fix(fullsend): bump to v0.39.0, pin grillme to Opus 4.6, allowlist claude.exe

[johnmcollier]

Summary

• Bump reusable-dispatch.yml to v0.39.0 (ISSUE_URL / tracker env for triage).
• Pin grillme to claude-opus-4-6; Opus 4.8 is not enabled on this Vertex project.
• Allowlist **/claude.exe so Claude Code 2.1+ can reach Vertex STS.

Test plan

• Merge, then comment /fs-triage on a small issue and confirm the pre-script gets past ISSUE_URL
• For customized agents, confirm the run uses claude-opus-4-6 and does not policy_denied on Vertex

Made with Cursor

[fullsend-ai-review]

🤖 Finished Review · ❌ Failure · Started 2:48 PM UTC · Completed 2:50 PM UTC

Commit: 7332a36 · View workflow run →

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 2:51 PM UTC · Completed 3:01 PM UTC

Commit: 7332a36 · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $2.82

[fullsend-ai-retro]

Retro: PR #13 — bump to v0.39.0, pin grillme to Opus 4.6, allowlist claude.exe

PR #13 was a human-authored configuration change (co-authored with Cursor) that bumped the fullsend reusable workflow from v0.36.0 to v0.39.0, pinned the grillme custom agent model to claude-opus-4-6, and added **/claude.exe to the sandbox network policy binary allowlist for Claude Code 2.1+ compatibility.

Timeline

Time (UTC)	Event
14:47:39	PR opened
14:47:43	Review workflow 33521688590 triggered
14:48:49	Review bot posts "Started" status comment
14:49:08	PR merged by author (~89s after opening)
14:49:53	Review sandbox pre-flight check fails — GitHub API unreachable (HTTP 403 from sandbox proxy)
14:50:07	Review bot updates to "Finished Review — Failure"
14:51:12	Retro workflow 33521841752 starts

Key findings

1. Review agent failed due to transient sandbox proxy issue. The sandbox successfully bootstrapped (agent resolved from fullsend-ai/agents@v0, skills uploaded, code copied), but the pre-flight connectivity check got HTTP 403 trying to reach api.github.com through the sandbox proxy. The agent never executed. This appears to be a one-off — no other failures in the last 30 workflow runs on this repo.

2. PR merged before review could complete. The author self-merged 89 seconds after opening, well before the review agent could finish even without the proxy failure (sandbox setup alone took ~60s). No branch protection or required status check gates the review.

3. No human reviews were submitted. The PR was merged without formal review approvals. As a maintainer-authored infrastructure config change (+5/−2 lines across 4 .fullsend and workflow files), this is understandable but means no review of any kind occurred.

Existing issues with new supporting evidence

• Add retry with backoff for openshell binary download in sandbox setup fullsend-ai/fullsend#6828 (retry with backoff for openshell binary download): This retro found a related gap — the pre-flight GitHub API connectivity check also lacks retry logic and fails immediately on a single transient 403. See proposal below for a distinct retry mechanism.
• Post-review script should post completed review results as informational when PR was merged during review fullsend-ai/fullsend#4509 (post-review should post informational results when PR was merged during review): This PR is a concrete example of the merge-before-review pattern that #4509 aims to address.
• Add CLAUDE.md to provide agent context for future tasks #5 (add CLAUDE.md): Still open. No CLAUDE.md or AGENTS.md exists in this repo, which limits the context available to agents for future runs.

Proposals filed

• Add retry with backoff to pre-flight GitHub API connectivity check (in fullsend-ai/fullsend)