pi-legion 3.0.1 → 3.0.3

package/CHANGELOG.md

@@ -11,6 +11,141 @@ summary so you can see **2.x vs 0.16** without leaving the tree.
 
 ## [Unreleased]
 
+## [3.0.3] — 2026-10-19
+
+### Changed
+
+- **`spawnChild` returns as soon as the child accepts the prompt, not when its
+  first turn finishes (#227/#228).** The in-process backend used to await the
+  whole first agent run, so for 8–18 minutes the launch produced no job: the
+  status board showed a single ledger row with no token rate, and the parent
+  turn was blocked for the duration. Acceptance is defined in the child
+  transcript (a user line past the spawn-time size, `turns.length > 0`);
+  `isLastTurnComplete` is deliberately not required, because an accepted prompt
+  legitimately has no assistant line yet. The prompt promise is drained in the
+  background inside the same ALS child scope, so depth/identity semantics are
+  unchanged. A 15 s start bound is not an error — the handle is returned anyway
+  and `onSpawnDelayed` fires once. See `docs/inprocess-supervisor.md`.
+- **`worktree: true` is centurion-only (#214).** Members stay in-process; only a
+  centurion gets an OS worktree for the squad it leads. Roles that declared
+  `worktree: true` on a leaf no longer pay the clone/bootstrap cost, and the
+  isolation they still need is the centurion's. See REV-17.
+
+### Fixed
+
+- **A first prompt rejected before acceptance no longer produces a ghost
+  child (#227 D1a / #230).** The non-blocking spawn above had a regression: when
+  the SDK threw before `appendMessage` (no model, no API key, mid-compaction)
+  the spawn still returned a handle, which disabled the `fallbackModels` chain —
+  `Orchestrator.launch` only retries the next model when `spawnChild` throws. A
+  reject **before** acceptance now throws `SubagentError(START_FAILED)` and takes
+  the existing rollback path; a reject **after** acceptance keeps the handle and
+  reports `SettleResult.reason = "child turn failed: …"`.
+- **An interrupted `vexillatio` keeps the work it already collected (#227/#229).**
+  The frozen head line stays first and verbatim, then the collected segments (in
+  handle order) and any `✗ <agent>: <error>` launch failures follow, with
+  `details.stillRunning` telling you what is still live. A reason-less abort is
+  now named (`ABORT_WITHOUT_REASON`) instead of surfacing as bare engine text,
+  and `Orchestrator.collect` no longer stamps a deadline on a child whose last
+  turn already finished.
+- **Foreign children are judged by their run-directory owner (#209-3 / #215).**
+  `status` reported a long-dead child as `working` and `stop` could not reach it,
+  because both consulted only the in-memory liveness registry, which is empty by
+  construction for a run this process never spawned. Both now resolve the owner
+  from the run directory.
+- **The rendered `supervisor=` is the arbitrated backend (#209-2 / #216).**
+  `supervisor=rpc` was a literal in three render paths, so an in-process child
+  was labeled `rpc`. `RunRecord.herdr.supervisor` widens to `SpawnMode`, handles
+  carry `spawnMode`, and legacy runs with no such field render `unknown`.
+- **Extensions that register providers work in the default backend
+  (#209-1 / #213).** The in-process session factory loaded additional extension
+  paths into `DefaultResourceLoader` but resolved models through a bare
+  `ModelRuntime.create()`, so non-builtin providers were unreachable. Registered
+  providers are now drained into the model runtime and their catalog refresh is
+  awaited before `resolveCliModel()` reads the synchronous snapshot.
+- **`legion-doctor` no longer implies RPC-only remedies for in-process
+  children (#212 / #222).** In-process spawns got a false-lead
+  `PI_LEGION_CHILD_EXTENSIONS` recommendation and RPC CLI-probe issues
+  (`CHILD_NO_MODELS` / `CHILD_FEWER_MODELS`); model gaps now point at
+  `models.json`, role extensions, or `inheritExtensions`.
+- **An `orphan detected` notice now carries its detection time.** The notice is
+  queued through the outbox and can be delivered tens of minutes later; it used
+  to look identical to a fresh scan, so the same dead node was triaged repeatedly.
+  It now leads with `detected at <ISO>`. `staleMs` is intentionally unchanged —
+  it is the value at scan time, matching the persisted `orphan_detected` event.
+- **`tok/s` is a decode rate again, and in-process children show it while they
+  run.** Two independent causes, both fixed:
+  - `generationMs` (the denominator) was derived from the nested
+    `message.timestamp`, which marks when generation *started*. It captured only
+    prompt→first-token latency — a real 1506s session accumulated 616ms — so the
+    `MIN_RATE_GENERATION_MS` floor always tripped and `tok/s` silently fell back
+    to a 60s wall-clock window that counted tool wait as generation, and the tree
+    row's `avg t/s` (gated on `generationMs >= 1000`) never rendered at all. It
+    now sums the measured call length (`message.durationMs`) or, when absent,
+    entry-to-entry write times. Nested timestamps stay a last resort for
+    synthetic sessions, paired so the two clock axes are never mixed.
+  - The railway header showed `0 tok/s` for an in-process child's entire
+    lifetime while `rpc` looked correct: `spawnChild` awaits the whole first run,
+    so no tracked job exists and the status board only had the ledger node row —
+    which has no `generationMs` and, because `pollUsage` skips non-`running`
+    nodes, no `tokens_out` until settle either. `statusEntryFromNode` now parses
+    `node.sessionFile` directly and falls back to the ledger.
+- A live agent stopped by a tool call longer than the 60s window still decays to
+  `0 tok/s`. That is the honest answer for the trailing window and is now pinned
+  by tests instead of being "fixed" into a stale number later.
+- **Completion notices no longer wait for the user's next message.** A child
+  finishing while the parent agent run was still active was sent as a
+  `followUp`, and the host only drains that queue inside a live run: the notice
+  sat invisible until the next prompt. Notices are now held by the extension and
+  flushed as one burst on `agent_settled` (hard cap `MAX_PENDING_AGE_MS`, 80 s,
+  after which they are force-sent). The join window no longer extends itself on a
+  busy parent — that turn-level signal was the wrong predictor and doubled the
+  delay.
+- **No more phantom `working` rows.** `awaiting`/re-armed states repaint
+  immediately and `agent_settled` repaints once, so the widget board never holds
+  a stale row past a state change.
+
+### Known limits (deliberately out of scope)
+
+- **Blocked prompts and root notices are still sent as bare `followUp`.**
+  `notifyBlocked` (`index.ts`, the blocked-approval prompt) and `notice()`
+  (`index.ts`, the root-level advisory) bypass the run-scoped outbox, so while
+  the parent run is active they sit in the host queue exactly like the
+  completion notices used to. Only completions are covered by this fix; the
+  same treatment for those two paths is a follow-up, not part of this change.
+
+## [3.0.2] — 2026-10-18
+
+### Changed
+
+- **Reverted the pnpm migration (#162): npm is the reference path again, and
+  the repository no longer enforces any package manager (#210).** Both previous
+  attempts to force contributors onto pnpm broke npm consumers: 3.0.0 shipped a
+  `preinstall` guard that failed every consumer install, and 3.0.1 replaced it
+  with `devEngines.packageManager` (`onFail: "error"`), which npm enforces for
+  *every* non-global command in the checkout. The practical effect was that
+  `npm run`, `npm pack`, `npm exec`, `npx` (and therefore MCP servers launched
+  with `npx`) all failed with `EBADDEVENGINES` inside the repo, and
+  `pi install git:<repo>` failed too, because that path is `git clone` +
+  `npm install --omit=dev --legacy-peer-deps`.
+- `package.json`: dropped `devEngines` (it makes npm reject the whole checkout),
+  `packageManager` (corepack then refuses to run npm at all) and `engines.pnpm`.
+  `engines.node` stays `>=22`.
+- `scripts.test:all` and `scripts.prepublishOnly` compose via `node --run`
+  instead of `pnpm run`, so `npm run` and `pnpm run` both work and no script
+  names a package manager binary. (`node --run` does not execute pre/post hooks;
+  there are none in this manifest.)
+- Lockfile: `pnpm-lock.yaml` is replaced by `package-lock.json`, and
+  `pnpm-lock.yaml` / `yarn.lock` are gitignored so two lockfiles cannot drift
+  apart again. `src/runs/provision.ts` still detects all four lockfiles, so
+  contributors who prefer pnpm can keep using it locally.
+- CI and release install with `npm ci` (no `pnpm/action-setup`, no `cache:
+  pnpm`), and the release workflow packs with `npm pack` again — the
+  `pnpm pack` → scratch-directory workaround for `devEngines` is gone.
+- `test/unit/publish-manifest.test.ts` now forbids `devEngines` and
+  `packageManager` in the manifest and asserts the scripts stay package-manager
+  neutral. `release.yml` fails the tag build if either field comes back.
+
 ## [3.0.1] — 2026-10-18
 
