Code agent should detect existing PRs for an issue and exit early

[fullsend-ai-retro]

What happened

For issue #834, the code agent was dispatched at 17:21 UTC. Meanwhile, the human (ralphbean) authored the fix using Claude Sonnet 4.5 and opened PR #835 at 17:26 UTC — roughly 5 minutes after dispatch. The code agent ran to completion (producing a 78KB artifact) but correctly did not create a competing PR.

While the agent avoided the worst outcome (a competing PR), it still consumed compute and tokens for work that was unnecessary. For a simple one-line fix, this waste is minimal, but for complex issues the cost could be significant.

What could go better

The code agent could check at startup whether a PR already exists for the issue it's working on (by searching for PRs that reference the issue number or checking for branches matching the agent/{issue}-* pattern or human PRs linked to the issue). If a PR is found, the agent could exit early with a log message.

Confidence: Medium. This optimization has a clear benefit (token savings) but the implementation needs care: (a) the human's PR might not exist yet when the agent starts but could appear mid-run, (b) the existing PR might be a draft or WIP that the agent should still work on, (c) race conditions between dispatch and PR creation. The code agent already handled this case gracefully by not creating a competing PR, so the current behavior is acceptable — this is an optimization, not a bug fix.

Proposed change

Add an early-exit check to the code agent's startup sequence (likely in the agent definition or harness pre-flight):

1. Before starting work, query gh pr list --search "issue:{ISSUE_NUMBER}" --repo {REPO} or check for open PRs whose body references the issue.
2. If a non-draft PR exists that addresses the issue, log a message and exit with success (no artifact, no PR).
3. If a draft PR exists, proceed with work (the draft may be abandoned).

This check should be in the code agent's harness configuration in .fullsend, not in the platform code, since different orgs may want different behavior.

Validation criteria

When a human opens a PR for an issue before the code agent starts its main work loop, the code agent exits within 30 seconds instead of running to completion. Validate on the next 3 instances where a human beats the code agent to a fix.

---

Generated by retro agent from #835

[github-actions]

fullsend triage is working on this — view logs

[fullsend-ai-triage]

This issue describes the same problem as #460 ("Code agent creates duplicate PR when a human PR already exists for the issue"). Both identify the need for the code agent to detect existing PRs linked to an issue and avoid unnecessary work.

The specific early-exit optimization proposed here — querying for existing PRs at startup and exiting within 30 seconds — is a subset of the solution being discussed in #460, which also covers triage-side detection, labeling, and override flags.

#460 has active discussion from multiple contributors (including architectural considerations about whether the fix belongs in the triage agent, the code agent, or both) and already has a triage summary and priority score.

The retro-specific context here (the incident where the agent ran to completion but didn't create a competing PR) is a useful data point — consider adding it as a comment on #460 to inform the implementation.