m.callId)).toContain(callId)
+    expect(host.ofType(WorkerToHostType.AgentEnd)[0]!.info.outcome).toBe('cancelled')
+    // No post-cancel narration left the runtime (the hooks threw at entry).
+    expect(host.ofType(WorkerToHostType.Phase).map(m => m.title)).toEqual(['before'])
+    expect(host.ofType(WorkerToHostType.Log)).toEqual([])
+    host.close()
+  })
+
+  it('cancellation between a queued waiter and its slot: the waiter rejects without a child-start', async () => {
+    const host = fakeHost({ go: true })
+    void runWorkerSession(host.port, init(
+      "return await parallel([() => agent('a'), () => agent('b')])",
+      undefined,
+      { maxConcurrentAgents: 1 },
+    ))
+    await vi.waitFor(() => { expect(host.ofType(WorkerToHostType.ChildStart).length).toBe(1) })
+    host.send({ type: HostToWorkerType.Cancel, reason: 'raced' })
+    const result = await host.result()
+    expect(result.stopReason).toBe('cancelled')
+    // Only the first agent ever reached the host.
+    expect(host.ofType(WorkerToHostType.ChildStart).length).toBe(1)
+    host.close()
+  })
+
+  it('a stray (never-awaited) agent is reaped after settlement: cancel + dispose RPCs flow, no unhandled rejection', async () => {
+    const unhandled: unknown[] = []
+    const onUnhandled = (reason: unknown): void => { unhandled.push(reason) }
+    process.on('unhandledRejection', onUnhandled)
+    try {
+      const host = fakeHost()
+      void runWorkerSession(host.port, init(`
+        agent('stray, never awaited')
+        return 'done without awaiting'
+      `))
+      const result = await host.result()
+      expect(result.stopReason).toBe('completed')
+      await vi.waitFor(() => { expect(host.ofType(WorkerToHostType.ChildStart).length).toBe(1) })
+      const callId = host.ofType(WorkerToHostType.ChildStart)[0]!.callId
+      host.send({ type: HostToWorkerType.ChildStarted, callId, childId: 'child-0' })
+      host.send({ type: HostToWorkerType.ChildSettled, callId, result: { output: [], stopReason: 'aborted' } })
+      await vi.waitFor(() => { expect(host.ofType(WorkerToHostType.ChildDispose).map(m => m.callId)).toContain(callId) })
+      await new Promise(resolve => setTimeout(resolve, 20))
+      expect(unhandled).toEqual([])
+      host.close()
+    } finally {
+      process.off('unhandledRejection', onUnhandled)
+    }
+  })
+
+  it('an unparseable body settles an error result instead of dying without one (host pre-parse skew guard)', async () => {
+    const host = fakeHost()
+    await runWorkerSession(host.port, init('return ((('))
+    const result = await host.result()
+    expect(result.stopReason).toBe('error')
+    expect(result.error).toContain('does not parse')
+    expect(result.agentsStarted).toBe(0)
+    host.close()
+  })
+
+  it('a synchronous spin in the initial slice dies by the in-worker vm timeout', async () => {
+    const host = fakeHost()
+    void runWorkerSession(host.port, init('while (true) {}', undefined, { syncTimeoutMs: 50 }))
+    const result = await host.result()
+    expect(result.stopReason).toBe('error')
+    expect(result.error?.toLowerCase()).toContain('timed out')
+    host.close()
+  })
+
+  it('a non-JSON return value fails loud as RESULT_UNSERIALIZABLE', async () => {
+    const host = fakeHost()
+    void runWorkerSession(host.port, init('return { when: new Date(0) }'))
+    const result = await host.result()
+    expect(result.stopReason).toBe('error')
+    expect(result.error).toContain('not plain JSON data')
+    host.close()
+  })
+
+  it('tolerates replies for unknown callIds (a teardown race): nothing crashes, the run completes', async () => {
+    const host = fakeHost({ reply: () => text('fine') })
+    void runWorkerSession(host.port, init("return await agent('p')"))
+    host.send({ type: HostToWorkerType.ChildStarted, callId: 999, childId: 'ghost' })
+    host.send({ type: HostToWorkerType.ChildStartError, callId: 999, rendered: 'ghost' })
+    host.send({ type: HostToWorkerType.ChildSettled, callId: 999, result: text('ghost') })
+    host.send({ type: HostToWorkerType.ChildFailed, callId: 999, rendered: 'ghost' })
+    host.send({ type: HostToWorkerType.ChildDisposed, callId: 999 })
+    const result = await host.result()
+    expect(result.stopReason).toBe('completed')
+    expect(result.value).toBe('fine')
+    host.close()
+  })
+
+  it('caps and malformed hook arguments reject loud (the runtime runs unchanged inside the session)', async () => {
+    const cases: [string, string][] = [
+      ['return await agent(42)', 'non-empty prompt string'],
+      ["return await agent('')", 'non-empty prompt string'],
+      ["return await agent('p', 'opts')", 'options must be an object'],
+      ["return await agent('p', { label: 3 })", '"label" must be a string'],
+      ["return await agent('p', { get label() { throw new Error('read failed') } })", 'options must be plain JSON data'],
+      ["return await agent('p', { bogus: true })", '"bogus" is not recognized'],
+      ["return await agent('p', { effort: 'high' })", '"effort" is deferred'],
+      ["return await agent('p', { schema: { type: 'object', oneOf: [] } })", 'outside the supported subset'],
+      ['return await parallel([() => 1, () => 2, () => 3])', 'over the per-call cap (2)'],
+      ['return await pipeline([1, 2, 3], (x) => x)', 'maxItemsPerCall'],
+      ["return await parallel('no')", 'parallel() requires an array'],
+      ['return await parallel([3])', 'item 0 is not a function'],
+      ["return await pipeline('no', () => 1)", 'pipeline() requires an items array'],
+      ['return await pipeline([1])', 'at least one stage'],
+      ["return await pipeline([1], 'x')", 'stage 0 is not a function'],
+      ["phase('')", 'phase() requires a non-empty title string'],
+      ['log(3)', 'log() requires a message string'],
+    ]
+    for (const [body, expected] of cases) {
+      const host = fakeHost({ reply: () => text('ok') })
+      void runWorkerSession(host.port, init(body, undefined, { maxItemsPerCall: 2 }))
+      const result = await host.result()
+      expect(result.stopReason).toBe('error')
+      expect(result.error).toContain(expected)
+      host.close()
+    }
+  })
+
+  it('combinator semantics: thunk/stage throws null the item; a forged fatal-shaped object stays null; real fatals propagate', async () => {
+    const host = fakeHost({ reply: () => text('fine') })
+    void runWorkerSession(host.port, init(`
+      const viaParallel = await parallel([
+        () => { throw new Error('boom') },
+        () => agent('fine'),
+        () => 'plain value',
+        () => { throw { name: 'WorkflowError', fatal: true, message: 'forged fatal' } },
+      ])
+      const viaPipeline = await pipeline([10, 20],
+        (prev, item, index) => { if (item === 10) throw new Error('ordinary failure'); return 'kept-' + item + '-' + index },
+      )
+      return { viaParallel, viaPipeline }
+    `))
+    const result = await host.result()
+    expect(result.stopReason).toBe('completed')
+    expect(result.value).toEqual({
+      viaParallel: [null, 'fine', 'plain value', null],
+      viaPipeline: [null, 'kept-20-1'],
+    })
+    host.close()
+  })
+
+  it('trips the total-agent cap with a message naming the config knob', async () => {
+    const host = fakeHost({ reply: () => text('ok') })
+    void runWorkerSession(host.port, init("await agent('1'); await agent('2'); await agent('3')", undefined, { maxTotalAgents: 2 }))
+    const result = await host.result()
+    expect(result.stopReason).toBe('error')
+    expect(result.error).toContain('total agent cap (2)')
+    expect(result.agentsStarted).toBe(2)
+    host.close()
+  })
+
+  it('queued agents proceed through the concurrency semaphore in FIFO order', async () => {
+    const host = fakeHost({ reply: request => text(`ok:${request.prompt}`) })
+    void runWorkerSession(host.port, init(
+      "return await parallel([1, 2, 3].map((n) => () => agent('job ' + n)))",
+      undefined,
+      { maxConcurrentAgents: 1 },
+    ))
+    const result = await host.result()
+    expect(result.value).toEqual(['ok:job 1', 'ok:job 2', 'ok:job 3'])
+    host.close()
+  })
+
+  it('labels default from the prompt first line, truncated; explicit label/phase options win', async () => {
+    const host = fakeHost({ reply: () => text('ok') })
+    void runWorkerSession(host.port, init(`
+      phase('Find')
+      await agent('a prompt that is quite long and will surely get truncated down to a display label\\n'
+        + 'with a second line the label must not include')
+      await agent('short', { label: 'named', phase: 'Custom' })
+      return null
+    `))
+    await host.result()
+    const starts = host.ofType(WorkerToHostType.AgentStart).map(m => m.info)
+    expect(starts[0]).toMatchObject({ seq: 1, phase: 'Find' })
+    expect(starts[0]!.label.length).toBeLessThanOrEqual(48)
+    expect(starts[0]!.label).not.toContain('second line')
+    expect(starts[1]).toMatchObject({ seq: 2, label: 'named', phase: 'Custom' })
+    host.close()
+  })
+
+  it('non-text output blocks are filtered out of the text result', async () => {
+    const host = fakeHost({
+      reply: () => ({
+        output: [
+          { type: 'text', text: 'first ' },
+          { type: 'tool_call', id: 'c1', name: 'x', arguments: {} } as never,
+          { type: 'text', text: 'second' },
+        ],
+        stopReason: 'completed',
+      }),
+    })
+    void runWorkerSession(host.port, init("return await agent('p')"))
+    const result = await host.result()
+    expect(result.value).toBe('first second')
+    host.close()
+  })
+
+  it('a cancel landing DURING the start round-trip winds the fresh child down (cancel + dispose) and dies cancelled', async () => {
+    const host = fakeHost({ manual: true })
+    void runWorkerSession(host.port, init("return await agent('p')"))
+    await vi.waitFor(() => { expect(host.ofType(WorkerToHostType.ChildStart).length).toBe(1) })
+    const callId = host.ofType(WorkerToHostType.ChildStart)[0]!.callId
+    // Cancel FIRST, then the (stale) started reply: the worker processes them
+    // in order, so the agent() continuation resumes already-cancelled — the
+    // window the real host cannot produce (it refuses starts once cancelled)
+    // but a teardown race can.
+    host.send({ type: HostToWorkerType.Cancel, reason: 'raced the start' })
+    host.send({ type: HostToWorkerType.ChildStarted, callId, childId: 'child-0' })
+    const result = await host.result()
+    expect(result.stopReason).toBe('cancelled')
+    await vi.waitFor(() => {
+      expect(host.ofType(WorkerToHostType.ChildCancel).map(m => m.callId)).toContain(callId)
+      expect(host.ofType(WorkerToHostType.ChildDispose).map(m => m.callId)).toContain(callId)
+    })
+    // The child never became an agent-start: it was wound down pre-lifecycle.
+    expect(host.ofType(WorkerToHostType.AgentStart)).toEqual([])
+    host.close()
+  })
+
+  it('a start refusal arriving after a cancel reads as the cancellation, not a broken seam', async () => {
+    const host = fakeHost({ manual: true })
+    void runWorkerSession(host.port, init(`
+      try { await agent('p'); return 'unreachable' } catch (e) { return { code: e.code } }
+    `))
+    await vi.waitFor(() => { expect(host.ofType(WorkerToHostType.ChildStart).length).toBe(1) })
+    const callId = host.ofType(WorkerToHostType.ChildStart)[0]!.callId
+    host.send({ type: HostToWorkerType.Cancel, reason: 'stopping' })
+    host.send({ type: HostToWorkerType.ChildStartError, callId, rendered: 'workflow run cancelled: stopping' })
+    const result = await host.result()
+    // The run reports cancelled (the script died of CANCELLED, not AGENT_START).
+    expect(result.stopReason).toBe('cancelled')
+    host.close()
+  })
+
+  it('a child result rejection while cancelled pairs a cancelled agent-end, and the run reports cancelled', async () => {
+    const host = fakeHost({ manual: true })
+    void runWorkerSession(host.port, init("return await agent('doomed')"))
+    await vi.waitFor(() => { expect(host.ofType(WorkerToHostType.ChildStart).length).toBe(1) })
+    const callId = host.ofType(WorkerToHostType.ChildStart)[0]!.callId
+    host.send({ type: HostToWorkerType.ChildStarted, callId, childId: 'child-0' })
+    await vi.waitFor(() => { expect(host.ofType(WorkerToHostType.AgentStart).length).toBe(1) })
+    host.send({ type: HostToWorkerType.Cancel, reason: 'user aborted' })
+    host.send({ type: HostToWorkerType.ChildFailed, callId, rendered: 'backend crashed on abort' })
+    const result = await host.result()
+    expect(result.stopReason).toBe('cancelled')
+    expect(host.ofType(WorkerToHostType.AgentEnd)[0]!.info.outcome).toBe('cancelled')
+    host.close()
+  })
+
+})
+
+describe('the worker bootstrap', () => {
+  it('requireParentPort narrows a real port and throws on the main thread', () => {
+    const channel = new MessageChannel()
+    expect(requireParentPort(channel.port1)).toBe(channel.port1)
+    channel.port1.close()
+    expect(() => requireParentPort(null)).toThrow(/inside a worker thread/)
+  })
+
+  it('the entry module itself throws when loaded on the main thread (no parentPort)', async () => {
+    // This import EXECUTES ../src/worker.ts on the main thread, which is what
+    // covers the bootstrap file: requireParentPort throws before
+    // runWorkerSession is reached.
+    await expect(import('../src/worker.ts')).rejects.toThrow(/inside a worker thread/)
+  })
+})

