solo-retro

/retro

This skill is self-contained — follow the phases below instead of delegating to other skills (/review, /audit, /build) or spawning Task subagents. Run all analysis directly.

Post-pipeline retrospective. Parses pipeline logs, counts productive vs wasted iterations, identifies recurring failure patterns, scores the pipeline run, and suggests concrete patches to skills/scripts to prevent the same failures next time.

Live Context

• Branch: !git branch --show-current 2>/dev/null
• Recent commits: !git log --oneline -10 2>/dev/null
• Modified files: !git diff --name-only HEAD~5..HEAD 2>/dev/null | head -20

When to use

After a pipeline completes (or gets cancelled). This is the process quality check — /review checks code quality, /retro checks pipeline process quality.

Can also be used standalone on any project — with or without pipeline logs.

MCP Tools (use if available)

• session_search(query) — find past pipeline runs and known issues
• codegraph_explain(project) — understand project architecture context
• codegraph_query(query) — query code graph for project metadata

If MCP tools are not available, fall back to Glob + Grep + Read.

Phase 1: Locate Artifacts

1. Detect project from $ARGUMENTS or CWD:

   • If argument provided: use it as project name
   • Otherwise: extract from CWD basename (e.g., ~/projects/my-app -> my-app)

2. Find pipeline state file: .solo/pipelines/solo-pipeline-{project}.local.md (project-local) or ~/.solo/pipelines/solo-pipeline-{project}.local.md (global fallback)

   • If it exists: pipeline is still running or wasn't cleaned up — read YAML frontmatter for project_root:
   • If not: pipeline completed — use CWD as project root

3. Verify artifacts exist (parallel reads):

   • Pipeline log: {project_root}/.solo/pipelines/pipeline.log
   • Iter logs: {project_root}/.solo/pipelines/iter-*.log
   • Progress file: {project_root}/.solo/pipelines/progress.md
   • Plan-done directory: {project_root}/docs/plan-done/
   • Active plan: {project_root}/docs/plan/

4. Determine analysis mode:

   • If pipeline log exists: proceed with full log-based analysis (Phases 2-4)
   • If NO pipeline log: switch to fallback mode (see Fallback Analysis below)

5. Count iter logs (if they exist): ls {project_root}/.solo/pipelines/iter-*.log | wc -l

   • Report: "Found {N} iteration logs"

Fallback Analysis (No Pipeline Logs)

If no pipeline logs exist, the retro can still provide value by analyzing:

1. Git history: git log --oneline --since="1 week ago" — commit frequency, patterns, conventional format
2. Test results: run test suite if configured in CLAUDE.md or package.json
3. Build status: run build if configured
4. CLAUDE.md changes: git log --oneline -- CLAUDE.md — how docs evolved
5. Code quality metrics: file counts, TODO/FIXME density, dead code indicators
6. Project structure: completeness of docs/, tests/, CI config

Skip Phases 2-4 and proceed directly to Phase 5 (Plan Fidelity) and Phase 6 (Git & Code Quality). Adjust Phase 7 scoring to weight available data more heavily.

Phase 2: Parse Pipeline Log (quantitative)

Read pipeline.log in full. Parse line-by-line, extracting structured data from log tags:

Log format: [HH:MM:SS] TAG | message

Extract by tag:

Tag	What to extract
START	Pipeline run boundary — count restarts (multiple START lines = restarts)
STAGE	iter N/M | stage S/T: {stage_id} — iteration count per stage
SIGNAL	<solo:done/> or <solo:redo/> — which stages got completion signals
INVOKE	Skill invoked — extract skill name, check for wrong names
ITER	commit: {sha} | result: {stage complete|continuing} — per-iteration outcome
CHECK	{stage} | {path} -> FOUND|NOT FOUND — marker file checks
FINISH	Duration: {N}m — total duration per run
MAXITER	Reached max iterations ({N}) — hit iteration ceiling
QUEUE	Plan cycling events (activating, archiving)
CIRCUIT	Circuit breaker triggered (if present)
CWD	Working directory changes
CTRL	Control signals (pause/stop/skip)

Compute metrics:

total_runs = count of START lines
total_iterations = count of ITER lines
productive_iters = count of ITER lines with "stage complete"
wasted_iters = total_iterations - productive_iters
waste_pct = wasted_iters / total_iterations * 100
maxiter_hits = count of MAXITER lines
plan_cycles = count of QUEUE lines with "Cycling"

per_stage = {
  stage_id: {
    attempts: count of STAGE lines for this stage,
    successes: count of ITER lines with "stage complete" for this stage,
    waste_ratio: (attempts - successes) / attempts * 100,
  }
}