 ### Fixed
@@ -258,11 +393,12 @@ branch — pin `pi-legion@0.16.x` if you still need that model. See
 Earlier 0.11–0.15 history (including the `pi-herdr-swarm` /
 `@zzjcool/pi-herdr-subagents` renames) lives in git tags and older Releases.
 
-[Unreleased]: https://github.com/zzjcool/pi-legion/compare/v3.0.0...HEAD
+[Unreleased]: https://github.com/zzjcool/pi-legion/compare/v3.0.2...HEAD
+[3.0.2]: https://github.com/zzjcool/pi-legion/compare/v3.0.1...v3.0.2
+[3.0.1]: https://github.com/zzjcool/pi-legion/compare/v3.0.0...v3.0.1
 [3.0.0]: https://github.com/zzjcool/pi-legion/compare/v2.0.4...v3.0.0
 [2.0.4]: https://github.com/zzjcool/pi-legion/compare/v2.0.3...v2.0.4
 [2.0.3]: https://github.com/zzjcool/pi-legion/compare/v2.0.2...v2.0.3
 [2.0.2]: https://github.com/zzjcool/pi-legion/compare/v2.0.1...v2.0.2
-[2.0.1]: https://github.com/zzjcool/pi-legion/compare/v2.0.0...v2.0.1
 [2.0.0]: https://github.com/zzjcool/pi-legion/releases/tag/v2.0.0
 [0.16.x]: https://github.com/zzjcool/pi-legion/releases

package/CONTRIBUTING.md

@@ -21,13 +21,13 @@ directly to `master`. Implementers do not merge their own pull requests.
 **Before you push.** Run:
 
 ```bash
-pnpm run typecheck
-pnpm test
-pnpm run test:integration   # or pnpm run test:all
+npm run typecheck
+npm test
+npm run test:integration   # or npm run test:all
 ```
 
 The required GitHub Actions check `ci` runs typecheck, unit, integration, and
-mock-llm. Run `pnpm run test:integration` (or `pnpm run test:all`) locally
+mock-llm. Run `npm run test:integration` (or `npm run test:all`) locally
 before you push, especially when the change touches `src/runs/`, `src/legion/`,
 or `src/extension/runtime.ts`.
 
@@ -45,8 +45,8 @@ timeout the caller is awaiting stays referenced, as with the wait timer in
 `src/extension/runtime.ts`.
 
 **npm release.** A `v*.*.*` tag runs the release workflow
-(`.github/workflows/release.yml`). Dependencies are installed with **pnpm**;
-staging still uses the npm CLI (`npm stage publish`) and a stage-only token.
+(`.github/workflows/release.yml`). Dependencies are installed with **npm**;
+staging uses the npm CLI (`npm stage publish`) and a stage-only token.
 A maintainer promotes the staged version on npm with 2FA. Do not use plain
 `npm publish` in GitHub Actions.
 
