F6a: push the rewritten head behind a durable marker

agents/process-group.ts

@@ -49,6 +49,13 @@ export interface ProcessGroupOptions {
    * Callers without durable ownership wait fail-closed instead of returning while unknown work is still live.
    */
   readonly allowUnsettledReturn?: boolean;
+  /**
+   * Bound only the wait for the output pipes once the whole group is gone: a process that left the group but still holds
+   * them makes the call return failed (`ESTDIOHELD`) after the pipe drain, instead of waiting for it. Members of the
+   * group are still waited for, fail-closed. For callers whose command changes nothing (a read), so a process that
+   * escaped the group cannot affect anything they decide.
+   */
+  readonly boundEscapedOutput?: boolean;
 }
 
 const DEFAULT_GRACE_MS = 5_000;
@@ -94,7 +101,8 @@ export const drainProcessGroup = (group: ProcessGroup): Promise<boolean> => drai
  * returns control to the event loop. On abort or at the deadline the whole group gets SIGTERM, then SIGKILL after the
  * grace period, so a child that ignores SIGTERM is still bounded. The promise settles only after the leader has
  * exited and every other member of its group is gone, unless a caller with durable recovery ownership explicitly
- * allows an unsettled return. Never rejects: failures are in the outcome, shaped like `runDocker`'s, so `status` is a
+ * allows an unsettled return. `boundEscapedOutput` bounds only the wait for a process outside the group that holds the
+ * output pipes. Never rejects: failures are in the outcome, shaped like `runDocker`'s, so `status` is a
  * number only when the command ran to completion.
  */
 export function runInProcessGroup(file: string, args: readonly string[],
@@ -175,7 +183,7 @@ export function runInProcessGroup(file: string, args: readonly string[],
       options.signal?.removeEventListener('abort', onAbort);
       void drainGroup(group, options.allowUnsettledReturn === true).then(async drained => {
         let stdioTimer: ReturnType<typeof setTimeout> | undefined;
-        if (options.allowUnsettledReturn) {
+        if (options.allowUnsettledReturn || options.boundEscapedOutput) {
           await Promise.race([closed, new Promise<void>(done => { stdioTimer = setTimeout(done, STDIO_CLOSE_MS); })]);
         } else await closed;
         // Cleared so a finished call never keeps the process alive.

