ai-review: model requests intermittently hang ~27.6 min against the Anthropic gateway

[eduly-haldarwish]

Summary

Model requests through ANTHROPIC_BASE_URL (https://ai.eduly.com) intermittently hang on the first turn of a stage and return after ~27.6 minutes having done nothing. Six occurrences measured, across three repos and two model IDs.

This is not an ai-review defect. Three fixes were attempted inside the action and all three were falsified by measurement. Filing so it reaches whoever operates the gateway.

Fingerprint

{"turns":1,"costUsd":0,"numToolCalls":0,"durationMs":1654722,"dominantStage":"review"}

The stage log shows a clean init, then silence, then an error result with no body:

06:40:26  {"subtype":"init","model":"claude-sonnet-5"}
          ← 27m35s, zero output
07:08:01  {"subtype":"success","is_error":true,"duration_ms":1654722,
           "num_turns":1,"total_cost_usd":0,"permission_denials_count":0}

The six observations

Run	Repo	Stage	duration_ms
31317613715	github-actions	context (haiku-4-5)	1,647,262
lagn A	Lagn-App/lagn	context + review + retry	1,666,947
lagn B	Lagn-App/lagn	context + review + retry	1,666,947
31356469199	github-actions	context (haiku-4-5)	1,656,307
31360398002	github-actions	context (haiku-4-5)	1,658,923
31362834124	github-actions	review (sonnet-5)	1,654,722

Spread: 1.2%. A ceiling that tight across different repos, days, and model IDs is a configured timeout expiring, not variable slowness.

What has been ruled out

Hypothesis	Ruled out by
HTTP 529 / model overload	No 529 appears in any log. The original attribution came from a grep matching the branch name perf/ai-qa-hotfix-529
--fallback-model would recover it	Live during run 31317613715; stall happened anyway. A fallback triggers on an overloaded response — this gateway never answers
CLI per-request timeout × SDK retries (~10 min × 3)	Run 31356469199 stalled 27m36s with API_TIMEOUT_MS=180000 (3 min). A 3-min timeout cannot produce a 27.6-min stall
Specific to the context stage	Run 31362834124 stalled in the review stage with context skipped. The earlier pattern was selection bias — context was simply always the first stage
Specific to a model or model ID	Observed on both claude-haiku-4-5 and claude-sonnet-5
Specific to one repo	Observed in EdulyCom/github-actions and Lagn-App/lagn

Impact

• Six killed or wasted jobs, each burning ~28 min of runner time for zero output
• The lagn runs stalled on three stages serially — 83-minute jobs producing nothing
• Contributed to the 32 timeout kills measured across 682 consumer review jobs

What is already handled

ai-review fails closed: a stalled stage is continue-on-error, and Publish posts an explicit "inconclusive — re-run required" verdict rather than a false pass. Verified on run 31362834124 — Publish succeeded and the gate job was correctly skipped.

So this does not produce wrong verdicts. It wastes wall-clock and runner minutes, and it blocks CI until a re-run.

What would help from the gateway side

1. What happens to a request that sits ~27.6 min and returns zero bytes — is there a proxy/LB idle timeout near 1,650,000 ms?
2. Are these requests reaching the upstream API at all, or terminating at the gateway?
3. Anything correlating with the trigger — concurrency, cold start, connection reuse?

Reproduction

Intermittent — roughly 6 occurrences in ~25 runs today. Not deterministic. Any ai-review run can hit it; the signature to grep for is ai-review-metrics with "turns":1 and "costUsd":0.