@@ -58,36 +58,49 @@ the maintainer user for a routine review or merge ask.
 ```bash
 git clone https://github.com/zzjcool/pi-legion.git
 cd pi-legion
-pnpm install
+npm install
 ```
 
-Requires Node.js 22+ and **pnpm** 9+ (`packageManager` in `package.json`;
-`node --experimental-strip-types` is how tests run; there is no compile step
-to ship). Prefer Corepack (`corepack enable`) or install pnpm 9.x. Do not use
-`npm install` in this repo.
+Requires Node.js 22+ (`node --experimental-strip-types` is how tests run; there
+is no compile step to ship). **npm is the reference path** — it is what Pi uses
+for `pi install` / `pi update --extensions` and what the git-source install
+path runs, so the repository must stay installable and usable with npm.
+
+**Do not add package-manager enforcement.** A `preinstall` guard (3.0.0)
+failed every npm consumer; `devEngines.packageManager` / `packageManager`
+(3.0.1) made every npm command in the checkout fail with `EBADDEVENGINES` and
+broke `pi install git:<repo>`. See issue #210. pnpm may be used locally by
+anyone who prefers it — `node --run` keeps the scripts neutral — but the
+repository does not demand it, and CI uses `npm ci`.
+
+> pnpm 12+ users: the manifest no longer pins a pnpm version, so a modern pnpm
+> refuses the install with `ERR_PNPM_IGNORED_BUILDS` (`@google/genai` and
+> `protobufjs` have build scripts). Either run `pnpm approve-builds`, or install
+> with `pnpm install --config.strict-dep-builds=false`. Do not add an `.npmrc`
+> for this — npm warns about the unknown key on every command.
 
 ## Checks
 
 The required GitHub Actions check `ci` (`.github/workflows/ci.yml`, on pull
 requests and pushes to `master`) runs, in order:
 
-1. `pnpm run typecheck`
-2. `pnpm test` (unit)
-3. `pnpm run test:integration`
-4. `pnpm run test:mock-llm`
+1. `npm run typecheck`
+2. `npm test` (unit)
+3. `npm run test:integration`
+4. `npm run test:mock-llm`
 
 A failing integration step fails `ci`, so nesting and orchestrator regressions
 are blocked before merge. `test:live` and `test:docker` stay out of that job.
 
 ```bash
-pnpm run typecheck
-pnpm test                 # test/unit/*.test.ts
-pnpm run test:integration # FakeSupervisor / nested fakes; no live `pi` binary
-pnpm run test:all         # typecheck + unit + integration
-pnpm run test:mock-llm    # also part of the required ci check
+npm run typecheck
+npm test                 # test/unit/*.test.ts
+npm run test:integration # FakeSupervisor / nested fakes; no live `pi` binary
+npm run test:all         # typecheck + unit + integration
+npm run test:mock-llm    # also part of the required ci check
 ```
 
-`pnpm run test:live` is a **pi RPC live** smoke (`test/live/rpc-live.test.ts`).
+`npm run test:live` is a **pi RPC live** smoke (`test/live/rpc-live.test.ts`).
 It skips itself when the `pi` binary is missing from `PATH` (see the probe at
 the top of that file). Do not fail CI on it.
 
@@ -105,8 +118,8 @@ For offline Pi / legion smoke without provider quota, start the scriptable
 OpenAI-compatible mock (see [`test/mock-llm/README.md`](test/mock-llm/README.md)):
 
 ```bash
-pnpm run mock-llm                 # default scenario=echo on :8765
-pnpm run test:mock-llm            # engine + HTTP + optional pi --print smoke
+npm run mock-llm                 # default scenario=echo on :8765
+npm run test:mock-llm            # engine + HTTP + optional pi --print smoke
 ```
 
 `test/docker/mock-llm.mjs` launches the same CLI (default scenario `tool-once`
@@ -145,12 +158,12 @@ bash test/docker/verify-11-nested-alias.sh    # nested + alias + completionGuard
 Or all at once (RPC mainline **1–11** only):
 
 ```bash
-pnpm run test:docker
+npm run test:docker
 ```
 
 **Legacy Herdr / Space / kind checks** live under `test/legacy/docker/` (need
 pi-legion **0.16.x** Herdr-era setup). They are **not** part of
-`pnpm run test:docker`:
+`npm run test:docker`:
 
 ```bash
 bash test/legacy/docker/verify-12-workspace.sh  # Herdr Space pin (legacy)

package/README.md

@@ -83,14 +83,16 @@ trades process isolation for spawn latency.
 
 **Automatic fallback to `rpc`:** when `spawnMode` is omitted *and* the resolved
 spawn cannot be in-process, Legion launches it over `pi --mode rpc` instead.
-That happens when the role needs an OS worktree (`worktree: true` — the writer
-and centurion default). A `kind !== "pi"` role is rejected earlier with
-`INVALID_PARAMS` (only `kind: pi` children are supported at all), so it never
-reaches this fallback.
+That happens when the role needs an OS worktree (`worktree: true` — only the
+`centurion` role may hold this; see Isolation below). A `kind !== "pi"` role is
+rejected earlier with `INVALID_PARAMS` (only `kind: pi` children are supported
+at all), so it never reaches this fallback.
 
 **Explicit is strict:** `spawnMode: "inprocess"` with `worktree: true` is
-refused with `INVALID_PARAMS` rather than silently downgraded, so a role that
-asks for in-process never gets a different transport behind its back.
+refused with `INVALID_PARAMS` (message prefers `set worktree: false`) rather
+than silently downgraded, so a role that asks for in-process never gets a
+different transport behind its back. Non-centurion requests for `worktree: true`
+are demoted to `false` with a once-per-run warning instead of throwing.
 
 Limits for the in-process backend are separate (`subagents.inprocess.*`) and are
 not shared with the `legion.*` tree budgets — see
@@ -488,9 +490,12 @@ another connection holds the db.
 ## Delivery and resume
 
 A centurion owns a Git worktree for its subtree; its worker/scout/reviewer
-children work in that subtree, and a nested centurion gets a nested worktree.
-Each subtree is delivered as a squash MR to its parent level. MRs move upward
-through the tree until the root reviews and merges into the default branch.
+children work in that subtree (`worktree: false`, typically in-process), and a
+nested centurion gets a nested worktree. **Parallel isolated writers:** dispatch
+N centurions from the root (each owns one worktree); do not give every worker
+its own worktree. Each subtree is delivered as a squash MR to its parent level.
+MRs move upward through the tree until the root reviews and merges into the
+default branch.
 
 Resuming follows the tree's saved sessions and node records. If a supervisor
 process ends, its child processes do not become detached workers; resuming the
@@ -522,11 +527,11 @@ another CLI, keep it on the 0.16.x line or replace it with a Pi role.
 ## Development
 
