Merge branch 'master' into agent/task-surface-rfc
This commit is contained in:
Ziya
@@ -1,6 +0,0 @@
|
||||
# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
|
||||
# side as of the last confirmed-consistent state. Both languages carry equal authority;
|
||||
# after editing either side, bring the other along and re-record with:
|
||||
# pnpm run verify-translation-pairing --write .agents/notes/proposed/feature/2026-07-07-claude-code-and-codex-subagent-backends.md
|
||||
2026-07-07-claude-code-and-codex-subagent-backends.md: ee8576f97a9fdef8c88dcad3a73f28b63ca3ebe1
|
||||
2026-07-07-claude-code-and-codex-subagent-backends.zh.md: bd76a8f34d86b704494b56331c47ea89bfc8a0aa
|
||||
@@ -1,89 +0,0 @@
|
||||
# Agent Note: Claude Code and Codex subagent backends (out-of-process delegation to external coding agents)
|
||||
|
||||
Status: proposed
|
||||
|
||||
English | [中文](2026-07-07-claude-code-and-codex-subagent-backends.zh.md)
|
||||
|
||||
## Problem
|
||||
|
||||
The subagent seam ([the seam Agent Note](../../implemented/feature/2026-06-21-subagent-capability-seam.md)) hosts multiple named providers on `ctx.subagents`, and the ACP backend ([the ACP backend Agent Note](../../implemented/feature/2026-06-22-acp-subagent-backend.md)) proved the seam generalizes across a process boundary; its Future-providers section explicitly named the Codex app-server and the Claude Code Agent SDK as mechanically similar siblings. Those two are the engines actually worth delegating to today: a harness turn should be able to hand a self-contained task to a real Claude Code or a real Codex — a separate product with its own model, tools, and sandbox — and get back one final answer, without the parent deployment leaking its secrets into the child or the child's behavior silently depending on whatever `~/.claude` / `~/.codex` state exists on the host machine.
|
||||
|
||||
## Proposal
|
||||
|
||||
Two sibling provider packages, structural variants of the ACP backend, plus one extraction:
|
||||
|
||||
- `@deepseek-ai/dsh-subagent-claude-code` — drives a Claude Code child through `@anthropic-ai/claude-agent-sdk`'s `query()` (the SDK runs in the parent process and spawns its bundled `claude` CLI as the subprocess). Provider name `claude-code`: the child is the Claude Code *product*, not an Anthropic model adapter — "claude" stays reserved for a future `dsh-llm` adapter.
|
||||
- `@deepseek-ai/dsh-subagent-codex` — spawns `codex app-server` and drives one thread/turn over its JSON-RPC-over-stdio protocol with a hand-rolled newline-JSON client (~200–300 lines) in the package.
|
||||
- `@deepseek-ai/dsh-subagent-process` — a pure library (the `subagent-inprocess` precedent) extracting what `dsh-subagent-acp` already carries and both new backends need: the credential env scrub (`buildChildEnv`), the EOF → SIGTERM → SIGKILL dispose ladder, and new isolated-config-dir helpers (`mkdtemp` create, best-effort remove). The ACP backend migrates onto it; `bash-local`'s sibling copy is left alone to bound the change.
|
||||
|
||||
Both providers copy the ACP backend's seam posture verbatim: fresh child per `start`, exactly one prompt round-trip, capabilities all `false`, `inheritsParentContext: false`, `request.parent`/`request.agentOptions` ignored, `id = SessionId(randomUUID())`, `result` never rejects — child-level failure flattens to a stop reason and the original error goes to `ctx.logger` via an `onError` spec callback. Model exposure is zero new code: `dsh-tool-subagent` is loaded once per provider with a distinct `toolName` (`subagent_claude_code`, `subagent_codex`). No new session events are needed — the only model-visible artifact is the tool result, so reconstructability holds exactly as it did for ACP. To be explicit about the boundary: the session log reconstructs the model-visible transcript, not workspace mutation history — a child granted write access mutates files as an ambient side effect outside the log, exactly as the bash tools and the ACP backend already do; replay reproduces requests, not the disk.
|
||||
|
||||
## Verified interface facts (pinned versions)
|
||||
|
||||
Both integration surfaces were verified against pinned implementations before this proposal — types and bundled source read, keyless spikes run — not from vendor docs alone. The pins are the verification baseline, not a runtime contract: the backends perform no runtime version probe (no `codex --version` gate, no SDK version sniffing). Compatibility is enforced at development time — every dependency bump re-runs the keyless suites against the real load path — and at runtime by failing loudly: a protocol-level surprise settles `error` via `onError`, never a silent misbehavior.
|
||||
|
||||
**`@anthropic-ai/claude-agent-sdk` 0.3.202.** `options.env` REPLACES the child environment (no merge with `process.env`), which is exactly what the scrub needs. `settingSources` defaults to loading ALL filesystem settings — isolation requires explicitly passing `[]`. Result subtypes are `success` | `error_during_execution` | `error_max_turns` | `error_max_budget_usd` | `error_max_structured_output_retries`. On abort the SDK escalates the CLI child itself: stdin EOF immediately, SIGTERM ~2s later if the child ignores it (observed; no leftover processes) — no bespoke kill fallback needed. `outputFormat: {type: 'json_schema'}` and an `agents` option exist, giving future landing points for the seam's `outputSchema` capability and named subagent types; both are out of scope here.
