fix(fullsend): bump dispatch workflow to v0.39.0

[johnmcollier]

Summary

• Pin reusable-dispatch.yml to fullsend v0.39.0 so triage pre-scripts receive ISSUE_URL / FULLSEND_TRACKER.
• Vanilla playgrounds were still on v0.36.0 while agents@v0 already required those env vars.

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]

🤖 Review · ❌ Terminated · Started 2:48 PM UTC · Ended 2:50 PM UTC

Commit: 7ee4d28 · View workflow run →

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 2:50 PM UTC · Completed 2:58 PM UTC

Commit: 7ee4d28 · View workflow run →

Runtime: claude · Model: opus → claude-opus-4-6 · Effort: high · Cost: $2.72

[fullsend-ai-review]

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

Commit: 7ee4d28 · View workflow run →

[fullsend-ai-retro]

Retro: PR #4 — bump dispatch workflow to v0.39.0

Outcome: The review agent failed before it could run. No review was produced.

Timeline

1. 14:47:05 UTC — Human opens PR #4: a 1-line change bumping reusable-dispatch.yml from v0.36.0 to v0.39.0.
2. 14:47:10 — Fullsend workflow triggered (run 33521629866). Route dispatches the review agent.
3. 14:48:21 — Agent resolved from fullsend-ai/agents@v0 (commit c4e059f9). Fullsend CLI v0.36.0 installed (from the old workflow pin on main).
4. 14:48:52 — Pre-script confirms PR is open and proceeds.
5. 14:48:53 — Human merges the PR — 1 second after the pre-script check.
6. 14:49:59 — Sandbox created after 66.8s of setup.
7. 14:50:17 — Preflight connectivity check fails: GitHub API unreachable from sandbox (HTTP 403 — proxy allowlist issue).
8. 14:50:34 — Review agent exits with error. Status comment posted: "Finished Review · ❌ Failure."

Analysis

Root cause: The sandbox proxy blocked HTTPS CONNECT to api.github.com, causing the preflight connectivity check to fail. The review agent never executed.

Wasted compute: ~3 minutes of wall time including 66.8s of sandbox creation, 16.6s of bootstrap, and skill uploads — all for an agent that couldn't connect to GitHub.

Race condition: The pre-script checked PR state 1 second before merge. Even without the proxy failure, the agent would have reviewed a merged PR after spending 80+ seconds on sandbox setup.

Human judgment was correct: This was a trivial infrastructure fix (bumping a SHA pin). The human merged quickly without waiting for agent review, which was the right call.

Existing issues — new evidence

• Sandbox proxy blocking GitHub API (fullsend-ai/fullsend #4124, #2402, #5280): These issues document the retro agent sandbox hitting the same proxy 403. This retro provides evidence that the review agent is also affected — the proxy issue is systemic across agent types, not specific to retro.
• PR state check before dispatch (fullsend-ai/fullsend #6725, #3010, #1439): This case shows a 1-second race window between the pre-script PR state check and merge, followed by 80+ seconds of sandbox setup. Checking PR state after sandbox creation (or at the routing layer) would avoid this waste.
• Automatic workflow version updates (fullsend-ai/fullsend #2454): The PR itself was fixing a stale scaffold pin (v0.36.0 → v0.39.0). The planned --upstream-ref mechanism would prevent repos from falling behind.

Proposals

No new proposals. All identified issues are well-covered by existing open issues. The novel data point — that the review agent sandbox also fails with the proxy issue, not just the retro agent — is noted above as evidence for the existing issues.