 ```bash
-pnpm install
-pnpm run typecheck
-pnpm test                  # unit tests
-pnpm run test:integration  # integration tests
-pnpm run test:all          # typecheck + unit + integration tests
+npm install
+npm run typecheck
+npm test                  # unit tests
+npm run test:integration  # integration tests
+npm run test:all          # typecheck + unit + integration tests
 ```
 
 ### Mock LLM (offline pi / legion smoke)
@@ -536,9 +541,9 @@ It registers with Pi like any local gateway (`models.json` → `api: "openai-com
 so spawn / stop / doctor / verify paths can run without real API keys:
 
 ```bash
-pnpm run mock-llm                 # http://127.0.0.1:8765/v1  scenario=echo
-pnpm run test:mock-llm            # unit + integration smoke for the mock
-MOCK_LLM_SCENARIO=spawn-scout pnpm run mock-llm
+npm run mock-llm                 # http://127.0.0.1:8765/v1  scenario=echo
+npm run test:mock-llm            # unit + integration smoke for the mock
+MOCK_LLM_SCENARIO=spawn-scout npm run mock-llm
 ```
 
 Use `writeMockPiHome()` from `test/mock-llm` to generate an isolated
@@ -547,7 +552,7 @@ Use `writeMockPiHome()` from `test/mock-llm` to generate an isolated
 
 See [CONTRIBUTING.md](CONTRIBUTING.md#project-workflow) for the project
 workflow: issues, pull requests, the merge gate, and releases
-(deps via pnpm; npm stage publish for staging).
+(deps via npm; npm stage publish for staging).
 
 Version history for consumers: [CHANGELOG.md](CHANGELOG.md). Tag-by-tag
 release bodies: [GitHub Releases](https://github.com/zzjcool/pi-legion/releases).

package/agents/worker.md

@@ -19,7 +19,7 @@ acceptance:
 kind: pi
 placement: split-down
 steer: true
-worktree: true
+worktree: false
 maxSubagentDepth: 3
 ---
 
@@ -38,7 +38,7 @@ maxSubagentDepth: 3
 - **跑不动的命令**（交互式、需要确认）走默认放行，不要挂起等待。
 - **禁止开新的 worker/写操作子 agent**。你唯一允许的嵌套派生是只读的 advisor，
   且仅限失败升级场景（见 AGENTS.md「失败止损」：重试 2 次仍失败时咨询，按裁决执行）。
-- **隔离工作区**：你在独立 git worktree / 分支上干活，不要写父会话的 checkout。改动通过 merge request / pull request 合回默认分支，不要本地 merge，不要 push 到 main/master。
+- **工作区**：在 centurion 子树 worktree（或父会话 cwd）**原地**改代码；隔离与 MR 由 centurion（或根会话）收队。不要本地 merge，不要 push 到 main/master。
 - **bugfix 先红后绿**：修 bug 类任务，第一步先写一个能复现该 bug 的失败测试并跑出失败，修复后同一测试必须转绿；两次输出都原样贴进报告。不复现就修复的，视为未验收。
 
 ## 工作流程
@@ -50,14 +50,14 @@ maxSubagentDepth: 3
    - 全部步骤完成、交活前才跑一次项目的**全量档**（完整测试套件/CI 等效命令）；交活后插件会自动再验证一遍判定验收，不需要中途反复跑全量。全量失败按「失败止损」处理：重试不超过 2 次，仍失败升级咨询 advisor。
    - 项目找不到任何验证入口时，在报告里写明依据（查过哪些地方），不要自己编命令冒充。
 3. 把最终那次 full 验证输出**原样**（含命令与结果）放进最终报告。
-4. push 前逐文件重读完整 diff（`git diff main...HEAD`）：检查遗留的调试代码、console.log/print、注释掉的老逻辑、无关改动混入；发现问题先修再 push。
-5. push 当前分支并打开 MR/PR；把 URL 写进报告和 verdict reason。MR 由主 agent 审查后统一合回，你不要自己 merge。
-6. 写报告到指定路径（任务里会给），格式：做了什么 / 测试覆盖 / 验证输出 / MR 链接 / 未决问题。
+4. 交活前逐文件重读完整 diff：检查遗留的调试代码、console.log/print、注释掉的老逻辑、无关改动混入；发现问题先修。
+5. 不要自己开独立 worktree / 不要自己开 MR；隔离与合回由 centurion（或根会话）收队。把改动说明写进报告。
+6. 写报告到指定路径（任务里会给），格式：做了什么 / 测试覆盖 / 验证输出 / 未决问题。
 
 最后输出 verdict：
 