|
||||
|
||||
**codex CLI 0.142.5, `codex app-server` (v2 vocabulary).** LF-delimited JSON, JSON-RPC 2.0 shapes with the `"jsonrpc"` header omitted.
|
||||
|
||||
- Lifecycle: `initialize{clientInfo}` + `initialized` → `thread/start` (accepts `cwd`, `model`, `sandbox`, `approvalPolicy`, `ephemeral`; succeeds unauthenticated) → `turn/start{threadId, input:[{type:'text',text}]}` returns an `inProgress` turn immediately; the terminal signal is the `turn/completed` notification carrying `Turn{status: completed|interrupted|failed|inProgress, error}`.
|
||||
- Approvals are server-initiated requests — `item/commandExecution/requestApproval`, `item/fileChange/requestApproval`, `item/permissions/requestApproval`, `item/tool/requestUserInput`, `mcpServer/elicitation/request` — answered with `accept`/`decline`-family decisions.
|
||||
- Auth: `account/login/start{type:'apiKey', apiKey}` is a first-class RPC and `account/read` reports `requiresOpenaiAuth` — and an unauthenticated `turn/start` does NOT fail fast (it hangs in retry), so the backend MUST pre-check auth and settle `error` loudly instead of waiting on the turn.
|
||||
- Isolation: `CODEX_HOME` redirection is honored (the `initialize` response echoes it, so tests can assert isolation), and `ephemeral: true` threads leave no session files at all.
|
||||
|
||||
## Isolation and credentials
|
||||
|
||||
Deployments authenticate with API keys only, and the child must not see the host user's Claude Code / Codex configuration: behavior has to be a function of `cordis.yml` alone. Each run gets a fresh `mkdtemp` config dir — `CLAUDE_CONFIG_DIR` for Claude Code (paired with an explicit `settingSources: []`), `CODEX_HOME` for Codex — removed best-effort on dispose; a config field can pin a persistent dir instead. The child env reuses the ACP backend's `buildChildEnv` semantics verbatim via the extraction: the ambient env is forwarded MINUS credential-shaped vars (`/KEY|SECRET|TOKEN/i`), with `config.env` layered on top — so `PATH`, `HOME`, `TMPDIR`, locale, and proxy vars survive and the CLIs run normally, while only credential-shaped ambient vars are scrubbed (`ANTHROPIC_API_KEY` enters explicitly through `config.env` for Claude Code), and the Codex key travels via the `account/login/start` RPC into the isolated `CODEX_HOME` rather than a hand-written `auth.json`.
|
||||
|
||||
## Permission and approval policy
|
||||
|
||||
Instead of collapsing to ACP's single `permission: allow|reject` knob, each backend exposes its engine's native vocabulary as config, with conservative defaults: Claude Code gets `permissionMode` (default `default`) plus `permission: allow|reject` (default `reject`) as the `canUseTool` auto-answer for whatever falls through; Codex gets `sandboxMode` (default `read-only`) and `approvalPolicy` (default `never`) plus the same `permission` fallback for approval requests that still arrive. Defaults are deliberately do-no-harm (the out-of-box child cannot write files); examples demonstrate opening up (`acceptEdits` / `workspace-write`). The mechanical rule: EVERY server-initiated request is settled programmatically and promptly — the enumerated approval/user-input/elicitation requests by the configured policy, an unknown request method with a JSON-RPC method-not-found error response (never left pending), unknown notifications consumed — so no child request can wedge a turn waiting on an answer that will never come. Prompts never reach a human in this cut, matching ACP.
|
||||
|
||||
## StopReason mapping
|
||||
|
||||
Claude Code: `success` → `completed`; `error_max_turns`, `error_during_execution`, `error_max_budget_usd`, `error_max_structured_output_retries` → `error` (aligning with the ACP call on `max_turn_requests`: an unfinished task is not success); generator abort → `aborted`; anything unknown → `error`. Codex: `Turn.status` `completed` → `completed`; `interrupted` → `aborted`; `failed` with `codexErrorInfo: 'contextWindowExceeded'` → `max-tokens`, any other `failed` → `error`; transport/spawn/auth-precheck failure → `error` (or `aborted` if cancel was requested). In both, `cancel()` is the ACP shape: flag + abort/interrupt + a cancel-settled race arm so an uncooperative child cannot stall the result.
|
||||
|
||||
Liveness posture, stated explicitly: teardown timing is config, turn duration is not. Both backends take the dispose ladder's grace periods as defaulted validated config fields (the ACP backend's `disposeEofGraceMs`/`disposeGraceMs` shape, carried by the extraction), but there is deliberately NO turn-duration or startup timeout — matching ACP, liveness during a turn belongs to the caller via `cancel()`/the abort signal, a subagent turn is legitimately minutes long, and the Codex auth precheck removes the one verified guaranteed-hang; a deployment wanting a wall-clock bound cancels from the parent.
|
||||
|
||||
## Testing
|
||||
|
||||
Named at every tier per the root AGENTS.md rule, and de-risked up front:
|
||||
|
||||
- **Keyless unit/integration**, mirroring the ACP spec list per backend (round-trip and output accumulation, every stop mapping, both cancel paths, already-aborted, permission auto-answer under both policies, unknown-message tolerance, bad-command spawn failure, HMR provider cleanup, export shape, isolation assertions on child env and temp-dir removal; Codex adds the auth-precheck failure path). Claude Code's harness is a scripted fake `claude` executable behind `pathToClaudeCodeExecutable` driven by the REAL SDK — a spike already passed end-to-end keyless in 24ms (the fake CLI answers one `control_request/initialize` and speaks plain stream-json, ~40 lines). Codex's harness is a scripted mock app-server subprocess speaking the verified wire protocol, the `mock-acp-server.ts` shape.