docs/architecture.md

@@ -42,9 +42,9 @@ Writing standard: plain language, ISO 24495-1:2023
 - **Build state:**
   - **In use:** plan and linking library, SQLite store, review screen, plan display/import/authoring screen, Ask (questions to a Claude agent), guarded merge with merge-queue support, ranked Issues screen, agent isolation boundary, and the single-runner lock.
   - **In use for a non-demo review with a `github` block:** planning. `/api/plan/suggestions` and `/api/plan/drafts` ask Claude, through lane D, for suggestion cards or a whole next plan revision (#117, #124). Nothing becomes a revision until you apply it.
-  - **In use with an opt-in `runner` block in `review.json`:** the whole loop after planning. `start` and `resume` on `/api/runner` run a task's plan item by item with a Claude agent (#91). They require current plan-item approvals, bind both the task state and review version, recheck approvals before each later item, and refuse when a completed prefix no longer ends at the task's current head (#107). After an out-of-scope pause, an approved amended plan can reconcile the completed prefix and resume only the unfinished suffix (#88, PR #134). When a run ends, codeboost publishes the task's pull request, a draft if the task needs a person (#103). `prepare-merge` refreshes the published PR's exact base/head, rebases through the durable F3/F4 path, recomputes attribution and approvals, and runs allowlisted `cmd:` checks in a credential-free read-only container bound to the resulting head (#22). Approve & merge merges that pull request (#121). Cancelling a task closes its pull requests (#111). These are API actions; the screen has no buttons for them yet.
-  - **Partly built for pre-merge automation:** the trusted local rebase engine records one-to-one commit mappings, verifies the resulting checkout byte for byte, owns its subprocesses durably, and recovers interrupted work (#22, PR #136). Its production conflict path gives a sandboxed agent only the bounded conflict files and imports only audited results for both foreign and plan-owned commits (PRs #138, #140 and #144). `prepare-merge` now invokes that path after current pull-request base/head reads, then refreshes attribution and approvals and runs exact-head `cmd:` checks; required-check refresh, the already-fixed rerun, rewritten-head push, and guarded merge handoff remain for F6.
-  - **Not built:** screens for the runner actions, remote push/check refresh, review rounds, and the Queue and Learning screens.
+  - **In use with an opt-in `runner` block in `review.json`:** the whole loop after planning. `start` and `resume` on `/api/runner` run a task's plan item by item with a Claude agent (#91). They require current plan-item approvals, bind both the task state and review version, recheck approvals before each later item, and refuse when a completed prefix no longer ends at the task's current head (#107). After an out-of-scope pause, an approved amended plan can reconcile the completed prefix and resume only the unfinished suffix (#88, PR #134). When a run ends, codeboost publishes the task's pull request, a draft if the task needs a person (#103). `prepare-merge` refreshes the published PR's exact base/head, rebases through the durable F3/F4 path, recomputes attribution and approvals, pushes a rewritten head whose review is still clear over exactly the PR head it rewrote, and runs allowlisted `cmd:` checks in a credential-free read-only container bound to the resulting head (#22). Approve & merge merges that pull request (#121). Cancelling a task closes its pull requests (#111). These are API actions; the screen has no buttons for them yet.
+  - **Partly built for pre-merge automation:** the trusted local rebase engine records one-to-one commit mappings, verifies the resulting checkout byte for byte, owns its subprocesses durably, and recovers interrupted work (#22, PR #136). Its production conflict path gives a sandboxed agent only the bounded conflict files and imports only audited results for both foreign and plan-owned commits (PRs #138, #140 and #144). `prepare-merge` now invokes that path after current pull-request base/head reads, refreshes attribution and approvals, pushes the rewritten head behind a durable push marker (F6a), and runs exact-head `cmd:` checks; required-check refresh, the already-fixed rerun, and the one-click guarded merge handoff remain for F6b.
+  - **Not built:** screens for the runner actions, the required-check refresh and one-click merge handoff (F6b), review rounds, and the Queue and Learning screens.
 
 ## Terms used
 
@@ -357,7 +357,7 @@ The diagram shows the designed transitions. With the `runner` block, running, ne
 
 ### Stored data
 
-All state is in one SQLite file, opened with WAL and full synchronization. Each write takes an immediate write lock. The schema version is `PRAGMA user_version`, and migrations run in explicit steps (version 17 today); an unknown version fails.
+All state is in one SQLite file, opened with WAL and full synchronization. Each write takes an immediate write lock. The schema version is `PRAGMA user_version`, and migrations run in explicit steps (version 18 today); an unknown version fails.
 