packages/workflow/workflow-workerthread/tests/workflow-workerthread.e2e.ts

@@ -0,0 +1,102 @@
+import { afterEach, describe, expect, it } from 'vitest'
+import { Context } from 'cordis'
+import LlmService from '@deepseek-ai/dsh-llm'
+import SessionStore from '@deepseek-ai/dsh-session'
+import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
+import ToolRegistry from '@deepseek-ai/dsh-tools'
+import AgentRegistry, { AgentId } from '@deepseek-ai/dsh-agent'
+import AgentLoop from '@deepseek-ai/dsh-agent-loop'
+import * as LlmDeepSeek from '@deepseek-ai/dsh-llm-deepseek'
+import SubagentService from '@deepseek-ai/dsh-subagent'
+import * as Spawn from '@deepseek-ai/dsh-subagent-spawn'
+import WorkerWorkflowEngine from '../src/index.ts'
+
+/**
+ * With-key e2e: a REAL script in a REAL worker thread
+ * drives REAL spawn children against the live DeepSeek API — one plain child
+ * and one schema'd child through the real structured-output runtime — and
+ * the run's value, events, and child sessions are asserted from the outside
+ * (never the script's self-report alone). Key-gated (self-skips without
+ * DEEPSEEK_API_KEY).
+ */
+
+let ctx: Context | undefined
+
+afterEach(async () => {
+  await ctx?.fiber.dispose()
+  ctx = undefined
+})
+
+async function harness(): Promise<Context> {
+  const built = new Context()
+  await built.plugin(LlmService)
+  await built.plugin(SessionStore)
+  await built.plugin(SystemPrompt)
+  await built.plugin(ToolRegistry)
+  await built.plugin(AgentRegistry)
+  await built.plugin(AgentLoop, { agents: [] })
+  await built.plugin(LlmDeepSeek, { models: ['deepseek-v4-flash'] })
+  await built.plugin(SubagentService)
+  await built.plugin(Spawn, { providerName: 'spawn' })
+  await built.plugin(WorkerWorkflowEngine, { provider: 'spawn' })
+  return built
+}
+
+const META = {
+  name: 'e2e-worker-arithmetic',
+  description: 'two real children through a worker thread: one prose, one structured',
+  phases: [{ title: 'Ask' }, { title: 'Judge' }],
+}
+const SCRIPT = `phase('Ask')
+log('asking the prose child')
+const prose = await agent('Reply with exactly one short sentence: what is 2 + 2?')
+phase('Judge')
+const judged = await agent(
+  'Here is an answer to the question "what is 2+2": ' + prose
+  + ' — report whether it contains the number 4 and your confidence between 0 and 1.',
+  { schema: { type: 'object', properties: { containsFour: { type: 'boolean' }, confidence: { type: 'number' } }, required: ['containsFour'] } },
+)
+return { prose, containsFour: judged === null ? null : judged.containsFour }`
+
+describe.skipIf(!process.env.DEEPSEEK_API_KEY)('worker workflow engine with-key e2e', () => {
+  it('runs a two-phase script in a worker thread over real children, one through the structured runtime', async () => {
+    ctx = await harness()
+    const parentHandle = ctx.agents.create({
+      agentId: AgentId('wf-worker-e2e-parent'),
+      sessionId: 'wf-worker-e2e-session' as never,
+      agentOptions: { model: 'deepseek-v4-flash' },
+    })
+
+    const events: string[] = []
+    const childIds: string[] = []
+    for (const name of ['workflow/start', 'workflow/phase', 'workflow/log', 'workflow/agent-start', 'workflow/agent-end', 'workflow/end'] as const) {
+      ctx.on(name, (...payload: unknown[]) => {
+        events.push(name)
+        if (name === 'workflow/agent-start') childIds.push((payload[1] as { childId: string }).childId)
+      })
+    }
+
+    const run = ctx.workflows.start({ script: SCRIPT, meta: META, parent: parentHandle.agent })
+    const result = await run.result
+    await run.dispose()
+
+    expect(result.stopReason).toBe('completed')
+    expect(result.agentsStarted).toBe(2)
+    const value = result.value as { prose: string; containsFour: boolean | null }
+    // World checks: the prose child really answered (a real completion), and
+    // the structured child judged it against the REAL schema-forced tool.
+    expect(value.prose.length).toBeGreaterThan(0)
+    expect(value.containsFour).toBe(true)
+
+    expect(events[0]).toBe('workflow/start')
+    expect(events.at(-1)).toBe('workflow/end')
+    expect(events.filter(name => name === 'workflow/phase').length).toBe(2)
+    expect(events.filter(name => name === 'workflow/agent-start').length).toBe(2)
+    expect(childIds.length).toBe(2)
+    // The children were disposed to quiescence after collection.
+    for (const childId of childIds) {
+      expect(ctx.agents.get(AgentId(childId))).toBeUndefined()
+    }
+    await parentHandle.dispose()
+  }, 240_000)
+})

