Skip retro dispatch on PRs with no agent involvement

[fullsend-ai-retro]

What happened

PR #32 (problem-cognitive-debt) was opened by a human contributor on 2026-03-18 and closed without merging on 2026-05-14. The branch name does not follow the agent/{issue}-{slug} pattern. No triage, code, review, or fix workflow runs were dispatched in fullsend-ai/.fullsend for this PR. The only workflow run was fullsend.yaml run 25871472428, triggered by the pull_request_target close event, which correctly skipped all agent dispatch jobs but still dispatched a retro. The retro agent then ran, consumed tokens, and found nothing to analyze.

What could go better

The retro dispatch should not fire on PRs where no agent participated. This retro run consumed tokens and compute time with no useful output — there was no agent workflow graph to trace, no agent decisions to evaluate, and no agent-generated code to review. Confidence: high. The evidence is unambiguous — zero agent runs were associated with this PR, and the branch naming clearly indicates human authorship. This pattern likely recurs on any human-authored PR that gets closed or merged in repos with the fullsend workflow installed.

Proposed change

Add a condition to the dispatch-retro job in fullsend.yaml that checks for agent involvement before dispatching. Possible heuristics (use one or combine):

1. Branch name check: Only dispatch retro if the head branch matches agent/* pattern.
2. Actor check: Only dispatch if at least one commit was authored by the agent bot account.
3. Workflow run check: Only dispatch if there were prior triage/code/review/fix dispatches for the same PR.

Option 1 (branch name) is the simplest and cheapest — it requires no API calls and can be evaluated in the job's if: condition. Example: if: startsWith(github.head_ref, 'agent/'). Options 2 and 3 are more robust but require additional API calls in a setup step.

Validation criteria

After the change, closing or merging a human-authored PR (branch not matching agent/*) should not trigger a retro dispatch. Verify by: (1) closing a test PR with a non-agent branch name and confirming the dispatch-retro job shows as skipped, and (2) closing a PR with an agent/* branch and confirming the retro still dispatches normally. Monitor for 2 weeks to ensure no false negatives (agent PRs that should get retros but don't).

---

Generated by retro agent from #32

[github-actions]

fullsend triage is working on this — view logs

[fullsend-ai-triage]

Triage Summary

Severity: Low | Category: Enhancement

Problem

The dispatch-retro job in fullsend.yaml fires on every closed PR, including human-authored ones where no agent participated. This wastes tokens and compute on retro runs that have nothing to analyze.

Root Cause

The job's if: condition (lines 316–318) only checks for the pull_request_target closed event with no filter for agent involvement.

Recommended Fix

Add a branch name guard to the dispatch-retro job condition — the simplest of the three options proposed in the issue:

if: >-
  github.event_name == 'pull_request_target'
    && github.event.action == 'closed'
    && startsWith(github.head_ref, 'agent/')

This requires no API calls and can be evaluated entirely in the job's if: expression. The dispatch-retro-command job (manual /retro trigger) should remain unaffected since it serves a different use case.

Proposed Test Plan

1. Close a test PR with a non-agent branch name (e.g., 'feature/my-change')
   → Expected: dispatch-retro job shows as skipped
2. Close a test PR with an agent branch name (e.g., 'agent/123-fix-bug')
   → Expected: dispatch-retro job runs and dispatches normally
3. Verify dispatch-retro-command job is unaffected by the change
4. Monitor for 2 weeks for false negatives (agent PRs missing retros)

Related Context

PR #906 takes a similar approach for the fix agent (disabling auto-run on human-authored PRs), establishing a pattern for agent-only dispatch filtering.