|
||||
- **With-key e2e** per backend: the real engine does real file work verified on disk, under a pinned opened-up config so acceptance and the do-no-harm defaults don't collide — `permissionMode: 'acceptEdits'` for Claude Code, `sandboxMode: 'workspace-write'` + `approvalPolicy: 'never'` for Codex; self-skips report exactly what is missing (binary vs key). CI has no secrets, so these run locally per the with-key policy.
|
||||
- **Snapshot**: deferred as `TODO(claude-code-subagent-replay)` / `TODO(codex-subagent-replay)` — the same distinct replay shape the ACP backend deferred ([the per-session replay Agent Note](../../implemented/testing/2026-06-22-subagent-snapshot-replay.md)); the keyless suites carry deterministic coverage meanwhile.
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
### Why not the official `@openai/codex-sdk` instead of a hand-rolled client?
|
||||
|
||||
The dispose ladder and env scrub require owning the child process (spawn args, env, signals, exit await); the SDK hides the process. The wire format is trivial to frame (LF JSON), the shapes are generatable per pinned version (`codex app-server generate-json-schema`), and the repo precedent (`hook-protocol`) is to own thin protocol cores rather than wrap someone's runtime. The SDK would save protocol-evolution maintenance but costs the exact control this backend exists to have.
|
||||
|
||||
### Why not a model-visible `subagent_type` parameter (one Task-style tool)?
|
||||
|
||||
Claude Code's own Task tool puts the subagent type in the model-facing schema, selecting a prompt-plus-toolset persona. Here the choice is between EXECUTION ENGINES, and only the deployer knows which engines have credentials configured — so selection stays deployment config, preserving `dsh-tool-subagent`'s documented one-provider-per-tool contract. A persona-style type selector would be a separate Agent Note against the tool, not the backends.
|
||||
|
||||
### Why not login-state credentials and the user's own config?
|
||||
|
||||
Inheriting `~/.claude` / `~/.codex` (subscription login, user settings, skills, MCP servers) would make child behavior depend on host-machine state and punch an implicit exception through the "credentials enter explicitly via `config.env`, never ambiently" rule the ACP backend and bash executor established. API-key-only plus forced config-dir isolation keeps runs reproducible; deployments wanting shared state can point the config-dir field at a persistent directory deliberately.
|
||||
|
||||
### Why not a driver-injection seam for the Claude Code keyless tests?
|
||||
|
||||
Injecting a fake `query()` would mock our own boundary and leave the real SDK load path untested (the real-over-mock policy in docs/testing.md). The risk that justified considering it — the SDK↔CLI stream-json control protocol being internal — was retired by the spike: the fake-CLI harness works against the real pinned SDK today. If an SDK upgrade breaks the mock, the keyless suite fails the upgrade PR, which is the gate working.
|
||||
|
||||
### Why not ACP adapters (e.g. `claude-code-acp`) reusing the existing backend?
|
||||
|
||||
Community shims wrap both engines in ACP, which would make them "just config" on `dsh-subagent-acp`. But that inserts an unofficial third-party layer between the harness and the engine, erases the native control surfaces this Agent Note exposes (permissionMode, sandboxMode/approvalPolicy, config-dir isolation, apiKey RPC), and trades first-party protocol stability for a shim's release cadence. First-party surfaces — the Agent SDK and the app-server — are the supported integration points.
|
||||
|
||||
## Acceptance criteria
|
||||
|
||||
On a machine with both engines and keys configured: a REPL-driven model completes one real file task through `subagent_claude_code` and one through `subagent_codex`, the tool result being the child's final answer, with only `tool/call` + `tool/result` in the parent session log. Keyless suites pass at 100% per-file coverage in a credential-less environment, asserting isolation (scrubbed child env, no temp config dirs left after dispose) and that child behavior is unchanged by the presence or absence of `~/.claude` / `~/.codex`. Cancelling a parent turn quiesces both backends in bounded time with no leftover child processes. E2e suites self-skip cleanly, naming the missing prerequisite.
|
||||
|
||||
## Risks
|
||||
|
||||
- `codex app-server` is CLI-flagged experimental and its v1/v2 vocabularies coexist; the client pins 0.142.5, implements v2 only, and consumes unknown methods/notifications without crashing, but a future codex bump can still force rework (regenerate schemas and re-run the keyless suite on every bump — the development-time enforcement behind the no-runtime-version-probe stance above).
|
||||
- The Claude Code fake-CLI mock rides an internal protocol: any SDK upgrade must go through the keyless suite, and a breaking control-protocol change means reworking the mock (fallback: the driver-injection seam rejected above becomes the escape hatch).
|
||||
- The SDK's optionalDependencies weigh ~280MB per platform — accepted, and confined to the one backend package.
|
||||
- The SDK's SIGKILL branch beyond EOF→SIGTERM was not observed and is trusted; e2e keeps a no-leftover-process assertion.
|
||||
- Codex is a deployment prerequisite (no npm-bundled binary); a missing or incompatible binary surfaces as a loud spawn/protocol `error`, not a version probe.