 ```json
-{"ok": true, "reason": "N files changed, typecheck+test green, mr: <url>"}
+{"ok": true, "reason": "N files changed, typecheck+test green"}
 ```
 
 无法完成时（如依赖缺失、计划矛盾、无法开 MR）：

package/docs/contracts/vexillatio-nonblocking-launch-v1.md

@@ -0,0 +1,161 @@
+# 冻结契约：非阻塞 spawn / 中断不丢结果 / 死子不空等（#227）
+
+日期：2026-10-19 ｜ 状态：frozen（唯一接口；worker 不得自行解释矛盾，发现矛盾上报修订）
+依据：issue #227；HEAD `c97cd27`；spawnMode 缺省 inprocess；`docs/contracts/esc-no-retire-v3.md`（仍在生效）
+
+## 0. 已确认事实（worker 直接引用，禁止重新侦察）
+
+| # | 事实 | 证据 |
+|---|---|---|
+| F1 | `InProcessSupervisor.spawnChild` 在 `await session.prompt(...)` 上阻塞到**整轮** agent run 结束 | `src/supervisor/inprocess-supervisor.ts:649`；SDK `agent-session.js:1380-1413`（`_runAgentPrompt` 才 resolve） |
+| F2 | `RpcSupervisor.spawnChild` 只等到 prompt **被受理**（`preflightResult` → disposition），不跑完 | `rpc-client.js:135-138`、`rpc-mode.js:299-316` |
+| F3 | `Orchestrator.launch` 在 `spawnChild` 返回**之后**才写 `node_launched` / `recordChild` | `src/runs/orchestrator.ts:765`、`:799`、`:904` |
+| F4 | vexillatio 的 abort 早退丢掉了已收集结果，且 details 是 `{}` | `index.ts:2416`、`index.ts:3206-3208`（`ok()` 的 details 恒为 `{}`） |
+| F5 | `interruptedCollect` 恒返回 `status:"running"` + `reason:"aborted by parent"` | `src/runs/orchestrator.ts:1064-1071` |
+| F6 | `collect()` 的 deadline 从**本次调用**起算 | `src/runs/orchestrator.ts:980` |
+| F7 | 会话文件里 user 行先于 `agent_start` 落盘 | `agent-session.js:1601`(appendMessage) → `:1630`(preflight) → `:1631`(`_runAgentPrompt`) |
+| F8 | `event.type === "error"`（`pi-ai` `lazyStream` 写入的 `stopReason:"error"` 记录）**今天被 inprocess supervisor 直接丢弃** | `inprocess-supervisor.ts:1005`（`turn_end|agent_end` 只吃 `event.message`）、`:1038`（`COARSE_EVENT_TYPES` 无 `error`） |
+| F9 | `FakeSupervisor.spawnChild` 是同步返回的（无 `await prompt`） | `test/helpers/fake-supervisor.ts:139-190` |
+| F10 | `test/unit/inprocess-supervisor.test.ts` 的 `FakeSession.prompt` 也是非阻塞语义 | 同文件 `:42-47` |
+| F11 | 门禁基线：`npm run typecheck` / `npm test`(1113) / `npm run test:integration`(81) 在 HEAD 全绿，各约 33s | 实测 |
+| F12 | `spawnedAt` 在 spawn 前后共写 2 次（`persistParentReservation` 与 `recordChild`/`previous`） | `orchestrator.ts:429/444`、`:777`、`:904` |
+
+## 1. 决策（D 编号，reviewer 必读）
+
+**D1 — inprocess `spawnChild` 改为「受理即返回 + 后台 drain」。**
+`session.prompt(...)` 不再 `await` 到整轮结束。`spawnChild` 在下面两个条件**任一**成立时返回 `ChildHandle`：
+- (a) 会话文件已出现属于**本轮**的 user 行（`F7`：prompt 已被受理并落盘）；或
+- (b) 时钟推进到 `spawnStartTimeoutMs`（默认 `DEFAULT_INPROCESS_SPAWN_START_TIMEOUT_MS = 15_000`）兜底。
+
+判定实现细则（冻结）：
+- 拉起点 = `spawnChild` 调用时刻的 `statSync(sessionFile).size`（文件不存在记为 0）。
+- 出现 user 行 = `parseSessionFile(file).turns.length > 0 && fileSize > baseline`。**不得**用 `isLastTurnComplete`（`F7` 已经证明 prompt resolve 时末轮为「有 user、无 assistant」，用它是自相矛盾的门）。
+- 轮询间隔复用 `POLL_INTERVAL_MS = 500` 的既有 cadence。**勘误（见 §7 v1.1）**：该常量实际是 `src/runs/orchestrator.ts:143` 的 module-private 常量，从未存在于 `src/shared/types.ts`；Worker A 在 `src/supervisor/spawn-mode.ts` 导出同名同值常量（与 `DEFAULT_INPROCESS_*` 并列），inprocess-supervisor 从那里 import，**不要**去改 `orchestrator.ts`。
+- 兜底超时**不是错误**：到点照常返回 handle（会话已创建即持久化了 transcript），只是记录 `onSpawnDelayed` 一次（`InProcessSupervisorOptions` 新增可选回调，用于诊断；无默认实现）。
+
+**D1a — 「受理前 reject」仍必须让 spawn 失败（fallback 链保命条款，父会话审查补入）。**
+`session.prompt(...)` 的 reject 分两种，**不得**合并处理：
+- **受理前 reject**（`firstTurnAccepted === false`：会话文件未长过 baseline、无 user turn。真实触发：无 model / 无 API key / 压缩中，见 `agent-session.js:1568-1581` —— 这些抛错发生在 `appendMessage` 与 `preflightResult("started")` **之前**）→ `spawnChild` **必须抛出** `SubagentError(原始 message, ErrorCodes.START_FAILED)`，走既有 `rollbackSpawn` + `onSpawnRefused` + rethrow，使 `Orchestrator.launch` 的 `canFallbackStart`（`orchestrator.ts:235-240`）能继续降级到下一个模型。**不得**返回 handle。
+- **受理后 reject**（`firstTurnAccepted === true`：首轮已开跑才炸）→ 按 D2 返回 handle，`lastTurnResult.reason = "child turn failed: <msg>"`，`onSpawnRefused` 上报。
+
+理由：非阻塞改造若把「受理前拒绝」也吞掉，会静默废掉 `fallbackModels`（已实测：HEAD 抛错 / 首版实现返回 handle，`canFallbackStart` 失效）。这是 #227 之外的回归，必须有先红后绿用例守住。
+
+**D2 — 后台 drain 必须保住 ALS 身份/深度与错误可见性（最大风险点）。**
+- 复用现有的 `runAtInProcessChildDepthAsync(...)` + `runInChildContext(registration.context, ...)` 作用域，把「settle 之后的收尾」放进去；`beginTurn` 保持在 prompt 之前（现状即是）。
+- **事件会脱出 ALS 作用域**：`agent_start` 在 `await session.prompt()` 之前同步 emit（`agent-session.js:1601→1631`），但 `agent_settled` 在 `_runAgentPrompt` 的 `finally` 里 emit，那时已在 drain 的后续微任务链上。因此 `updateTurnState` 里做 settle 状态流转时**不得依赖** `currentChildContext()`；`publishEvent` 已有 `runOutsideChildContext`（REV2-7），不要动。
+- drain 的 `catch`：**不得**调用 `rollbackSpawn`（会 `children.delete` 把已返回的 handle 抹掉）。必须：`turnActive=false`、`state="idle"`、`lastTurnResult={settled:false,abnormal:false,reason:"child turn failed: <msg>"}`、`resolveWaiters(...)`，并把 `<msg>` 交给 `onSpawnRefused`（该回调语义扩为「child 运行期失败诊断」，文档同步）。
