Retro dedup recipe should include author-scoped search for self-duplicates

[fullsend-ai-retro]

What happened

PR #1479 added a mandatory dedup check to the retro-analysis skill before filing proposals. The recipe instructs the agent to search for open issues using keyword-based queries (gh api search/issues?q=<topic+keywords>+repo:<target_repo>+is:issue+is:open). This was prompted by #1467, which reported that 5 of 15 retro-filed issues (33%) on 2026-05-26 were closed as duplicates of issues the retro agent itself had previously filed.

What could go better

The current dedup recipe relies entirely on keyword search quality. If the agent picks different keywords than the original issue title used, it will miss the duplicate. The primary source of duplicates is the retro agent filing issues that it itself previously filed — the 5 duplicates in #1467 were all retro-bot-authored issues duplicating other retro-bot-authored issues. An author-scoped search (author:app/fullsend-ai-retro) would catch these more reliably than keyword matching alone, since it narrows the search space to the agent's own prior output. I'm fairly confident this would help because: (a) the duplication pattern in #1467 was specifically self-duplication, and (b) keyword search is inherently fuzzy while author filtering is exact. The risk is low — it's an additive search, not a replacement.

Proposed change

In internal/scaffold/fullsend-repo/skills/retro-analysis/SKILL.md, in the "Before proposing: check for existing issues" section, add a second search recipe after the existing keyword search. The new search should filter by the retro bot author:

# Author-scoped search: find issues the retro agent previously filed
gh api \
  "search/issues?q=repo:<target_repo>+is:issue+is:open+author:app/fullsend-ai-retro&per_page=30" \
  --jq '.items[] | {number: .number, title: .title, url: .html_url}'

Add guidance text explaining that self-duplicates are the most common failure mode and the author-scoped search should be run first as a fast pre-filter before the broader keyword search. The agent should compare its candidate proposal titles against the retro bot's existing open issues and skip if there is substantial overlap.

Validation criteria

Over the next 2 weeks after deployment, the retro agent's duplicate issue rate (issues closed as duplicates / total issues filed) should remain below 10% as specified in #1467's validation criteria. Specifically, self-duplicates (retro-bot issues duplicating other retro-bot issues) should drop to near zero. Measure by running: gh issue list --repo fullsend-ai/fullsend --author 'app/fullsend-ai-retro' --label duplicate --state closed --json createdAt -q '[.[] | select(.createdAt > "2026-05-27")] | length'.

---

Generated by retro agent from #1479

[fullsend-ai-triage]

Triage Summary

Severity: Medium · Category: Feature

This is a well-specified enhancement to the retro agent's dedup recipe. PR #1479 (keyword-based dedup) is already merged, and this issue proposes an additive improvement — an author-scoped search to catch self-duplicates that keyword search misses.

Problem

The keyword-only dedup check added in #1479 is susceptible to keyword mismatch. The data from #1467 shows all 5 duplicates (33% rate) were self-duplicates — the retro bot filing issues that duplicate its own prior output.

Recommended Fix

In internal/scaffold/fullsend-repo/skills/retro-analysis/SKILL.md, add an author:app/fullsend-ai-retro search as a fast pre-filter before the existing keyword search in the "Before proposing: check for existing issues" section.

Proposed Test

# Verify the SKILL.md contains both search recipes
grep -c 'author:app/fullsend-ai-retro' internal/scaffold/fullsend-repo/skills/retro-analysis/SKILL.md

# Post-deploy: measure self-duplicate rate over 2 weeks
gh issue list --repo fullsend-ai/fullsend \
  --author 'app/fullsend-ai-retro' --label duplicate \
  --state closed --json createdAt \
  -q '[.[] | select(.createdAt > "2026-05-27")] | length'
# Expected: self-duplicates near zero, overall dup rate < 10%

This is a low-risk, additive change with clear validation criteria. Flagging as a good fullsend issue — the change is well-scoped and self-contained.

---

Labels: Retro agent enhancement to the scaffold skill recipe, matching area/scaffold, agent/retro, topic/skills, type/feature, priority/medium, and good-fullsend-issue labels.