Skip to content

[aw-failures] P1: Daily BYOK Ollama Test — every request returns transient_bad_request, all 4 attempts fail (6-day 100% outage) #39476

Description

@github-actions

Recommendation

Fix the Ollama BYOK request path — every inference call returns transient_bad_request, taking this test red for 6+ consecutive days. The BYOK Ollama endpoint is rejecting the request shape (or the backend is down), so the agent never produces a turn.

Problem statement

The Daily BYOK Ollama Test agent job fails on every run. The Copilot CLI spawns, issues a model request to the Ollama BYOK backend, and the request immediately fails with Request failed (transient_bad_request). The harness retries with --continue across all 4 attempts; each dies in ~3s. Result: turns=0, exitCode=1, no agent work performed.

Affected workflow and run IDs

  • Workflow: Daily BYOK Ollama Test (.github/workflows/daily-byok-ollama-test.lock.yml)
  • Representative failed run: §27581947650 (2026-06-15)
  • Comparator (prior failed, same signature): §27447521382 (2026-06-12)
  • Chronic: failure on 2026-06-08, 06-09, 06-10, 06-11, 06-12, 06-15 (every scheduled run sampled).

Probable root cause

Requests to the Ollama BYOK backend are rejected before any tokens are generated. Two correlated signals: (1) all 4 agent attempts fail in ~3s with transient_bad_request (a 400-class rejection, not a timeout), and (2) the post-run awf-reflect models probe to (apiproxy/redacted) returns 503after 5 retries. Together these indicate the Ollama model/endpoint behind the BYOK api-proxy (port 10002) is unavailable or the request payload (model name / API shape) is incompatible with the backend. NoteisModelNotSupportedError=falseandisCAPIError400=false`, so the harness does not currently classify this as a model/credit error — it is silently treated as a generic transient failure.

Proposed remediation

  1. Verify the Ollama BYOK backend reachable behind api-proxy:10002 is healthy and serving the configured model; the 503 on /v1/models suggests the upstream Ollama service is down or unconfigured in CI.
  2. If the backend is intentionally external/flaky, classify transient_bad_request + /v1/models 503 as an infrastructure-unavailable condition and emit a non-failing noop/skip rather than a hard failure, so a missing BYOK backend does not generate daily false-failure noise.
  3. Add a pre-flight model-availability check (probe /v1/models before spawning the agent) and short-circuit with a clear diagnostic when the BYOK endpoint is unavailable.

Success criteria / verification

  • A scheduled run either (a) completes with the agent producing ≥1 turn against a healthy Ollama backend, or (b) cleanly skips with a non-failure conclusion and a logged reason when the backend is unavailable.
  • No transient_bad_request-only runs marked failure for 3 consecutive scheduled runs.

Analyzed run IDs: 27581947650, 27447521382. Parent: #29109.
Related to #29109

Generated by 🔍 [aw] Failure Investigator (6h) · 183.4 AIC · ⌖ 12.6 AIC · ⊞ 4.5K ·

  • expires on Jun 22, 2026, 5:50 PM UTC-08:00

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions