feat(agents): add retro agent and supporting skills

[ralphbean]

Summary

• Adds the retro agent definition (agents/retro.md) — a retrospective analyst that investigates agent failures, rework, and missed opportunities
• Adds three supporting skills: trace-gathering, forming-hypotheses, localizing-fixes
• Adds .claude/agents symlink (matches pattern from feat(agents): add review agent and pr-review skill #171)

The retro agent has three triggers:

1. /retro <explanation> — human-initiated investigation
2. Triage follow-up on retro-filed issues — responds to clarifying questions
3. PR merged/closed — proactive opportunity detection against configured goals

It follows a consistent workflow: gather trace → form hypotheses → localize fix → file issue. It does not propose or implement fixes — it files issues that flow through triage → implementation asynchronously.

Design decisions

• Retro agent files issues, not fixes. Keeps diagnosis separate from implementation. The agent with the trace context diagnoses; the agent with the code context implements.
• Fix localization prefers broad scope. "Would every fullsend adopter want this fix?" If yes, recommend upstream. Aligns with open source ethos over defensive scoping.
• Trace bundle attached to filed issues. Triage and implementation agents get the raw evidence without re-gathering it.
• Skills kept concise. Reviewed against obra/superpowers writing-skills guidance — all under 400 words, descriptions use "Use when..." trigger format.

Test plan

• Review agent definition for completeness against Story 8 acceptance criteria
• Review skill conciseness — do they teach only what Claude doesn't already know?
• Validate that .claude/agents symlink resolves correctly
• Skills will need refinement once real agent runs produce traces in GitHub Actions

Closes #131

🤖 Generated with Claude Code

[ralphbean]

Keeping this in draft until the other three agent defs land.

[rh-hemartin]

I tested it (with Opus 4.6) locally leading to:

 Analysis 1: Slack ADR Notification Chain (#183 → #191 → #198)

  Rework pattern detected. The initial Slack notification workflow needed two follow-up fixes within 30 hours:

  - PR #191 fixed error-swallowing patterns (|| true masking gh pr diff failures) and a missing --repo flag exposed when actions/checkout was removed per review feedback
  - PR #198 switched to pull_request_target because fork PRs can't access repo secrets under the pull_request trigger

  Root causes (high confidence):
  1. Review identified the silent failure issue but marked it "non-blocking" — the fix for it introduced a second-order bug (missing --repo flag)
  2. Fork PR secret access is a non-obvious GitHub Actions behavior that neither author nor reviewer caught

  Recommended fix locations:
  - Upstream — document the pull_request vs pull_request_target security model and the fork secret gotcha
  - Upstream — add linting/skill for detecting || true error-swallowing in workflow scripts
  - Org-level — consider requiring workflow validation before merge for infra changes

  ---
  Analysis 2: Closed PRs #190 and #192

  No issue warranted. These were human-driven (with AI assistance), not agent-driven. The developer followed a normal incremental pattern:

  1. PR #190: regression tests (passed CI)
  2. PR #192: consolidated tests + Makefile fixes (e2e revealed a new bug)
  3. PR #193: all work + org secret visibility fix (still open)

  Each closure was intentional consolidation, not failure. Total time was 75 minutes with minimal waste. The pattern actually helped discover a production bug (org secret visibility) that might have been missed otherwise.
  

[rh-hemartin]

retro-closed-prs-190-192.md

retro-slack-notify-chain.md

Same tests but more output.

[ralphbean]

I'm going to drop this so that it's not in your way, @rh-hemartin .