|
||||
- Every run pays a fresh child process and only the final answer surfaces — thoughts, tool cards, and usage are consumed and dropped; pooling, intermediate-progress surfacing, `sendMessage`/`resume`, `outputSchema` via the SDK's `outputFormat`, and named subagent types via the SDK's `agents` option are all deliberate deferrals.
|
||||
@@ -1,89 +0,0 @@
|
||||
# Agent Note: Claude Code 与 Codex subagent 后端（向外部编码 agent 的进程外委派）
|
||||
|
||||
Status: proposed
|
||||
|
||||
[English](2026-07-07-claude-code-and-codex-subagent-backends.md) | 中文
|
||||
|
||||
## 问题
|
||||
|
||||
subagent seam（[seam Agent Note](../../implemented/feature/2026-06-21-subagent-capability-seam.md)）在 `ctx.subagents` 上托管多个命名提供方，ACP（Agent Client Protocol）后端（[ACP 后端 Agent Note](../../implemented/feature/2026-06-22-acp-subagent-backend.md)）证明了该 seam 能跨越进程边界泛化；其「未来提供方」一节明确将 Codex app-server 与 Claude Code Agent SDK 列为实现机制相似的同类方案。如今真正值得委派的就是这两个引擎：harness 的一个轮次应能把一个自包含任务交给真实的 Claude Code 或真实的 Codex——一个拥有自身模型、工具与沙箱的独立产品——并取回一个最终答案，同时父部署不向子进程泄漏密钥，子进程行为也不静默依赖宿主机上碰巧存在的 `~/.claude` / `~/.codex` 状态。
|
||||
|
||||
## 提案
|
||||
|
||||
两个兄弟提供方包，作为 ACP 后端的结构变体，另加一次提取：
|
||||
|
||||
- `@deepseek-ai/dsh-subagent-claude-code`：通过 `@anthropic-ai/claude-agent-sdk` 的 `query()` 驱动一个 Claude Code 子进程（SDK 在父进程中运行，并将其内置的 `claude` CLI（命令行界面）作为子进程 spawn）。提供方名称为 `claude-code`：子进程是 Claude Code 这个*产品*，而非 Anthropic 模型适配器——「claude」保留给未来的 `dsh-llm` 适配器。
|
||||
- `@deepseek-ai/dsh-subagent-codex`：spawn `codex app-server`，通过其 JSON-RPC-over-stdio 协议驱动一个 thread 及其中的一个轮次，使用包内一个手写的换行 JSON 客户端（约 200–300 行）。
|
||||
- `@deepseek-ai/dsh-subagent-process`：纯库（沿用 `subagent-inprocess` 的先例），提取 `dsh-subagent-acp` 已有且两个新后端都需要的内容：凭证环境清洗（`buildChildEnv`）、EOF → SIGTERM → SIGKILL 的 dispose（资源释放）阶梯，以及新的隔离配置目录辅助函数（`mkdtemp` 创建、尽力删除）。ACP 后端迁移到该库上；`bash-local` 的兄弟副本保持不动以限制变更范围。
|
||||
|
||||
两个提供方逐字复制 ACP 后端的 seam 姿态：每次 `start` 创建全新子进程、恰好一次提示词往返、所有能力均为 `false`、`inheritsParentContext: false`、忽略 `request.parent`/`request.agentOptions`、`id = SessionId(randomUUID())`，且 `result` 从不 reject——子进程级失败扁平化为 stop reason，原始错误则通过 `onError` spec 回调送到 `ctx.logger`。模型暴露无需新代码：每个提供方各加载一次 `dsh-tool-subagent`，使用不同的 `toolName`（`subagent_claude_code`、`subagent_codex`）。无需新的会话事件——唯一的模型可见产物是工具结果，因此可重建性与 ACP 完全相同。明确边界：会话日志重建模型可见的 transcript（文本记录），而不是工作区变更历史——获准写入的子进程将文件作为日志之外的环境副作用进行修改，与 bash 工具和 ACP 后端现有行为完全一致；回放复现请求，而非磁盘。
|
||||
|
||||
## 已验证的接口事实（固定版本）
|
||||
|
||||
两个集成面在本提案之前均已针对固定版本进行了验证——阅读类型与打包源码、运行无需密钥的 spike——而非仅依赖厂商文档。固定版本是验证基线，不是运行时契约：后端不执行运行时版本探测（无 `codex --version` 门禁、无 SDK 版本嗅探）。兼容性在开发时强制执行——每次依赖升级都会针对真实加载路径重跑无密钥套件——在运行时则通过显式失败来保障：协议层面的意外通过 `onError` 结算为 `error`，绝不静默异常。
|
||||
|
||||
**`@anthropic-ai/claude-agent-sdk` 0.3.202。** `options.env` 会替换子进程环境（不与 `process.env` 合并），恰好满足清洗需求。`settingSources` 默认加载所有文件系统设置——隔离要求显式传入 `[]`。结果子类型为 `success` | `error_during_execution` | `error_max_turns` | `error_max_budget_usd` | `error_max_structured_output_retries`。中止时 SDK 自行逐级加强对 CLI 子进程的终止措施：立即关闭 stdin，约 2 秒后若子进程未退出则发送 SIGTERM（已观察到；无残留进程）——无需自定义 kill 回退。`outputFormat: {type: 'json_schema'}` 和 `agents` 选项已存在，为 seam 的 `outputSchema` 能力和命名 subagent 类型提供了未来着陆点；两者均不在本 Agent Note 范围内。
|
||||
|
||||
**codex CLI 0.142.5，`codex app-server`（v2 词汇）。** LF 分隔的 JSON，JSON-RPC 2.0 形状但省略 `"jsonrpc"` 头。
|
||||
|
||||