packages/workflow/workflow-workerthread/tests/workflow-workerthread.spec.ts

@@ -0,0 +1,904 @@
+import { describe, expect, it, vi } from 'vitest'
+import { fileURLToPath } from 'node:url'
+import { Context } from 'cordis'
+import Loader from '@cordisjs/plugin-loader'
+import { AgentId } from '@deepseek-ai/dsh-agent'
+import type { Agent } from '@deepseek-ai/dsh-agent'
+import SubagentService from '@deepseek-ai/dsh-subagent'
+import type { SubagentCapabilities, SubagentProvider, SubagentResult, SubagentRun, SubagentStartRequest } from '@deepseek-ai/dsh-subagent'
+import type { WorkflowMeta, WorkflowResult, WorkflowResultInfo, WorkflowRunInfo } from '@deepseek-ai/dsh-workflow'
+import * as workerEngineModule from '../src/index.ts'
+import WorkerWorkflowEngine, { type Config } from '../src/index.ts'
+
+/** A minimal parent stand-in: the engine only threads it through to the provider. */
+function fakeParent(): Agent {
+  return { id: AgentId('workflow-parent'), options: {} } as unknown as Agent
+}
+
+/** The vm-context escape hatch, spelled once: real Worker tests use it to make the WORKER misbehave. */
+const ESCAPE = "globalThis.constructor.constructor('return process')()"
+
+/** One controllable child run: the test (or auto mode) settles it. */
+interface ControlledRun {
+  request: SubagentStartRequest
+  settle(result: SubagentResult): void
+  cancelled: string | undefined
+  disposed: boolean
+  disposeCalls: number
+}
+
+/**
+ * A scripted in-test provider over the REAL SubagentService registry: `auto`
+ * settles each run via the reply function on a microtask; `manual` piles runs
+ * up in `runs` for the test to settle. A run aborts (settles `aborted`) when
+ * the request signal fires, like the real in-process backends.
+ */
+class StubProvider implements SubagentProvider {
+  readonly capabilities: SubagentCapabilities = { outputSchema: true, depthLimit: true, toolFilter: true }
+  readonly inheritsParentContext = false
+  readonly runs: ControlledRun[] = []
+
+  constructor(
+    readonly name: string,
+    private readonly reply?: (request: SubagentStartRequest, index: number) => SubagentResult,
+    private readonly disposeDelayMs = 0,
+  ) {}
+
+  start(request: SubagentStartRequest): SubagentRun {
+    let settle!: (result: SubagentResult) => void
+    const result = new Promise<SubagentResult>((resolve) => { settle = resolve })
+    const controlled: ControlledRun = { request, settle, cancelled: undefined, disposed: false, disposeCalls: 0 }
+    this.runs.push(controlled)
+    const index = this.runs.length - 1
+    request.signal?.addEventListener('abort', () => { settle({ output: [], stopReason: 'aborted' }) }, { once: true })
+    if (this.reply) {
+      const reply = this.reply
+      queueMicrotask(() => { settle(reply(request, index)) })
+    }
+    return {
+      id: AgentId(`stub-child-${index}`),
+      result,
+      cancel: (reason?: string) => {
+        controlled.cancelled = reason ?? 'cancelled'
+        settle({ output: [], stopReason: 'aborted' })
+      },
+      dispose: () => {
+        controlled.disposeCalls += 1
+        if (this.disposeDelayMs === 0) {
+          controlled.disposed = true
+          return Promise.resolve()
+        }
+        return new Promise<void>((resolve) => {
+          setTimeout(() => {
+            controlled.disposed = true
+            resolve()
+          }, this.disposeDelayMs)
+        })
+      },
+    }
+  }
+}
+
+/** Text-reply helper for auto providers. */
+function text(reply: string): SubagentResult {
+  return { output: [{ type: 'text', text: reply }], stopReason: 'completed' }
+}
+
+interface SetupOptions {
+  config?: Config
+  reply?: (request: SubagentStartRequest, index: number) => SubagentResult
+  manual?: boolean
+  disposeDelayMs?: number
+}
+
+async function setup(options?: SetupOptions) {
+  const ctx = new Context()
+  await ctx.plugin(SubagentService)
+  const provider = new StubProvider(
+    'stub',
+    options?.manual ? undefined : options?.reply ?? (() => text('stub reply')),
+    options?.disposeDelayMs ?? 0,
+  )
+  ctx.subagents.registerProvider(provider)
+  // A fixed concurrency ceiling: the auto-resolved default is machine-derived
+  // (cores - 2, floored at 1), so tests that expect N children in flight
+  // would wedge on small CI runners.
+  await ctx.plugin(WorkerWorkflowEngine, { provider: 'stub', maxConcurrentAgents: 8, ...options?.config })
+  return { ctx, provider, parent: fakeParent() }
+}
+
+/** The standard test meta plus a body, spread into a start request. */
+function scripted(body: string, metaExtra?: Partial<WorkflowMeta>): { script: string; meta: WorkflowMeta } {
+  return { script: body, meta: { name: 'test-flow', description: 'a test workflow', ...metaExtra } }
+}
+
+/** Start + await one run, disposing on the way out. */
+async function run(ctx: Context, parent: Agent, source: { script: string; meta: WorkflowMeta }, args?: unknown): Promise<WorkflowResult> {
+  const handle = ctx.workflows.start({ ...source, parent, ...args !== undefined ? { args } : {} })
+  try {
+    return await handle.result
+  } finally {
+    await handle.dispose()
+  }
+}
+
+describe('dsh-workflow-workerthread', () => {
+  describe('script execution over a real worker thread', () => {
+    it('runs a script end-to-end: agent() text results, phases, log, args, return value, events', async () => {
+      const { ctx, parent, provider } = await setup({ reply: (_request, index) => text(`answer-${index}`) })
+      const events: [string, unknown[]][] = []
+      for (const name of ['workflow/start', 'workflow/phase', 'workflow/log', 'workflow/agent-start', 'workflow/agent-end', 'workflow/end'] as const) {
+        ctx.on(name, (...payload: unknown[]) => { events.push([name, payload]) })
+      }
+      const result = await run(ctx, parent, scripted(`
+        phase('Scan')
+        log('starting with ' + args.files.length + ' files')
+        const answers = await pipeline(args.files, (prev, item) => agent('read ' + item))
+        phase('Report')
+        return { answers, count: args.files.length }
+      `, { phases: [{ title: 'Scan' }, { title: 'Report' }] }), { files: ['a.ts', 'b.ts'] })
+
+      expect(result.stopReason).toBe('completed')
+      expect(result.agentsStarted).toBe(2)
+      expect(result.value).toEqual({ answers: ['answer-0', 'answer-1'], count: 2 })
+      expect(provider.runs.every(r => r.disposed)).toBe(true)
+
+      const names = events.map(([name]) => name)
+      expect(names[0]).toBe('workflow/start')
+      expect(names).toContain('workflow/phase')
+      expect(names).toContain('workflow/log')
+      expect(names.at(-1)).toBe('workflow/end')
+      const info = events[0]![1][0] as WorkflowRunInfo
+      expect(info.meta.name).toBe('test-flow')
+      const end = events.at(-1)![1][1] as Record<string, unknown>
+      expect(end).toEqual({ stopReason: 'completed', agentsStarted: 2 })
+      expect('value' in end).toBe(false)
+    })
+
+    it('agent({schema, model}) forwards outputSchema and agentOptions to the provider across the thread', async () => {
+      const { ctx, parent, provider } = await setup({
+        reply: () => ({ output: [], structured: { files: ['x.ts', 'y.ts'] }, stopReason: 'completed' }),
+      })
+      const result = await run(ctx, parent, scripted(`
+        const found = await agent('list files', { model: 'deepseek-v4-pro', schema: { type: 'object', properties: { files: { type: 'array', items: { type: 'string' } } }, required: ['files'] } })
+        return { first: found.files[0], count: found.files.length }
+      `))
+      expect(result.value).toEqual({ first: 'x.ts', count: 2 })
+      expect(provider.runs[0]!.request.outputSchema).toEqual({
+        type: 'object',
+        properties: { files: { type: 'array', items: { type: 'string' } } },
+        required: ['files'],
+      })
+      expect(provider.runs[0]!.request.agentOptions).toEqual({ model: 'deepseek-v4-pro' })
+      expect(provider.runs[0]!.request.parent).toBeDefined()
+    })
+
+    it('a fatal hook error inside the worker kills the script and reports the error', async () => {
+      const { ctx, parent } = await setup()
+      const result = await run(ctx, parent, scripted("return await parallel([() => agent('x', { isolation: 'worktree' })])"))
+      expect(result.stopReason).toBe('error')
+      expect(result.error).toContain('"isolation" is deferred')
+    })
+
+    it('a provider start failure crosses back as a fatal AGENT_START error', async () => {
+      const { ctx, parent } = await setup({ config: { provider: 'nonexistent' } })
+      const result = await run(ctx, parent, scripted("return await pipeline([1], () => agent('p'))"))
+      expect(result.stopReason).toBe('error')
+      expect(result.error).toContain('agent() could not start a child')
+    })
+
+    it('a child result REJECTION crosses back as a fatal AGENT_RESULT error (a broken provider is not a failed child)', async () => {
+      const ctx = new Context()
+      await ctx.plugin(SubagentService)
+      const provider: SubagentProvider = {
+        name: 'rejecting',
+        capabilities: { outputSchema: true, depthLimit: true, toolFilter: true },
+        inheritsParentContext: false,
+        start: () => ({
+          id: AgentId('reject-child'),
+          result: Promise.reject(new Error('backend exploded')),
+          cancel: () => { /* nothing in flight */ },
+          dispose: () => Promise.resolve(),
+        }),
+      }
+      ctx.subagents.registerProvider(provider)
+      await ctx.plugin(WorkerWorkflowEngine, { provider: 'rejecting', maxConcurrentAgents: 2 })
+      const result = await run(ctx, fakeParent(), scripted(`
+        try { await agent('p'); return 'unreachable' } catch (e) { return { name: e.name, code: e.code, fatal: e.fatal, message: e.message } }
+      `))
+      expect(result.value).toMatchObject({ name: 'WorkflowError', code: 'AGENT_RESULT', fatal: true })
+      expect((result.value as { message: string }).message).toContain('backend exploded')
+    })
+
+    it('a child whose dispose() rejects cannot wedge the script (the host acks anyway)', async () => {
+      const ctx = new Context()
+      await ctx.plugin(SubagentService)
+      const provider: SubagentProvider = {
+        name: 'bad-dispose',
+        capabilities: { outputSchema: true, depthLimit: true, toolFilter: true },
+        inheritsParentContext: false,
+        start: () => ({
+          id: AgentId('bad-dispose-child'),
+          result: Promise.resolve({ output: [{ type: 'text', text: 'fine' }], stopReason: 'completed' }),
+          cancel: () => { /* settled already */ },
+          dispose: () => Promise.reject(new Error('dispose exploded')),
+        }),
+      }
+      ctx.subagents.registerProvider(provider)
+      await ctx.plugin(WorkerWorkflowEngine, { provider: 'bad-dispose', maxConcurrentAgents: 2 })
+      const result = await run(ctx, fakeParent(), scripted("return await agent('p')"))
+      expect(result.stopReason).toBe('completed')
+      expect(result.value).toBe('fine')
+    })
+
+    it('a child dispose() rejecting an UNRENDERABLE value still acks — the containment warn is total', async () => {
+      const ctx = new Context()
+      await ctx.plugin(SubagentService)
+      const provider: SubagentProvider = {
+        name: 'coercion-trap-dispose',
+        capabilities: { outputSchema: true, depthLimit: true, toolFilter: true },
+        inheritsParentContext: false,
+        start: () => ({
+          id: AgentId('trap-child'),
+          result: Promise.resolve({ output: [{ type: 'text', text: 'fine' }], stopReason: 'completed' }),
+          cancel: () => { /* settled already */ },
+          // The rejection VALUE's own coercion throws: a warn built with bare
+          // String(error) would itself throw, skipping the ChildDisposed ack
+          // and wedging the script's finally until the grace/terminate path.
+          // eslint-disable-next-line @typescript-eslint/prefer-promise-reject-errors -- the non-Error rejection IS the scenario under test
+          dispose: () => Promise.reject({ toString: () => { throw new Error('coercion trap') } }),
+        }),
+      }
+      ctx.subagents.registerProvider(provider)
+      await ctx.plugin(WorkerWorkflowEngine, { provider: 'coercion-trap-dispose', maxConcurrentAgents: 2 })
+      const result = await run(ctx, fakeParent(), scripted("return await agent('p')"))
+      expect(result.stopReason).toBe('completed')
+      expect(result.value).toBe('fine')
+    })
+
+    it('the worker spawns with an EMPTY environment: an escaped script finds no ambient credentials', async () => {
+      const { ctx, parent } = await setup()
+      // A canary in the HARNESS process's env: with an inherited environment
+      // the escape below would read it back (exactly how DEEPSEEK_API_KEY
+      // would leak); env: {} in the spawn options is what keeps it out.
+      process.env.WORKFLOW_ENV_CANARY = 'leak me'
+      try {
+        const result = await run(ctx, parent, scripted(`
+          const proc = ${ESCAPE}
+          return { canary: proc.env.WORKFLOW_ENV_CANARY ?? null, keys: Object.keys(proc.env).length }
+        `))
+        expect(result.stopReason).toBe('completed')
+        expect(result.value).toEqual({ canary: null, keys: 0 })
+      } finally {
+        delete process.env.WORKFLOW_ENV_CANARY
+      }
+    })
+
+    it('the unbuilt worker forwards exactly TSX_TSCONFIG_PATH through the scrub: the paths-map pin survives, secrets do not', async () => {
+      const { ctx, parent } = await setup()
+      // The ACP snapshot harness runs the parent with its cwd OUTSIDE the
+      // repo and pins the repo tsconfig through this variable; the worker
+      // must inherit the pin (or its dsh-* imports silently resolve to
+      // unbuilt lib/ bundles) while every other variable stays scrubbed.
+      const tsconfig = fileURLToPath(new URL('../../../../tsconfig.json', import.meta.url))
+      process.env.TSX_TSCONFIG_PATH = tsconfig
+      process.env.WORKFLOW_ENV_CANARY = 'leak me'
+      try {
+        const result = await run(ctx, parent, scripted(`
+          const proc = ${ESCAPE}
+          return { keys: Object.keys(proc.env), tsconfig: proc.env.TSX_TSCONFIG_PATH }
+        `))
+        expect(result.stopReason).toBe('completed')
+        expect(result.value).toEqual({ keys: ['TSX_TSCONFIG_PATH'], tsconfig })
+      } finally {
+        delete process.env.TSX_TSCONFIG_PATH
+        delete process.env.WORKFLOW_ENV_CANARY
+      }
+    })
+  })
+
+  describe('lifecycle: parse errors, cancellation, termination, disposal', () => {
+    it('start() throws synchronously for invalid meta data or an unparseable body (host-side pre-checks)', async () => {
+      const { ctx, parent } = await setup()
+      // Meta is DATA — shape violations reject loud, every one named.
