Retire the readline front door and the repl-agent example

Delete packages/ui/stdio and examples/repl-agent; rename stdio-demo to
@deepseek-ai/dsh-tui-demo (TUI-only, refuses pipes before Loader boot).
tui-agent owns the coding composition inline; echo-agent and the CI demo
smoke move to the one-shot cli-demo bin, which gains -p/--prompt. The
UI-independent with-key e2es move verbatim to tui-agent. SDK wizard's
'stdio' interface becomes 'tui'. PTY testing stays confined to TUI
surfaces; all other subprocess tests ride pipes.

See .agents/notes/implemented/simplification/2026-07-20-retire-readline-front-door.md

packages/ui/app-boot/README.md

@@ -5,10 +5,14 @@ Shared boot glue for the app bins ([`dsh-tui-demo`](../../examples/tui-demo/READ
 | Export | Role |
 |---|---|
 | `resolveConfigPath(path, snapshotMode, cwd?)` | Absolute config path; `snapshotMode === 'replay'` swaps a `cordis.yml`/`.yaml` basename for its sibling `cordis.snapshot.yml` |
+| `parseResumeArg(argv)` | Split the `--resume <id>` / `--resume=<id>` flag out of the arguments, returning `{ resumeSessionId, rest }`; a valueless, empty, or repeated flag throws so a mistyped resume fails loud instead of silently starting fresh |
 | `loadEnv(binName, dir?, warn?)` | Load the gitignored `.env` (Node `process.loadEnvFile`); absent file is fine, an unloadable one warns a single labelled line (default: stderr) |
 | `installFailLoud(binName, proc?)` | Turn a post-`boot()` unhandled Loader rejection into one labelled stderr line + `exit(1)`; returns the uninstaller (for tests) |
 | `assertEntriesLoaded(ctx, binName)` | Throw when a settled tree holds an enabled entry with no fiber (a plugin module that failed to import) |
-| `boot(binName, absoluteConfigPath)` | Mount the Loader, mount the statically imported include plugin as the `cordis:include` builtin (so the config may live outside `node_modules` reach), include the config by absolute `file://` URL, await the whole tree, assert entries loaded, return the root context |
+| `loadPersonalPatches(binName, dir?)` | Parse the optional `config.yaml` in the Harness home (default [`resolveDshHome()`](../../util/paths/README.md): `$DSH_HOME`, else `~/.dsh`) — a top-level YAML array of include `PatchOptions` (id-targeted config overrides, `insert` lists, `!!js` allowed); absent file → `undefined`, an unreadable/unparsable/non-array file throws |
+| `boot(binName, absoluteConfigPath, patches?)` | Mount the Loader, mount the statically imported include plugin as the `cordis:include` builtin (so the config may live outside `node_modules` reach), include the config by absolute `file://` URL with the optional overlay patches, await the whole tree, assert entries loaded, return the root context |
+| `addHarnessSourceSection(ctx, sourceRoot)` | Add a global `harness:source` prompt section (ordered just after the harness identity, before the persona) telling the agent the on-disk path to its own source checkout; a no-op returning `undefined` when the booted tree has no `systemPrompt` service. The section is registered against that service's fiber, so a dev HMR reload of the system prompt drops it until the next boot |
+| `HARNESS_SOURCE_SECTION` | The `'harness:source'` section name `addHarnessSourceSection` registers under |
 
 Two failure classes the guards handle: `loader.await()` swallows init rejections (`Promise.allSettled`) — Node still exits non-zero on the resulting unhandled rejection, and `installFailLoud` replaces the noisy dump with one labelled line and a guaranteed `exit(1)`; a failed plugin IMPORT is only logged by the Loader (the process would otherwise exit 0 on a usable config typo), leaving a fiber-less entry that `assertEntriesLoaded` turns into a `boot()` rejection.
 
@@ -16,16 +20,27 @@ Bare plugin specifiers in a config (`@deepseek-ai/dsh-*`, npm packages) resolve
 
 This package carries no loader hooks and no dev-mode surface: the `dsh-scripts` launcher ([`sdk/scripts`](../../sdk/scripts/README.md), with the shared project model in [`sdk/helper`](../../sdk/helper/README.md)) owns process startup, tsx registration, and local-plugin source resolution, and consumes these helpers for the boot sequence itself.
 
