reproduce-pi-threshold-compaction-stall

Metadata

Status	done
Agent identity	`02e879681e52e0a384106169be043416c4d946e850ab26b2269c57681b52a6e7`
Created	2026-08-05T20:52:42.199180862+00:00
Started	2026-08-06T05:24:25.293352780+00:00
Completed	2026-08-06T06:13:00.165144711+00:00
Tags	`pi`, `compaction`, `regression`, `research`
Tokens	16463075 in / 32153 out

Description

Objective

Produce a falsifiable, credential-free reproducer for the specific Pi behavior reported as upstream issue earendil-works/pi#6424: a long agentic turn reaches threshold auto-compaction, compaction succeeds, the compacted context still records unfinished work, no steering/follow-up is queued, and Pi emits agent_settled/exits without taking another concrete turn.

Use the installed @earendil-works/pi-coding-agent version and its real JSON event/session semantics. Distinguish this from overflow compaction (willRetry=true), compaction failure, manual /compact, and the independent bug where an extension queues a follow-up during agent_end. Record exact event ordering and session entries. Prefer a deterministic fake provider or upstream-compatible fixture; no paid credential may be required. Do not implement the fix.

Write docs/research/pi-threshold-compaction-stall-reproducer.md with commands, fixture design, expected/actual event trace, installed version, and the exact red assertion the implementation task should inherit. Any fixture files must live under a uniquely named tests/fixtures/fake-pi-compaction-stall/ path.

Validation

A scripted invocation exercises Pi JSON mode and produces compaction_start(reason=threshold), successful compaction_end(willRetry=false), then idle/settled without a subsequent recovery turn.
The compacted summary/session evidence contains an explicit unfinished next step, so the test does not infer incompletion from prose such as “continue”.
Control cases prove overflow retry continues automatically and manual/failed compaction is not the same condition.
The reproducer fails only on the missing post-threshold continuation behavior and requires no network credential.
git diff --check passes and all commands/output needed to rerun the reproducer are captured in the report.

## Objective
Produce a falsifiable, credential-free reproducer for the specific Pi behavior reported as upstream issue `earendil-works/pi#6424`: a long agentic turn reaches threshold auto-compaction, compaction succeeds, the compacted context still records unfinished work, no steering/follow-up is queued, and Pi emits `agent_settled`/exits without taking another concrete turn.

Use the installed `@earendil-works/pi-coding-agent` version and its real JSON event/session semantics. Distinguish this from overflow compaction (`willRetry=true`), compaction failure, manual `/compact`, and the independent bug where an extension queues a follow-up during `agent_end`. Record exact event ordering and session entries. Prefer a deterministic fake provider or upstream-compatible fixture; no paid credential may be required. Do not implement the fix.

Write `docs/research/pi-threshold-compaction-stall-reproducer.md` with commands, fixture design, expected/actual event trace, installed version, and the exact red assertion the implementation task should inherit. Any fixture files must live under a uniquely named `tests/fixtures/fake-pi-compaction-stall/` path.

## Validation
- [ ] A scripted invocation exercises Pi JSON mode and produces `compaction_start(reason=threshold)`, successful `compaction_end(willRetry=false)`, then idle/settled without a subsequent recovery turn.
- [ ] The compacted summary/session evidence contains an explicit unfinished next step, so the test does not infer incompletion from prose such as “continue”.
- [ ] Control cases prove overflow retry continues automatically and manual/failed compaction is not the same condition.
- [ ] The reproducer fails only on the missing post-threshold continuation behavior and requires no network credential.
- [ ] `git diff --check` passes and all commands/output needed to rerun the reproducer are captured in the report.

Depends on

(none)

Required by

done design-wg-pi-compaction-kick

✉ Messages 4 messages (4 unread)

#1user2026-08-05T20:53:16.713331553+00:00delivered

Initial evidence: upstream issue https://github.com/earendil-works/pi/issues/6424 gives the exact threshold-compaction idle case and a local workaround. Installed Pi is 0.83.0. Installed dist/core/agent-session.js lines ~1583-1585 triggers threshold compaction; ~1698 emits compaction_end; ~1707-1709 continues only when agent.hasQueuedMessages(). Capture this exact empty-queue path and distinguish willRetry overflow.

#2reproduce-pi-threshold-compaction-stall2026-08-06T05:51:40.664152570+00:00read

Done — initial evidence incorporated and the credential-free Pi 0.83.0 reproducer/report are in commit 1f8cea28.

#3reproduce-pi-threshold-compaction-stall2026-08-06T06:06:23.735373318+00:00read

Completion is reviewed and landed at main 1b2ca643, but my first premature wg done (before submission) cached a deterministic 'missing completion candidate' response for this attempt; later wg done replays it even though graph now has accepted FLIP+eval receipts. Operator repair/retry of the DoneHandoff request is needed.