+      expect(() => ctx.workflows.start({ script: 'return 1', meta: { name: '', description: 'd' }, parent })).toThrow(/meta\.name must be a non-empty string/)
+      expect(() => ctx.workflows.start({ script: 'return 1', meta: { name: 'x', description: 'd', extra: 1 } as unknown as WorkflowMeta, parent })).toThrow(/META_INVALID|not a recognized field/)
+      expect(() => ctx.workflows.start({ ...scripted('return ((('), parent })).toThrow(/does not parse/)
+      // The likeliest authoring slip — a Claude Code-style meta header in the
+      // body — gets a pointed message, not a bare SyntaxError.
+      expect(() => ctx.workflows.start({ ...scripted("export const meta = { name: 'x', description: 'd' }\nreturn 1"), parent })).toThrow(/meta rides the `meta` request field/)
+    })
+
+    it('cancel() aborts in-flight children (signal AND cancel RPC) and settles the run cancelled', async () => {
+      const { ctx, parent, provider } = await setup({ manual: true })
+      const ends: unknown[] = []
+      ctx.on('workflow/agent-end', (_info, agent) => { ends.push(agent) })
+      const runEnds: WorkflowResultInfo[] = []
+      ctx.on('workflow/end', (_info, result) => { runEnds.push(result) })
+      const handle = ctx.workflows.start({ ...scripted("return await agent('long job')"), parent })
+      await vi.waitFor(() => { expect(provider.runs.length).toBe(1) })
+      handle.cancel('user stopped it')
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+      expect(result.error).toContain('user stopped it')
+      await handle.dispose()
+      expect(provider.runs[0]!.disposed).toBe(true)
+      expect(ends).toEqual([expect.objectContaining({ seq: 1, outcome: 'cancelled' })])
+      // workflow/end is an observer's only death signal: it fires for a
+      // cancelled run too, mirroring the settled outcome data.
+      expect(runEnds).toEqual([{ stopReason: 'cancelled', error: result.error, agentsStarted: result.agentsStarted }])
+    })
+
+    it('an already-aborted request signal cancels before the body ever runs (the go handshake holds it)', async () => {
+      const { ctx, parent, provider } = await setup()
+      const controller = new AbortController()
+      controller.abort()
+      const logs: string[] = []
+      ctx.on('workflow/log', (_info, message) => { logs.push(message) })
+      const handle = ctx.workflows.start({ ...scripted("log('ran')\nreturn 123"), parent, signal: controller.signal })
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+      expect(result.value).toBeNull()
+      expect(logs).toEqual([])
+      expect(provider.runs.length).toBe(0)
+      await handle.dispose()
+    })
+
+    it('cancel() right after start() cancels before the body runs; the signal aborting mid-run cancels like cancel()', async () => {
+      const { ctx, parent, provider } = await setup({ manual: true })
+      const first = ctx.workflows.start({ ...scripted("return await agent('never')"), parent })
+      // No-reason cancel: the canonical default reason must ride the result.
+      first.cancel()
+      const firstResult = await first.result
+      expect(firstResult.stopReason).toBe('cancelled')
+      expect(firstResult.error).toContain('workflow cancelled')
+      expect(provider.runs.length).toBe(0)
+      await first.dispose()
+
+      const controller = new AbortController()
+      const second = ctx.workflows.start({ ...scripted("return await agent('job')"), parent, signal: controller.signal })
+      await vi.waitFor(() => { expect(provider.runs.length).toBe(1) })
+      controller.abort()
+      expect((await second.result).stopReason).toBe('cancelled')
+      await second.dispose()
+    })
+
+    it('a child-start racing the host cancel is refused: no child starts after cancellation', async () => {
+      const { ctx, parent, provider } = await setup({ manual: true })
+      // Cancel from INSIDE the log listener: the worker has already posted
+      // its child-start (queued right behind the log message), so the host
+      // processes it with cancelReason set — the refusal arm no real-world
+      // timing can hit reliably. (The closure runs only after `handle` below
+      // is initialized — the listener fires on the worker's first message.)
+      ctx.on('workflow/log', () => { handle.cancel('cancelled from the log listener') })
+      const handle = ctx.workflows.start({ ...scripted("log('mark')\nreturn await agent('late')"), parent })
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+      expect(provider.runs.length).toBe(0)
+      await handle.dispose()
+    })
+
+    it('post-cancel narration is suppressed host-side, and completion racing a cancel reports cancelled', async () => {
+      const { ctx, parent } = await setup()
+      const narration: string[] = []
+      ctx.on('workflow/log', (_info, message) => { narration.push(message) })
+      ctx.on('workflow/phase', (_info, title) => { narration.push(`phase:${title}`) })
+      const handle = ctx.workflows.start({
+        // The sync spin keeps the worker's loop busy so the cancel message
+        // cannot be processed before the script settles `completed` — the
+        // worker posts a completed result that must LOSE to the in-flight
+        // host cancellation. The trailing narration exercises host-side
+        // suppression: posted pre-cancel-processing worker-side, arriving
+        // post-cancel host-side.
+        ...scripted(`
+          log('started')
+          const end = Date.now() + 1000
+          while (Date.now() < end) {}
+          phase('late phase')
+          log('late log')
+          return 'done'
+        `),
+        parent,
+      })
+      await vi.waitFor(() => { expect(narration).toContain('started') })
+      handle.cancel('raced the completion')
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+      expect(result.error).toContain('raced the completion')
+      expect(narration).toEqual(['started'])
+      await handle.dispose()
+    }, 15_000)
+
+    it('cancel() force-settles a script parked on a promise no hook owns, and TERMINATES its worker', async () => {
+      const { ctx, parent } = await setup({ config: { provider: 'stub', disposeGraceMs: 50 } })
+      const runEnds: WorkflowResultInfo[] = []
+      ctx.on('workflow/end', (_info, result) => { runEnds.push(result) })
+      const handle = ctx.workflows.start({
+        ...scripted("await new Promise(() => {})\nreturn 'unreachable'"),
+        parent,
+      })
+      handle.cancel('user aborted')
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+      expect(result.error).toContain('user aborted')
+      // The grace force-settle fires workflow/end exactly like an ordinary
+      // settlement — a terminated script's death still reaches observers.
+      expect(runEnds).toEqual([{ stopReason: 'cancelled', error: result.error, agentsStarted: 0 }])
+      await handle.dispose()
+    })
+
+    it('dispose() on a stuck script returns within the grace instead of hanging (result settles cancelled)', async () => {
+      const { ctx, parent } = await setup({ config: { provider: 'stub', disposeGraceMs: 50 } })
+      const handle = ctx.workflows.start({
+        ...scripted("await new Promise(() => {})\nreturn 'unreachable'"),
+        parent,
+      })
+      const before = Date.now()
+      await handle.dispose()
+      expect(Date.now() - before).toBeLessThan(2000)
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+    })
+
+    it('dispose() is idempotent and settles cleanly after a completed run', async () => {
+      const { ctx, parent } = await setup()
+      const handle = ctx.workflows.start({ ...scripted('return 1'), parent })
+      await handle.result
+      await handle.dispose()
+      await handle.dispose()
+    })
+
+    it('a settled run arms NO grace timer: disposing a completed run must not pin it for disposeGraceMs', async () => {
+      // A distinctive grace so the spy can tell the cancel-path grace timer
+      // apart from every other timeout in flight.
+      const GRACE = 44_444
+      const { ctx, parent } = await setup({ config: { provider: 'stub', disposeGraceMs: GRACE } })
+      const handle = ctx.workflows.start({ ...scripted('return 1'), parent })
+      await handle.result
+      const spy = vi.spyOn(globalThis, 'setTimeout')
+      try {
+        await handle.dispose()
+        // dispose()'s own bounded-wait sleep is the ONLY grace-sized timer
+        // allowed here; before the settled guard, cancel() armed a second one
+        // that nothing would ever clear (the run was already settled), keeping
+        // the WorkerRun/Worker closure alive until the grace expired.
+        const graceTimers = spy.mock.calls.filter(call => call[1] === GRACE)
+        expect(graceTimers.length).toBe(1)
+      } finally {
+        spy.mockRestore()
+      }
+    })
+
+    it('strays: children fired without await are aborted once the script settles, and dispose() waits for their disposal', async () => {
+      const { ctx, parent, provider } = await setup({ manual: true, disposeDelayMs: 40 })
+      const handle = ctx.workflows.start({
+        ...scripted(`
+          agent('stray')
+          return 'done without awaiting'
+        `),
+        parent,
+      })
+      const result = await handle.result
+      expect(result.stopReason).toBe('completed')
+      await vi.waitFor(() => { expect(provider.runs.length).toBe(1) })
+      await handle.dispose()
+      // Not a waitFor: by the time dispose() returns, the slow child disposal
+      // must already be complete (host-side registry quiescence).
+      expect(provider.runs[0]!.disposed).toBe(true)
+    })
+
+    it('the settle-reap fires the request signal too: a provider honoring ONLY the signal winds its stray down promptly', async () => {
+      const ctx = new Context()
+      await ctx.plugin(SubagentService)
+      const aborted: string[] = []
+      const provider: SubagentProvider = {
+        name: 'signal-only',
+        capabilities: { outputSchema: true, depthLimit: true, toolFilter: true },
+        inheritsParentContext: false,
+        start: (request) => {
+          let settle!: (result: SubagentResult) => void
+          const result = new Promise<SubagentResult>((resolve) => { settle = resolve })
+          request.signal?.addEventListener('abort', () => {
+            aborted.push(String(request.signal?.reason))
+            settle({ output: [], stopReason: 'aborted' })
+          }, { once: true })
+          return {
+            id: AgentId('signal-only-child'),
+            result,
+            // The seam leaves a provider free to honor EITHER cancel channel;
+            // this one deliberately ignores run.cancel() — only the request
+            // signal can wind it down.
+            cancel: () => { /* signal-only by design */ },
+            dispose: () => Promise.resolve(),
+          }
+        },
+      }
+      ctx.subagents.registerProvider(provider)
+      await ctx.plugin(WorkerWorkflowEngine, { provider: 'signal-only', maxConcurrentAgents: 2 })
+      const handle = ctx.workflows.start({
+        ...scripted(`
+          agent('stray, never awaited')
+          return 'done'
+        `),
+        parent: fakeParent(),
+      })
+      const result = await handle.result
+      expect(result.stopReason).toBe('completed')
+      // BEFORE dispose(): the settlement itself must have aborted the signal —
+      // without it this child would stay live until dispose's terminate.
+      await vi.waitFor(() => { expect(aborted).toEqual(['workflow settled']) })
+      await handle.dispose()
+    })
+
+    it("cancel() drives each child's explicit cancel() host-side: a wedged worker cannot delay it", async () => {
+      const ctx = new Context()
+      await ctx.plugin(SubagentService)
+      let starts = 0
+      const cancelled: string[] = []
+      const provider: SubagentProvider = {
+        name: 'cancel-only',
+        capabilities: { outputSchema: true, depthLimit: true, toolFilter: true },
+        inheritsParentContext: false,
+        start: () => {
+          starts += 1
+          return {
+            id: AgentId('cancel-only-child'),
+            result: new Promise(() => { /* only cancel() ends this child */ }),
+            // Deliberately ignores the request signal — the seam leaves a
+            // provider free to honor ONLY the explicit cancel() channel.
+            cancel: (reason?: string) => { cancelled.push(reason ?? 'cancelled') },
+            dispose: () => Promise.resolve(),
+          }
+        },
+      }
+      ctx.subagents.registerProvider(provider)
+      // A deliberately huge grace: if only the grace/terminate reap could
+      // reach this child, the assertion below would time out first.
+      await ctx.plugin(WorkerWorkflowEngine, { provider: 'cancel-only', maxConcurrentAgents: 2, disposeGraceMs: 30_000 })
+      const handle = ctx.workflows.start({
+        // The stray child's start RPC reaches the host, then the script wedges
+        // its own worker in a synchronous spin: the worker cannot process the
+        // Cancel message, so it can relay NO ChildCancel RPC — only the host's
+        // own children loop can deliver the explicit cancel in time. The
+        // microtask yields let the agent() continuation POST its child-start
+        // before the spin seizes the worker's loop (the posted message needs
+        // no further worker-loop turns to reach the host).
+        ...scripted(`
+          agent('wedged child')
+          for (let i = 0; i < 20; i++) await null
+          const end = Date.now() + 1500
+          while (Date.now() < end) {}
+          return 'raced'
+        `),
+        parent: fakeParent(),
+      })
+      await vi.waitFor(() => { expect(starts).toBe(1) })
+      handle.cancel('stop now')
+      await vi.waitFor(() => { expect(cancelled).toEqual(['stop now']) }, { timeout: 800 })
+      // The wedged worker's own completion loses to the in-flight cancel.
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+      await handle.dispose()
+    }, 15_000)
+
+    it('dispose() on a wedged worker host-drives child disposal inside the grace: it returns with the children DISPOSED, not with their teardown still in flight', async () => {
+      const { ctx, parent, provider } = await setup({
+        manual: true,
+        disposeDelayMs: 40,
+        config: { provider: 'stub', maxConcurrentAgents: 8, disposeGraceMs: 400 },
+      })
+      const handle = ctx.workflows.start({
+        // Same shape as the wedged-cancel test above: the child's start RPC
+        // reaches the host, then the script seizes its worker's loop, so the
+        // worker can relay NO dispose RPC — the host's own dispose() drive is
+        // the only thing that can start (and finish) this child's disposal
+        // before the grace runs out.
