fix(agents): child background tools, slice 2 tracker - #2575

Current evidence update — 2026-11-02

Refs #2266. Existing draft scope and source boundaries are unchanged; no merge. This section supersedes the pending-status claims in the preserved historical description below.

Candidate-bound native disposition

All three source heads are unchanged. Child 1 (817cd5370f7a05fef193fda0688df64327db1a2a..1bd87f203fb03c91af687d494b06f120bc0fa930) and child 2 (1bd87f203fb03c91af687d494b06f120bc0fa930..73cc0e8ff08ef7a2915b323fa80d6ee43487d115) were assessed medium/under_budget: native review was not due; neither is described as approved. Child 3 (73cc0e8ff08ef7a2915b323fa80d6ee43487d115..223ed83e7fc7c9bb81b99a9cef87deb75fd54f55) received native approval; exact acknowledgement consumed review-ab8426c0bd78958a. No correction was required; nonblocking finding count was not retrieved. Approval covers that range only, not the combined validation candidate, the whole feature or merge. Existing independent focused checks remain 119/127/140; they are not replaced by native review.

Child 2 CI update — one authorized Windows job repeat

Current source CI is SUCCESS for all three children. Child 2 run https://github.com/rubenpenap/gentle-shell/actions/runs/41568820369 completed attempt 2: SUCCESS, on unchanged head 73cc0e8ff08ef7a2915b323fa80d6ee43487d115.

The owner explicitly authorized one repeat of the failed Windows installer job. Exactly one job-rerun request was submitted for job 125696704914; new Windows job https://github.com/rubenpenap/gentle-shell/actions/runs/41568820369/job/125701920708 passed (2026-11-03 03:09:23–03:19:04 UTC). The other five records in attempt 2 retain their original 02:53–02:57 execution timestamps: their successful results were inherited, not newly executed. No source changes, additional reruns or merge.

The initial 14-second guard intervention and cleanup EPERM remain preserved below. The passing repeat does not establish root cause, prove flakiness or mean a code fix was applied. Bench remains unrun and Slice 3 remains out of scope. This update supersedes the child-2-red status in the initial-attempt table below, not the failure evidence.

Initial source CI attempts — history preserved

These runs followed owner opt-in to Actions and one close/reopen event per draft; no source commits, reruns or merge. In all three, verify (tests, typecheck, runtime modules, package contents and packed install), macOS session transport, Windows repository checks and Ubuntu/macOS installer jobs passed. Windows installer also passed in children 1 and 3.

Child 2 failed at Run installer suites, test 449: native Windows entry: spaces, Unicode, CMD metacharacters and early missing bundle stop. The 14-second outer guard fired with readyObserved:false, recordExists:false, diagnosticSeen:false; probe result status:null, guardKilled:true, expectedDiagnostic:false. The reported TAP error was EPERM / Permission denied in rmSync during fixture cleanup (tests/installer-windows-bootstrap.test.ts:208, finally at :676), which can obscure an earlier assertion. Totals: 525 tests / 460 passed / 1 failed / 64 skipped / 0 cancelled, exit 1. The later native-Windows requirement step was skipped.

The root cause is unresolved. These observations establish neither a Slice 2 regression nor a baseline/flaky classification. No timeout increase, guard weakening, automatic rerun or installer fix was applied. Exact head CI evidence supersedes the earlier no-checks status, not the historical local combined-candidate limitations. Bench remains unrun and merge is not authorized.

Additional local validation (not the published source heads)

Combined local candidate e56e2d2ededfec3ac42ea046d11da14f942023cb contains published child 3 plus the separate numeric-ownership installer fix and its fixture correction. Neither installer commit has been added to this source chain.

• Ordinary private pnpm test: actual exit 0, unit result 7,560 passed / 0 failed / 85 skipped, provider-contract and runtime-harness stages passed. The supervisor wrapper separately exited 1 after a cleanup-budget error; observed-PID cleanup was not proved. This is not a clean wrapper result or global orphan proof.
• Typecheck, runtime-module and package-file structural checks: exit 0.
• Native-required targeted checks: 15/15 passed, no skips, exit 0, pinned provider 4.2.0. These are functional tests, not a new review approval.
• Default packed-package E2E: exit 0 (gentle-pi 4.3.0 / provider 4.2.0); disposable consumer used ranged peers, not a frozen consumer lock.
• Bounded public differential against 817cd5370f7a05fef193fda0688df64327db1a2a: six child-off cases and two fake-runner cases per lane passed. Only the requested removal of stale GENTLE_PI_BACKGROUND_JOBS differed after declared generated-ID/path normalization. Parent-tool/real-shell behavior and all-byte identity are not established.

