This was generated by AI during triage.
Summary
fetchPreviousRun (the CI baseline fetch, src/reporters/site-api.ts / driven from src/main.ts) can hang indefinitely. Its 10s AbortSignal.timeout(10000) does not survive a half-open socket during body read, so a transient Site API stall becomes a full CI-job hang that only ends at the job's 20-minute timeout.
Evidence
Observed in GitHub Actions run 28626422344 (repo veksen-enterprises/d2armory, PR #907, test job 84893821211, attempt 1). Tests passed in ~70s; the job then hung on the final query-doctor/analyzer@main step and was killed by the job's timeout-minutes: 20 ("The operation was canceled" at 23:08:21, exactly 20 min after job start 22:48:09).
Log trace:
22:50:24.09 Site API ingestion successful: {"id":"019f2506-..."} # POST /ci/runs = HTTP 200, ~1s
22:50:24.09 Fetching previous run from https://api.querydoctor.com/ci/runs/latest?repo=veksen-enterprises%2Fd2armory&branch=main&excludeId=...
22:50:25.85 [main] API connection broken during CI run: Error: RPC session was shut down by disposing the main stub
...total silence until the 23:08:21 kill (~18 min)
Key point: none of fetchPreviousRun's own failure/retry log lines appeared — no Failed to fetch previous run, no Retrying baseline fetch (attempt 2/3). So the 10s timeout never took effect. Execution was sitting in await response.json() on the multi-MB /ci/runs/latest response after headers had already arrived, and the abort did not interrupt the stalled body read once the server (capnweb / Cloudflare-Workers "main stub") was disposed mid-stream and left a half-open TCP socket (no FIN/RST).
The ingest POST succeeded first — this is not the dup-hash ingest 500.
Fix
Enforce a hard wall-clock deadline covering the entire fetch + parse, not just the initial request:
Promise.race([fetchAndParse(), rejectAfter(deadlineMs)]), or a total-deadline AbortSignal that also aborts body consumption, so await response.json() on a half-open socket is guaranteed to reject.- On deadline, fall through to the existing
{ kind: "error" } → comparisonUnavailable path ("comparison temporarily unavailable — re-run") instead of hanging.
Result: a stalled baseline fetch degrades in ~10s instead of hanging 18+ minutes.
Severity
Medium. Not currently firing (the run self-resolved on re-run with no client change, consistent with a Site API restart), but it recurs on any api.querydoctor.com blip during the baseline fetch, and each recurrence is a 20-minute hung job + red PR check.
Related
- Root-cause counterpart (server side): Query-Doctor/Site — RPC "main stub" disposal leaving a half-open response socket.
- Optional CI guardrail: veksen-enterprises/d2armory — step-level timeout on the analyzer step.
Summary
fetchPreviousRun(the CI baseline fetch,src/reporters/site-api.ts/ driven fromsrc/main.ts) can hang indefinitely. Its 10sAbortSignal.timeout(10000)does not survive a half-open socket during body read, so a transient Site API stall becomes a full CI-job hang that only ends at the job's 20-minute timeout.Evidence
Observed in GitHub Actions run
28626422344(repoveksen-enterprises/d2armory, PR #907,testjob84893821211, attempt 1). Tests passed in ~70s; the job then hung on the finalquery-doctor/analyzer@mainstep and was killed by the job'stimeout-minutes: 20("The operation was canceled" at 23:08:21, exactly 20 min after job start 22:48:09).Log trace:
Key point: none of
fetchPreviousRun's own failure/retry log lines appeared — noFailed to fetch previous run, noRetrying baseline fetch (attempt 2/3). So the 10s timeout never took effect. Execution was sitting inawait response.json()on the multi-MB/ci/runs/latestresponse after headers had already arrived, and the abort did not interrupt the stalled body read once the server (capnweb / Cloudflare-Workers "main stub") was disposed mid-stream and left a half-open TCP socket (no FIN/RST).The ingest POST succeeded first — this is not the dup-hash ingest 500.
Fix
Enforce a hard wall-clock deadline covering the entire fetch + parse, not just the initial request:
Promise.race([fetchAndParse(), rejectAfter(deadlineMs)]), or a total-deadlineAbortSignalthat also aborts body consumption, soawait response.json()on a half-open socket is guaranteed to reject.{ kind: "error" }→comparisonUnavailablepath ("comparison temporarily unavailable — re-run") instead of hanging.Result: a stalled baseline fetch degrades in ~10s instead of hanging 18+ minutes.
Severity
Medium. Not currently firing (the run self-resolved on re-run with no client change, consistent with a Site API restart), but it recurs on any
api.querydoctor.comblip during the baseline fetch, and each recurrence is a 20-minute hung job + red PR check.Related