+        ...scripted(`
+          agent('wedged child')
+          for (let i = 0; i < 20; i++) await null
+          const end = Date.now() + 1500
+          while (Date.now() < end) {}
+          return 'raced'
+        `),
+        parent,
+      })
+      await vi.waitFor(() => { expect(provider.runs.length).toBe(1) })
+      const before = Date.now()
+      await handle.dispose()
+      // Bounded by the grace (plus the terminate), never by the 1.5s spin.
+      expect(Date.now() - before).toBeLessThan(1200)
+      // Not a waitFor: dispose() resolving IS the quiescence claim — the slow
+      // child disposal must be complete, not merely started (before the
+      // host-driven drive, disposal only STARTED at the post-terminate reap,
+      // so dispose() returned with it still in flight).
+      expect(provider.runs[0]!.disposed).toBe(true)
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+    }, 15_000)
+
+    it('a live child disposed by the dispose() drive is disposed ONCE, and the worker\'s late dispose RPC still gets its ack (the script settles, not the grace)', async () => {
+      const { ctx, parent, provider } = await setup({ manual: true })
+      const handle = ctx.workflows.start({
+        ...scripted(`
+          await agent('long child')
+          return 'unreachable'
+        `),
+        parent,
+      })
+      await vi.waitFor(() => { expect(provider.runs.length).toBe(1) })
+      const handleDispose = handle.dispose()
+      const result = await handle.result
+      // The script itself settled (the wrapper's own dispose RPC found the
+      // child already reaped host-side and was acked) — a missing ack would
+      // wedge the wrapper's finally until the 5s default grace force-settle.
+      expect(result.stopReason).toBe('cancelled')
+      expect(result.error).toContain('workflow disposed')
+      await handleDispose
+      expect(provider.runs[0]!.disposed).toBe(true)
+      // The memo: the host drive and the worker's RPC share one disposal.
+      expect(provider.runs[0]!.disposeCalls).toBe(1)
+    })
+
+    it('the grace force-settle pairs every stranded start: a host-synthesized cancelled agent-end lands before workflow/end', async () => {
+      const { ctx, parent, provider } = await setup({ manual: true, config: { provider: 'stub', maxConcurrentAgents: 8, disposeGraceMs: 300 } })
+      const ends: { seq: number; outcome: string }[] = []
+      const order: string[] = []
+      ctx.on('workflow/agent-start', (_info, agent) => { order.push(`start:${agent.seq}`) })
+      ctx.on('workflow/agent-end', (_info, agent) => {
+        ends.push({ seq: agent.seq, outcome: agent.outcome })
+        order.push(`end:${agent.seq}`)
+      })
+      ctx.on('workflow/end', () => { order.push('run-end') })
+      const handle = ctx.workflows.start({
+        // 'slow' starts and its agent-start crosses to observers (the awaited
+        // 'fast' call keeps the worker loop turning), then the script seizes
+        // the loop: the wedged worker can never author slow's agent-end —
+        // only the host's ledger can close the pair.
+        ...scripted(`
+          const p = agent('slow')
+          await agent('fast')
+          const end = Date.now() + 1500
+          while (Date.now() < end) {}
+          return 'raced'
+        `),
+        parent,
+      })
+      await vi.waitFor(() => { expect(order.filter(entry => entry.startsWith('start:')).length).toBe(2) })
+      const fast = provider.runs.find(run => (run.request.prompt[0] as { text?: string }).text === 'fast')!
+      fast.settle(text('fast done'))
+      handle.cancel('stop now')
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+      // fast's end is the worker's own report; slow's is host-synthesized at
+      // the force-settle — exactly one end per started seq, no third event.
+      expect(ends).toEqual([
+        { seq: 2, outcome: 'completed' },
+        { seq: 1, outcome: 'cancelled' },
+      ])
+      // Both ends reached observers BEFORE workflow/end: a progress consumer
+      // can finalize its state at run-end without dangling agents.
+      expect(order.indexOf('run-end')).toBe(order.length - 1)
+      await handle.dispose()
+    }, 15_000)
+
+    it('graceful cancellation keeps pairing worker-authored: exactly one agent-end per start, nothing synthesized on top', async () => {
+      const { ctx, parent, provider } = await setup({ manual: true })
+      const ends: { seq: number; outcome: string }[] = []
+      const order: string[] = []
+      ctx.on('workflow/agent-end', (_info, agent) => {
+        ends.push({ seq: agent.seq, outcome: agent.outcome })
+        order.push(`end:${agent.seq}`)
+      })
+      ctx.on('workflow/end', () => { order.push('run-end') })
+      const handle = ctx.workflows.start({
+        ...scripted("await parallel([() => agent('a'), () => agent('b')])\nreturn 'unreachable'"),
+        parent,
+      })
+      await vi.waitFor(() => { expect(provider.runs.length).toBe(2) })
+      handle.cancel('user stop')
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+      // The live worker reported both pairs itself; the ledger must not add
+      // a synthesized duplicate on any path that settles inside the grace.
+      expect(ends.map(end => end.outcome)).toEqual(['cancelled', 'cancelled'])
+      expect(new Set(ends.map(end => end.seq)).size).toBe(2)
+      expect(order.indexOf('run-end')).toBe(order.length - 1)
+      await handle.dispose()
+    })
+  })
+
+  describe('worker death', () => {
+    it('a worker that exits before settling reports an error result and reaps its children', async () => {
+      const ctx = new Context()
+      await ctx.plugin(SubagentService)
+      // The child's dispose() REJECTS on top of the worker death: the reap
+      // must contain it (warn, not crash) while still emptying the registry.
+      const cancelled: string[] = []
+      const provider: SubagentProvider = {
+        name: 'doomed',
+        capabilities: { outputSchema: true, depthLimit: true, toolFilter: true },
+        inheritsParentContext: false,
+        start: () => ({
+          id: AgentId('doomed-child'),
+          result: new Promise(() => { /* never settles; the reap is the teardown */ }),
+          cancel: (reason?: string) => { cancelled.push(reason ?? 'cancelled') },
+          dispose: () => Promise.reject(new Error('dispose exploded during reap')),
+        }),
+      }
+      ctx.subagents.registerProvider(provider)
+      await ctx.plugin(WorkerWorkflowEngine, { provider: 'doomed', maxConcurrentAgents: 2 })
+      const runEnds: WorkflowResultInfo[] = []
+      ctx.on('workflow/end', (_info, result) => { runEnds.push(result) })
+      const handle = ctx.workflows.start({
+        // The stray child's start RPC reaches the host, then the script kills
+        // its own worker through the documented vm escape — the host must
+        // settle `error` with the exit diagnostics and wind the child down.
+        ...scripted(`
+          agent('doomed')
+          const proc = ${ESCAPE}
+          const st = globalThis.constructor.constructor('return setTimeout')()
+          await new Promise(resolve => st(resolve, 200))
+          proc.exit(7)
+        `),
+        parent: fakeParent(),
+      })
+      const result = await handle.result
+      expect(result.stopReason).toBe('error')
+      expect(result.error).toContain('exit code 7')
+      expect(result.agentsStarted).toBe(1)
+      // A worker death is a stop reason like any other: workflow/end fires
+      // with the error outcome — for a bus observer it is the only obituary.
+      expect(runEnds).toEqual([{ stopReason: 'error', error: result.error, agentsStarted: 1 }])
+      await vi.waitFor(() => { expect(cancelled.length).toBe(1) })
+      await handle.dispose()
+    }, 15_000)
+
+    it('an uncaught exception inside the worker surfaces as an error result and reaps the in-flight child', async () => {
+      const { ctx, parent, provider } = await setup({ manual: true })
+      const handle = ctx.workflows.start({
+        ...scripted(`
+          agent('in flight when the worker dies')
+          const proc = ${ESCAPE}
+          const st = globalThis.constructor.constructor('return setTimeout')()
+          await new Promise(resolve => st(resolve, 200))
+          proc.nextTick(() => { throw new Error('worker blew up') })
+          await new Promise(() => {})
+        `),
+        parent,
+      })
+      const result = await handle.result
+      expect(result.stopReason).toBe('error')
+      expect(result.error).toContain('worker blew up')
+      // The reap wound the stray child down (cancel + a CLEAN dispose).
+      await vi.waitFor(() => {
+        expect(provider.runs.length).toBe(1)
+        expect(provider.runs[0]!.disposed).toBe(true)
+      })
+      await handle.dispose()
+    }, 15_000)
+
+    it('a worker death pairs every stranded start: the synthesized cancelled agent-end precedes the error workflow/end', async () => {
+      const { ctx, parent, provider } = await setup({ manual: true })
+      const ends: { seq: number; outcome: string }[] = []
+      const order: string[] = []
+      ctx.on('workflow/agent-start', (_info, agent) => { order.push(`start:${agent.seq}`) })
+      ctx.on('workflow/agent-end', (_info, agent) => {
+        ends.push({ seq: agent.seq, outcome: agent.outcome })
+        order.push(`end:${agent.seq}`)
+      })
+      ctx.on('workflow/end', () => { order.push('run-end') })
+      const handle = ctx.workflows.start({
+        // Same choreography as the force-settle pairing test, but the worker
+        // DIES (the documented vm escape) instead of being terminated: the
+        // exit path must close slow's pair from the ledger too. The escaped
+        // setTimeout lets the already-posted messages flush before the kill.
+        ...scripted(`
+          const p = agent('slow')
+          await agent('fast')
+          const proc = ${ESCAPE}
+          const st = globalThis.constructor.constructor('return setTimeout')()
+          await new Promise(resolve => st(resolve, 150))
+          proc.exit(7)
+        `),
+        parent,
+      })
+      await vi.waitFor(() => { expect(order.filter(entry => entry.startsWith('start:')).length).toBe(2) })
+      const fast = provider.runs.find(run => (run.request.prompt[0] as { text?: string }).text === 'fast')!
+      fast.settle(text('fast done'))
+      const result = await handle.result
+      expect(result.stopReason).toBe('error')
+      expect(result.error).toContain('exit code 7')
+      expect(ends).toEqual([
+        { seq: 2, outcome: 'completed' },
+        { seq: 1, outcome: 'cancelled' },
+      ])
+      expect(order.indexOf('run-end')).toBe(order.length - 1)
+      await handle.dispose()
+    }, 15_000)
+
+    it('a dispose ack racing the worker death is dropped, not crashed (post after exit)', async () => {
+      // Slow child disposal: the ack resolves only AFTER the worker died, so
+      // it has nowhere to go and must be dropped silently (the workerGone
+      // guard in post()).
+      const { ctx, parent, provider } = await setup({ disposeDelayMs: 300 })
+      const handle = ctx.workflows.start({
+        // The STRAY child settles instantly, so its wrapper starts the slow
+        // host-side disposal concurrently while the script goes on to kill
+        // its own worker — the ack then resolves into a dead thread.
+        ...scripted(`
+          agent('stray, never awaited')
+          const proc = ${ESCAPE}
+          const st = globalThis.constructor.constructor('return setTimeout')()
+          await new Promise(resolve => st(resolve, 150))
+          proc.exit(5)
+        `),
+        parent,
+      })
+      const result = await handle.result
+      expect(result.stopReason).toBe('error')
+      expect(result.error).toContain('exit code 5')
+      await vi.waitFor(() => { expect(provider.runs[0]!.disposed).toBe(true) })
+      await handle.dispose()
+    }, 15_000)
+
+    it('a worker death AFTER a cancel reports cancelled, not error', async () => {
+      const { ctx, parent } = await setup({ config: { provider: 'stub', disposeGraceMs: 60_000 } })
+      const handle = ctx.workflows.start({
+        ...scripted(`
+          const proc = ${ESCAPE}
+          const st = globalThis.constructor.constructor('return setTimeout')()
+          log('armed')
+          await new Promise(resolve => st(resolve, 400))
+          proc.exit(3)
+        `),
+        parent,
+      })
+      const logs: string[] = []
+      ctx.on('workflow/log', (_info, message) => { logs.push(message) })
+      await vi.waitFor(() => { expect(logs).toContain('armed') })
+      handle.cancel('stop it')
+      // The grace is deliberately huge: only the worker's own death (exit 3,
+      // unreachable by the cancel — the script ignores hooks) settles this.
+      const result = await handle.result
+      expect(result.stopReason).toBe('cancelled')
+      expect(result.error).toContain('stop it')
+      await handle.dispose()
+    }, 15_000)
+  })
+
+  describe('service surface', () => {
+    it('run ids are unique per start; the run handle and event payloads hold SEPARATE meta clones', async () => {
+      const { ctx, parent } = await setup()
+      let eventMeta: WorkflowRunInfo | undefined
+      ctx.on('workflow/start', (info) => { eventMeta = info })
+      const first = ctx.workflows.start({ ...scripted('return 1'), parent })
+      const second = ctx.workflows.start({ ...scripted('return 2'), parent })
+      expect(first.id).not.toBe(second.id)
+      eventMeta!.meta.name = 'corrupted'
+      expect(second.meta.name).toBe('test-flow')
+      await Promise.all([first.result, second.result])
+      await first.dispose()
+      await second.dispose()
+    })
+
+    it('unregisters ctx.workflows when the engine fiber is disposed (HMR safety), and default config runs (auto concurrency)', async () => {
+      const ctx = new Context()
+      await ctx.plugin(SubagentService)
+      const fiber = await ctx.plugin(WorkerWorkflowEngine, {})
+      expect(ctx.get('workflows')).toBeDefined()
+      // A zero-agent run through the DEFAULT config exercises the auto
+      // concurrency resolution (cores - 2, capped) in start().
+      const result = await run(ctx, fakeParent(), scripted('return 6 * 7'))
+      expect(result.value).toBe(42)
+      await fiber.dispose()
+      expect(ctx.get('workflows')).toBeUndefined()
+    })
+
+    it('has the class-plugin export shape (default = the engine service class)', () => {
+      expect(workerEngineModule.default).toBe(WorkerWorkflowEngine)
+      const loader = Object.create(Loader.prototype) as Loader
+      const unwrapped: unknown = loader.unwrapExports(workerEngineModule)
+      expect(unwrapped).toBe(WorkerWorkflowEngine)
+    })
+  })
+})

