Its AIOS setup tested in parts (wiki ingest, query, lint, audits)

Run was interrupted at ~18:03 by a usage limit and resumed. State found on resume (from files in outputs/):

  • A main (48 runs): DONE (outputs/a/results-main.jsonl, raw/V0_*, raw/V1_*).
  • A3 staleness: NOT started (no outputs/a/results-stale.jsonl).
  • B ingest: DONE (outputs/b/wiki, outputs/logs/b_ingest.log).
  • B query: W and R arms DONE (36 each); L arm partial (Q1-Q9 trial 1 done; Q9 t2, Q10-Q12 missing; crashed with "retries exhausted").
  • B lint: NOT done (results-lint.jsonl empty, crash "retries exhausted"). No lint grader script exists yet.
  • C1: 2 of 6 runs saved (grill_1, control_0); results-c1.jsonl empty (crash).
  • C2: crashed in the first baseline audit (1 raw audit saved, start_2).
  • C3: repo HTTP status done (outputs/logs/c3-repo-status.txt); table not written.

Resume log (new entries below)

Resume session

  • Cause of earlier crashes (outputs/logs/b_query.log, b_lint.log, c_grill.log, c_audit.log all end "retries exhausted"): the Gemini endpoint returns intermittent empty-body HTTP 404 and HTTP 429 (probes at resume: gemini-3.5-flash returned 200,429,429,200,429,429 on 6 consecutive identical calls). The original client gave up after 12 tries.
  • Fix (harness only, original kept as outputs/harness/llm_original_pre_resume.py): llm.py now retries 404-with-empty-body, 429, 5xx up to 40 times with capped backoff. run_a.py, b_query.py, c_grill.py, c_audit.py skip any run whose raw file already exists. New result jsonl files are written as *-resumed.jsonl; earlier outputs are untouched. outputs/summary.json is rebuilt from raw/ by harness/aggregate.py.
  • Launched in background (3 chains, one per model quota): A stale (gemini-3.1-flash-lite), B L-arm remainder then B lint (gemini-3.5-flash-lite), C1 then C2 (gemini-3.5-flash). Concurrency: 3 per chain max, under the 3 concurrent claude -p / 1 local model limits (no claude -p or local model used).
  • Status: A main DONE; A3 stale IN PROGRESS; B L-arm remainder IN PROGRESS; B lint NOT STARTED (queued); C1 IN PROGRESS (4 runs left); C2 queued; C3 table to write.
  • 2026-10-03 resume: ALL Gemini routes dead. v1 API for gemini-3.5-flash returns 429 "limit: 20 ... GenerateRequestsPerDayPerProjectPerModel-FreeTier ... retry in 5h28m"; other models return empty-body 404 on both v1/v1beta. Background Gemini jobs killed by PID (A stale produced 9 files V0-moved T1..., kept; they are Gemini, not mixed in with the Haiku stale arm). Alternative route chosen: claude -p (Haiku 4.5 worker, Sonnet helper/judge) via outputs/harness/cc.py. Details in insights/DEVIATIONS.md.