- 生命周期：`initialize{clientInfo}` + `initialized` → `thread/start`（接受 `cwd`、`model`、`sandbox`、`approvalPolicy`、`ephemeral`；未认证即可成功）→ `turn/start{threadId, input:[{type:'text',text}]}` 立即返回一个 `inProgress` 的轮次；终止信号是携带 `Turn{status: completed|interrupted|failed|inProgress, error}` 的 `turn/completed` 通知。
|
||||
- 审批是服务端发起的请求——`item/commandExecution/requestApproval`、`item/fileChange/requestApproval`、`item/permissions/requestApproval`、`item/tool/requestUserInput`、`mcpServer/elicitation/request`——以 `accept`/`decline` 系列决策应答。
|
||||
- 认证：`account/login/start{type:'apiKey', apiKey}` 是一等 RPC，`account/read` 报告 `requiresOpenaiAuth`——且未认证的 `turn/start` 不会快速失败（它会挂在重试中），因此后端必须预检认证状态，并在失败时明确结算为 `error`，而非等待轮次。
|
||||
- 隔离：`CODEX_HOME` 重定向被尊重（`initialize` 响应会回显它，测试可据此断言隔离），`ephemeral: true` 的 thread 不留任何会话文件。
|
||||
|
||||
## 隔离与凭证
|
||||
|
||||
部署只使用 API key 认证，子进程不得看到宿主用户的 Claude Code / Codex 配置：行为必须只由 `cordis.yml` 决定。每次运行获得一个全新的 `mkdtemp` 配置目录——Claude Code 使用 `CLAUDE_CONFIG_DIR`（并显式设置 `settingSources: []`），Codex 使用 `CODEX_HOME`——dispose 时尽力删除；配置字段也可以固定一个持久目录。子进程环境通过提取逐字复用 ACP 后端的 `buildChildEnv` 语义：转发环境变量，但移除凭证形态的变量（`/KEY|SECRET|TOKEN/i`），再叠加 `config.env`——因此 `PATH`、`HOME`、`TMPDIR`、locale 和代理变量保留，CLI 正常运行；只有环境中的凭证形态变量被清洗（Claude Code 的 `ANTHROPIC_API_KEY` 通过 `config.env` 显式进入），Codex key 则通过 `account/login/start` RPC 进入隔离的 `CODEX_HOME`，而非手写 `auth.json`。
|
||||
|
||||
## 权限与审批策略
|
||||
|
||||
每个后端不压缩为 ACP 单一的 `permission: allow|reject` 旋钮，而把引擎原生词汇作为配置暴露，并采用保守默认值：Claude Code 获得 `permissionMode`（默认 `default`）以及 `permission: allow|reject`（默认 `reject`），后者作为所有未被前者处理的请求的 `canUseTool` 自动应答；Codex 获得 `sandboxMode`（默认 `read-only`）和 `approvalPolicy`（默认 `never`），以及同一个 `permission` 后备值，用来应答仍然到达的审批请求。默认值刻意做到不造成损害（开箱即用的子进程无法写文件）；示例演示如何开放权限（`acceptEdits` / `workspace-write`）。机械规则是：每一个服务端发起的请求都由程序迅速结算——枚举出的审批/用户输入/elicitation 请求按配置策略应答，未知请求方法用 JSON-RPC method-not-found 错误响应（绝不保持 pending），未知通知被消费——因此任何子进程请求都不会因等待永远不会到来的应答而卡住轮次。这一版中提示词不会到达人类，与 ACP 一致。
|
||||
|
||||
## StopReason 映射
|
||||
|
||||
Claude Code：`success` → `completed`；`error_max_turns`、`error_during_execution`、`error_max_budget_usd`、`error_max_structured_output_retries` → `error`（与 ACP 对 `max_turn_requests` 的处理对齐：未完成的任务不是成功）；生成器中止 → `aborted`；未知值 → `error`。Codex：`Turn.status` 为 `completed` → `completed`；`interrupted` → `aborted`；`failed` 且 `codexErrorInfo: 'contextWindowExceeded'` → `max-tokens`，其他 `failed` → `error`；传输/spawn/认证预检失败 → `error`（若已请求取消则为 `aborted`）。两者中，`cancel()` 采用 ACP 形状：标志位 + abort/interrupt + 一个 cancel-settled 竞争分支，使不合作的子进程无法阻塞结果。
|
||||
|
||||
活性姿态，明确声明：teardown 时序是配置项，轮次时长不是。两个后端将 dispose 阶梯的宽限期作为带默认值的已验证配置字段（ACP 后端的 `disposeEofGraceMs`/`disposeGraceMs` 形状，由提取库承载），但刻意不设轮次时长或启动超时——与 ACP 一致：轮次期间的活性由调用方通过 `cancel()`/abort signal 掌控，subagent 轮次持续数分钟也属合理，而 Codex 认证预检消除了唯一已验证的必然挂起场景；需要墙钟上限的部署从父侧取消即可。
|
||||
|
||||
## 测试
|
||||
|
||||
依照根 AGENTS.md 规则在每个层级明确命名，并预先消除风险：
|
||||
|
||||
- **无密钥单元/集成测试**：每个后端都镜像 ACP spec 清单（往返和输出累积、每种 stop 映射、两条取消路径、已中止、两种策略下的权限自动应答、未知消息容错、错误命令的 spawn 失败、HMR（热模块替换）提供方清理、导出形状、对子进程环境隔离和临时目录删除的断言；Codex 另加认证预检失败路径）。Claude Code harness 是通过 `pathToClaudeCodeExecutable` 接入真实 SDK 的脚本化假 `claude` 可执行文件——一个 spike 已在 24 ms 内完成端到端无密钥验证（假 CLI 应答一次 `control_request/initialize`，并使用普通 stream-json 通信，约 40 行）。Codex harness 是讲已验证协议格式的脚本化 mock app-server 子进程，沿用 `mock-acp-server.ts` 形状。
|
||||
- **有密钥 e2e 测试**：每个后端的真实引擎执行真实文件操作，并通过磁盘状态进行验证，固定使用开放后的配置，以免验收与不造成损害的默认值冲突——Claude Code 使用 `permissionMode: 'acceptEdits'`，Codex 使用 `sandboxMode: 'workspace-write'` + `approvalPolicy: 'never'`；自跳过会准确报告缺失的是二进制还是 key。CI 没有密钥，因此依照有密钥策略在本地运行。
