retrospective

name: retrospective description: This skill should be used when the user wants to analyze evaluation patterns across completed plans and evolve checklists. Triggered by asking to "run a retrospective", "analyze evaluation patterns", "evolve checklists", or "/superpowers:retrospective". For autonomous multi-turn runs, invoke wrapped in /goal. argument-hint: [plan-path-2] [--across-all] user-invocable: true allowed-tools: ["Read", "Glob", "Grep", "Write", "Edit", "Bash(python3:)", "Bash(git:)", "Bash(${CLAUDE_PLUGIN_ROOT}/lib/seed-checklists.sh:)", "Bash(${CLAUDE_PLUGIN_ROOT}/lib/post-plan-diff.sh:)", "Bash(${CLAUDE_PLUGIN_ROOT}/lib/jsonl-emit.sh:*)"]

Retrospective

Analyze evaluation patterns across completed plans, identify recurring failures, and auto-apply checklist evolution. The user reviews post-commit via git show docs/retros/checklists/.

Chain position: This skill is the downstream consumer of executing-plans Phase 4 "Checklist Evolution Candidates". It aggregates signals across plans and produces versioned checklist updates.

Recommended: run wrapped in /goal

A retrospective aggregates signals across plans and can span multiple turns. Launch it under Claude Code's built-in /goal (v2.1.139+):

/goal "Claude has narrated a successful checklist-evolution commit (with commit hash) and stated the retrospective is complete" /superpowers:retrospective <plan-paths>

/goal is a user-typed outer wrapper — it must prefix the invocation; a skill cannot enable it for itself mid-run. The evaluator judges only what Claude narrates in the transcript (it does NOT read files or run commands) — phrase the condition against narrated output (the commit-hash line, an explicit completion statement), never filesystem state, which is unverifiable and will time out. Full semantics and condition phrasing: ../../skills/references/goal-wrapper.md.

Pre-Check (run first, in order)

A. INSUFFICIENT-POST-PLAN advisory (informational)

Read the most recent plan_completed event from docs/retros/plans-completed.jsonl. If hours_since_completion < 24h AND bash "${CLAUDE_PLUGIN_ROOT}/lib/post-plan-diff.sh" summary <completion_commit> <files...> returns total == 0, output the INSUFFICIENT-POST-PLAN reminder verbatim (see ./references/post-plan-diff.md §Pre-Check A) and proceed to Phase 0 — do NOT pause. Skip silently when completion_commit is missing.

B. Recall persistent memory (calibration priors)

CRITICAL — do this before Phase 0, do NOT skip it because it feels like setup. A retrospective calibrates checklist evolution against prior decisions, and persistent memory holds the human judgments the evolution-log cannot capture (rejected directions, debt/gate trackers, harness-design stance, working-style feedback). Recall is a SECONDARY calibration signal — docs/retros/evolution-log.jsonl (Phase 1 step 5) stays authoritative; when memory and the log disagree, the log wins.

Claude Code injected the MEMORY.md index at session start, so its one-line hooks are ALREADY in your context — there is nothing to read from disk. Scan those hooks now as prior-decision evidence and select the ones bearing on this run:

• Prior evolution / rejection decisions and debt trackers — anti-add-bias / "rejected" / "gate" / "pending" notes. A memory-recorded rejection is the same evidence type as a Phase 1 step 5 item_removed log row.
• Harness-design principles — e.g. simplify-don't-add stance.
• Working-style feedback — e.g. "auto-produce, never pause", "L2 must carry CRITICAL".

Carry these forward as priors into Phase 1 step 5 (calibration history), Phase 3 (REMOVE-is-load-bearing suppression of weakly-justified ADDs), and Phase 4 (self-reject a proposal that contradicts a recalled prior, citing the memory entry). If no calibration-relevant hook exists, log Pre-Check B: no calibration-relevant memory and proceed. Never resolve a memory path, never read a topic file, never block, never ask, never read "to make sure" — the index hooks carry the signal on their own.

Phase 0: Bootstrap (run only when no checklists exist)

Before Phase 1, check whether docs/retros/checklists/ contains {mode}-v1.md for each mode (design / plan / code).