+## Personal config
+
+A developer's machine-local preferences live outside every repository in the Harness home (default `~/.dsh`, overridable via `$DSH_HOME`; the single root [`resolveDshHome`](../../util/paths/README.md) resolves), consumed by the `dsh` CLI's TUI surface ([`apps/cli`](../../../apps/cli/README.md)); the demo bins boot their committed trees verbatim. Two optional files:
+
+- **`.env`** — loaded after the invoking directory's `.env`; `process.loadEnvFile` never overrides, so precedence is ambient environment > project `.env` > personal `.env`.
+- **`config.yaml`** — loader overlay patches applied over the shipped default config, with the same semantics as an include entry's `patches` (the committed Code Mode overlay is the template): an id-targeted patch replaces the named entry's whole `config` (restate unchanged fields), `insert` adds entries, and `!!js` expressions interpolate at mount — so a personal `apiKey` can reference the personal `.env`. A patch naming an entry id absent from the booted tree is skipped with a loader warning. An empty or comments-only file throws (it parses to nothing, not to a list); disable the overlay with `[]` or by deleting the file.
+
+Subprocess test launchers point `DSH_HOME` at an isolated per-test directory so a developer's personal overlay can never leak into fixtures.
+
 ## Model Experience
 
-Indirectly, through the plugin tree it loads, which determines the prompts, schemas, messages, and model adapter in the resulting application.
+Indirectly, through the plugin tree it loads, which determines the prompts, schemas, messages, and model adapter in the resulting application; the one export that contributes model-visible text, `addHarnessSourceSection`, does so only when a consumer calls it after boot.
 
 #### KV Cache effect
 
-No direct invalidation; the named consumer owns any request-prefix changes.
+No direct invalidation from `boot()`; a consumer that calls `addHarnessSourceSection` places one short line near the system prompt's head, before per-request content, so it does not invalidate the cache across turns, and any other request-prefix change is owned by the named consumer.
 
 ## Known Limitations and Deferred Work
 
 - **Bare package specifiers depend on Loader internals** — production bins need `node --expose-internals` or the Loader's optional native fallback; an in-process caller without either must use resolvable relative/file specifiers or tsx path mapping.
 - **Snapshot replay swapping is basename-specific** — only a config ending in `cordis.yml` or `cordis.yaml` maps to the sibling `cordis.snapshot.yml`; custom config names require caller-managed selection.
 - **Environment loading is cwd-scoped and optional** — the helper loads one `.env` file and warns on failure; it does not search parents, merge profiles, or validate required variables.
+- **Personal config is patch-shaped** — an id-targeted patch replaces the entry's whole `config` rather than deep-merging, so a personal override restates the base fields it keeps.
+- **Personal patches see only the booted file's own entries** — an overlay leaf that reaches its base through a nested include entry (the Code Mode configs) resolves personal patch ids against the overlay's top-level entries, not the included subtree.

packages/ui/app-boot/src/index.ts

@@ -1,15 +1,21 @@
 /**
  * Shared boot glue for the app bins (`dsh-tui-demo`, `dsh-cli-demo`, `dsh-acp-demo`): load the gitignored
- * `.env`, install the fail-loud Loader guards, resolve the config path (snapshot-aware), and
- * drive the cordis Loader against a leaf `cordis.yml` until the whole tree has settled.
+ * `.env`, install the fail-loud Loader guards, resolve the config path (snapshot-aware), load the
+ * optional personal overlay patches from the Harness home (`~/.dsh`), and drive the cordis Loader
+ * against a leaf `cordis.yml` until the whole tree has settled.
  * @module @deepseek-ai/dsh-app-boot
  */
 
 import { pathToFileURL } from 'node:url'
-import { basename, dirname, resolve } from 'node:path'
+import { readFileSync } from 'node:fs'
+import { basename, dirname, join, resolve } from 'node:path'
+import * as yaml from 'js-yaml'
 import { Context } from 'cordis'
 import Loader from '@cordisjs/plugin-loader'
-import Include from '@cordisjs/plugin-include'
+import Include, { type PatchOptions } from '@cordisjs/plugin-include'
+import { resolveDshHome } from '@deepseek-ai/dsh-paths'
+// Side-effect type import: resolves `ctx.get('systemPrompt')` to the service.
+import type {} from '@deepseek-ai/dsh-system-prompt'
 