packages/workflow/workflow-workerthread/tsconfig.json

@@ -0,0 +1,39 @@
+{
+  "extends": "../../../tsconfig.base.json",
+  "compilerOptions": {
+    "rootDir": "src",
+    "outDir": "lib/types"
+  },
+  "include": [
+    "src"
+  ],
+  "references": [
+    {
+      "path": "../../../vendor/cosmokit"
+    },
+    {
+      "path": "../../../vendor/cordis"
+    },
+    {
+      "path": "../../../vendor/schemastery"
+    },
+    {
+      "path": "../../core/agent"
+    },
+    {
+      "path": "../../util/brand"
+    },
+    {
+      "path": "../../llm/llm"
+    },
+    {
+      "path": "../../subagent/subagent"
+    },
+    {
+      "path": "../../core/tools"
+    },
+    {
+      "path": "../workflow"
+    }
+  ]
+}

packages/workflow/workflow-workerthread/tsdown.config.ts

@@ -0,0 +1,32 @@
+import { defineConfig } from 'tsdown'
+
+/**
+ * The engine ships two runtime entries: the engine service (index) and the
+ * worker-thread entry (worker) the engine spawns via `new Worker`. The
+ * entries are JS emitted by tsc under lib/types and are bundled as two
+ * single-entry passes so shared modules (realm, runtime, session) are inlined
+ * into each instead of split into a hash-named chunk (the worker entry must
+ * be a self-contained file the Worker constructor can load by path).
+ */
+export default defineConfig([
+  {
+    entry: ['lib/types/index.js'],
+    outDir: 'lib',
+    format: ['esm'],
+    platform: 'node',
+    target: 'es2024',
+    fixedExtension: false,
+    dts: false,
+    clean: false,
+  },
+  {
+    entry: ['lib/types/worker.js'],
+    outDir: 'lib',
+    format: ['esm'],
+    platform: 'node',
+    target: 'es2024',
+    fixedExtension: false,
+    dts: false,
+    clean: false,
+  },
+])

packages/workflow/workflow/README.md

@@ -0,0 +1,29 @@
+# @deepseek-ai/dsh-workflow
+
+The **workflow seam** (`ctx.workflows`): an abstract service defining WHAT a workflow engine does — execute a model-written orchestration script that fans out subagents — without saying HOW. The bash-shaped third of the [workflow family](../README.md): implementations subclass `WorkflowService` and register as the `workflows` service (one per context); [`dsh-workflow-workerthread`](../workflow-workerthread/README.md) (one worker thread per run) is the implementation, and [`dsh-tool-workflow`](../tool-workflow/README.md) is the model-facing consumer.
+
+## Service: `WorkflowService` (abstract)
+
+`start(request: WorkflowStartRequest): WorkflowRun` — parse and execute a script. Throws synchronously (`SCRIPT_PARSE`/`META_INVALID`) for a script that cannot begin; once a run is returned, its `result` NEVER rejects — every failure resolves with `stopReason: 'error'` (or `'cancelled'`) — and once the run is cancelled, `result` settles within the implementation's bounded grace even if the script itself never settles (a consumer awaiting `result` must never be wedged past a cancellation). `dispose()` must reach quiescence within a bounded grace (cancel → wait for the script to settle and its children to finish disposing → abandon), never hanging its caller. Runs are HOLDER-owned: the engine does not track its live runs, so disposing the engine's fiber mid-run leaves each run to its holder's teardown.
+
+The protected `emitWorkflowEvent` helper dispatches the `workflow/*` events with PER-LISTENER containment and PER-LISTENER payload snapshots (a throwing subscriber is logged, never propagated, and cannot starve later listeners; each subscriber gets its own clone of the payload, so mutating it corrupts neither the engine nor other listeners) — the same containment guarantee as the subagent seam's lifecycle emits.
+
+## Vocabulary
+
+- `WorkflowStartRequest` — `{ script, args?, parent: Agent, signal? }`. `parent` is REQUIRED: every child the script spawns is attributed to it. `args` must be plain host-realm JSON data.
+- `WorkflowMeta` / `WorkflowPhase` — the workflow's identity block, carried as plain JSON data on the start request (Claude Code meta vocabulary: required `name`/`description`, optional `whenToUse`/`phases`) and shape-validated by the engine.
+- `WorkflowRun` — `{ id, meta, result, cancel(reason?), dispose() }`; the consumer awaits `result` and MUST `dispose` on every path.
+- `WorkflowResult` — `{ value, stopReason: 'completed'|'cancelled'|'error', error?, agentsStarted }`; `value` is the script's materialized return (plain JSON data; `null` for no return).
+- `WorkflowError` — `HarnessError` with a `WorkflowErrorCode` and a `fatal` flag driving the combinator discipline: a fatal error (bad hook arguments, unsupported options/schemas, tripped caps, seam start failures, cancellation) always propagates through `parallel()`/`pipeline()` instead of dissolving into a per-item `null`. `isFatalWorkflowError(error)` is the catch-site predicate.
+
+## Events
+
+All observe-only emits carrying DATA SNAPSHOTS (`WorkflowRunInfo` = id + meta) — never the live `WorkflowRun`, so a listener cannot gain `cancel`/`dispose`; control stays with the `start()` caller:
+
+- `workflow/start`(info) / `workflow/end`(info, resultInfo) — run lifecycle; `resultInfo` deliberately omits the value.
+- `workflow/phase`(info, title) / `workflow/log`(info, message) — script narration.
+- `workflow/agent-start`(info, agent) / `workflow/agent-end`(info, agent + outcome) — one pair per `agent()` call that STARTED a child run (a call rejected at validation or caps, refused at start, or cancelled while queued for a slot emits no pair), correlated by `seq`.
+
+## Non-goals (this cut)
+
+Background collection, journaling/resume, saved workflows, nested `workflow()`, token budgets — see the [RFC's deferred section](../../../docs/rfc/implemented/feature/2026-07-05-dynamic-workflows.md).

packages/workflow/workflow/package.json

@@ -0,0 +1,37 @@
+{
+  "name": "@deepseek-ai/dsh-workflow",
+  "description": "Workflow capability seam: ctx.workflows service, run vocabulary, and workflow/* events",
+  "version": "0.0.1",
+  "private": true,
+  "type": "module",
+  "main": "lib/index.js",
+  "types": "lib/types/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./lib/types/index.d.ts",
+      "default": "./lib/index.js"
+    },
+    "./src/*": "./src/*",
+    "./package.json": "./package.json"
+  },
+  "files": [
+    "lib/index.js",
+    "lib/types/**/*.d.ts",
+    "lib/types/**/*.d.ts.map",
+    "src"
+  ],
+  "license": "BSD-3-Clause",
+  "peerDependencies": {
+    "@deepseek-ai/dsh-agent": "^0.0.1",
+    "@deepseek-ai/dsh-brand": "^0.0.1",
+    "@deepseek-ai/dsh-llm": "^0.0.1",
+    "cordis": "^4.0.0-rc.6"
+  },
+  "devDependencies": {
+    "@deepseek-ai/dsh-agent": "workspace:^",
+    "@deepseek-ai/dsh-brand": "workspace:^",
+    "@deepseek-ai/dsh-llm": "workspace:^",
+    "@deepseek-ai/dsh-session": "workspace:^",
+    "cordis": "^4.0.0-rc.6"
+  }
+}