If all three modes already have a v1 file, log Phase 0: all checklists present, skipping seed and proceed to Phase 1.

Phase 0 runs per-mode independently — only modes missing a v{N} file are seeded. Do not skip the entire phase because one mode already has a checklist.

Path A — Completed plans or evaluation reports exist

Seed the generic template and proceed to Phase 1:

bash "${CLAUDE_PLUGIN_ROOT}/lib/seed-checklists.sh" <mode> docs/retros/checklists/<mode>-v1.md

Log Seeded initial checklist: {mode}-v1.md. Skip the Full History Analysis below — Phase 1 has real evaluation data to work with.

Path B — Cold start (no completed plans, no evaluation reports)

When docs/retros/plans-completed.jsonl is absent or empty AND no evaluation-round-*.md files exist anywhere under docs/plans/, the retrospective has no evaluation data. Instead of producing a zero-signal run, perform a Full History Bootstrap: analyze the project's entire git history to extract project-specific failure patterns and augment the generic template with tailored checklist items.

Step 1 — Seed the generic template (same command as Path A):

bash "${CLAUDE_PLUGIN_ROOT}/lib/seed-checklists.sh" <mode> docs/retros/checklists/<mode>-v1.md

Step 2 — Git history gate: count commits via git rev-list --count HEAD. If < 50, log Phase 0: insufficient git history ({N} commits, need 50+) for bootstrap analysis, skip Step 3, and proceed to Phase 1 with the generic template only.

Step 3 — Full History Analysis (see ./references/analysis-patterns.md §Bootstrap Analysis for the detailed methodology):

1. git log --oneline --all — collect all commits
2. Classify each commit by conventional-commit prefix into feedback (fix:, refactor:, style:, perf:) or evolution (feat:, docs:, chore:, build:, ci:, test:)
3. Group feedback commits by scope+type combination, rank by frequency
4. For the top clusters, git show <sha> the diffs and extract recurring failure patterns
5. Classify each pattern into a mode layer:
   • code: dead code, lint violations, i18n gaps, duplicate definitions, stub implementations
   • design: stale references, missing BDD scenarios, references to deleted features
   • plan: oversized tasks, missing cleanup tasks, batch ordering violations
6. Generate one checklist item per failure pattern using the evolution-protocol.md New Item Template format (ID + description + check method + evidence format + rework format)

Step 4 — Append project-specific items: for each mode that received items, insert a new ## Project-Specific Items (Bootstrap Analysis) section into the seeded {mode}-v1.md immediately before the existing ## Evaluation Protocol section. Each item gets a unique ID following {MODE}-{CATEGORY}-{NN} naming (e.g., CODE-I18N-01, DESIGN-STALE-01, PLAN-SCOPE-01).

Step 5 — Report: log the analysis statistics (total commits, feedback/evolution split, top clusters, items generated per mode) in the retrospective report.

Exit code handling: the seed script refuses to clobber an existing checklist (exit code 3) — treat that as "already seeded, proceed". Real failures (exit 1 = unknown mode, exit 2 = usage error) abort the phase. To genuinely reset an existing checklist (e.g., after a major harness change), append --force after the output path.

The canonical v1 template content lives in lib/seed-checklists.sh. To inspect or modify the seed bodies, edit that script — do NOT re-inline templates here.

Phase 1: Data Collection

1. Resolve inputs: Parse $ARGUMENTS for plan paths. If --across-all, scan docs/plans/ for all *-plan/ directories with evaluation reports. If no argument is given, read docs/retros/plans-completed.jsonl and auto-scope to plans completed after the most recent retrospective_run event in docs/retros/evolution-log.jsonl.

2. Resolve evals: For each plan path, look for evaluation reports in the plan directory (evaluation-round-*.md, evaluation-design-round-*.md, evaluation-plan-round-*.md). If a sibling *-evals/ directory exists, read from there instead.

3. Read checklists: Scan docs/retros/checklists/ for latest versions of each mode ({mode}-v{N}.md, highest N).

4. Read reports: For each plan, read all evaluation report files. Extract per-item results (Item ID, Result, Evidence) and rework items.

