implement-wg-pi-compaction-kick — WorksGood live mirror

Metadata

Status	in-progress
Assigned	`agent-16`
Agent identity	`02e879681e52e0a384106169be043416c4d946e850ab26b2269c57681b52a6e7`
Created	2026-08-05T20:52:42.575689353+00:00
Started	2026-08-06T12:47:35.948396021+00:00
Tags	`pi`, `compaction`, `watchdog`, `regression`
Tokens	48466290 in / 94761 out

Description

Objective

Implement the ratified docs/design-pi-compaction-kick.md so an unattended WG Pi task worker that becomes idle after a successful threshold compaction receives one safe continuation kick and resumes the exact unresolved task. Start from the credential-free reproducer and make it RED before production changes.

Requirements:

Reuse the existing Pi watchdog/lifecycle continuation authority and first-terminal-wins boundary; do not create a second retry controller.
Detect only the designed qualifying threshold-compaction gap. Deliver through Pi’s native extension/session API only after a durable WG permit, with exact action idempotency and acknowledgement.
Preserve generation, attempt/fence, worktree lease, exact session/branch, frozen handler/provider/model/reasoning/endpoint, and same process epoch for an in-process nudge.
Suppress on terminal/wait receipt, queued messages, active/non-idle agent, overflow retry, manual or failed/aborted compaction, unsafe effects, proof mismatch, and exhausted finite budget.
Expose bounded diagnostics in existing wg show / wg pi-watchdog status surfaces.
If worksgood-pi/src/** or compat changes, run make embed-worksgood-pi and commit worksgood-pi/embedded/; do not depend on a globally installed extension.
Add a permanent grow-only smoke scenario owned by this task which drives the installed WG binary, generated task wrapper, embedded plugin, and Pi JSON/session event path. The pre-fix behavior must fail because no post-compaction turn occurs; the fixed behavior must show exactly one kick followed by a concrete continuation/terminal receipt.

Validation

RED-first focused test reproduces successful threshold compaction_end(willRetry=false) -> idle/settled with unresolved WG task and no queued work.
Fixed flow emits/delivers/acks exactly one durable kick and the next Pi turn sees the compacted summary and same task/session/route/worktree proof.
Crash/restart tests at permit, session-send, acknowledgement, settle, and wrapper-exit boundaries never duplicate prompt, process, budget charge, side effect, owner, attempt, or session.
Must-not-trigger tests cover manual, overflow, aborted/failed compaction, queued message, active agent, accepted done/fail/wait, non-WG Pi console/chat, unsafe tool/effect, and budget/proof failures.
Existing Pi watchdog, continuation-epoch, native progress/session proof, candidate finalization, and wrapper/convergence tests remain green.
Live installed-binary smoke is registered in tests/smoke/manifest.toml with owner implement-wg-pi-compaction-kick and passes without credentials.
npm test/typecheck for worksgood-pi, cargo fmt, cargo fmt --check, cargo clippy, focused/full tests, make embed-worksgood-pi when applicable, cargo install --path . --locked, and owned smoke pass.

## Objective
Implement the ratified `docs/design-pi-compaction-kick.md` so an unattended WG Pi task worker that becomes idle after a successful threshold compaction receives one safe continuation kick and resumes the exact unresolved task. Start from the credential-free reproducer and make it RED before production changes.

Requirements:
- Reuse the existing Pi watchdog/lifecycle continuation authority and first-terminal-wins boundary; do not create a second retry controller.
- Detect only the designed qualifying threshold-compaction gap. Deliver through Pi’s native extension/session API only after a durable WG permit, with exact action idempotency and acknowledgement.
- Preserve generation, attempt/fence, worktree lease, exact session/branch, frozen handler/provider/model/reasoning/endpoint, and same process epoch for an in-process nudge.
- Suppress on terminal/wait receipt, queued messages, active/non-idle agent, overflow retry, manual or failed/aborted compaction, unsafe effects, proof mismatch, and exhausted finite budget.
- Expose bounded diagnostics in existing `wg show` / `wg pi-watchdog status` surfaces.
- If `worksgood-pi/src/**` or compat changes, run `make embed-worksgood-pi` and commit `worksgood-pi/embedded/`; do not depend on a globally installed extension.
- Add a permanent grow-only smoke scenario owned by this task which drives the installed WG binary, generated task wrapper, embedded plugin, and Pi JSON/session event path. The pre-fix behavior must fail because no post-compaction turn occurs; the fixed behavior must show exactly one kick followed by a concrete continuation/terminal receipt.

