Disable fix agent by default on human-created PRs

[fullsend-ai-triage]

Note

This issue was automatically generated from meeting notes by the scribe agent.
Please review, edit, and add any missing context before prioritizing.

Problem

The fix agent triggers on human-created PRs by default. When a human opens a PR and CI fails, the fix agent may attempt automated fixes that conflict with the author's intent or create noise. The team agreed this should be opt-in for human PRs rather than automatic.

Options considered

1. Filter by PR author at dispatch time — the shim workflow checks whether the PR was created by a bot or human and skips dispatch for human PRs. Simple but requires maintaining author detection logic.
2. Configuration flag in harness or config.yaml — add fix_agent.human_prs: false as a default, allowing orgs to opt in. More flexible but adds config surface area.
3. Label-based opt-in — require a specific label (e.g., fullsend-fix) on human PRs to trigger the fix agent. Gives per-PR control but adds manual steps.

Acceptance criteria

• Fix agent does not trigger automatically on PRs authored by humans
• Fix agent still triggers on agent-authored PRs (e.g., from fullsend-coder)
• Mechanism exists for humans to opt in to fix agent on their PRs if desired
• Existing behavior for agent-created PRs is unchanged
• Documentation updated to describe the default behavior

Related

• Related: #473 — fix: skip code agent when human PR already exists for issue (similar pattern for code agent)
• Related: #604 — Support agent-level enablement with deferred app provisioning
• Related: #615 — Fix agent fails when HUMAN_INSTRUCTION contains shell metacharacters
• Source: Meeting notes

[github-actions]

fullsend triage is working on this — view logs

[fullsend-ai-triage]

Triage Summary

Severity: Medium | Category: Enhancement

Problem

The fix agent currently triggers on all PRs when CI fails, including human-authored PRs. This creates noise and can conflict with the author's ongoing work.

Root Cause

The shim workflow dispatching the fix agent does not distinguish between human and bot PR authors.

Recommended Approach

Add author-type filtering in the fix agent dispatch path. PR #473 already implements an analogous pattern for the code agent (skipping when a human PR exists for an issue) and can serve as a reference for the filtering logic.

The three options presented (dispatch-time filtering, config flag, label-based opt-in) are complementary. A practical approach would combine dispatch-time author detection (option 1) with a label-based opt-in (option 3), giving per-PR control without adding config surface area. The config flag (option 2) could be added later if orgs want a global override.

Proposed Test Case

Scenario: Fix agent dispatch filtering by PR author type

1. Bot-authored PR + CI failure → fix agent TRIGGERS (unchanged)
2. Human-authored PR + CI failure → fix agent DOES NOT trigger
3. Human-authored PR + CI failure + opt-in label → fix agent TRIGGERS
4. Non-fullsend bot PR + CI failure → verify expected behavior
5. Edge: PR authorship changes (transfer) → verify filtering uses current author

Related Work

• PR fix: skip code agent when human PR already exists for issue #473 (open) implements similar author-based filtering for the code agent — coordinate to share logic
• Issue Support agent-level enablement with deferred app provisioning #604 tracks agent-level enablement config — the opt-in mechanism here should align with that design
• Issue shim: Review→Fix dispatch uses changes_requested event instead of label-based trigger #551 tracks shim dispatch patterns — this change touches the same dispatch layer