5. Read evolution history (calibration input): Read docs/retros/evolution-log.jsonl if present. Build a history table keyed by item_id with: most recent event (item_added|item_removed|item_modified|item_promoted), timestamp, rationale. This history feeds Phase 3 — do NOT re-propose an ADD for an item REMOVEd in the most recent retrospective unless the new evidence is materially different from the original removal rationale. Cite the historical entry in any such proposal.

6. Read post-plan diff: For each plan with a completion_commit field, run bash "${CLAUDE_PLUGIN_ROOT}/lib/post-plan-diff.sh" list <completion_commit> <completion_modified_files...> and pass the classified commit list to Phase 5a. See ./references/post-plan-diff.md for classification rules and skip conditions.

7. Minimum data check: If only 1 plan provided, warn that ADD proposals require 2+ plans (except the post-plan-diff 1-plan ADD override in Phase 5a). REMOVE proposals need 3+ reports with zero failures.

Phase 2: Pattern Analysis

Aggregate data across all plans. See ./references/analysis-patterns.md for detailed logic.

1. Failure frequency: Count distinct plans where each checklist item FAILed. Rank by frequency descending.
2. Plateau tasks: Identify tasks that were REWORK across 2+ consecutive evaluation rounds in any plan. Extract the root cause from rework items.
3. Never-failing items: Find items with 0 FAILs across 10+ evaluation reports. These are REMOVE candidates.
4. Variety gaps: From executing-plans completion summaries, find batches where all items PASS but 2+ rework rounds occurred -- the checklist missed the failure mode.

Output a structured analysis report with tables for each category.

Phase 3: Evolution Proposals

Generate proposals from analysis results. See ./references/evolution-protocol.md for thresholds and format.

Rate limit (EVO-6): Max 3 proposals per mode per retrospective run. Defer excess with evidence for future runs.

Counter monotonic growth (REMOVE is load-bearing): ADD is cheap to trigger (even a 1-plan post-plan-diff override) while REMOVE used to require 10+ reports/item — a volume real single-project usage never reaches, so checklists only ever grew. The 3+ reports/item REMOVE threshold above is deliberately reachable. Each run, actively scan for never-firing items and propose REMOVE; a checklist that only grows is a calibration failure, not success.

Each proposal includes: type, target checklist, item ID, description, rationale with plan evidence.

Phase 4: Auto-Apply Proposals

Apply every Phase 3 proposal (ordered by priority: regression breaks first, then by frequency). No per-proposal approval gate — EVO-6 (max 3/mode/run) + Phase 3 thresholds + post-commit git show docs/retros/checklists/ are the quality surface. proposals_rejected is reserved for self-rejection at apply time: when a proposal duplicates a recent removal (Phase 1 step 5 history) or contradicts a recalled memory prior (Pre-Check B) without materially new evidence, log to the report under "Self-Rejected Proposals" with the cited historical or memory entry, increment proposals_rejected, and skip the checklist row. All other proposals advance.

Apply steps:

1. Pre-edit snapshot: Write current checklist content to the retrospective report under "Pre-Edit Snapshot" with rollback instructions
2. Create new version: Write {mode}-v{N+1}.md with all applied changes. Version increments once per run (not per proposal). Original version preserved unchanged.
3. Log evolution — CRITICAL: a proposal is NOT "applied" until its evolution-log row exists. Immediately after writing the new version file, append one row to docs/retros/evolution-log.jsonl per applied proposal via lib/jsonl-emit.sh with <channel>=evolution-log — emit per-proposal here, do NOT defer to the end of the run. The event arg is one of item_added | item_removed | item_modified | item_promoted. The full canonical bash invocation (every required field and --arg pair) lives in ./references/evolution-protocol.md §"Canonical Emit Invocations" — substitute the event arg per applied proposal. These rows feed Phase 1 step 5's re-proposal guard; a dropped item_removed row silently re-adds the just-removed item next run. The Stop hook (hooks/stop-state-sync.sh) backfills item_added/item_removed from the checklist version diff only when every row for this version is missing (all-or-nothing) and carries no rationale — so the in-skill emit is authoritative and must run.
4. Verify the log — CRITICAL self-check, do NOT skip: before leaving Phase 4, count evolution-log rows whose checklist_version equals the version(s) written this run and confirm the count equals proposals_approved. Emit any missing rows now. This is the guard against a partial drop the hook's all-or-nothing backfill will not catch.

