[kadel]

Upgrade Fullsend to v0.38.0 and pin model selection for the repository-customized agents.

Assisted-by: Codex

[fullsend-ai-review]

🤖 Finished Review · ❌ Failure · Started 11:37 AM UTC · Completed 11:39 AM UTC

Commit: 3040066 · View workflow run →

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 11:51 AM UTC · Completed 12:02 PM UTC

Commit: 3040066 · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $4.22

[fullsend-ai-retro]

Retro: PR fullsend-ai/fullsend#6 — Pin Fullsend agents to Claude Opus 4.8

What happened

PR fullsend-ai/fullsend#6 was a human-authored config change by @kadel to pin all four repository-customized Fullsend agents (code, fix, review, e2e-triage) to claude-opus-4-8 and upgrade the Fullsend reusable workflow from v0.36.0 to v0.38.0. The changes touched two files: .fullsend/config.yaml (added model: claude-opus-4-8 to each agent) and .github/workflows/fullsend.yaml (bumped the workflow ref).

Timeline:

• 11:36 UTC — PR opened
• 11:37 UTC — Review agent dispatched (run 33503365421)
• 11:39 UTC — Review agent failed with Error code policy_denied — the opus model alias resolved to claude-opus-5, which is not enabled on the org's Vertex AI project. The harness classified the error as "unknown" and retried twice before giving up. Zero tokens consumed.
• 11:49 UTC — PR merged by the author without human review (durandom was requested but did not review)

Root cause

This was a bootstrap chicken-and-egg problem. The PR intended to fix the model config, but the review agent ran against the base branch (via pull_request_target), which had no model field. The harness defaulted to the upstream opus alias, which since ~Aug 27 resolves to claude-opus-5 — a model not available on this org's Vertex AI project. The policy_denied error is deterministic and non-retryable, but the harness retried it anyway.

Now that the PR is merged, the base branch has the claude-opus-4-8 pin, so future reviews will use the correct model.

Assessment

The author's decision to self-merge was appropriate — this was a low-risk infrastructure config change, and the review failure was itself caused by the problem the PR was fixing. No code logic was changed.

Existing issues cover all improvements

All improvement opportunities identified in this retro are already tracked by open issues:

• fullsend-ai/agents#1117 — Directly covers the opus → claude-opus-5 → policy_denied outage. Proposes both enabling the model and adding a model-availability pre-check in the CLI. This retro provides additional evidence: the same failure pattern reproduces on the rhdh-parasol fork, confirming the issue is systemic across orgs sharing the same Vertex AI constraints.

• fullsend-ai/fullsend#2058 — Covers failing fast on non-retryable auth/policy errors instead of wasting time on retries. The policy_denied error in this retro is a case where this improvement would have saved ~30 seconds of wasted retry time.

• fullsend-ai/fullsend#5382 — Covers classifying infrastructure failures in status comments. The generic "Failure" status posted by the review bot gave no indication that the cause was a model policy denial.

No new proposals are warranted.