These local results do not prove exact-published-head full-suite success, resolution of all 85 skips, full Ubuntu/Windows CI, or post-settlement completion.

Remaining gates

• Offline core-corpus bench has not run. The supported build/entrypoint that makes --binary load each distinct gentle-shell TypeScript revision is still unresolved. The question is already posted at bug(agents): declared bash_background is never registered in verifier child processes #2266 (comment); no newer decode2 answer was present at this update. Existing before/after/compare commands and default-off acceptance remain unchanged. No identical-provider comparison or cost/performance claim.
• Source CI on the unchanged published heads is now SUCCESS for all three children. Children 1 and 3 passed initially; child 2 passed after one explicitly authorized Windows job repeat, retaining the other five results. The initial failure remains documented above; its cause is unresolved. This is not bench completion or merge authorization.
• Slice 3 remains out of scope: drain/hold/timers, post-settlement completion and automatic child-model delivery are not implemented or claimed. Keep test(review): intermittent materialized-symlink ENOENT in review-candidate-view tests — observed once in a full-suite run, not reproducible #2552 separate. Draft/no-merge remains in force.
• This update transcribes retained local evidence; no tests or native reviews were rerun. A new read-only verifier could not start because its declared bash_background conflicts with background-subagents policy off; that policy was not changed. Parent structural readback is not a new independent verification.

Historical publication description (superseded status; failures retained)

Linked Issue

Refs #2266. Keep open for Slice 3. PR URL: #2575.
No repository PR template is present; approved mechanical fallback used.

PR Type

• Bug fix

Issue #2266 has both status:approved and type:bug upstream; verified on 2026-11-02.

Summary

• Coordination-only tracker until reviewed children integrate; no source tools are in this initial head.
• All four PRs remain draft/no-merge. Source chain, review, bench and Slice 3 coordination must resolve before tracker merge.
• Publication is authorized through the contributor fork; no merge is authorized. Source PRs target immediate parent branches in that fork.

Chain Context

main (Gentleman-Programming/gentle-shell)
└── 📍 upstream #2575 · feat/2266-slice2-tracker
    └── fork #1 · feat/2266-slice2-01-dispatch
        └── fork #2 · feat/2266-slice2-02-job-ipc
            └── fork #3 · feat/2266-slice2-03-registration

Starts at verified main cb0c5a2e548e278bd65fc503552b10862c4efa11.
Target: Gentleman-Programming/gentle-shell, base main, head rubenpenap:feat/2266-slice2-tracker.
Ends with coordination now; eventual source integration requires separate gate resolution.
Source transplant and bounded test-fixture reconciliation are complete; depends on all child reviews. Follow-up: children 1–3 then Slice 3 coordination.
Out of scope: drain/hold/timers, post-settlement completion, automatic child-model delivery and #2552 fixes.
No size exception is requested or approved; source counts exclude coordination docs and need current-base confirmation.

Changes Table

Test Plan

Current upstream-base aggregate (child 03): parent and independent focused checks 140/140 passed, typecheck 0 diagnostics. Child 01 runner 119/119; child 02 runner/messaging 127/127. Parent topology and isolated 46/194/310-line diffs verified.
Current-base pnpm test exited 1 before executing tests: ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY during an automatic dependency-install attempt. No reinstall, retry, CI override or substitute full-suite run. A fresh current-base off-path historical differential remains unverified. CI/packed install: pending.
Historical old-base only: focused 136/136, typecheck 0, differentials without observable regression.
Old full suite: 7,385 total, 7,290 passed, 22 failed, 73 skipped; baseline environmental failures diagnosed, not a current-base pass.
Native review unavailable on original retained-selection mismatch; no lineage/approval. New candidates unreviewed.
Bench per decode2 comment6424231029: deterministic offline core corpus, no new scenarios, not executed:

gentle-ai-bench run --binary <main-build> --out before.json
gentle-ai-bench run --binary <candidate-build-with-your-slice-2> --out after.json
gentle-ai-bench compare --before before.json --after after.json

Preserve both run exits. Default policy off: no journey status flip; friction within baseline noise.
No paid LLM bench or measured cost/performance claim. Functional post-yield acceptance belongs to future Slice 3.

Contributor Checklist

• Nonclosing reference and chain boundaries documented.
• Issue approval and exactly one available type:bug label confirmed/applied for publication.
• Children integrated; fresh verification, native review, offline bench, packed install and required CI complete.
• Slice 3 coordination resolved before tracker merge; no merge authorization implied.