Phase 5: Harness Health (advisory)

Surface components that may no longer earn their cost and turn post-plan corrections into checklist proposals. Everything here is advisory: it feeds Phase 3 proposals and the Phase 6 report, and is reviewed by the user via the post-commit diff. Phase 5 never disables a component or mutates the harness.

Removed in v2.9.0 — the automated disable-test loop. The one-at-a-time harness-config.json disable protocol and its harness-observations.jsonl telemetry were deleted. Empirical reason: across every real project those channels stayed empty, and the single disable test that ever ran (user-simulation, 2026-05-08, recurring_failure_patterns) was wrong and had to be reverted by hand the same day. Assumption-testing-by-auto-disable imported an industrial-harness pattern that never closed a cycle at single-project scale. Component changes now go through ordinary REMOVE/MODIFY proposals (Phase 3) with human review.

5a. Post-Plan Correction Mining (highest-value signal)

This is the load-bearing input — it produced the most valuable checklist evolution in practice. For each plan's post-plan-diff (Phase 1 step 6): when feedback-classified commits (refactor/fix/style/perf on plan-modified files) cluster around a pattern no batch evaluator flagged, that is a real evaluator coverage gap. Render the corrections table (./references/post-plan-diff.md §Phase 5a) and graduate each missed pattern to a Phase 3 ADD proposal at 1-plan evidence. This catches what grep-based checks cannot: consistency, API-contract, and coverage gaps. (This is exactly how user-simulation's CODE-CONTRACT/CONS/COV items were added.)

5b. Usage-Driven Recommendations (report notes only)

• If all tasks pass first round (no REWORK) across 3+ plans, note that per-batch evaluation may be reducible — a candidate for a future Phase 3 MODIFY/REMOVE, not an automatic change.
• If a mode's checklist has only regression items all passing across 3+ plans, recommend a spot-check cadence in the report.

These are report notes. Never disable a component from this phase. Never-failing items are handled by Phase 3 REMOVE proposals (3+ reports threshold) — do not duplicate.

Phase 6: Output

Write the retrospective report to docs/retros/retro-{date}-{topic}.md:

1. Analysis tables (failure frequency, plateaus, never-failing, variety gaps)
2. Proposals with approval status
3. Checklist versions updated (if any)
4. Harness Health notes (5a post-plan corrections mined into ADD proposals; 5b informational recommendations)
5. Summary: N proposals approved, M rejected, checklists updated to version X

Close the calibration loop — CRITICAL: do this before you stop, not after the report "feels done." Append one retrospective_run row to docs/retros/evolution-log.jsonl via the canonical emit pattern in ./references/evolution-protocol.md §"Canonical Emit Invocations", recording proposals_approved and proposals_rejected. This row is the closure marker the next run's auto-scope (Phase 1 step 1) reads to avoid re-analyzing these plans — skip it and the next retrospective silently re-analyzes already-analyzed plans, re-proposing the same changes. Do not skip it even when zero proposals were approved. The Stop hook (hooks/stop-state-sync.sh) backfills a minimal watermark from this run's retro-*.md report if you drop this, but only your emit carries proposals_approved / proposals_rejected / plans_analyzed — so write the rich row here.

References

• ./references/analysis-patterns.md - Failure frequency, plateau detection, never-failing analysis
• ./references/evolution-protocol.md - Proposal types, thresholds, version management, evolution log schema, pre-edit snapshot
• ../../skills/references/goal-wrapper.md - /goal wrapper semantics and condition phrasing (shared)
• ${CLAUDE_PLUGIN_ROOT}/lib/post-plan-diff.sh - classifies post-plan commits as feedback (refactor/fix/style/perf — user correcting superpowers output) or evolution (feat/chore/docs/build/ci/test — user adding new requirements). Used by the Pre-Check and Phase 1 step 6 to mine the post-plan correction signal (Phase 5a)