feat(retro): retro-filing-policy skill (#525)

[rh-hemartin]

Summary

• Add retro-filing-policy skill that the retro agent reads before deciding which proposals to file as GitHub issues. Each proposal gets a filing_decision field (file: bool, reason: string); the post-script skips proposals with file=false and folds them into the summary comment.
• The evidence-for gate remains as a hard safety net after the filing policy gate; filing_decision.file=true cannot override it.
• Extract sanitize_for_gha() in post-retro.sh to deduplicate GHA workflow command sanitization across 6 call sites.
• Backward compatible: when filing_decision is absent, all proposals are filed (current default).

Closes #525
Related: fullsend-ai/fullsend#5723

Test plan

• 10 new post-retro tests covering skip, count, summary folding, mixed proposals, explicit approval, backward compat, injection sanitization, and evidence-for gate override
• 5 new schema validation tests covering file=true, file=false, absent field, missing reason rejected, extra field rejected
• All 37 post-retro tests pass
• All 55 schema validation tests pass

🤖 Generated with Claude Code

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 1:55 PM UTC · Completed 2:12 PM UTC
Commit: c310276 · View workflow run →

[fullsend-ai-review]

Review

Findings

High

• [stale-schema-documentation] skills/retro-analysis/SKILL.md:232 — The retro-analysis skill documentation states proposal objects allow "ONLY the six fields shown above" and lists the six required fields (target_repo, title, what_happened, what_could_go_better, proposed_change, validation_criteria). This PR adds filing_decision as a seventh optional field to schemas/retro-result.schema.json, making the "ONLY six fields" claim incorrect. Since the retro agent reads retro-analysis as its primary output format reference, the explicit prohibition will likely prevent the agent from populating filing_decision, rendering the new feature inoperative. The example JSON at line 214 also needs updating.
  Remediation: Update skills/retro-analysis/SKILL.md — change line 232 to reference seven fields (or "six required fields plus one optional field") and add filing_decision to the example JSON at line 214.

Medium

• [protected-path] agents/retro.md, harness/retro.yaml, scripts/post-retro.sh, scripts/post-retro-test.sh, scripts/validate-output-schema-test.sh, skills/retro-filing-policy/SKILL.md — Six of eight changed files are under protected paths (agents/, harness/, scripts/, skills/). The PR links to issue feat(retro): accept a skill to control proposal filing behavior #525 and provides clear rationale for the changes. Human approval is always required for protected-path changes, regardless of context.

Low

• [edge-case] scripts/post-retro.sh — When a proposal has both filing_decision.file=false AND a title matching the evidence-for pattern, the filing policy gate (which runs first) catches the proposal via continue, so the evidence-for gate never evaluates it. The functional outcome is correct (not filed), but the proposal is classified under "Proposals skipped by filing policy" rather than "Evidence notes." No test covers this combination.
  Remediation: Consider moving the evidence-for gate before the filing policy gate, or add a test documenting this ordering as intentional.

• [GHA-workflow-command-injection] scripts/post-retro.sh:97 — The sanitize_for_gha() function handles ::, %0A/%0D, and newlines but does not strip ANSI escape sequences or other control characters. Not exploitable in GHA Actions UI (rendered literally), but a defense-in-depth gap for raw log consumers.

• [scope-overshoot] scripts/post-retro.sh — The retro-filing-policy skill text says filtering applies "after you have generated candidate proposals," meaning full proposal content is generated before the filing decision is made. Issue feat(retro): accept a skill to control proposal filing behavior #525 envisioned filtering "before token-expensive proposal writing." The post-script gate is a reasonable safety net, but the agent still generates full proposals for filtered items.

---

Labels: PR modifies the retro agent subsystem (agent definition, harness, skill, schema, post-script) and adds a new feature

---

Next steps:

• /fs-fix — agent addresses review findings automatically
• /fs-fix <your instruction> — agent fixes with your specific guidance
• Push commits directly — review re-runs automatically on push
• /fs-fix-stop — disable automatic fix runs for this PR

[fullsend-ai-review]

See the review comment for full details.

[fullsend-ai-review]

[low] GHA-workflow-command-injection

sanitize_for_gha() does not strip ANSI escape sequences or control characters. Not exploitable in GHA Actions UI but a defense-in-depth gap for raw log consumers.