Phase 3: Parse Progress.md (qualitative)

Read progress.md and scan for error patterns:

1. Unknown skill errors: grep for Unknown skill: — extract which skill name was wrong
2. Empty iterations: iterations where "Last 5 lines" show only errors or session header (no actual work done)
3. Repeated errors: same error appearing in consecutive iterations -> spin-loop indicator
4. Doubled signals: <solo:done/><solo:done/> in same iteration -> minor noise (note but don't penalize)
5. Redo loops: count how many times build->review->redo->build cycles occurred

For each error pattern found, record:

• Pattern name
• First occurrence (iteration number)
• Total occurrences
• Consecutive streak (max)

Phase 4: Analyze Iter Logs (sample-based)

Do NOT read all iter logs — could be 60+. Use smart sampling:

1. First failed iter per pattern: For each failure pattern found in Phase 3, read the first iter log that shows it

   • Strip ANSI codes when reading: sed 's/\x1b\[[0-9;]*m//g' < iter-NNN-stage.log | head -100

2. First successful iter per stage: For each stage that eventually succeeded, read the first successful iter log

   • Look for <solo:done/> in the output

3. Final review iter: Read the last iter-*-review.log (the verdict)

4. Extract from each sampled log:

   • Tools called (count of tool_use blocks)
   • Errors encountered (grep for Error, error, Unknown, failed)
   • Signal output (<solo:done/> or <solo:redo/> present?)
   • First 5 and last 10 meaningful lines (skip blank lines)

Phase 5: Plan Fidelity Check

For each track directory in docs/plan-done/ and docs/plan/:

1. Read spec.md (if exists):

   • Count acceptance criteria: total - [ ] and - [x] checkboxes
   • Calculate: criteria_met = checked / total * 100

2. Read plan.md (if exists):

   • Count tasks: total - [ ] and - [x] checkboxes
   • Count phases (## headers)
   • Check for SHA annotations (<!-- sha:... -->)
   • Calculate: tasks_done = checked / total * 100

3. Compile per-track summary:

   • Track ID, criteria met %, tasks done %, has SHAs

Phase 6: Git & Code Quality (lightweight)

Quick checks only — NOT a full /review:

1. Commit count and format:

   git -C {project_root} log --oneline | wc -l
   git -C {project_root} log --oneline | head -30
   

   • Count commits with conventional format (feat:, fix:, chore:, test:, docs:, refactor:, build:, ci:, perf:)
   • Calculate: conventional_pct = conventional / total * 100

2. Committer breakdown:

   git -C {project_root} shortlog -sn --no-merges | head -10
   

3. Test status (if test command exists in CLAUDE.md or package.json):

   • Run test suite, capture pass/fail count
   • If no test command found, skip and note "no tests configured"

4. Build status (if build command exists):

   • Run build, capture success/fail
   • If no build command found, skip and note "no build configured"

Phase 6.5: Context Degradation Analysis

Check for signs of context window problems during the pipeline run:

1. Iteration quality curve: Compare early iterations vs late iterations.

   • Did error rates increase over time? (sign of context degradation)
   • Did the agent start repeating itself or losing track of the plan?

2. Observation masking usage: Check if scratch/ directory exists in project root.

   • If yes: good — agent was offloading large outputs
   • If no but iter logs show >100-line tool outputs: flag as waste source

3. Plan recitation evidence: In sampled iter logs, check if the agent re-read plan.md at task boundaries.

   • Absent recitation + task drift = context engineering gap

4. CLAUDE.md bloat: wc -c {project_root}/CLAUDE.md

   • 40,000 chars: WARN — attention dilution likely

   • 60,000 chars: RED — severe context budget pressure

Add findings to the report under ## Context Health:

## Context Health
- Iteration quality trend: {STABLE / DEGRADING / N/A}
- Observation masking: {USED / NOT USED / N/A}
- Plan recitation: {OBSERVED / ABSENT / N/A}
- CLAUDE.md size: {N} chars — {OK / WARN / BLOATED}

Phase 7: Score & Report

Load scoring rubric from ${CLAUDE_PLUGIN_ROOT}/skills/retro/references/eval-dimensions.md. If plugin root not available, use the embedded weights:

Scoring weights:

• Efficiency (waste %): 25%
• Stability (restarts): 20%
• Fidelity (criteria met): 20%
• Quality (test pass rate): 15%
• Commits (conventional %): 5%
• Docs (plan staleness): 5%
• Signals (clean signals): 5%
• Speed (duration per track, normalized): 5%

Note: In fallback mode (no pipeline logs), redistribute Efficiency and Stability weights to Fidelity, Quality, and Commits.