+- drain 的 `finally` 必须落一件机器事实：把 prompt reject 的**异常**写成一条 `event.type === "error"` 的事件送进 `handleSessionEvent`，好让 D3/D3b 的映射生效（见 D3b）。
+
+**D3 — `error` 事件要能进 settle 判定（缺陷 3 的 inprocess 半边）。**
+`updateTurnState` 增加分支：`event.type === "error"` → 记 `child.pendingFailure = <error.message 或 error 的字符串化>`。当随后的 `agent_settled` 到达时，若 `lastTurnResult` 会判成 `aborted:false` 而 `pendingFailure` 非空，则用 `F8` 已存在的事件把它变成可见事实（不得凭空造 stopReason）。
+
+**D3b — 中断被写成 `stopReason:"error"` 的读法修正（本仓半边）。**
+`src/shared/session.ts` 的 `deriveOutcome`：当 `stopReason === "error"` 且 `errorMessage` 同时匹配「abort 语义」与「**无 reason 的 AbortController**」时，`reason` 必须给出可判别文案，禁止再与真正的 provider 失败同形。冻结文案：
+- 现有 `/abort/i` 分支保持（`"This operation was aborted"` 已命中，`test/unit/session.test.ts:59-63` 锁死，**不得改动该断言**）。
+- 新增：当 `errorMessage` 匹配 `/^(this operation was aborted|the operation was aborted|operation aborted)$/i` 且**不含** `timeout` 时，`reason` 置为 `ABORT_WITHOUT_REASON`（新增导出常量，值 `"parent turn ended (abort signal without a reason)"`）。`status` 仍为 `aborted`（不变），只改 `reason` 的可读性。
+- 目的：让 issue #227 里「`stopReason:"error"` 被读成 provider 失败」的分析不再成立，同时不改任何 `status` 语义（避免回归 `deriveOutcome` 的既有 511 处断言面）。上游 `pi-ai` 的 `createSetupErrorMessage` 不在本仓，**不在本契约范围**（见 §5）。
+
+**D4 — vexillatio 中断不丢结果。**
+`index.ts:2416` 与 execute 层 `index.ts:743-745` 都不得丢弃已算出的结果。冻结返回形状：
+
+```ts
+// 文本：interrupted 头行 + 已收集的每一段（顺序 = handles 顺序）
+`interrupted (parent turn ended; children still running)\n\n${results.join("\n\n")}`
+// details
+{ runId, handles: [...{name, agent, sessionFile}], async: false, interrupted: true,
+  ...(stillRunning.length ? { stillRunning: string[] } : {}) }
+```
+
+- 头行**必须**保持冻结文案 `interrupted (parent turn ended; children still running)`（`esc-no-retire-v3.md:22`），因为 `test/unit/teams.test.ts:1298` 用 `/interrupted \(parent turn ended; children still running\)/` 匹配——把结果段追加在头行之后不破坏该正则。
+- `ok()` 新增可选 `details` 参数（`ok(text, details?)`，缺省仍是 `{}`），**不要**引入 `isError`（`pi-agent-core` 的 `AgentToolResult` 虽有该字段，但本仓 511 处断言与 TUI 渲染均按「无 isError」编写，改它属于越界）。
+- launch 失败行（`✗ <agent>: <error>`，`index.ts:2364`）**也**必须在中断时返回（现在是先早退、后 push，一并丢）。
+- 真机上 `collect` 内的 `interruptedCollect` 会给 `status:"running"`：渲染时照旧打印 `execution: running (aborted by parent)`，**不要**把它改写成 success（`F5` + `orchestrator.test.ts:153-168` 锁死）。
+
+**D5 — 死子 / 早退子不得空等到 collect deadline（缺陷 1 的收尾）。**
+`Orchestrator.collect`：当 `initial` 里**本轮**已有终态 assistant（`isLastTurnComplete(initial) === true`）却因 `liveConnectionBlip` 走进 `awaitSupervisor` 的路径之外——**新增**：若 `wait === "timeout"` 且 `!this.supervisor.isAlive(name)` 且 `isLastTurnComplete(parsed)`，直接按已完成的 turn 收尾（走现有 `deriveOutcome` + `finishCollect`），不得再套 `COLLECT_DEADLINE_REASON`。现有 `collectArtifactGrowing` 的 grace 循环保持不动。
+
+**D6 — `collect` 的 deadline 不因本轮非阻塞 launch 而变（显式不做）。**
+`F6` 标明 deadline 从 collect 调用起算；issue 建议「timeoutMs 覆盖 launch」是**另一件事**，本契约**不改** deadline 起点，只删掉 `COLLECT_DEADLINE_REASON` 注释里那句与实现不符的「budget from launch」（`orchestrator.ts:1007`）。理由：改起点会让 `runtime` 的 re-arm watch 预算翻倍（`runtime.ts:1017-1032`），属于 #227 之外的语义变更。
+
+**D7 — notify 的 `batch N` 文案（显式不做）。**
+`notify.ts:351` 的 `batch ${entries.length}` 只是 TUI 折叠行的可读摘要（`notice-renderer.ts:69`），不是机器事实，且没有测试断言它。issue 把它列为「顺带观察」，本契约**不修改**，只在 §6 记为已知表述。
+
+## 2. 文件所有权（两 worker 并行，零共享文件）
+
+### Worker A（`spawn` 非阻塞 + drain + error 事件）
+
+| 文件 | 动作 |
+|---|---|
+| `src/supervisor/inprocess-supervisor.ts` | D1 / D2 / D3 |
+| `src/supervisor/spawn-mode.ts` | 新增导出常量 `DEFAULT_INPROCESS_SPAWN_START_TIMEOUT_MS = 15_000` 与勘误后的 `POLL_INTERVAL_MS = 500`（均与 `DEFAULT_INPROCESS_MAX_ACTIVE` 并列） |
+| `src/api.ts` | 若导出面需要（`InProcessSupervisorOptions` 已导出）；仅在新增类型被外部引用时改 |
+| `docs/inprocess-supervisor.md` | 改写 `spawnChild awaits the whole first agent run` 那段（现 `:130-137`），改为新语义；同步 `runtime.ts:1478-1482` 注释所指事实变化（**只改文档**，`runtime.ts` 属 Worker B 边界外、也不改） |
+| `test/unit/inprocess-supervisor.test.ts` | 新增 D1/D2/D3 用例；**必须保住**既有 `child context is active during the first prompt`(`:569`) 与 `maxDepth holds across generations driven from the root chain (REV2-1)`(`:700`) 与 `onEvent listeners run at root identity`(`:1071`) 三条 |
+
+### Worker B（中断不丢结果 + collect 收尾）
+
+| 文件 | 动作 |
+|---|---|
+| `index.ts` | D4（`ok()` 可选 details、`launchUtilityFamily` 返回、execute 两层早退） |
+| `src/runs/orchestrator.ts` | D5、D6 的注释修正 |
+| `src/shared/session.ts` | D3b（`ABORT_WITHOUT_REASON` + `deriveOutcome` 新分支） |
+| `test/unit/session.test.ts` | D3b 新用例（**不得改动** `:59-63` 既有断言） |
+| `test/unit/teams.test.ts` | D4 用例：abort 时 vexillatio 返回已收集结果 + details 带 `interrupted`/`stillRunning`/`runId`（既有 `:1254` 用例的断言不得删减） |
+| `test/integration/orchestrator.test.ts` | D5 用例（死子不空等）；`:153` 既有断言不得改动 |
+| `docs/vexillatio-contract.md` | §2.2 步骤 3 补一句中断返回语义（append-only 修订记录 REV-17.9） |