packages/workflow/workflow/src/index.ts

@@ -0,0 +1,269 @@
+/**
+ * The workflow capability seam (`ctx.workflows`): an abstract service defining
+ * WHAT a workflow engine does — execute a model-written orchestration script
+ * that fans out subagents — without saying HOW. Implementations subclass
+ * {@link WorkflowService} and register as the `workflows` service (one
+ * implementation per context, cordis' standard duplicate-service behavior);
+ * the implementation is `@deepseek-ai/dsh-workflow-workerthread`, which runs each
+ * script in its own worker thread. Hardened engines (an isolated-vm or
+ * separate-process sandbox) swap in without touching the model-facing tool
+ * that consumes them (`@deepseek-ai/dsh-tool-workflow`).
+ *
+ * The `workflow/*` lifecycle events are OBSERVE-ONLY data snapshots: they
+ * carry {@link WorkflowRunInfo} (id + meta), never the live {@link WorkflowRun}
+ * — a listener must not gain `cancel`/`dispose`; control stays with the
+ * `start()` caller holding the run. Every emit is per-listener contained (a
+ * throwing subscriber is logged, never propagated) and every listener gets its
+ * own payload clone (mutating it corrupts nothing), so one bad observer can
+ * neither strand a live run, starve later listeners, nor poison another
+ * listener's view.
+ *
+ * @module @deepseek-ai/dsh-workflow
+ */
+
+import { Context, Service } from 'cordis'
+import { HarnessError } from '@deepseek-ai/dsh-llm'
+import type {
+  WorkflowAgentEndInfo,
+  WorkflowAgentInfo,
+  WorkflowResultInfo,
+  WorkflowRun,
+  WorkflowRunInfo,
+  WorkflowStartRequest,
+} from './types.ts'
+
+export { WorkflowRunId } from './types.ts'
+export type {
+  WorkflowAgentEndInfo,
+  WorkflowAgentInfo,
+  WorkflowAgentOutcome,
+  WorkflowMeta,
+  WorkflowPhase,
+  WorkflowResult,
+  WorkflowResultInfo,
+  WorkflowRun,
+  WorkflowRunInfo,
+  WorkflowStartRequest,
+  WorkflowStopReason,
+} from './types.ts'
+
+declare module 'cordis' {
+  interface Context {
+    workflows: WorkflowService
+  }
+
+  interface Events {
+    /**
+     * A workflow run started — the script's meta block validated, the body
+     * about to execute. Paired with {@link Events['workflow/end']}.
+     * @param info - the run's identity snapshot (id + meta).
+     * @mode emit
+     */
+    'workflow/start'(info: WorkflowRunInfo): void
+    /**
+     * The script entered a phase (a `phase(title)` call) — progress grouping
+     * for observers; no execution semantics.
+     * @param info - the run's identity snapshot.
+     * @param title - the phase title, verbatim.
+     * @mode emit
+     */
+    'workflow/phase'(info: WorkflowRunInfo, title: string): void
+    /**
+     * The script emitted a narration line (a `log(message)` call).
+     * @param info - the run's identity snapshot.
+     * @param message - the logged message, verbatim.
+     * @mode emit
+     */
+    'workflow/log'(info: WorkflowRunInfo, message: string): void
+    /**
+     * One `agent()` call started a child run. Paired with
+     * {@link Events['workflow/agent-end']} by `agent.seq`.
+     * @param info - the run's identity snapshot.
+     * @param agent - the call's sequence number, label, phase, and child id.
+     * @mode emit
+     */
+    'workflow/agent-start'(info: WorkflowRunInfo, agent: WorkflowAgentInfo): void
+    /**
+     * One `agent()` call settled (clean result, child failure, or run
+     * cancellation). Paired with {@link Events['workflow/agent-start']} by
+     * `agent.seq`, exactly once per started call on every stop path — on an
+     * engine termination path (a worker killed past its grace) the end is
+     * engine-synthesized with outcome `'cancelled'`.
+     * @param info - the run's identity snapshot.
+     * @param agent - the call identity plus its outcome.
+     * @mode emit
+     */
+    'workflow/agent-end'(info: WorkflowRunInfo, agent: WorkflowAgentEndInfo): void
+    /**
+     * A workflow run settled (any stop reason). Fired when
+     * {@link WorkflowRun.result} resolves. Paired with
+     * {@link Events['workflow/start']}.
+     * @param info - the run's identity snapshot.
+     * @param result - the outcome data (stop reason, error, agent count) —
+     *   deliberately WITHOUT the result value (see {@link WorkflowResultInfo}).
+     * @mode emit
+     */
+    'workflow/end'(info: WorkflowRunInfo, result: WorkflowResultInfo): void
+  }
+}
+
+/** The full set of `workflow/*` event names {@link WorkflowService.emitWorkflowEvent} dispatches. */
+export type WorkflowEventName =
+  | 'workflow/start'
+  | 'workflow/phase'
+  | 'workflow/log'
+  | 'workflow/agent-start'
+  | 'workflow/agent-end'
+  | 'workflow/end'
+
+/**
+ * The workflow-seam error codes. Every one of these is FATAL when it reaches
+ * a script (see {@link WorkflowError.fatal}): the combinators re-throw it
+ * instead of dissolving it into an ordinary per-item `null`.
+ *
+ * - `SCRIPT_PARSE` — the script (or its meta statement) does not parse.
+ * - `META_INVALID` — the meta block evaluated but fails the shape contract.
+ * - `INVALID_ARGUMENT` — a hook was called with malformed arguments.
+ * - `UNSUPPORTED_OPTION` — an `agent()` option this engine does not support
+ *   (deferred: `effort`/`isolation`/`agentType`) or does not know.
+ * - `UNSUPPORTED_SCHEMA` — an `agent()` schema outside the structured-output
+ *   subset (see dsh-tools).
+ * - `AGENT_CAP` / `ITEM_CAP` — the run/agent caps tripped.
+ * - `AGENT_START` — the subagent seam refused to start a child.
+ * - `AGENT_RESULT` — a child's `result` REJECTED: an infrastructure fault at
+ *   the subagent seam, distinct from a child that failed and resolved (which
+ *   is the per-item `null`, never an error).
+ * - `RESULT_UNSERIALIZABLE` — a value crossing the script/host value boundary
+ *   is not plain JSON data.
+ * - `CANCELLED` — the run was cancelled; pending and future hooks reject
+ *   with this (the script-kill mechanism).
+ */
+export type WorkflowErrorCode =
+  | 'SCRIPT_PARSE'
+  | 'META_INVALID'
+  | 'INVALID_ARGUMENT'
+  | 'UNSUPPORTED_OPTION'
+  | 'UNSUPPORTED_SCHEMA'
+  | 'AGENT_CAP'
+  | 'ITEM_CAP'
+  | 'AGENT_START'
+  | 'AGENT_RESULT'
+  | 'RESULT_UNSERIALIZABLE'
+  | 'CANCELLED'
+
+/**
+ * Typed error for workflow-seam failures. Extends {@link HarnessError}, so the
+ * `code` is machine-routable taxonomy. `fatal` drives the combinator
+ * discipline: `parallel()`/`pipeline()` re-throw a fatal error (a typo'd
+ * option or a tripped cap must kill the script loudly), and reserve the
+ * per-item `null` for child-run failures and ordinary in-stage script errors.
+ * Every {@link WorkflowErrorCode} is fatal in this cut; the flag exists so the
+ * distinction is explicit at every catch site rather than implied.
+ */
+export class WorkflowError extends HarnessError {
+  /** Whether combinators must propagate this error instead of nulling the item. */
+  readonly fatal: boolean
+
+  constructor(message: string, code: WorkflowErrorCode, options?: ErrorOptions & { fatal?: boolean }) {
+    super(message, code, options)
+    this.name = 'WorkflowError'
+    this.fatal = options?.fatal ?? true
+  }
+}
+
+/**
+ * Whether combinators must re-throw `error` instead of mapping the item to `null`.
+ * @param error - any thrown value; fatality is host `instanceof` (unforgeable from a script realm).
+ * @returns true iff `error` is a {@link WorkflowError} whose `fatal` flag is set.
+ */
+export function isFatalWorkflowError(error: unknown): boolean {
+  return error instanceof WorkflowError && error.fatal
+}
+
+/**
+ * Abstract workflow execution service. Subclass, implement {@link start}, and
+ * load the subclass as a plugin — it registers as `ctx.workflows` (one
+ * implementation per context; loading a second throws, cordis' standard
+ * duplicate-service behavior).
+ *
+ * Semantics every implementation must honor:
+ * - {@link start} throws synchronously for a request that cannot begin (an
+ *   unparseable script, an invalid meta block). Once it returns a
+ *   {@link WorkflowRun}, `result` NEVER rejects — every failure resolves with
+ *   `stopReason: 'error'` (or `'cancelled'`) — and once the run is cancelled,
+ *   `result` SETTLES within the implementation's bounded grace even if the
+ *   script itself never settles (a consumer awaiting `result` must never be
+ *   wedged past a cancellation).
+ * - The `workflow/*` events fire through {@link emitWorkflowEvent} (data
+ *   snapshots, per-listener containment); `workflow/end` fires exactly once
+ *   per started run, after `result` is settled or as it settles.
+ * - `dispose()` reaches quiescence within a bounded grace: it cancels, waits
+ *   for the script to settle AND its started children to finish disposing,
+ *   and abandons whatever is left rather than hanging its caller (the engine
+ *   documents what abandonment leaves behind).
+ * - Runs are HOLDER-OWNED: the engine hands control (`cancel`/`dispose`) to
+ *   the `start()` caller and does not track its live runs — disposing the
+ *   engine's own fiber mid-run deliberately leaves those runs to their
+ *   holders' teardown, so an engine reload cannot yank a run out from under
+ *   the consumer awaiting it.
+ */
+export abstract class WorkflowService extends Service {
+  constructor(ctx: Context) {
+    super(ctx, 'workflows')
+  }
+
+  /**
+   * Parse and execute a workflow script.
+   * @param request - the script, its `args`, the parent agent, and an
+   *   optional cancel signal.
+   * @returns the live run; its `result` resolves when the script settles.
+   */
+  abstract start(request: WorkflowStartRequest): WorkflowRun
+
+  /**
+   * Emit one `workflow/*` lifecycle event with PER-LISTENER containment and
+   * PER-LISTENER payload snapshots: each subscriber is dispatched individually
+   * with its OWN structural clone of the payload (the payloads are plain JSON
+   * data by the seam contract), so a listener mutating what it received can
+   * corrupt neither the engine's live state nor any other listener's or later
+   * event's view; a thrown listener is logged (never propagated — the logging
+   * itself is total, even for a thrown value whose own string coercion
+   * throws), so one bad subscriber can neither fail the engine mid-run,
+   * surface as an unhandled rejection on a detached settle hook, nor starve
+   * the listeners registered after it (cordis `emit` halts on the first throw
+   * — same guarantee as the subagent seam's lifecycle emits).
+   * @param name - the `workflow/*` event to dispatch.
+   * @param args - the event's payload, matching its declared signature.
+   */
+  protected emitWorkflowEvent(name: WorkflowEventName, ...args: unknown[]): void {
+    for (const callback of this.ctx.events.dispatch('emit', [name, ...args])) {
+      try {
+        // The declared workflow/* signatures are all void-returning emits; the
+        // dispatch callback applies the payload tuple.
+        ;(callback as (...payload: unknown[]) => void)(...structuredClone(args))
+      } catch (error: unknown) {
+        this.ctx.logger.warn(`workflow: ${name} listener threw: ${renderListenerError(error)}`)
+      }
+    }
+  }
+}
+
+/**
+ * Total renderer for a listener-thrown value: the containment catch must never
+ * itself throw, and `String(error)` does when the value's own `toString` /
+ * `Symbol.toPrimitive` throws. Local rather than an engine package's renderer
+ * — the seam sits below every engine and cannot import one.
+ * @param error - any thrown value.
+ * @returns `String(error)`, or a fixed label when even coercion throws.
+ */
+function renderListenerError(error: unknown): string {
+  try {
+    return String(error)
+  } catch {
+    // Only a throwing toString/Symbol.toPrimitive lands here; the fixed label
+    // keeps the containment guarantee total.
+    return '[unrenderable thrown value]'
+  }
+}
+
+export default WorkflowService