Generate report at {project_root}/docs/retro/{date}-retro.md:

# Pipeline Retro: {project} ({date})

## Overall Score: {N}/10

## Pipeline Efficiency

| Metric | Value | Rating |
|--------|-------|--------|
| Total iterations | {N} | |
| Productive iterations | {N} ({pct}%) | {emoji} |
| Wasted iterations | {N} ({pct}%) | {emoji} |
| Pipeline restarts | {N} | {emoji} |
| Max-iter hits | {N} | {emoji} |
| Total duration | {time} | {emoji} |
| Tracks completed | {N} | |
| Duration per track | {time/tracks} | {emoji} |

## Per-Stage Breakdown

| Stage | Attempts | Successes | Waste % | Notes |
|-------|----------|-----------|---------|-------|
| scaffold | | | | |
| setup | | | | |
| plan | | | | |
| build | | | | |
| deploy | | | | |
| review | | | | |

## Failure Patterns

### Pattern 1: {name}
- **Occurrences:** {N} iterations
- **Root cause:** {analysis}
- **Wasted:** {N} iterations
- **Fix:** {concrete suggestion with file reference}

### Pattern 2: ...

## Plan Fidelity

| Track | Criteria Met | Tasks Done | SHAs | Rating |
|-------|-------------|------------|------|--------|
| {track-id} | {N}% | {N}% | {yes/no} | {emoji} |

## Code Quality (Quick)

- **Tests:** {N} pass, {N} fail (or "not configured")
- **Build:** PASS / FAIL (or "not configured")
- **Commits:** {N} total, {pct}% conventional format

## Three-Axis Growth

| Axis | Score | Evidence |
|------|-------|----------|
| **Technical** (code, tools, architecture) | {0-10} | {what changed} |
| **Cognitive** (understanding, strategy, decisions) | {0-10} | {what improved} |
| **Process** (harness, skills, pipeline, docs) | {0-10} | {what evolved} |

If only one axis is served — note what's missing.

## Recommendations

1. **[CRITICAL]** {patch suggestion with file:line reference}
2. **[HIGH]** {improvement}
3. **[MEDIUM]** {optimization}
4. **[LOW]** {nice-to-have}

## Suggested Patches

### Patch 1: {file} — {description}

**What:** {one-line description}
**Why:** {root cause reference from Failure Patterns}

\```diff
- old line
+ new line
\```

Rating guide (use these emojis):

• GREEN = excellent
• YELLOW = acceptable
• RED = needs attention

Phase 8: Interactive Patching

After generating the report:

1. Show summary to user: overall score, top 3 failure patterns, top 3 recommendations

2. For each suggested patch (if any), use AskUserQuestion:

   • Question: "Apply patch to {file}? {one-line description}"
   • Options: "Apply" / "Skip" / "Show diff first"

3. If "Show diff first": display the full diff, then ask again (Apply / Skip)

4. If "Apply": use Edit tool to apply the change directly

5. After all patches processed:

   • If any patches were applied: suggest committing with fix(retro): {description}
   • Do NOT auto-commit — just suggest the command

Phase 9: CLAUDE.md Revision

After patching, revise the project's CLAUDE.md to keep it lean and useful for future agents.

Steps:

1. Read CLAUDE.md and check size: wc -c CLAUDE.md
2. Add learnings from this retro:
   • Pipeline failure patterns worth remembering (avoid next time)
   • New workflow rules or process improvements
   • Updated commands or tooling changes
   • Architecture decisions that emerged during the pipeline run
3. If over 40,000 characters — trim ruthlessly:
   • Collapse completed phase/milestone histories into one line each
   • Remove verbose explanations — keep terse, actionable notes
   • Remove duplicate info (same thing explained in multiple sections)
   • Remove historical migration notes, old debugging context
   • Remove examples that are obvious from code or covered by skill/doc files
   • Remove outdated troubleshooting for resolved issues
4. Verify result <= 40,000 characters — if still over, cut least actionable content
5. Write updated CLAUDE.md, update "Last updated" date

Priority (keep -> cut):

1. ALWAYS KEEP: Tech stack, directory structure, Do/Don't rules, common commands, architecture decisions
2. KEEP: Workflow instructions, troubleshooting for active issues, key file references
3. CONDENSE: Phase histories (one line each), detailed examples, tool/MCP listings
4. CUT FIRST: Historical notes, verbose explanations, duplicated content, resolved issues

Rules:

• Never remove Do/Don't sections — critical guardrails
• Preserve overall section structure and ordering
• Every line must earn its place: "would a future agent need this to do their job?"
• Commit the update: git add CLAUDE.md && git commit -m "docs: revise CLAUDE.md (post-retro)"

Phase 10: Factory Critic (optional)

Run this phase only if ${CLAUDE_PLUGIN_ROOT} is available (i.e., solo-factory is installed). Skip if running as a standalone skill without the factory context.

