operator
14 lines
2.1 KiB
Plaintext
14 lines
2.1 KiB
Plaintext
# 081-completion-funnel: Completion funnel triage
|
|
|
|
Goal: Find worst completion funnel family and fix or specify remediation
|
|
|
|
Steps:
|
|
1. Read latest completion-audit JSON report under logs/ and logs/completion-audit-state.json. 2. Identify worst job family by stall/degrade rate. 3. Fix root cause if small and safe, else write exact remediation steps. Done criteria: result notes name worst family with numbers plus fix commit description or precise remediation.
|
|
|
|
Done criteria: result notes appended below; file moved to done/.
|
|
|
|
Result notes (append below before moving to done/):
|
|
|
|
Completed 2026-10-08T05:41:53Z via box tasks done:
|
|
WORST FAMILY: 'muse tasks' — 10 results / 10 fail / 0 ok in 24h window (100pct fail; stable across last 8 reports). NUMBERS: totals results=636 ok=618 fail=18; 'muse tasks' = 10 of 18 fails. Linked: auto-work-queue-f03/f07/f11/f15/f19 (muse scavengers) sent+dispatched 12-13 each, results 0 — all 5 in silent_families (101 silent total). Swarm slots failed=216 static (not growing; done 161->166), stale_running empty. ROOT CAUSE (traced, not fixed): muse posts plain-text DECLINEs in 'muse tasks' thread 3881dae3 ('Input-wait to sweep/scavenge auto-work queue — not an authorized timer. Declining.', hourly 17:06 Oct7->04:34 Oct8). Sweeps dispatch to dedicated sidechats with 30m followups; at harvest no pending followup matches, so response-harvester.py:966-978 falls back to job-sidechats.json alias 'muse tasks' as job_id, logs success=False. One dispute counted twice: bogus 100pct-fail family + 5 silent families. NO CODE CHANGE (needs operator policy call). REMEDIATION: (1) Operator decides: authorize queue sweeps for muse OR stop dispatching f03/f07/f11/f15/f19 to muse (box job archive each, or retarget agent in jobs/auto-work-queue-f*.json). (2) Optional metric fix: harvester synthesized-DECLINE path logs outcome=declined; completion-audit.py compute_funnel buckets DECLINE snippets as declined not fail; add tests, suite green. (3) Verify: next audit shows 'muse tasks' fails->0 and queue-f families flowing or gone from silent list. NOTE: first claim was swept mid-work; re-claimed to file.
|