|
||||
- **快照测试**：以 `TODO(claude-code-subagent-replay)` / `TODO(codex-subagent-replay)` 推迟——即 ACP 后端也推迟的独立回放形状（[按会话回放 Agent Note](../../implemented/testing/2026-06-22-subagent-snapshot-replay.md)）；在此期间由无密钥套件提供确定性覆盖。
|
||||
|
||||
## 曾考虑的替代方案
|
||||
|
||||
### 为什么不用官方 `@openai/codex-sdk` 而手写客户端？
|
||||
|
||||
dispose 阶梯和环境清洗要求拥有子进程（spawn 参数、env、信号、exit 等待）；SDK 隐藏了进程。协议格式（wire format）极其简单（LF JSON），形状可按固定版本生成（`codex app-server generate-json-schema`），仓库先例（`hook-protocol`）是拥有薄协议核心而非包装他人的运行时。SDK 能节省协议演进的维护成本，但代价是失去本后端存在的意义所在的精确控制。
|
||||
|
||||
### 为什么不用模型可见的 `subagent_type` 参数（单一 Task 风格工具）？
|
||||
|
||||
Claude Code 自身的 Task 工具将 subagent 类型放在模型可见的 schema 中，选择一个提示词 + 工具集人格。这里的选择是在执行引擎之间做出的，而只有部署者知道哪些引擎配置了凭证——因此选择留在部署配置层，保持 `dsh-tool-subagent` 文档中的「一个提供方对应一个工具」契约。人格风格的类型选择器应是针对工具的另一个 Agent Note，而非针对后端。
|
||||
|
||||
### 为什么不用登录态凭证和用户自身的配置？
|
||||
|
||||
继承 `~/.claude` / `~/.codex`（订阅登录、用户设置、skill（技能）、MCP 服务器）会使子进程行为依赖宿主机状态，并在 ACP 后端和 bash 执行器确立的「凭证通过 `config.env` 显式进入，绝不隐式继承」规则上打开一个隐式例外。仅 API key 加强制配置目录隔离使运行可复现；需要共享状态的部署可以有意将配置目录字段指向一个持久目录。
|
||||
|
||||
### 为什么不为 Claude Code 无密钥测试注入驱动层 seam？
|
||||
|
||||
注入假的 `query()` 会 mock 我们自己的边界，使真实 SDK 加载路径未被测试（docs/testing.md 中的 real-over-mock 策略）。曾考虑此方案的风险——SDK↔CLI 的 stream-json 控制协议是内部实现——已被 spike 消除：假 CLI harness 今天能对真实固定版本的 SDK 正常工作。如果 SDK 升级破坏了 mock，无密钥套件会让升级 PR（Pull Request）失败，这正是门禁在发挥作用。
|
||||
|
||||
### 为什么不用 ACP 适配器（如 `claude-code-acp`）复用既有后端？
|
||||
|
||||
社区 shim 将两个引擎包装为 ACP，这会使它们在 `dsh-subagent-acp` 上变成「仅配置」。但这在 harness 与引擎之间插入了一个非官方的第三方层，抹去了本 Agent Note 暴露的原生控制面（permissionMode、sandboxMode/approvalPolicy、配置目录隔离、apiKey RPC），并以 shim 的发布节奏替换了第一方协议的稳定性。第一方接口——Agent SDK 和 app-server——才是受支持的集成点。
|
||||
|
||||
## 验收标准
|
||||
|
||||
在两个引擎和密钥均已配置的机器上：一个 REPL 驱动的模型通过 `subagent_claude_code` 完成一个真实文件任务，通过 `subagent_codex` 完成另一个，工具结果为子进程的最终答案，父会话日志中仅有 `tool/call` + `tool/result`。无密钥套件在无凭证环境下以逐文件 100% 覆盖率通过，断言隔离（清洗后的子进程环境、dispose 后无残留临时配置目录），并断言 `~/.claude` / `~/.codex` 的存在与否不影响子进程行为。取消父轮次后，两个后端在有界时间内完全停稳，无残留子进程。e2e 套件干净地自跳过，命名缺失的前置条件。
|
||||
|
||||
## 风险
|
||||
|
||||
- `codex app-server` 被 CLI 标记为实验性，其 v1/v2 词汇共存；客户端固定 0.142.5、仅实现 v2、对未知方法/通知消费而不崩溃，但未来 codex 升级仍可能迫使返工（每次升级重新生成 schema 并重跑无密钥套件——这是上述「不做运行时版本探测」立场背后的开发时强制执行）。
|
||||
- Claude Code 假 CLI mock 依赖一个内部协议：任何 SDK 升级都必须通过无密钥套件，控制协议的破坏性变更意味着返工 mock（回退方案：上面否决的驱动注入 seam 成为逃生舱口）。
|
||||
- SDK 的 optionalDependencies 每平台约 280 MB——已接受，限制在单个后端包内。
|
||||
- SDK 的 SIGKILL 分支（EOF→SIGTERM 之后）未被观察到，信任其实现；e2e 保留无残留进程断言。
|
||||
- Codex 是部署前置条件（无 npm 内置二进制）；缺失或不兼容的二进制会明确报出 spawn/协议 `error`，而非版本探测。
|
||||
- 每次运行付出一个全新子进程的代价，且仅最终答案浮出——思考、工具卡片和用量被消费后丢弃；池化、中间进度浮出、`sendMessage`/`resume`、通过 SDK 的 `outputFormat` 实现 `outputSchema`、以及通过 SDK 的 `agents` 选项实现命名 subagent 类型，均为刻意推迟。
|
||||
@@ -2,5 +2,5 @@
|
||||
# side as of the last confirmed-consistent state. Both languages carry equal authority;
|
||||
# after editing either side, bring the other along and re-record with:
|
||||
# pnpm run verify-translation-pairing --write .agents/notes/proposed/feature/2026-08-01-windows-pwsh-default.md
|
||||
2026-08-01-windows-pwsh-default.md: a310174b6864bb880070280835ccfd8623e26342
|
||||
2026-08-01-windows-pwsh-default.zh.md: 079c1e3cac789a5e3fa4d0bb889026b3fb69f78c
|
||||
2026-08-01-windows-pwsh-default.md: 1c3ccef23bb5cd9bc37237bd69aac2e2c56649a8
|
||||
