fix(fullsend): use available Claude Opus 4.6 model

[kadel]

Pin customized Fullsend agents to Claude Opus 4.6.

Opus 4.8 is not enabled in the Fullsend Vertex deployment, causing review agents to exit before processing the PR.

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 1:43 PM UTC · Completed 1:54 PM UTC

Commit: 8faac08 · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $4.43

[fullsend-ai-review]

🤖 Finished Review · ❌ Failure (validation failed after 2 iteration(s)) · Started 1:44 PM UTC · Completed 1:46 PM UTC

Commit: 8faac08 · View workflow run →

Runtime: claude · Model: claude-opus-4-8

[fullsend-ai-review]

Review skipped — this PR is already merged.

The /fs-review command only reviews open pull requests.

_{Posted by fullsend pre-review check}

[fullsend-ai-retro]

Retro: PR fullsend-ai/fullsend#14 — Model availability cascade failure

Timeline

PR fullsend-ai/fullsend#14 was a single-commit human fix by kadel that changed the model for all four fullsend agents from claude-opus-4-8 to claude-opus-4-6 in .fullsend/config.yaml. It was created and merged within 18 seconds (13:42:11–13:42:29 UTC on 2026-09-01) with no human or agent review.

The full cascade:

1. 11:49 UTC — PR fullsend-ai/fullsend#6 merged, pinning all agents to claude-opus-4-8 and bumping the fullsend workflow to v0.38.0.
2. 13:20 UTC — PR fullsend-ai/fullsend#13 opened to fix a separate claude.exe sandbox policy issue. Its review dispatch (run 33512830596) failed with policy_denied because claude-opus-4-8 was not available on the Vertex deployment.
3. 13:35 UTC — PR Add problem document for strategic risks fullsend-ai/fullsend#13 merged by the human without agent review.
4. 13:36 UTC — An /fs-review command on PR fullsend-ai/fullsend#10 triggered run 33514404069, which also failed with an explicit error: "The model claude-opus-4-8 is not available on your vertex deployment."
5. 13:42 UTC — PR Add testing-agents problem document fullsend-ai/fullsend#14 created and instantly self-merged, downgrading all agents to claude-opus-4-6.
6. 13:44 UTC — The review agent dispatched for PR Add testing-agents problem document fullsend-ai/fullsend#14 (run 33514987540) correctly detected the PR was already merged and skipped review. However, it still attempted to start the agent with claude-opus-4-8 (from the base-branch config) and hit the same model unavailability error.

Root cause

PR fullsend-ai/fullsend#6 pinned all agents to claude-opus-4-8, a model not enabled on this org's Vertex AI deployment. This broke every subsequent agent dispatch for approximately 2 hours until PR fullsend-ai/fullsend#14 fixed it. The chicken-and-egg problem was compounded by pull_request_target using the base-branch config — so even the fix PR's review agent inherited the broken model.

What went well

• The human identified and resolved the issue quickly (two fix PRs within 20 minutes of the first failure).
• The review agent's pre-review check correctly detected PR Add testing-agents problem document fullsend-ai/fullsend#14 was already merged and posted a skip notification.
• The retro agent dispatched successfully on the fixed config.

Existing issues that cover the improvement opportunities

All improvement opportunities identified by this retro are already tracked by existing open issues upstream. This retro provides additional real-world evidence for them:

• Review sub-agents should fall back to an available model when the primary model is unavailable fullsend-ai/fullsend#1771 — Sub-agents should fall back to an available model when the primary is unavailable. This retro demonstrates a ~2-hour total agent outage caused by a single unavailable model pin, with no fallback path.
• fullsend-ai/fullsend#6795 — Review agent outage from opus alias resolving to a policy-denied model on Vertex AI. Same class of failure observed here.
• Implement pi fullsend-models.js: cross-provider fallback chain, phases, and budget fullsend-ai/fullsend#6531 — Cross-provider fallback chain. Would have prevented the cascade by trying alternative models.
• Review sub-agent model config should use deployment-agnostic aliases to prevent dispatch failures fullsend-ai/agents#1116 — Sub-agent model config should use deployment-agnostic aliases. The repo pinned to a version-specific model identifier (claude-opus-4-8) instead of an alias, which was the direct cause of the failure.
• Review sub-agents should handle model unavailability with fallback or explicit alerting fullsend-ai/agents#379 — Sub-agents should handle model unavailability with fallback or explicit alerting. The first failure showed an ambiguous policy_denied error; only the second failure gave the clear "model not available" message.
• Extend retro skip logic to merged PRs with zero fullsend agent involvement fullsend-ai/fullsend#6297 — Retro skip logic for merged PRs with zero agent involvement. This retro dispatched on a PR that had zero successful agent activity (review was skipped post-merge), spending tokens on a workflow with limited agent decisions to analyze.

No new proposals

All identified improvements are covered by existing open issues. No new proposals are warranted.