packages/workflow/workflow/src/types.ts

@@ -0,0 +1,173 @@
+/**
+ * Workflow seam vocabulary: the request/run/result types a workflow engine
+ * consumes and produces, plus the payload shapes of the `workflow/*` events.
+ * Types only (plus the id-brand factory), per the package convention.
+ *
+ * @module @deepseek-ai/dsh-workflow/types
+ */
+
+import type { Branded } from '@deepseek-ai/dsh-brand'
+import type { Agent, AgentId } from '@deepseek-ai/dsh-agent'
+
+/** Identifies one workflow run. */
+export type WorkflowRunId = Branded<'WorkflowRunId'>
+
+/**
+ * Brand a string as a {@link WorkflowRunId}.
+ * @param id - the raw id string (the engine mints UUIDs; tests may pass fixtures).
+ * @returns the same string, branded.
+ */
+export function WorkflowRunId(id: string): WorkflowRunId {
+  return id as WorkflowRunId
+}
+
+/**
+ * One phase declared in a script's `meta.phases` (progress vocabulary only —
+ * phases group agents in observers/UIs; they impose no execution structure).
+ */
+export interface WorkflowPhase {
+  /** The phase title; `phase()` calls match against it by exact string. */
+  title: string
+  /** Optional one-line description of what the phase does. */
+  detail?: string
+  /** Optional model override this phase is expected to use (informational). */
+  model?: string
+}
+
+/**
+ * The script's identity block, provided as plain JSON data alongside the
+ * script body (the model-facing tool carries it as its `meta` parameter) and
+ * validated by the engine before the body runs. `name`/`description` are
+ * required; the rest is optional annotation. The field vocabulary matches the
+ * Claude Code dynamic-workflows meta block.
+ */
+export interface WorkflowMeta {
+  /** Short kebab-case workflow name (display + persistence key). */
+  name: string
+  /** One-line description of what the workflow does. */
+  description: string
+  /** Optional guidance on when this workflow applies (shown in listings). */
+  whenToUse?: string
+  /** Optional phase declarations matched by `phase()` calls. */
+  phases?: WorkflowPhase[]
+}
+
+/**
+ * What a caller asks for when starting a workflow run. `meta` and `args` are
+ * plain JSON DATA by the seam contract (the tool builds both from the model's
+ * schema-validated call; the engine validates `meta`'s shape and rejects loud
+ * before anything runs) — an engine never evaluates script text to obtain
+ * them. `parent` is REQUIRED — every `agent()` the script spawns is
+ * attributed to it (cwd, lineage, depth flow through the subagent seam).
+ */
+export interface WorkflowStartRequest {
+  /** The plain-JS script body (top-level await allowed; ends with `return <json-value>`). */
+  script: string
+  /** The workflow's identity block, as plain JSON data (shape-validated by the engine). */
+  meta: WorkflowMeta
+  /** Optional input exposed verbatim to the script as the `args` global. */
+  args?: unknown
+  /** The agent on whose behalf the run executes (parent of every child). */
+  parent: Agent
+  /** Cancels the run when aborted (the tool's `exec.signal`). */
+  signal?: AbortSignal
+}
+
+/**
+ * Why a run settled. CLOSED union (engine-owned, consumers may exhaust):
+ * `completed` = the script ran to its final `return`; `cancelled` = the run
+ * was cancelled (caller `cancel()`/signal); `error` = the script threw, a
+ * fatal `WorkflowError` propagated, or the result failed materialization.
+ */
+export type WorkflowStopReason = 'completed' | 'cancelled' | 'error'
+
+/**
+ * The outcome of one run, resolved by {@link WorkflowRun.result}. `value` is
+ * the script's materialized return value (plain host-realm JSON data; `null`
+ * when the script returned `undefined`) — meaningful only for `completed`.
+ * A non-`completed` reason carries the failure in `error`; the consumer maps
+ * it to an `isError` tool result rather than reporting partial output.
+ */
+export interface WorkflowResult {
+  /** The script's return value (host JSON data; `null` for no return). */
+  value: unknown
+  /** Why the run settled. */
+  stopReason: WorkflowStopReason
+  /** The failure message (present iff `stopReason` is not `completed`). */
+  error?: string
+  /**
+   * How many `agent()` calls the run accepted over its whole lifetime. On a
+   * graceful settlement this is the script-side count (calls still queued for
+   * a concurrency slot included); on a termination path (grace force-settle,
+   * worker death) it degrades to the host-observed count — calls queued
+   * inside a terminated script are unknowable then.
+   */
+  agentsStarted: number
+}
+
+/**
+ * The handle the consumer holds while a script executes. The consumer awaits
+ * `result`, may `cancel` mid-flight, and MUST `dispose` on every path.
+ * `result` does NOT reject — a script failure resolves with `stopReason:
+ * 'error'` — and once the run is cancelled it SETTLES within the engine's
+ * bounded grace even if the script itself never settles (the engine
+ * force-settles `cancelled`; what becomes of the script is engine-documented
+ * — the worker-thread engine terminates its worker), so a consumer awaiting
+ * `result` is never wedged past a cancellation. `dispose()` = cancel + that
+ * bounded settle + child quiescence; it never hangs on a stuck script and is
+ * safe to call on every path (idempotent).
+ */
+export interface WorkflowRun {
+  readonly id: WorkflowRunId
+  /** The validated meta block (available before the body runs). */
+  readonly meta: WorkflowMeta
+  readonly result: Promise<WorkflowResult>
+  /** Cancel the run: children abort, pending hooks reject, the script dies at its next await (or is force-settled at the grace). */
+  cancel(reason?: string): void
+  /** Cancel + bounded-grace settle; safe to call on every path (idempotent). */
+  dispose(): Promise<void>
+}
+
+/** Identifying detail for a run, carried by every `workflow/*` event (a data snapshot, never the live run). */
+export interface WorkflowRunInfo {
+  /** The run's id. */
+  id: WorkflowRunId
+  /** The run's validated meta block. */
+  meta: WorkflowMeta
+}
+
+/** One `agent()` call's identity within a run (the `workflow/agent-start` payload). */
+export interface WorkflowAgentInfo {
+  /** 1-based sequence number of this `agent()` call within the run. */
+  seq: number
+  /** The display label (the `label` option, or a prompt snippet). */
+  label: string
+  /** The phase this agent belongs to (the `phase` option, else the current `phase()` title). */
+  phase?: string
+  /** The child agent's id on the subagent seam. */
+  childId: AgentId
+}
+
+/** How one `agent()` call settled: clean result, child failure (script sees `null`), or run cancellation. */
+export type WorkflowAgentOutcome = 'completed' | 'failed' | 'cancelled'
+
+/** One `agent()` call's settlement (the `workflow/agent-end` payload). */
+export interface WorkflowAgentEndInfo extends WorkflowAgentInfo {
+  /** How the call settled. */
+  outcome: WorkflowAgentOutcome
+}
+
+/**
+ * A settled run's outcome as event data (the `workflow/end` payload): the
+ * {@link WorkflowResult} minus `value` (a listener observing outcomes must not
+ * receive a mutable alias of the caller's result value; a consumer that needs
+ * the value holds the run and awaits `result`).
+ */
+export interface WorkflowResultInfo {
+  /** Why the run settled. */
+  stopReason: WorkflowStopReason
+  /** The failure message (present iff `stopReason` is not `completed`). */
+  error?: string
+  /** How many `agent()` calls the run accepted (see {@link WorkflowResult.agentsStarted}). */
+  agentsStarted: number
+}

packages/workflow/workflow/tests/workflow.spec.ts

@@ -0,0 +1,124 @@
+import { describe, expect, it, vi } from 'vitest'
+import { Context } from 'cordis'
+import WorkflowServiceDefault, {
+  isFatalWorkflowError,
+  WorkflowError,
+  WorkflowRunId,
+  WorkflowService,
+} from '../src/index.ts'
+import type { WorkflowRun, WorkflowRunInfo, WorkflowStartRequest } from '../src/index.ts'
+
+/** A minimal concrete subclass exposing the protected emit helper for tests. */
+class StubEngine extends WorkflowService {
+  start(request: WorkflowStartRequest): WorkflowRun {
+    void request
+    throw new Error('not under test')
+  }
+
+  emit(name: Parameters<WorkflowService['emitWorkflowEvent']>[0], ...args: unknown[]): void {
+    this.emitWorkflowEvent(name, ...args)
+  }
+}
+
+const INFO: WorkflowRunInfo = { id: WorkflowRunId('run-1'), meta: { name: 'w', description: 'd' } }
+
+describe('dsh-workflow (interface)', () => {
+  it('WorkflowRunId brands a string (identity at runtime)', () => {
+    expect(WorkflowRunId('abc')).toBe('abc')
+  })
+
+  it('WorkflowError carries code + fatal (default true) and reads as a HarnessError', () => {
+    const error = new WorkflowError('cap hit', 'AGENT_CAP')
+    expect(error.code).toBe('AGENT_CAP')
+    expect(error.fatal).toBe(true)
+    expect(error.name).toBe('WorkflowError')
+    const soft = new WorkflowError('advisory', 'ITEM_CAP', { fatal: false })
+    expect(soft.fatal).toBe(false)
+  })
+
+  it('isFatalWorkflowError: true only for a fatal WorkflowError', () => {
+    expect(isFatalWorkflowError(new WorkflowError('x', 'CANCELLED'))).toBe(true)
+    expect(isFatalWorkflowError(new WorkflowError('x', 'CANCELLED', { fatal: false }))).toBe(false)
+    expect(isFatalWorkflowError(new Error('plain'))).toBe(false)
+    expect(isFatalWorkflowError('string')).toBe(false)
+  })
+
+  it('registers as ctx.workflows and unregisters when its fiber is disposed (HMR safety)', async () => {
+    const ctx = new Context()
+    const fiber = await ctx.plugin(StubEngine)
+    expect(ctx.get('workflows')).toBeInstanceOf(StubEngine)
+    await fiber.dispose()
+    expect(ctx.get('workflows')).toBeUndefined()
+  })
+
+  it('emitWorkflowEvent dispatches to every listener with the payload tuple', async () => {
+    const ctx = new Context()
+    await ctx.plugin(StubEngine)
+    const seen: unknown[][] = []
+    ctx.on('workflow/log', (info, message) => { seen.push([info, message]) })
+    ctx.on('workflow/agent-start', (info, agent) => { seen.push([info, agent]) })
+    const engine = ctx.workflows as StubEngine
+    engine.emit('workflow/log', INFO, 'hello')
+    engine.emit('workflow/agent-start', INFO, { seq: 1, label: 'l', childId: 'c' })
+    expect(seen).toEqual([
+      [INFO, 'hello'],
+      [INFO, { seq: 1, label: 'l', childId: 'c' }],
+    ])
+  })
+
+  it('gives each listener its OWN payload snapshot: mutation corrupts neither peers nor the caller', async () => {
+    const ctx = new Context()
+    await ctx.plugin(StubEngine)
+    const seen: string[] = []
+    ctx.on('workflow/agent-start', (info, agent) => {
+      agent.label = 'HACKED'
+      info.meta.name = 'HACKED'
+      seen.push('mutator')
+    })
+    ctx.on('workflow/agent-start', (info, agent) => {
+      seen.push(`${info.meta.name}/${agent.label}`)
+    })
+    const engine = ctx.workflows as StubEngine
+    const info: WorkflowRunInfo = { id: WorkflowRunId('run-2'), meta: { name: 'w', description: 'd' } }
+    const payload = { seq: 1, label: 'original', childId: 'c' }
+    engine.emit('workflow/agent-start', info, payload)
+    expect(seen).toEqual(['mutator', 'w/original'])
+    // The caller's own objects are pristine too — no listener ever saw them.
+    expect(info.meta.name).toBe('w')
+    expect(payload.label).toBe('original')
+  })
+
+  it('contains a throwing listener PER LISTENER: later listeners still run, nothing propagates', async () => {
+    const ctx = new Context()
+    await ctx.plugin(StubEngine)
+    const warn = vi.spyOn(ctx.logger, 'warn').mockImplementation(() => ctx.logger)
+    const reached: string[] = []
+    ctx.on('workflow/phase', () => { throw new Error('bad listener') })
+    ctx.on('workflow/phase', (_info, title) => { reached.push(title) })
+    const engine = ctx.workflows as StubEngine
+    expect(() => { engine.emit('workflow/phase', INFO, 'Scan') }).not.toThrow()
+    expect(reached).toEqual(['Scan'])
+    expect(warn).toHaveBeenCalledOnce()
+    expect(String(warn.mock.calls[0]![0])).toContain('workflow/phase listener threw')
+  })
+
+  it('containment is total: a listener throwing a value whose coercion throws neither propagates nor starves later listeners', async () => {
+    const ctx = new Context()
+    await ctx.plugin(StubEngine)
+    const warn = vi.spyOn(ctx.logger, 'warn').mockImplementation(() => ctx.logger)
+    const reached: string[] = []
+    ctx.on('workflow/phase', () => {
+      throw { toString: () => { throw new Error('coercion trap') } }
+    })
+    ctx.on('workflow/phase', (_info, title) => { reached.push(title) })
+    const engine = ctx.workflows as StubEngine
+    expect(() => { engine.emit('workflow/phase', INFO, 'Scan') }).not.toThrow()
+    expect(reached).toEqual(['Scan'])
+    expect(warn).toHaveBeenCalledOnce()
+    expect(String(warn.mock.calls[0]![0])).toContain('[unrenderable thrown value]')
+  })
+
+  it('has the expected export surface (default = the abstract service class)', () => {
+    expect(WorkflowServiceDefault).toBe(WorkflowService)
+  })
+})

packages/workflow/workflow/tsconfig.json

@@ -0,0 +1,27 @@
+{
+  "extends": "../../../tsconfig.base.json",
+  "compilerOptions": {
+    "rootDir": "src",
+    "outDir": "lib/types"
+  },
+  "include": [
+    "src"
+  ],
+  "references": [
+    {
+      "path": "../../../vendor/cosmokit"
+    },
+    {
+      "path": "../../../vendor/cordis"
+    },
+    {
+      "path": "../../core/agent"
+    },
+    {
+      "path": "../../util/brand"
+    },
+    {
+      "path": "../../llm/llm"
+    }
+  ]
+}