2026-08-01-windows-pwsh-default.zh.md: 3958d21eb8a9d306009b11d6e9806a1654a8958e
|
||||
|
||||
@@ -13,9 +13,9 @@ The harness's shipped execution profile is bash-first on every platform. Windows
|
||||
Two follow-up stages, each independently shippable. The former stage 2 (bash-tool parity twin) shipped with the [pwsh tool bash parity decision](../../implemented/feature/2026-08-02-pwsh-tool-bash-parity.md): `tool-pwsh` now mirrors `tool-bash` for foreground and background work minus the sandbox surface, shares the `DSH_*` environment through `dsh-bash-env`, and carries a keyless application snapshot of its assembled surface.
|
||||
|
||||
1. **Windows default composition** — the shipped CLI compositions mount `dsh-pwsh-local` as the `ctx.bash` executor and `dsh-tool-pwsh` as the model-facing shell tool on Windows hosts (bash unmounted there), while POSIX hosts keep the bash stack. This is a composition/roster decision in `base.cordis.yml` and the surface overlays, gated by platform; it makes the shipped Windows experience PowerShell-native end to end.
|
||||
2. **pwsh TUI/GUI rendering** — the TUI and Web surfaces render pwsh output with PowerShell-aware presentation (native path display, `$env:` facts), the counterpart of the bash terminal cards. This is where terminal/console rendering conventions get a PowerShell twin.
|
||||
2. **pwsh GUI rendering** — the Web surface renders pwsh calls with the bash-shaped terminal presentation (terminal card with exit-status pill), the counterpart of the bash terminal cards. Shipped in the [pwsh UI presentation matches bash decision](../../implemented/feature/2026-08-05-pwsh-ui-bash-parity.md) with a keyless web lane; the TUI was removed, so no terminal twin remains. A PowerShell-aware presentation beyond bash parity (native path display, `$env:` facts) remains unclaimed.
|
||||
|
||||
The stages are deliberately sequenced: composition first (a Windows user gets PowerShell without choosing), then rendering. Nothing in this proposal changes POSIX behavior.
|
||||
The stages are ordered by dependency only where one exists: the rendering stage shipped first with the [pwsh UI presentation matches bash decision](../../implemented/feature/2026-08-05-pwsh-ui-bash-parity.md) because it is platform-independent and its keyless web lane runs on any host, while the Windows default composition remains the only unshipped stage. Nothing in this proposal changes POSIX behavior.
|
||||
|
||||
## Alternatives considered
|
||||
|
||||
@@ -30,10 +30,10 @@ The stages are deliberately sequenced: composition first (a Windows user gets Po
|
||||
- A Windows host running the shipped `dsh` TUI/Web gets `pwsh` as its shell tool and PowerShell as the `ctx.bash` executor without configuration, and `bash` is absent from the model-visible roster there.
|
||||
- POSIX hosts are byte-for-byte unaffected (same roster, same executor).
|
||||
- The shipped-composition e2es assert the platform-gated roster on both families.
|
||||
- Stage 1 lands with the keyless pwsh-tool snapshot already in place from the parity change; stage 2 lands with TUI/Web rendering snapshots for pwsh output.
|
||||
- Stage 1 lands with the keyless pwsh-tool snapshot already in place from the parity change; stage 2 landed with the web `pwsh-terminal` rendering lane (the TUI's removal left no terminal surface to snapshot).