#4reproduce-pi-threshold-compaction-stall2026-08-06T06:07:52.897282887+00:00sent

Final delivered state is commit 1b2ca643 on main (reviewed candidate b3:0ee0e70c... accepted and landed).  is still replaying cached request intent:4e5820... from the pre-candidate attempt; no code/review issue remains.

Log

2026-08-05T20:52:42.154374490+00:00 Task paused
2026-08-05T20:52:42.761699210+00:00 Task published
2026-08-06T05:24:22.935461414+00:00 Operator repair: removed failed quality-pass dependency after reviewing the gate artifact/log. The gate confirmed the batch satisfied all requested quality criteria but could not complete because task-scoped worker capability forbade cross-task wg show/edit. Failure remains visible on .quality-pass-pi-compaction-recovery.
2026-08-06T05:24:23.671908887+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.
2026-08-06T05:24:25.600914281+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-08-06T05:27:41.768651275+00:00 Inspected Pi 0.83.0 docs and installed agent-session implementation; building a real custom-provider JSON/RPC fixture with threshold, overflow, manual, failure, and agent_end-follow-up cases.
2026-08-06T05:33:26.775480357+00:00 Fixture now reproduces Pi 0.83.0: threshold success -> agent_settled with no recovery; overflow/manual/failure/agent_end-follow-up controls pass; default exact assertion is red.
2026-08-06T05:51:00.501953453+00:00 Validated task criteria: capture-only invocation passes all controls and observes threshold compaction_start -> successful compaction_end(willRetry=false) -> agent_settled; persisted summary/details explicitly mark next action unexecuted; default invocation exits 1 only at exact red continuation assertion; git diff --cached --check passes.
2026-08-06T05:51:00.830236231+00:00 Validation: cargo build passes. Full cargo test is not clean on this branch/environment (unrelated existing route/worker-authority/UI expectation failures); targeted rerun of the first flaky library failure passes. Fixture itself was rerun successfully credential-free.
2026-08-06T05:51:18.404687899+00:00 Committed: 1f8cea28 — pushed to remote
2026-08-06T05:51:54.618881032+00:00 Completion check: final inbox read was attempted again but worker-control returned 'No response from service'; sent completion reply successfully as message_id=2. Working tree is clean and commit is pushed.
2026-08-06T05:53:34.497903728+00:00 Selected immutable completion candidate; prior review receipts invalidated
2026-08-06T05:53:59.362268056+00:00 Manifest b3:bf1107827a7581977d3c8ee62009ca2a0e86214f16e262856a99ec3afbd4e878 review outcome: FlipRejected
2026-08-06T05:56:39.573211079+00:00 Selected immutable completion candidate; prior review receipts invalidated
2026-08-06T05:58:01.898928632+00:00 Manifest b3:52855d24d418d8d78020d497587c1d0578f2fbd324d4e1ab40b7bd852fe79bad review outcome: Accepted
2026-08-06T05:58:25.216544749+00:00 Addressed FLIP feedback: child env now credential allowlisted; immutable validation evidence includes captured stdout/status, raw JSON events, raw session JSONL, and git diff --check=0. Revised candidate b3:52855d24... accepted by FLIP and eval.
2026-08-06T05:58:38.989018717+00:00 Selected immutable completion candidate; prior review receipts invalidated
2026-08-06T05:59:12.496285104+00:00 Manifest b3:52855d24d418d8d78020d497587c1d0578f2fbd324d4e1ab40b7bd852fe79bad review outcome: FlipRejected
2026-08-06T06:02:22.983839873+00:00 Selected immutable completion candidate; prior review receipts invalidated
2026-08-06T06:03:29.415214075+00:00 Manifest b3:0ee0e70c376294fd273de43afbe50b0d198400b2675d807bca61e88f88b9811e review outcome: Accepted
2026-08-06T06:05:46.026363264+00:00 Reviewed manifest b3:0ee0e70c376294fd273de43afbe50b0d198400b2675d807bca61e88f88b9811e published to integration ref
2026-08-06T06:06:24.263854512+00:00 Completion candidate b3:0ee0e70c... accepted by FLIP+eval and wg land published main=1b2ca643. wg done remains blocked by replay of this attempt's earlier pre-candidate cached 'missing completion candidate' response; candidate is present in graph. Requested operator repair.
2026-08-06T06:13:00.165144711+00:00 Done derived from exact reviewed manifest b3:0ee0e70c376294fd273de43afbe50b0d198400b2675d807bca61e88f88b9811e and contract publication