 /**
  * Resolve the config to boot. Replay swaps a `cordis.yml` basename for
@@ -30,6 +36,48 @@ export function resolveConfigPath(
   return resolve(dir, replayName)
 }
 
+/** CLI flag the interactive surface accepts to resume a persisted session by id. */
+const RESUME_FLAG = '--resume'
+
+/**
+ * Split a leading `--resume <id>` / `--resume=<id>` flag out of a CLI argument
+ * vector, returning the resumed session id (when the flag is present) and the
+ * remaining arguments with the flag and its value removed — so a positional
+ * config path stays readable regardless of the flag's position. A `--resume`
+ * with no following id, an empty id (`--resume=`), or a repeated `--resume`
+ * throws: a mistyped resume must fail loud, never silently start a fresh
+ * session. The id is not validated here; an unknown id fails loud downstream
+ * when the session cannot load.
+ * @param argv - the CLI arguments after subcommand dispatch.
+ * @returns the parsed resume id (or `undefined`) and the flag-stripped arguments.
+ */
+export function parseResumeArg(
+  argv: readonly string[],
+): { resumeSessionId: string | undefined; rest: string[] } {
+  const rest: string[] = []
+  let resumeSessionId: string | undefined
+  let skipNext = false
+  for (const [i, arg] of argv.entries()) {
+    if (skipNext) {
+      skipNext = false
+      continue
+    }
+    const inlineValue = arg.startsWith(`${RESUME_FLAG}=`)
+    if (arg === RESUME_FLAG || inlineValue) {
+      if (resumeSessionId !== undefined) throw new Error(`${RESUME_FLAG} may be given only once`)
+      const value = inlineValue ? arg.slice(RESUME_FLAG.length + 1) : argv[i + 1]
+      if (value === undefined || value === '') {
+        throw new Error(`${RESUME_FLAG} requires a session id (e.g. ${RESUME_FLAG} <session-id>)`)
+      }
+      resumeSessionId = value
+      skipNext = !inlineValue // the space form consumed the following token as its value
+      continue
+    }
+    rest.push(arg)
+  }
+  return { resumeSessionId, rest }
+}
+
 /**
  * Load the optional gitignored `.env` from `dir`. Missing files fall back to the
  * ambient environment; other read failures are reported through `warn`.
@@ -51,6 +99,62 @@ export function loadEnv(
   }
 }
 
