daily-fix: ownership-probe hardening (3 probed gaps)

Workflow gap

Three same-day incidents share one root: an orchestrator acted on a stale liveness/state read where a cheap probe of durable + process state would have prevented it.

1. Pre-stop probe missing (session kill with live implementer). 2026-08-05 15:21–15:52Z, task #2054: on a "make sure it's properly running" ask, the fleet-ops orchestrator concluded the #2054 session's r9 dispatch "never happened" (marker v79 read as phantom), stopped the session at 15:32:46Z — killing implementer impl-2054-r9, live since 15:22:38Z — then had to post TWO correction markers ("the r9 dispatch never happened. Ownership transferred." → retracted; "the idle figure was overstated"). Evidence: #2054 events (deliberate-stop 15:32:46Z, corrections 15:33:07Z/15:38:52Z, probed) + transcript 291f5901 row 304 ("was stopped by user" kill of the in-flight agent). A stage-dispatch marker was 10 minutes old at stop time.
2. Detached-children probe missing after a subagent death. 2026-08-05 ~10:52–10:59Z, task #1491: a 529-killed analyzer had already launched a detached cap-hit run; its death notification carried no mention of it and the orchestrator "did not probe for it", launched its own run into the same worktree, and the scratch-tree collision killed the duplicate (incident marker epm:progress v55, self-diagnosed: "exactly the ownership check CLAUDE.md mandates… and then skipped it for a VM-local job").
3. Ops statistics carried forward stale. 2026-08-05 07:13–07:16Z: the fleet-ops session quoted a stale cumulative failure ratio and misattributed the thrash fix; Thomas corrected twice ("but they're not thrashing they're overloaded right?", "no i think we fixed it by changing some setting related to max context"). The compose-time re-grep rule (CLAUDE.md § ad-hoc results summaries) covers experiment numbers but ops/ fleet statistics slipped through.

verified-at-filing: incidents 1–2 are probed from task events + the sessions' own incident/correction markers (kinds+ts cited above, read at compose time via task.py view); incident 3 is a transcript full-text read (session 4966e56e rows 1334–1382). grep -n "Ownership check before any resume/launch" CLAUDE.md → 1 hit (the bullet exists; none of the three sub-cases is currently named in it).

Proposed change

Extend the CLAUDE.md § "Orchestrator vs subagent re-invocation" ownership-check bullet (and mirror where the stop paths live) with the three probed sub-cases:

• Before stopping a session (operator stop, PM stop, watcher-adjacent manual stop): probe for live subagent children AND a stage-dispatch marker younger than ~10 min — either one ⇒ presume live, do not stop on an idle-looking self-report.
• After ANY subagent death (529/refusal/kill): probe for detached children it may have launched (pgrep -af on the job's distinctive invocation, bracketed pattern) BEFORE launching same-path work — a death notification never enumerates detached children.
• Ops/fleet statistics (failure rates, fixed-vs-not claims) in monitoring turns are recomputed from events.jsonl/live state in the same turn, never carried from a prior cycle — the ad-hoc-summaries compose-time re-grep rule extended to ops stats.

Planner decides exact placement (single bullet extension vs a short rules file) — CLAUDE.md is under active size-ratchet compaction, so the wording must be tight.

Provenance

• fingerprint: a571ee198618

• workflow_fix_target: CLAUDE.md

• origin: /daily 2026-08-05 problem sweep — miners 1 (P4/P5), 3 (P3), 5 (P3), 2 (P6); incidents independently mined from four transcripts.