+
+**共享文件：无。** 两 worker 各在自己的 worktree，冲突面为零；`test/unit/inprocess-supervisor.test.ts` 只属 A，`test/unit/teams.test.ts` 只属 B。
+
+## 3. 验收（机器可检查，父会话在 collect 时执行）
+
+每个 worker 的 `verify`：
+```bash
+npm run typecheck
+npm test
+npm run test:integration
+```
+首个非零退出即 rejected。另需人工核对的断言锚点（reviewer 用）：
+
+1. 新增 inprocess 用例：`prompt` 立即 reject 且**不写会话文件** → `assert.rejects(spawnChild)`，`error.code === START_FAILED`，`children.has(name) === false`（无幽灵子）、`isAlive === false`。**必须在修复前是红的**（D1a）。
+2. 新增 inprocess 用例（对照）：`prompt` 先写一条 user 行再 reject → `spawnChild` 正常返回 handle、`children.has(name) === true`、`waitSettled` 的 `reason` 含 `child turn failed`。
+3. 新增 inprocess 用例：用一个**故意永不 resolve** 的 `prompt()`（模拟 8~18 分钟的子轮），断言 `spawnChild` 在 `spawnStartTimeoutMs` 量级内返回（用注入的短超时，如 50ms，不真等 15s）。
+4. 新增 inprocess 用例：`onEvent` 仍收到 `agent_start` / `agent_settled`（后台 drain 不丢事件）。
+5. 新增 teams 用例：abort 后 vexillatio 文本同时含冻结头行**和** `── <name> ──` / `execution:` / 子结果正文；`details.interrupted === true`；`details.handles.length > 0`；`details.runId` 非空；child 未被 retire。
+6. 新增 orchestrator 用例：子进程已死且本轮已有终态 assistant → `collect` 不返回 `COLLECT_DEADLINE_REASON`（注：Worker B 已披露该分支在 HEAD 上不可观测，属防御性收尾，非行为修复——reviewer 不应据此判「假覆盖」）。
+7. 新增 session 用例：`stopReason:"error"` + `"This operation was aborted"` → `reason === ABORT_WITHOUT_REASON`；`+ "The operation was aborted due to timeout"` → 仍是 timeout 语义（不得被新分支吞掉）。
+8. 既有绿线不回退：`inprocess-supervisor.test.ts` 33 例、`teams.test.ts`、`orchestrator.test.ts` 25 例、`session.test.ts` 全部。
+
+## 4. 明确不做（reviewer 判「缺失」前先读本节）
+
+| 不做 | 理由 |
+|---|---|
+| 改 `collect()` 的 deadline 起点为 launch 时刻 | D6：会翻倍 runtime re-arm 预算，属 #227 之外的语义变更 |
+| 给 inprocess 加 `preflightResult`（SDK 有该选项） | D1 用会话文件判定，**不扩 `InProcessSession` 窄接口**：该接口是所有测试 fake 的契约面（`inprocess-supervisor.test.ts:29`），扩参数会让 20 个 fake 全部要改且没有一处需要它 |
+| 改 `notify.ts` 的 `batch N` 文案 | D7：TUI 摘要，非机器事实，无断言 |
+| 引入 `isError` | D4：`AgentToolResult` 有该字段，但本仓渲染链/TUI/511 处断言按「无 isError」工作，改它属越界 |
+| 上游 `pi-ai` 的 `createSetupErrorMessage` | 不在本仓（issue 也建议单独 file） |
+| RPC 侧 `spawnChild` 的 `await client.prompt` | `F2` 已经是非阻塞；**不改** |
+| `previous.spawnedAt` 的 resume 重写 | `F12` 是既有语义（resume 重置计时），与 #227 无关 |
+| `test/unit/legion-integration.test.ts:1595` 那条「因为 spawnChild 阻塞所以直读 sessionFile」的测试 | 断言本身仍成立（台账节点仍可无 job 而直读文件），**不改**；`runtime.ts` 的 `statusEntryFromNode` 注释由 Worker A 在文档层说明，代码不动 |
+
+## 5. 风险与回滚
+
+- **R1（最高）**：`F1` 的阻塞是 `child context is active during the first prompt` 与 REV2-1 深度测试的**隐式前提**（ALS 只在被 await 的链上存活）。改为非阻塞后，首轮 turn 的 tool 执行若脱出 ALS，这两个测试会红。D2 要求把 drain 挂在原作用域内；若测试仍红，**停下来上报**，不要改测试断言。
+- **R2**：`agent_settled` 在 `_runAgentPrompt` 的 `finally` 里 emit，可能晚于 drain 的 resolve；`beginTurn` 与 `lastTurnResult` 的时序必须保持「先清后置」，否则 `waitSettled` 会立刻拿到陈旧结果。
+- **R3**：`rollbackSpawn` 在新语义下不能用于「drain 失败」（会删号），见 D2。
+- **R4**：vexillatio 中断返回文本变长，`display`/preview 上限（`PREVIEW_CHARS=4000`）在 notify 侧，不在 tool result 侧；tool result 不截断（`index.ts:2400` 现状即不截断）。
+- 回滚点：两 PR 独立，可分别 `git revert`；D3b 是纯 reason 文案，回滚无副作用。
+
+## 6. 已知但不在本次范围
+
+- 上游 `pi-ai` 把 abort 写成 `stopReason:"error"`（issue 缺陷 3 的根因）——建议单独 file 给 `pi-ai`，本契约只做本仓可判别的 reason（D3b）。
+- `notify` 的 `batch N` 表述（D7）。
+- `subagent` 工具（非 vexillatio）的中断路径已保留 `runId`（`index.ts:1751-1754`），本次不动。
+
+## 7. 修订记录
+
+| 日期 | 版本 | 修订 | 原因 |
+|---|---|---|---|
+| 2026-10-19 | v1 | 首次冻结 | #227 三缺陷拆分：D1/D2/D3(A)、D4/D3b/D5(B)；D6/D7 显式排除 |
+| 2026-10-19 | v1.1 | 勘误 §1 D1 轮询常量位置：`src/shared/types.ts` → 实为 `src/runs/orchestrator.ts:143` 私有常量，改由 `src/supervisor/spawn-mode.ts` 导出同值常量；§2 Worker A 文件表同步 | Worker A 指出契约原文事实错误；`grep` 确认 `POLL_INTERVAL_MS` 从未在 `src/shared/types.ts` 出现过。属笔误型而非设计型矛盾，advisor 裁决按勘误继续（不回退重派），父会话在此落库 |
+| 2026-10-19 | v1.2 | 新增 **D1a**：首次 prompt 在「受理之前」reject 时 `spawnChild` 仍须抛出（保 `fallbackModels`）；§3 验收清单同步为 8 条 | 父会话审查 PR #228 实测发现回归：非阻塞改造把受理前 reject 吞成「已 spawn」，`canFallbackStart` 失效。HEAD 抛错 / 首版实现返回 handle，属 #227 之外的回归，必须以先红后绿用例守住 |

package/docs/design.md

@@ -219,18 +219,48 @@ launch (async) ──→ 子 pane 跑
        │
        ├─ TUI：输入框上方 aboveEditor widget（live agent 面板）
        │        （pi-subagents fleet-status；footer 不再重复计数，见 #192）
-       └─ collect 在后台等 turn 结束
+       └─ collect 在后台等 run 结束
               │
-              └─ pi.sendMessage({ customType: "subagent-notify" },
-                    { triggerTurn: true, deliverAs: "followUp" })
-                 父会话空闲则立刻唤醒；若父会话还在跑当前 turn（工具循环），
-                 等这一轮结束后再投递，避免 steer 打断正在干的事。
-                 所有状态都 display:true（transcript 留下完成/失败记录）；
-                 成功由 notice-renderer 压成单行并按 pane 宽度硬截断
-                 （控制字符已消毒，ctrl+o 展开），
-                 失败/中止保留完整块，details 只供 TUI 不进 LLM 上下文。
+              └─ runtime 完成通道（唯一的延迟源，按 **run** 级判定）
 ```
 