+/** File inside the Harness home holding the personal loader overlay patches. */
+export const PERSONAL_CONFIG_FILENAME = 'config.yaml'
+
+// The include's YAML dialect: `!!js` scalars become expression nodes the
+// Loader interpolates against each entry's context at mount time. Personal
+// patches are parsed with the same schema so they may reference `process.env`.
+// Load-only: this schema never dumps, so no `predicate`/`represent`.
+const jsExprType = new yaml.Type('tag:yaml.org,2002:js', {
+  kind: 'scalar',
+  resolve: data => typeof data === 'string',
+  construct: data => ({ __jsExpr: String(data) }),
+})
+const personalPatchesSchema = yaml.JSON_SCHEMA.extend(jsExprType)
+
+/**
+ * Load the optional personal overlay patches (`config.yaml` under the Harness
+ * home). The file is a top-level YAML array of loader patch entries
+ * (`@cordisjs/plugin-include`'s `PatchOptions`): id-targeted config overrides
+ * and `insert` lists, with `!!js` expressions allowed. A missing file means
+ * "no personal overlay"; an unreadable, unparsable, or non-array file throws —
+ * a present personal config that cannot apply is a misconfiguration and must
+ * fail loud at boot, never be silently skipped.
+ * @param binName - the diagnostic prefix on the thrown error.
+ * @param dir - the Harness home; defaults to {@link resolveDshHome} (`$DSH_HOME` or `~/.dsh`).
+ * @returns the parsed patches, or `undefined` when the file does not exist.
+ */
+export function loadPersonalPatches(
+  binName: string, dir: string = resolveDshHome(),
+): PatchOptions[] | undefined {
+  const file = join(dir, PERSONAL_CONFIG_FILENAME)
+  let content: string
+  try {
+    content = readFileSync(file, 'utf8')
+  } catch (error) {
+    if ((error as NodeJS.ErrnoException | null)?.code === 'ENOENT') return undefined
+    throw new Error(`${binName}: failed to read personal patches ${file}: ${String(error)}`)
+  }
+  let parsed: unknown
+  try {
+    parsed = yaml.load(content, { schema: personalPatchesSchema })
+  } catch (error) {
+    throw new Error(`${binName}: failed to parse personal patches ${file}: ${String(error)}`)
+  }
+  if (!Array.isArray(parsed)) {
+    throw new Error(`${binName}: personal patches ${file} must be a top-level YAML array of loader patch entries`)
+  }
+  // A present personal config that cannot apply is a misconfiguration and must
+  // fail loud here — the include only warns per entry at mount.
+  parsed.forEach((entry, index) => {
+    if (typeof entry !== 'object' || entry === null || Array.isArray(entry)) {
+      throw new Error(`${binName}: personal patches entry ${index + 1} in ${file} must be a mapping (a loader patch entry)`)
+    }
+  })
+  return parsed as PatchOptions[]
+}
+
 /**
  * The slice of `process` {@link installFailLoud} needs — injectable so tests
  * exercise the handler without registering on (or exiting) the real process.
@@ -109,18 +213,52 @@ export function assertEntriesLoaded(ctx: Context, binName: string): void {
  * @param binName - the diagnostic prefix for load-failure errors.
  * @param absoluteConfigPath - the config to include; must already be absolute
  * (see {@link resolveConfigPath}).
+ * @param patches - optional overlay patches applied over the included tree
+ * (see {@link loadPersonalPatches}); an empty list mounts none.
  * @returns the root context once every entry has started.
  */
-export async function boot(binName: string, absoluteConfigPath: string): Promise<Context> {
+export async function boot(
+  binName: string, absoluteConfigPath: string, patches?: PatchOptions[],
+): Promise<Context> {
   const ctx = new Context()
   ctx.baseUrl = pathToFileURL(dirname(absoluteConfigPath)).href + '/'
   await ctx.plugin(Loader)
   ctx.loader.builtins.include = Include
   await ctx.loader.create({
     name: 'cordis:include',
-    config: { path: pathToFileURL(absoluteConfigPath).href },
+    config: {
+      path: pathToFileURL(absoluteConfigPath).href,
+      ...patches !== undefined && patches.length > 0 ? { patches } : {},
+    },
   })
   await ctx.loader.await()
   assertEntriesLoaded(ctx, binName)
   return ctx
 }
+
+/** Prompt-section name for the harness-source location line an app bin adds after boot. */
+export const HARNESS_SOURCE_SECTION = 'harness:source'
+
+/**
+ * Add a global prompt section naming the on-disk path to the harness source
+ * checkout the running bin was launched from, so the agent knows where its own
+ * source lives (the self-referential `dsh-tool-cordis` toolset reads and edits
+ * it). Call once on the settled boot context ({@link boot}); the section orders
+ * just after the harness identity opener (`-100`) and before the deployment
+ * persona (`0`). A booted tree with no `systemPrompt` service has no prompt to
+ * augment, so this is then a no-op that returns `undefined`. The section is
+ * registered against the `systemPrompt` service's fiber, so a dev HMR reload of
+ * that plugin drops it until the next boot.
+ * @param ctx - the settled boot context whose global system prompt to augment.
+ * @param sourceRoot - the absolute path to the harness checkout root.
+ * @returns the section disposer, or `undefined` when no `systemPrompt` service is mounted.
+ */
+export function addHarnessSourceSection(ctx: Context, sourceRoot: string): (() => void) | undefined {
+  const systemPrompt = ctx.get('systemPrompt')
+  if (systemPrompt === undefined) return undefined
+  return systemPrompt.section({
+    name: HARNESS_SOURCE_SECTION,
+    order: -99,
+    text: `Your own source code is the checkout at ${sourceRoot}; you can read it there to learn how dsh works and how to extend it.`,
+  })
+}