 | Group | Tables | Notes |
 |---|---|---|
@@ -410,7 +410,7 @@ These are the decisions that shape the structure. The full list, with the reason
 | C, K | Guarded merge, merge queue | Done |
 | D | Agent isolation boundary | Done. Since the isolation gate: bounded diff export (#76), task-change inspection and the runner's commit as a bounded `git bundle` (#66, PRs #85 and #90), owner-scoped recovery for Ask (#95), the plan schema for Claude (#97), gitlink mounts and the pre-launch tree check (#81, #99, PR #100), and Codex refused in every phase (#93, PR #106). The runner-contract follow-ups (#51) are closed: the last one tests export and removal through a recovery handle after a real restart (PR #109). |
 | E | Planning library and suggestions | Done. E4 (PR #45) added the acceptance fixtures and recorded Claude output that replays through validation and the Store. |
-| F | Runner | Merged: F1 (lifecycle store, coordinator, shutdown wiring, startup recovery and the lock, planning API, feedback events, and stops that wait for their commit: #53, #74, #57, #59, #60, #96); F2a–F2b (execute prompts, audit, per-item execution, real workspace, durable safety findings: #67, #68, #92, #94, #128); F2d (already-fixed check, PR opening, branch pusher, production publishing, closing PRs on cancel: #84, #101, #110, #115, #116, #119, #123); production wiring and `start`/`resume` (#91, PRs #102, #105); planning through lane D (#117, #124, PRs #118, #120, #125); merging the task's own PR (#121, PR #126); start/resume approval, review-version, recovery and completed-prefix guards (#107, PR #130); continuation after a scope pause (#88, PR #134); the F3 durable trusted local-rebase foundation (#22, PR #136); the F4 foreign-conflict engine plus production sandbox adapter (#22, PRs #138 and #140); and sandboxed owned-commit conflict resolution (#22, PR #144). This branch adds F5: current base/head refresh, rebase admission, post-rebase attribution and approval refresh, and head-bound `cmd:` execution. F6 owns rewritten-head push, fresh required-check refresh, the already-fixed rerun, and exact-pair guarded-merge handoff. |
+| F | Runner | Merged: F1 (lifecycle store, coordinator, shutdown wiring, startup recovery and the lock, planning API, feedback events, and stops that wait for their commit: #53, #74, #57, #59, #60, #96); F2a–F2b (execute prompts, audit, per-item execution, real workspace, durable safety findings: #67, #68, #92, #94, #128); F2d (already-fixed check, PR opening, branch pusher, production publishing, closing PRs on cancel: #84, #101, #110, #115, #116, #119, #123); production wiring and `start`/`resume` (#91, PRs #102, #105); planning through lane D (#117, #124, PRs #118, #120, #125); merging the task's own PR (#121, PR #126); start/resume approval, review-version, recovery and completed-prefix guards (#107, PR #130); continuation after a scope pause (#88, PR #134); the F3 durable trusted local-rebase foundation (#22, PR #136); the F4 foreign-conflict engine plus production sandbox adapter (#22, PRs #138 and #140); and sandboxed owned-commit conflict resolution (#22, PR #144). F5 (PR #149) added current base/head refresh, rebase admission, post-rebase attribution and approval refresh, and head-bound `cmd:` execution. This branch adds F6a: the durable push of the rewritten head. F6b owns fresh required-check refresh, the already-fixed rerun, and the one-click exact-pair guarded-merge handoff; F6c owns common CI. |
 | Hardening | Allowlisted subprocess environments | Done: every Git command (#82, #83) and every `gh` runner (#84, #86). |
 | Tests | Test suite reliability | Done. The full suite runs ordinary tests first and Docker-backed files serially; CI and the real-Docker gate passed on PR #132 (#129). |
 | G | Planning screen | Done (#145). The screen displays/imports the current plan, requests and renders a next-revision draft or suggestion cards, and applies results with revision-bound exact ambiguous replay. G4 composes the browser with the production planning setup, D-backed provider boundary, F endpoints and SQLite persistence; focused adapter, recorded-output and exact-head real-Docker planning suites cover the lower boundary. |
@@ -419,7 +419,7 @@ These are the decisions that shape the structure. The full list, with the reason
 
 ## Known limits
 
-- `prepare-merge` invokes the trusted rebase and sandboxed conflict machinery, refreshes the review, and runs head-bound `cmd:` checks. Until F6 of #22 lands, it does not push a rewritten head, wait for fresh required checks, rerun the already-fixed check, or hand the exact pair to guarded merge.
+- `prepare-merge` invokes the trusted rebase and sandboxed conflict machinery, refreshes the review, pushes the rewritten head, and runs head-bound `cmd:` checks. Until F6b of #22 lands, it does not wait for fresh required checks, rerun the already-fixed check, or hand the exact pair to guarded merge; a person clicks Merge once GitHub's checks on the pushed head pass.
 - Linking is line-based. It cannot see an unrelated edit inside a declared file; only the review agent and you can catch that.
 - The container's guarantees are those of Docker. The agent can always reach its own vendor account.
 - `cmd:` acceptance checks run during `prepare-merge`, and their passing evidence is bound to the exact reviewed head. A plan item without current passing evidence still blocks merging.