ci(#5503): increase dispatch detection window to 4 minutes

[fullsend-ai-coder]

Summary

• Increase the dispatch detection window from 2 minutes (24 × 5s) to 4 minutes (48 × 5s) by raising dispatchMaxTry from 24 to 48 in pkg/behaviourtest/drivers/ci/githubactions/githubactions.go
• Add a regression test asserting the dispatch detection window is at least 4 minutes
• Addresses flaky behaviour test failures ("was not dispatched") caused by GitHub webhook delivery latency exceeding the 2-minute polling window, as observed on PRs feat(#5492): skip negative harness settle when positive wait elapsed #5497 and test(#5493): fold PR-to-issue negative into PR-label positive scenario #5498

Changes

The only production change is the dispatchMaxTry constant: 24 → 48. This extends the initial dispatch detection polling loop from 120s to 240s. The existing dispatchWait (12 minutes for workflow completion) is unchanged, so the new detection window is well within the overall timeout budget.

A new test (TestDispatchDetectionWindow_AtLeast4Minutes) validates that dispatchMaxTry × dispatchPoll ≥ 4 minutes, preventing future regressions if someone lowers the constant.

Testing

• All 16 tests in pkg/behaviourtest/drivers/ci/githubactions/ pass, including the new regression test
• go vet passes
• Validation: track dispatch-timing failures over the next 20 behaviour test runs on main and PR branches — the rate should drop from the observed pattern (2+ per day) to zero or near-zero

---

Closes #5503

Post-script verification

• Branch is not main/master (agent/5503-increase-dispatch-timeout)
• Secret scan passed (gitleaks — 481aab26ab5d62c1b026c22ff6b19b0012e41a23..HEAD)
• PR body secret scan passed (gitleaks — no-git)
• Pre-commit hooks passed (authoritative run on runner)
• Tests ran inside sandbox

[github-actions]

Site preview

Preview: https://424af7f4-site.fullsend-ai.workers.dev

Commit: abc4e163e52c3f4ba49cf0b93dd2cd4bec97cbdc

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 7:02 AM UTC · Completed 7:11 AM UTC
Commit: abc4e16 · View workflow run →

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[fullsend-ai-review]

Looks good to me

---

Labels: PR modifies behaviour test e2e driver for GitHub Actions dispatch detection.

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 8:33 AM UTC · Completed 8:44 AM UTC
Commit: abc4e16 · View workflow run →

[fullsend-ai-retro]

Retro: PR #5506 — ci(#5503): increase dispatch detection window to 4 minutes

This PR demonstrates an exemplary fully autonomous pipeline execution — from pattern detection to merged fix with zero rework.

Timeline

Time (UTC)	Event
06:47	Retro agent detected flaky dispatch timing pattern across PRs #5497 and #5498, filed issue fullsend-ai/fullsend#5503
06:48–06:52	Triage agent classified as medium bug, added ready-to-code
06:52–06:59	Code agent implemented fix: dispatchMaxTry 24→48 + regression test ($2.03, 45 turns, 7 min)
06:59–07:11	Review agent ran 4 sub-agents (correctness, style, intent, docs) + challenger pass. One low-severity style finding dismissed by challenger as false positive. Final verdict: approve with zero findings (12 min)
07:06	First human approval (ifireball, no comments)
08:08	Second human approval (rh-hemartin, no comments)
08:31	PR merged
08:37	Issue #5503 auto-closed

What went well

1. Full autonomous loop worked end-to-end: retro detection → issue filing → triage → code → review → merge. Total automated pipeline time: 24 minutes (issue to reviewed PR).
2. Code agent quality: Chose the simplest correct approach (Option 1 from 3 proposed in the issue), added a regression test, ran all 16 tests + vet. Zero rework iterations.
3. Review agent thoroughness: Despite the brief "Looks good to me" output, the review agent dispatched 4 specialized sub-agents in parallel (skipping security and cross-repo-contracts as irrelevant), ran an adversarial challenger pass, and correctly dismissed a false-positive style finding. The analysis depth matched the change appropriately.
4. Issue quality: The retro agent's issue Behaviour test "Sandbox blocks disallowed outbound URL" is flaky due to webhook delivery timing #5503 was well-structured with specific failure evidence from 2 PRs, root cause analysis, 3 ranked solution options, and validation criteria.
5. Cost efficiency: $2.03 for the code agent. Clean single-iteration pipeline.

Existing issues covering minor improvement areas

• Autonomy-readiness methodology produces false evidence when human review has no comments #5156 (autonomy-readiness methodology): Both humans approved without comments, matching the review agent's verdict. However, per Autonomy-readiness methodology produces false evidence when human review has no comments #5156's analysis, silent human approval is not strong evidence for autonomy expansion — the absence of human comments cannot be interpreted as evidence the agent covered more ground.
• #5157 (review detail scaling): The review agent's brief output ("Looks good to me") doesn't convey the internal analysis depth (4 sub-agents + challenger). However, Review agent should scale review detail to change complexity instead of applying uniform verbose template agents#439 argues for less verbose output on simple changes, and this output is arguably calibrated correctly for the change scope.

No new proposals

This workflow executed cleanly across all stages. The code was correct on the first attempt, the review was internally thorough, and the fix was appropriate for the problem. The minor improvement areas identified are already tracked by existing open issues.