packages/ui/tui/README.md

@@ -1,12 +1,12 @@
 # @deepseek-ai/dsh-tui
 
-The interactive terminal front door for DeepSeek Harness agents, built on [`@earendil-works/pi-tui`](https://www.npmjs.com/package/@earendil-works/pi-tui). It requires stdin and stdout TTYs; scripts and Loader pipes should use the headless [`@deepseek-ai/dsh-cli-demo`](../../examples/cli-demo/README.md) app instead.
+The interactive terminal front door for DeepSeek Harness agents, built on [`@earendil-works/pi-tui`](https://www.npmjs.com/package/@earendil-works/pi-tui). It requires stdin and stdout TTYs; scripts and Loader pipes should use the one-shot [`@deepseek-ai/dsh-cli-demo`](../../examples/cli-demo/README.md) app instead.
 
 The implemented [TUI feature Agent Note](../../../.agents/notes/implemented/feature/2026-07-17-dedicated-full-screen-tui-front-door.md) owns the front-door decision; the [terminal-state snapshot Agent Note](../../../.agents/notes/implemented/testing/2026-07-18-tui-terminal-state-snapshots.md) owns its verification strategy.
 
 Interactive terminals on macOS, Linux, and Windows are supported. Windows uses pi-tui's native console VT-input handling, and the [Windows support Agent Note](../../../.agents/notes/implemented/feature/2026-07-20-windows-tui-support.md) owns the platform decision and ConPTY process verification.
 
-This package owns interactive terminal presentation and input only. It injects `agents`, [`commands`](../commands/README.md), `llm`, `systemPrompt`, `tokenMeter`, `tools`, and `userInteraction`, then drives an agent created or resumed by app or developer code. Agent lifecycle, persistence, and the model-facing [`ask_user_question`](../tool-ask-user/README.md) tool remain separate composition entries.
+This package owns interactive terminal presentation and input only. It injects `agents`, [`commands`](../commands/README.md), `llm`, `systemPrompt`, `tokenMeter`, `tools`, and `userInteraction`, optionally reads a `skills` service (present only when one is mounted), then drives an agent created or resumed by app or developer code. Agent lifecycle, persistence, and the model-facing [`ask_user_question`](../tool-ask-user/README.md) tool remain separate composition entries.
 
 The TUI rebuilds resumed history from the active session surface, renders Markdown responses and reasoning, applies each tool's `presentCall` / `presentResult` intent to terminal, diff, or generic cards, keeps the latest `todo/write` plan above the editor, and presents `ctx.userInteraction` questions in a wide bottom-left keyboard panel with progress, numbered options, and aligned descriptions. The latest logged session title becomes the header subtitle, with `welcome` before a title exists, and the terminal window title becomes `<session title> — <configured title>`. A durable `llm/retry` event retracts the failed step's live chunks and renders the scheduled retry count, delay, and failure in the transcript; success, exhaustion, and cancellation then settle through ordinary session events. The footer totals each logged model step's usage once, including failed attempts, while treating committed-message usage as a fallback for logs without a usage chunk. Its idle view compares token-meter pressure with `ctx.llm.resolveModelContext()` for the current route, displays `context unknown` when the adapter has no capacity metadata, and also shows tool-card mode and the current model with reasoning state; while the agent runs, an elapsed working indicator and `esc interrupt` replace that summary. Surface replacement events rebuild the transcript so compacted history does not reappear.
 
@@ -14,15 +14,21 @@ An embedding may provide `TuiRuntime.formatCwd` when its logical workspace label
 
 Before model output, session events, tool presenters, questions, configuration, or diagnostics reach pi-tui's ANSI-aware renderers or the terminal title, the TUI renders C0 and C1 controls other than line feeds as visible `\xNN` text. Those sources cannot add terminal control sequences; the TUI and pi-tui retain ownership of terminal rendering and styling.
 
