What happened
Starting 2026-08-27, every retro workflow run in quay/.fullsend failed immediately with: The model claude-sonnet-4-6 is not available on your vertex deployment. Try --model to switch to claude-sonnet-4-5@20250929, or ask your admin to enable this model. Thirteen consecutive failures occurred through 2026-08-31 (representative run: 33208673719). The harness fetches harness/retro.yaml and agents/retro.md from fullsend-ai/agents@v0.37.0. Each run attempts two retries, both fail identically — zero tokens consumed, no agent-result.json produced. The triggered retro for PR #6876 itself (run 33418071633) was in-progress at analysis time and may also be affected.
What could go better
The retro harness configures claude-sonnet-4-6, which is not enabled on the Vertex deployment used by the runner. The error message itself identifies a valid alternative (claude-sonnet-4-5@20250929). Model unavailability is not a transient error, so retrying is wasted time. Confidence is very high — the error is explicit, consistent across 13 runs, and the suggested fallback model is named in the error text. This is a distinct issue from fullsend-ai/fullsend#6795 (opus alias, review agent) and #6666 (cost inflation) because it affects the retro agent on a different model, and no open issue currently covers it.
Proposed change
In fullsend-ai/agents, update harness/retro.yaml to replace the claude-sonnet-4-6 model specification with a model confirmed available on the Vertex deployment — the error suggests claude-sonnet-4-5@20250929. Before updating, verify the target model appears in the Vertex model allowlist for the quay org. If model-fallback logic is added to the harness runner (as proposed in fullsend-ai/fullsend#1771), this could also be resolved there — but the immediate fix is to pin a working model version in harness/retro.yaml.
Validation criteria
After the update, the next 3 retro workflow runs triggered for quay/quay should complete without a model-unavailability error and each produce a valid output/agent-result.json. The runs should show non-zero token counts in the logs.
Generated by retro agent from quay/quay#6876
What happened
Starting 2026-08-27, every retro workflow run in quay/.fullsend failed immediately with:
The model claude-sonnet-4-6 is not available on your vertex deployment. Try --model to switch to claude-sonnet-4-5@20250929, or ask your admin to enable this model.Thirteen consecutive failures occurred through 2026-08-31 (representative run: 33208673719). The harness fetchesharness/retro.yamlandagents/retro.mdfromfullsend-ai/agents@v0.37.0. Each run attempts two retries, both fail identically — zero tokens consumed, noagent-result.jsonproduced. The triggered retro for PR #6876 itself (run 33418071633) was in-progress at analysis time and may also be affected.What could go better
The retro harness configures
claude-sonnet-4-6, which is not enabled on the Vertex deployment used by the runner. The error message itself identifies a valid alternative (claude-sonnet-4-5@20250929). Model unavailability is not a transient error, so retrying is wasted time. Confidence is very high — the error is explicit, consistent across 13 runs, and the suggested fallback model is named in the error text. This is a distinct issue from fullsend-ai/fullsend#6795 (opus alias, review agent) and #6666 (cost inflation) because it affects the retro agent on a different model, and no open issue currently covers it.Proposed change
In
fullsend-ai/agents, updateharness/retro.yamlto replace theclaude-sonnet-4-6model specification with a model confirmed available on the Vertex deployment — the error suggestsclaude-sonnet-4-5@20250929. Before updating, verify the target model appears in the Vertex model allowlist for thequayorg. If model-fallback logic is added to the harness runner (as proposed in fullsend-ai/fullsend#1771), this could also be resolved there — but the immediate fix is to pin a working model version inharness/retro.yaml.Validation criteria
After the update, the next 3 retro workflow runs triggered for quay/quay should complete without a model-unavailability error and each produce a valid
output/agent-result.json. The runs should show non-zero token counts in the logs.Generated by retro agent from quay/quay#6876