After evaluating the project pipeline, step back and evaluate the factory itself — the skills, scripts, and pipeline logic that produced this result. Be a harsh critic.

What to evaluate:

1. Read the skills that were invoked in this pipeline run (from INVOKE lines in pipeline.log):

   • For each skill: ${CLAUDE_PLUGIN_ROOT}/skills/{stage}/SKILL.md
   • Did the skill have the right instructions for this project's needs?
   • Did it miss context it should have had?

2. Read pipeline script signal handling and stage logic:

   • ${CLAUDE_PLUGIN_ROOT}/scripts/solo-dev.sh
   • Were there structural issues (wrong stage order, missing re-exec, broken redo)?

3. Cross-reference with failure patterns from Phase 3:

   • For each failure: was the root cause in the skill, the script, or the project?
   • Skills that caused waste = factory defects

Score the factory (not the project):

Factory Score: {N}/10

Skill quality:
- {skill}: {score}/10 — {why}
- {skill}: {score}/10 — {why}

Pipeline reliability: {N}/10 — {why}

Missing capabilities:
- {what the factory couldn't do that it should have}

Top factory defects:
1. {defect} → {which file to fix} → {concrete fix}
2. {defect} → {which file to fix} → {concrete fix}

Harness Evolution — think about the bigger picture

After scoring the factory, step back further and think about the harness — the entire system that guides agents (CLAUDE.md, docs/, linters, skills, templates). Ask:

1. Context engineering: Did the agent have everything it needed in-repo? Or did it struggle because knowledge was missing / scattered / stale?

   • Missing docs -> add to docs/ or CLAUDE.md
   • Stale docs -> flag for doc-gardening
   • Knowledge only in your head -> encode it

2. Architectural constraints: Did the agent break module boundaries, produce inconsistent patterns, or ignore conventions?

   • Repeated boundary violations -> need a linter or structural test
   • Inconsistent patterns -> need golden principle in CLAUDE.md
   • Data shape errors -> need parse-at-boundary enforcement

3. Decision traces: What worked well that future agents should reuse? What failed that they should avoid?

   • Good patterns -> capture as precedent in docs or CLAUDE.md
   • Bad patterns -> encode as anti-pattern or lint rule
   • Think: "if another agent hits this same problem tomorrow, what should it find?"

4. Skill gaps: Which skills need better instructions? Which new skills should exist?

   • Skill that caused waste -> concrete SKILL.md patch
   • Missing capability -> new skill idea for evolution log

Write to evolution log:

Append findings to {project_root}/docs/evolution.md (create if not exists). If ~/.solo/evolution.md exists, append there as well for cross-project tracking.

## {YYYY-MM-DD} | {project} | Factory Score: {N}/10

Pipeline: {stages run} | Iters: {total} | Waste: {pct}%

### Defects
- **{severity}** | {skill/script}: {description}
  - Fix: {concrete file:change}

### Harness Gaps
- **Context:** {what knowledge was missing or stale for the agent}
- **Constraints:** {what boundary violations or inconsistencies occurred}
- **Precedents:** {patterns worth capturing for future agents — good or bad}

### Missing
- {capability the factory lacked}

### What worked well
- {skill/pattern that performed efficiently}

Rules:

• Be brutally honest — if a skill is broken, say so
• Every defect must have a concrete fix (file + what to change)
• Track what works well too — don't regress good patterns
• Keep entries compact — this file accumulates over time

Signal Output

Output signal: <solo:done/>

Important: /retro always outputs <solo:done/> — it never needs redo. Even if pipeline was terrible, the retro itself always completes.

Edge Cases

• No pipeline.log and no git history: abort with clear message — "No pipeline log or git history found. Nothing to analyze."
• No pipeline.log but git history exists: switch to fallback mode (see Fallback Analysis)
• Empty pipeline.log: report "Pipeline log is empty — was the pipeline cancelled before any iteration?"
• No iter logs: skip Phase 4 sampling, note in report
• No plan-done: skip Phase 5, note "No completed plans found"
• No test/build commands: skip those checks in Phase 6, note in report
• Pipeline still running: warn user — "State file exists, pipeline may still be running. Retro on partial data."

Reference Files

• ${CLAUDE_PLUGIN_ROOT}/skills/retro/references/eval-dimensions.md — scoring rubric (8 axes, weights)
• ${CLAUDE_PLUGIN_ROOT}/skills/retro/references/failure-catalog.md — known failure patterns and fixes