-While the agent is running, ordinary editor submissions call `agent.steer()`; otherwise they call `agent.send()`. A slash at the start of the submitted line enters `ctx.commands` instead: known commands execute directly, unknown commands produce a warning, and neither path reaches the model. The TUI registers `/help`, `/model`, `/clear`, `/cancel`, `/reasoning`, `/tools`, `/redraw`, and `/exit` as agent-scoped definitions; every other effective command joins autocomplete and `/help` dynamically. Ctrl+C or Escape cancels a running turn. Tool cards collapse long bodies into a configurable head/tail preview; Ctrl+O toggles every card between its preview and full output. Ctrl+R toggles reasoning, Ctrl+L redraws, and Ctrl+D exits while idle.
+While the agent is running, ordinary editor submissions call `agent.steer()`; otherwise they call `agent.send()`. A slash at the start of the submitted line enters `ctx.commands` instead: known commands execute directly, unknown commands produce a warning, and neither path reaches the model. The TUI registers `/help`, `/model`, `/clear`, `/reasoning`, `/tools`, `/redraw`, `/reload`, `/resume`, and `/exit` as agent-scoped definitions; every other effective command joins autocomplete and `/help` dynamically, as do `/skill:` completions. A status line above the editor reports the turn phase the TUI derives from session events — waiting for the first token, thinking, responding, or executing tools — with the elapsed time in that phase and the running step total, refreshed each second, and ends with the `Enter sends steering, Esc cancels` hint; while steering messages wait to reach the model it inserts a `N queued ·` badge before the hint that clears as each drains. Ctrl+C or Escape cancels a running turn. Tool cards collapse long bodies into a configurable head/tail preview; Ctrl+O toggles every card between its preview and full output. Ctrl+R toggles reasoning, Ctrl+L redraws, and Ctrl+D exits while idle.
 
 `/model` opens the advisory `ctx.llm` catalog as a keyboard selector: Up/Down moves, Enter selects, and Escape closes it. `/model <model>` still selects an unambiguous model id directly, while `/model <provider>/<model>` selects an exact target. The configured target or latest logged request header initializes the selector, and an unlisted current model remains visible because catalogs are advisory. Selection is local to this TUI session. Prompt assembly snapshots the target for one step, replaces `{{provider}}` and `{{model}}`, and applies the same pair through `agent/request`; a switch during assembly therefore starts with a later step. The request header durably records targets that reach the model, while an unused selection remains process-local.
 
+`/reload` (EXPERIMENTAL, dev-only) re-reads every file-backed loader config tree and applies the diff to the running app — the HMR watcher's config path, invoked manually; it needs the cordis Loader in the context and degrades to a warning without one, runs only while the agent is idle, and refuses re-entry while a reload is in flight. Module-source hot reload remains watcher-owned. When a `skills` service is mounted, `/skill:<name> [instructions]` loads that skill's instructions into the conversation as a user turn; autocomplete lists the model-invocable skills, and any skill (including a model-disabled one) is loadable by its exact name.
+
+The footer sums the session's reported usage as `↑<uncached input> ↓<output>`, followed by `cache <rate>%` once any input has been billed — the share of billed prompt tokens (uncached input plus cache reads and writes) served from the provider cache, rounded to a percent. It also compares token-meter pressure with `ctx.llm.resolveModelContext()` for the current route (omitting the context share when the adapter has no capacity metadata) and shows the current model and tool-card mode; the right side clips first when the footer is narrow.
+
+When `resumeCommand` is set and a `sessionPersistence` backend is mounted, exiting prints the resume command for the current session (once it has been persisted, so an abandoned session yields no hint), and `/resume` lists this workspace's persisted sessions newest-first, each with its resume command and a marker on the current one. `{session}` in the template expands to the session id; the TUI only prints commands to copy and never resumes in place.
+
 ## Config
 