+完成通知的投递分两段，**只有一处 run 级延迟**：join 窗口是纯批处理延迟，二者串行叠加
+（最坏串联约 10s + 80s）——除此之外不存在第二个"等父会话忙完"的延迟源。
+
+1. **join 批处理窗口**（`src/extension/join.ts`，默认 10s）：同一 runId 的多个 child
+   一起启动、先后结束，窗口把 straggler 合成一条消息（`Background tasks completed
+   (N of M)`）。窗口**只做批处理**，到期无条件 flush 到下一个环节，**不**因为父会话忙
+   而延长——turn 级的 busy 是错误预测器（工具循环里父会话按定义就是忙的），而且会在
+   outbox 之外叠加第二层延迟。
+2. **run 级 outbox**（`src/extension/runtime.ts`，`MAX_PENDING_AGE_MS = 80_000`）：
+   父 **run** 活跃时不 `pi.sendMessage`，通知入 FIFO 队列；`agent_settled` 到达后
+   一次性 burst 发出（一条通知一条消息），并重绘 UI。硬上限 80s（= 2× 旧的
+   `(1+3)×10s`）到点强制发送并计入 `noticeQueueStats().forcedFlushes`，宁可回到旧
+   行为也不无限隐藏。父 run 不活跃时**立即**发送（含 `mode: "each"`、单成员批次、
+   `triggerTurn:false` 项）。
+
+   上限是 **per-notice** 而不是 per-run：到点判定用 notice 自己的 `enqueuedAt` 算
+   （`runtime.ts` 的 `armOutboxTimer` 从 `outbox[0]` 起算，
+   `noticeQueueStats().oldestAgeMs` 也报最老一条）。所以父 run 持续 10 分钟时，不是
+   给整个 run 记一个 80s 预算、也不是因为 run 还在跑就无限顺延：队列里最老一条到点就
+   触发（`flushOutbox` 一次 `splice(0)`，连同队列里更年轻的 notice 一起按 FIFO burst
+   发出）。T3d 把「arm 与 oldestAgeMs 都跟最老一条」钉成测试。
+
+为什么不能只发 `followUp`：宿主只在 **live run 内部** drain `agent.followUpQueue`，
+所以父 run 活跃时发出去的 followUp 会一直躺在队列里，直到用户再发一条消息才出现。
+这正是「完成通知看起来没来」的根因。发送时形状不变——
+`pi.sendMessage({ customType: "subagent-notify" }, { triggerTurn: true, deliverAs: "followUp" })`
+——只是**时机**改到 run 结束时。
+
+所有状态都 display:true（transcript 留下完成/失败记录）；
+成功由 notice-renderer 压成单行并按 pane 宽度硬截断
+（控制字符已消毒，ctrl+o 展开），
+失败/中止保留完整块，details 只供 TUI 不进 LLM 上下文。
+
+> print/json 模式：`session_shutdown` → `dispose()` 丢弃未投递的 pending。child 的
+> session 文件仍在，可用 `collect` 从磁盘恢复；但 `finished` 缓存**不**保留
+> （`dispose()` 开头就 `jobs.clear()` / `finished.clear()`，此后才丢弃 outbox）。
+
 父 agent 的正确用法：`subagent` 工具 launch 之后把控制权交回用户，等 completion
 message，而不是轮询，也不是让孩子来拍自己。