## Validation
- [ ] RED-first focused test reproduces successful threshold `compaction_end(willRetry=false)` -> idle/settled with unresolved WG task and no queued work.
- [ ] Fixed flow emits/delivers/acks exactly one durable kick and the next Pi turn sees the compacted summary and same task/session/route/worktree proof.
- [ ] Crash/restart tests at permit, session-send, acknowledgement, settle, and wrapper-exit boundaries never duplicate prompt, process, budget charge, side effect, owner, attempt, or session.
- [ ] Must-not-trigger tests cover manual, overflow, aborted/failed compaction, queued message, active agent, accepted done/fail/wait, non-WG Pi console/chat, unsafe tool/effect, and budget/proof failures.
- [ ] Existing Pi watchdog, continuation-epoch, native progress/session proof, candidate finalization, and wrapper/convergence tests remain green.
- [ ] Live installed-binary smoke is registered in `tests/smoke/manifest.toml` with owner `implement-wg-pi-compaction-kick` and passes without credentials.
- [ ] `npm test`/typecheck for `worksgood-pi`, `cargo fmt`, `cargo fmt --check`, `cargo clippy`, focused/full tests, `make embed-worksgood-pi` when applicable, `cargo install --path . --locked`, and owned smoke pass.

Depends on

done design-wg-pi-compaction-kick

Required by

open harden-wg-pi-compaction-kick

✉ Messages 6 messages (6 unread)

#1user2026-08-05T20:53:17.566193817+00:00delivered

Implementation seed only: current Pi 0.83.0 already includes PR #5115 queue draining, so do not patch that. The missing behavior is threshold compaction with an empty queue. Preserve manual/overflow/failed-compaction controls and re-embed plugin if changed.

#2implement-wg-pi-compaction-kick2026-08-06T08:29:58.209727274+00:00read

Acknowledged implementation seed — I will target only empty-queue successful threshold compaction, preserve manual/overflow/failed controls, and re-embed the plugin after source changes.

#3user2026-08-06T11:59:57.927819619+00:00read

Recovery context: commit 059c5468 is preserved; the nine dirty files in your exact worktree are the post-FLIP fixes for failed/aborted suppression, crash boundaries, settle persistence, replay refresh fail-closed, summary proof, and must-not-trigger coverage. Inspect/preserve them, complete validation, commit, and resubmit. Prior exit was schema infrastructure, now repaired.

#4implement-wg-pi-compaction-kick2026-08-06T12:48:05.259048888+00:00read

Acknowledged recovery context — preserving commits 059c5468/4022184b and all current post-review dirty fixes; I will audit, validate, commit, install, and run the owned live smoke.

#5implement-wg-pi-compaction-kick2026-08-06T13:25:14.245040100+00:00read

Final validation complete in 0f5cbe82; checking the task inbox before handoff as required.

#6implement-wg-pi-compaction-kick2026-08-06T13:25:44.702090088+00:00read

Done — final validation and the owned installed-flow smoke pass on commit 0f5cbe82; proceeding to the smoke-gated handoff.

