daily-fix: healthcheck auto-launches /daily backfill

Workflow gap

scripts/cron_daily_healthcheck.sh (06:00 PT, task #711 heartbeat + #1189 husk arm) detects a missing or never-enriched (husk) nightly /daily file and ALERTS with the exact backfill command in the Telegram message — but never runs it. Recovery depends on a human copy-pasting the command the same day. When nobody does, the day's problem sweep is permanently skipped and the sentinel (sent-<date>.flag) suppresses any re-alert, so the gap goes silent after one buzz.

Live incident (2026-08-05): the 2026-08-04 nightly (session start 2026-08-05T06:27Z) died at 07:03:25Z on repeated API 529 Overloaded errors ~36 min in — stub + miner brief written, nothing else. The 06:00 healthcheck correctly detected the husk and pushed the alert (sent-2026-08-04.flag written, per logs/daily_healthcheck/2026-08-05.log), but no backfill ran all day; the 2026-08-05 nightly found the 08-04 file still an all-empty husk ~24 h later, and the 08-04 transcript window is still unmined.

verified-at-filing: grep -n 'claude -p' scripts/cron_daily_healthcheck.sh → 2 hits, both inside alert MESSAGE strings (lines 88, 90) — no execution path; grep -c 'backfill' → 2, message-only. Run at 2026-08-06T07:00Z on main. The 529 death is read from the dead session transcript (~/.claude/projects/-home-thomasjiralerspong-explore-persona-space/970a1dd6-aad6-425a-b99d-d0350698db33.jsonl, final assistant row 2026-08-05T07:03:25Z: "API Error: Repeated 529 Overloaded errors").

Proposed change

On a missing/husk detection, the healthcheck ADDITIONALLY launches the backfill it already names, detached and single-flight, keeping the alert:

• cd $PROJECT_DIR && CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS=10800000 $HOME/.local/bin/claude -p "/daily $YESTERDAY", detached (setsid + log breadcrumb under logs/daily_healthcheck/backfill-<date>.log), guarded by a per-date attempt sentinel (one auto-attempt per missed day; a failed auto-backfill re-alerts rather than looping) and a flock (never two /daily processes).
• Timing is safe by construction: 06:00 PT is far outside the skill's "no backfill within 60 min of the 23:27 nightly" rule, and the skill's Edit-in-place husk-recovery branch is the designed re-entry (.claude/skills/daily/SKILL.md § Output).
• The alert message changes from "run this command" to "auto-backfill launched (attempt 1); log: " — Thomas keeps visibility, loses the manual step.

Planner should decide the failure policy (e.g. if the auto-backfill itself dies, the next day's healthcheck sees the still-husk file but the sentinel must not suppress a second attempt + re-alert) and whether the mygoat-side retrospective log check stays.

Provenance

• fingerprint: 83734a70935b

• workflow_fix_target: scripts/cron_daily_healthcheck.sh

• origin: /daily 2026-08-05 nightly run — orchestrator observation while investigating the 2026-08-04 husk (529-killed nightly, session 970a1dd6, died 2026-08-05T07:03:25Z).