 | Key | Default | Meaning |
 |---|---|---|
-| `welcome` | `ready.` | Header subtitle until the session has a logged title. |
+| `welcome` | — | Banner subtitle line until the session has a logged title; unset, the banner sweeps in with no subtitle |
 | `sessionId` | `main` | Exact shared agent/session identity driven by the terminal |
 | `showReasoning` | `true` | Render reasoning blocks |
 | `maxToolOutputLines` | `6` | Output lines retained across a collapsed tool card's head/tail preview |
@@ -35,6 +41,8 @@ While the agent is running, ordinary editor submissions call `agent.steer()`; ot
 | `showHardwareCursor` | `false` | Show the hardware cursor at pi-tui's IME marker |
 | `color` | `true` | Apply the built-in ANSI palette (see [Color](#color)) |
 | `title` | `DeepSeek Harness` | Product suffix for the terminal window title. |
+| `resumeCommand` | — | Shell command template for the exit hint and `/resume`, with `{session}` expanded to the session id; unset disables both. Needs a `sessionPersistence` backend |
+| `autoTitle` | `true` | Replace `title` with a short model-made title derived from the session's first user message; a resumed session re-derives it from that stored message on mount (needs an `llm` service and an agent provider/model) |
 
 ```yaml
 - id: terminal
@@ -82,6 +90,20 @@ The selector adds no messages. A target change may alter interpolated system-pro
 
 Changing provider or model enters that target's cache domain; no cache reuse across distinct targets is assumed.
 
+### Manual skill invocation
+
+#### What the model sees
+
+A `/skill:<name> [instructions]` submission loads the named skill and delivers one text block: a `<skill name="…">` element wrapping the skill's instructions — preceded, when the provider exposes a resource base, by a line locating the skill's relative resources — followed by any trailing instructions the user typed. Delivery follows the same send-while-idle / steer-while-running rule as ordinary input. The command, not the model, chooses the skill; model-disabled skills are omitted from autocomplete but stay loadable by exact name.
+
+#### Token effect
+
+The rendered skill block and trailing instructions are retained as one user turn under the agent loop's normal session-history and compaction rules; a repeated invocation appends the body again.
+
+#### KV Cache effect
+
+Append-only; newly visible content follows the reusable request prefix and does not invalidate existing KV-cache entries.
+
 ### Interactive user-question answers
 
 #### What the model sees
@@ -100,4 +122,5 @@ Append-only; newly visible content follows the reusable request prefix and does
 
 - **One configured session owns the transcript and editor** — questions from other agents can still use the shared overlay provider, but session rendering and prompt input remain bound to `sessionId`.
 - **Tool cards are text terminal presentations** — terminal, diff, and generic cards use tool-owned titles/content, but session content currently has no image block for inline image rendering.
-- **Non-TTY operation is intentionally unsupported** — automation must use the headless app rather than expecting an internal fallback.
+- **Non-TTY operation is intentionally unsupported** — app bundles that need automation must compose a one-shot or server front door (`dsh-cli-demo`, `dsh-acp`) rather than expecting an internal fallback.
+- **Manual `/skill:` invocation always reloads the full skill body** — the TUI does not detect a skill already present in the conversation, so repeated invocations append its instructions again.

packages/ui/tui/tests/tui.spec.ts

@@ -1335,7 +1335,7 @@ describe('terminal mounting', () => {
     expect(terminal.output).toBe('')
     expect(exit).not.toHaveBeenCalled()
     ctx.emit('agent-loop/config-start-failed', SessionId('main-session'), new Error('resume \u001b]2;failure-controlled\u0007'))
-    expect(terminal.output).toBe('ui-tui: session "main-session" failed to start: resume \\x1b]2;failure-controlled\\x07\n')
+    expect(terminal.output).toBe('ui-tui: session "main-session" failed to start: Error: resume \\x1b]2;failure-controlled\\x07\n')
     expect(exit).toHaveBeenCalledWith(1)
 
     const session = ctx.sessions.create(SessionId('main-session'))
@@ -1365,7 +1365,7 @@ describe('terminal mounting', () => {
     })
 
     expect(terminal.started).toBe(0)
-    expect(terminal.output).toBe('ui-tui: session "main-session" failed to start: <unrenderable value>\n')
+    expect(terminal.output).toBe('ui-tui: session "main-session" failed to start: <unrenderable thrown value>\n')
     expect(exit).toHaveBeenCalledWith(1)
     await ctx.fiber.dispose()
   })