Log

2026-08-05T20:52:42.527739893+00:00 Task paused
2026-08-05T20:52:42.761701010+00:00 Task published
2026-08-06T05:24:23.857053166+00:00 Operator repair: assignment action successfully selected Careful Programmer before its wrapper failed on missing completion candidate; removed the failed plumbing edge without hiding the failure.
2026-08-06T08:29:40.279801500+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-08-06T08:29:58.601060866+00:00 Starting implementation: inspect prior WIP, ratified design, Pi lifecycle/watchdog, plugin, and smoke harness; will build RED reproducer before production changes.
2026-08-06T08:30:46.871744642+00:00 RED-first reproducer confirmed before production changes: credential-free installed Pi trace ended compaction_end(threshold success willRetry=false) -> agent_settled with no recovery turn; exact RED assertion failed (exit 1). Capture: /tmp/pi-compaction-kick-red-agent13.
2026-08-06T08:34:04.470185629+00:00 Focused plugin regression added RED-first; npm test failed before production changes because pi-worksgood/continuation.js does not exist (exit 127).
2026-08-06T10:49:24.093140795+00:00 Implementation and decisive installed-binary smoke are green; completing final validation and commit.
2026-08-06T10:49:47.389120621+00:00 Validated: RED-first non-WG control remains inert; 36 Pi plugin tests pass; 23 focused watchdog/lifecycle tests pass; cargo check/build/fmt/clippy pass; installed Pi 0.83 smoke passed three consecutive times for one- and two-occurrence same-process flows. Full cargo test reached 3105 passes with 7 unrelated parallel profile-global-state failures, all 7 passed on serial focused rerun.
2026-08-06T10:56:53.810531942+00:00 Committed and pushed: 059c5468 — authoritative Pi threshold-compaction permit/delivery/effect protocol, compat 0.3.0 embedded bundle, and installed-binary single+double occurrence smoke.
2026-08-06T10:56:54.053188586+00:00 Final validation: npm worksgood-pi 36/36; cargo fmt/check/clippy/build; focused watchdog 23/23; lib suite 3112 passed, 36 ignored; embedded rebuild byte-identical; non-WG reproducer remains RED/inert; installed smoke passed three consecutive runs.
2026-08-06T10:59:10.367611465+00:00 Selected immutable completion candidate; prior review receipts invalidated
2026-08-06T10:59:46.123635358+00:00 Manifest b3:4d9ca9e681e03ad03d002cfc49621d10f526d87e5532e64d5a847bfe6b956ebd review outcome: FlipRejected
2026-08-06T11:28:55.859343450+00:00 Task marked as failed: Pi worker exited with code 1 before reviewed completion
2026-08-06T11:59:18.693717538+00:00 Task reset for retry from failed (attempt #2) — reason: Infrastructure/schema failure after FLIP; preserve 059c5468 and nine in-place files addressing all six review findings
2026-08-06T11:59:21.086060478+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-08-06T11:59:40.318268528+00:00 Resuming prior attempt: inspecting existing branch/WIP, design, and focused tests before changing production code.
2026-08-06T12:20:11.987027700+00:00 Recovered and committed prior review repairs as 4022184b: all six FLIP findings now have production fail-closed fixes or explicit tests; npm 44/44, focused watchdog 28/28, cargo check/fmt, and regenerated embedded bundle pass. Continuing audit against the full ratified design before installed smoke.
2026-08-06T12:47:30.955152834+00:00 Task fenced for retry from in-progress (attempt #2) — waiting for exact owner release — reason: Both recovered workers hit ENOSPC concurrently during full validation; caches were reaped and 134 GiB is free. Resume exact worktree alone to avoid duplicate full builds.
2026-08-06T12:47:36.355000929+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-08-06T12:48:05.603357756+00:00 Resumed exact preserved worktree after ENOSPC recovery. Auditing committed implementation plus current dirty crash/suppression/proof fixes before validation; staying inline because changes share lifecycle/plugin files and are already integrated.
2026-08-06T13:19:26.482273003+00:00 Recovered hardening is committed as 0f5cbe82. Fixed a live-smoke-discovered cancellation deadlock: the bounded terminal subscription is capability-authenticated then detached from the single-threaded daemon lane, remains read-only/un-journaled, is armed before send, and converges post-ack terminal abort state. Candidate-binary Pi 0.83 single+double installed-flow smoke now passes.
2026-08-06T13:20:52.789563923+00:00 Message-read completion probe; validation still in progress.
2026-08-06T13:24:47.104427649+00:00 Validated final commit 0f5cbe82: worksgood-pi 46/46 tests + typecheck; integration_pi_watchdog 29/29; cargo fmt/check, clippy, build; full cargo suite 3105 passed with only the known 7 parallel profile-global-state failures, all 8 matching tests pass serially; embedded rebuild byte-identical; cargo install --locked succeeded; installed global-wg + native Pi 0.83 owned smoke passed single and two-occurrence flows.
2026-08-06T13:24:48.349736765+00:00 Committed and pushed: 0f5cbe82 — cancellation subscription, host-version fail-close, settlement/abort convergence, bounded diagnostics, exact summary proof, and permanent installed-flow hardening.
2026-08-06T13:28:02.160684480+00:00 Selected immutable completion candidate; prior review receipts invalidated
2026-08-06T13:28:32.091641697+00:00 Manifest b3:c8aaca3a3d30e6671222eb671f381a82c909f465f7ebc61fe2044726d13c9187 review outcome: FlipRejected