|
||||
|
||||
## Risks
|
||||
|
||||
- **Bash-dependent composition rows** — any shipped plugin that assumes `bash` semantics (hook bridges executing shell hooks, workspace tooling) must be audited per stage; the audit may force a staged rollout rather than one switch.
|
||||
- **Windows CI coverage gap** — unit coverage runs on Linux; Windows-only regressions in the pwsh stack surface through the Windows build/static lane and e2es, which must be extended per stage rather than assumed.
|
||||
- **Rendering conventions** — a PowerShell twin for terminal cards is a UI design decision with snapshot surface; deferring it (stage 2) keeps stage 1 shippable without UI churn.
|
||||
- **Rendering conventions** — the bash-shaped terminal twin shipped with the Web lane; a PowerShell-aware presentation beyond bash parity (native path display, `$env:` facts) remains a UI design decision with snapshot surface, deferred with stage 1.
|
||||
|
||||
@@ -13,9 +13,9 @@ harness 交付的执行画像在每个平台都是 bash 优先。Windows 主机
|
||||
两个阶段，各自可独立交付。原阶段 2（bash 工具对等孪生）已随 [pwsh 工具与 bash 对齐决策](../../implemented/feature/2026-08-02-pwsh-tool-bash-parity.md) 交付：`tool-pwsh` 现在在前台与后台工作（减 sandbox 面）上镜像 `tool-bash`，通过 `dsh-bash-env` 共享 `DSH_*` 环境，并携带其组装表面的 keyless 应用快照。
|
||||
|
||||
1. **Windows 默认组合**——交付的 CLI 组合在 Windows 主机上挂载 `dsh-pwsh-local` 作为 `ctx.bash` 执行器、`dsh-tool-pwsh` 作为面向模型的 shell 工具（那里不挂载 bash），POSIX 主机保持 bash 栈。这是 `base.cordis.yml` 与 surface 覆盖层里按平台门控的组合/清单决策；它让交付的 Windows 体验端到端 PowerShell 原生。
|
||||
2. **pwsh TUI/GUI 渲染**——TUI 与 Web 表面以 PowerShell 感知的呈现渲染 pwsh 输出（原生路径显示、`$env:` 实情），即 bash 终端卡片的对应物。这是终端/控制台渲染约定获得 PowerShell 孪生的地方。
|
||||
2. **pwsh GUI 渲染**——Web 表面以 bash 形状的终端呈现渲染 pwsh 调用（带退出状态 pill 的 terminal 卡），即 bash 终端卡片的对应物。已随 [pwsh UI 呈现与 bash 对齐决策](../../implemented/feature/2026-08-05-pwsh-ui-bash-parity.md) 及 keyless web 通道交付；TUI 已移除，不再有终端孪生。超出 bash 对齐的 PowerShell 感知呈现（原生路径显示、`$env:` 实情）仍无人认领。
|
||||
|
||||
各阶段刻意排序：先组合（Windows 用户无需选择即获得 PowerShell），再渲染。本提案不改变任何 POSIX 行为。
|
||||
各阶段仅在有依赖关系时排序：渲染阶段已随 [pwsh UI 呈现与 bash 对齐决策](../../implemented/feature/2026-08-05-pwsh-ui-bash-parity.md) 先行交付（平台无关，其 keyless web 通道可在任意宿主运行），而 Windows 默认组合仍是唯一未交付的阶段。本提案不改变任何 POSIX 行为。
|
||||
|
||||
## 备选方案
|
||||
|
||||
@@ -30,10 +30,10 @@ harness 交付的执行画像在每个平台都是 bash 优先。Windows 主机
|
||||
- 运行交付版 `dsh` TUI/Web 的 Windows 主机无需配置即获得 `pwsh` 作为其 shell 工具、PowerShell 作为 `ctx.bash` 执行器，且那里的模型可见清单中没有 `bash`。
|
||||
- POSIX 主机逐字节不受影响（清单相同，执行器相同）。
|
||||
- 交付组合 e2e 在两个平台族上断言按平台门控的清单。
|
||||
- 阶段 1 落地时，parity 变更带来的 keyless pwsh 工具快照已经就位；阶段 2 附带 pwsh 输出的 TUI/Web 渲染快照落地。
|
||||
- 阶段 1 落地时，parity 变更带来的 keyless pwsh 工具快照已经就位；阶段 2 已随 web `pwsh-terminal` 渲染通道落地（TUI 的移除让终端表面无快照可做）。
|
||||
|
||||
## 风险
|
||||
|
||||
- **依赖 bash 的组合行**——任何假设 bash 语义的交付插件（执行 shell hooks 的 hooks 桥、工作区工具）必须按阶段审计；审计可能迫使分阶段推出而非一次切换。
|
||||
- **Windows CI 覆盖缺口**——单元覆盖在 Linux 上运行；pwsh 栈里仅 Windows 的回归通过 Windows 构建/静态通道与 e2e 浮出，必须按阶段扩展而不是想当然。
|
||||
- **渲染约定**——终端卡片的 PowerShell 孪生是带快照表面的 UI 设计决策；把它延期（阶段 2）让阶段 1 无需 UI 翻动即可交付。
|
||||
- **渲染约定**——bash 形状的终端孪生已随 web 通道交付；超出 bash 对齐的 PowerShell 感知呈现（原生路径显示、`$env:` 实情）仍是带快照表面的 UI 设计决策，随阶段 1 一起延期。
|
||||
|
||||
Reference in New Issue
Block a user