diff --git a/agents/duckworth/work/row269/REVIEW-design-early.md b/agents/duckworth/work/row269/REVIEW-design-early.md index 74a39ade6..8574c6670 100644 --- a/agents/duckworth/work/row269/REVIEW-design-early.md +++ b/agents/duckworth/work/row269/REVIEW-design-early.md @@ -141,3 +141,16 @@ DEFERRED.md:474-480 gets its closing line when 269 lands. - N1, doc only: a survivor still falls through to bindLaunch and `endRun(host-lost)` while its runner lives. That's unchanged from aca16404c, where a failed kill also fell through. So the broker sees the run ended and its claim free, and only the launcher's Q2 check holds the instance. That check reads only sessionless survivors, and only for threaded starts. Say this in BUILD.md's Q2 section, so the claim release isn't read as proof the run exited. - N2: recover() now also waits up to 5s for the runner after the bwrap kill. With `--die-with-parent` that should be short, but a host start can take up to 5s per leftover whose runner lingers. - At the round: M10-M12 kill rules against this test, and the refusal read from the event (`refused` with thread A). + +## recover() change, design read (duckworth-0159, sage-2750) + +- Input: gizmoduck-0247. BUILD.md section "recover() change, for a design read" at local commit 922178622. Since 3d29876df the code diff is the survivor log line (launcher.mjs:1028, which now names both pids) and the matching test literal. I read it with git against aca16404c:947-976 and ran nothing. +- Verdict: accepted, no blockers. A CLEAR can cover it. The fallback isn't needed. +- What I checked: + - **Every leftover is still ended.** The survivor push (:1026-1029) comes before bindLaunch, with no `continue`, so the bind and the host-lost end (:1030-1040) run as in the base. The kill condition (:1018) is the base's. + - **The wait.** The longer wait sits inside the bwrap kill branch only, so a gone outer pid takes no wait. That's the test's case. + - **launcher.test.mjs:917-918.** These assert the outer pid and the runner are gone right after setup. Waiting on the runner can only make :918 hold sooner. A survivor log there would mean :917-918 already fail, so no old expected value changes. + - **stack-release.test.mjs:1928.** It writes `{pid: process.pid}` from the stand-in host script and has no start times. That file is never read by this recover(), and liveRun is false for it anyway. + - **Log-pinning tests.** launcher.test.mjs:923/:941/:964 use `includes`. None of them pins the survivor line. +- N3, doc only: a sessioned survivor is kept but counts nowhere, because :309 filters `!s.session`. The same session could start again while the old runner lives. That's the base's exposure, unchanged, since the base also ended the run and let the runner go unseen. Name it as a residual in the Q2 section, next to my 0158 N1 (the claim release isn't proof the run exited). +- N4, doc only: the section's last bullet says the refusedOnce edit waits for Mos under "mos-1290 term 5". mos-1292 (sage-2747) has already allowed it on mos-1285's terms, and a81b3446e holds it. Fix the bullet so the record doesn't show it as pending. diff --git a/docs/SESSIONS.md b/docs/SESSIONS.md index 38c968f38..bae182c2e 100644 --- a/docs/SESSIONS.md +++ b/docs/SESSIONS.md @@ -1493,3 +1493,4 @@ are never rewritten or removed; corrections are new entries. 2026-10-22T15:07Z | Duckworth (Opus 5.5) | row 253 M21 static lead (0155) and row 269 legacy-run-live early warning (0156) | lead sent to Sage for Mos; no-test is a code-round blocker | 2026-10-22T15:09Z | Duckworth (Opus 5.5) | mos-1292 rulings recorded: row 253 M21 path, row 269 refusedOnce terms (sage-2747) | recorded, no verdict | 2026-10-22T15:14Z | Duckworth (Opus 5.5) | row 269 legacy-run-live early read on local 3d29876df (duckworth-0158) | no blockers, 0156 blocker answered pending round, N1 survivor still ended host-lost, N2 5s runner wait | +2026-10-22T15:14Z | Duckworth (Opus 5.5) | row 269 recover() design read on local 922178622 (duckworth-0159, sage-2750) | accepted, no blockers; N3 sessioned survivor residual, N4 stale mos-1290 bullet |