maestro-flow-one 0.1.0

package/claude/maestro-flow/commands/learn/follow.md

@@ -0,0 +1,167 @@
+---
+name: learn-follow
+description: Guided follow-along reading of code or wiki entries, extracting patterns and building understanding
+argument-hint: "<path|wiki-id|topic> [--depth shallow|deep] [--save-wiki]"
+allowed-tools:
+  - Read
+  - Write
+  - Bash
+  - Glob
+  - Grep
+  - Agent
+  - AskUserQuestion
+---
+<purpose>
+Guided reading experience for code files, wiki entries, or topics. Instead of just reading code, this command walks through content section by section using forcing questions (inspired by gstack `/office-hours` brainstorming) to extract patterns, identify assumptions, and build a structured understanding map.
+
+Outputs reading notes with extracted patterns, open questions, and connection points. Insights persist to `lessons.jsonl` and optionally become wiki note entries for the knowledge graph.
+</purpose>
+
+<context>
+Arguments: $ARGUMENTS
+
+**Target resolution (auto-detected from first argument):**
+- File path (e.g., `src/commands/wiki.ts`) → Read source file
+- Wiki ID (e.g., `spec-auth`, `phase-planning`) → Fetch via `maestro wiki get`
+- Topic string (e.g., `"authentication flow"`) → Search via `maestro wiki search`, use top result
+
+**Flags:**
+- `--depth shallow` — Quick pass: key patterns and structure only (default)
+- `--depth deep` — Thorough: every function, every branch, every assumption
+- `--save-wiki` — Create a wiki note entry with the reading notes via `maestro wiki create --type note`
+
+**Storage written:**
+- `.workflow/learning/follow-{slug}-{YYYY-MM-DD}.md` — Reading notes with understanding map
+- `.workflow/learning/lessons.jsonl` — Appended pattern/technique insights
+- `.workflow/learning/learning-index.json` — Updated index
+- If `--save-wiki`: new wiki note entry
+
+**Storage read:**
+- Target source file or wiki entry
+- `maestro wiki backlinks <id>` / `maestro wiki forward <id>` — Relationship context
+- `.workflow/specs/coding-conventions.md` — Convention reference for pattern matching
+- `.workflow/learning/lessons.jsonl` — Prior insights for dedup and cross-reference
+</context>
+
+<execution>
+
+### Stage 1: Resolve Target
+- If argument looks like a file path (contains `/` or `\`, or matches a glob): verify file exists via Read
+- If argument matches wiki ID pattern (`<type>-<slug>`): fetch via `maestro wiki get <id>` (offline mode)
+- Otherwise: treat as topic string, run `maestro wiki search "<topic>"`, take the top result. If no result, fall back to Grep across `src/` for the topic.
+- If target cannot be resolved, AskUserQuestion with suggestions.
+
+### Stage 2: Load Context Web
+For the resolved target, build a 1-hop context neighborhood:
+
+**If wiki entry:**
+- `maestro wiki forward <id>` — What this entry references
+- `maestro wiki backlinks <id>` — What references this entry
+- Read the body of top 3 related entries for context
+
+**If code file:**
+- Parse imports/requires to identify dependency files
+- Grep for exports used by other files (reverse dependencies)
+- Read the first 50 lines of top 3 dependent files for context
+
+**If directory:**
+- List files, identify entry points (index.ts, main.ts, cli.ts)
+- Build reading order: entry point → core modules → utilities → tests
+
+### Stage 3: Build Reading Order
+- For a single file: split into logical sections (function boundaries, class boundaries, export groups)
+- For a wiki entry: split by markdown headings
+- For a directory: order files by dependency (entry points first, leaf modules last)
+- For `--depth shallow`: limit to top-level structure (function signatures, section headers)
+- For `--depth deep`: include every function body, every branch
+
+### Stage 4: Guided Reading with Forcing Questions
+Walk through each section in reading order. For each section, apply 4 forcing questions:
+
+1. **"What pattern is being used here?"** — Identify design patterns, idioms, conventions. Compare against `coding-conventions.md`.
+2. **"Why this approach instead of alternatives?"** — What trade-offs were made? What was the simpler approach not chosen?
+3. **"What assumption does this depend on?"** — What must be true for this code/content to be correct? External state? Input shape? Ordering?
+4. **"What would break if this changed?"** — Fragility analysis. What downstream effects would a change have?
+
+Record answers as structured annotations per section.
+
+### Stage 5: Extract Patterns
+From the forcing question answers, extract:
+- **Design patterns**: named patterns with code anchors (file:line)
+- **Naming conventions**: how things are named and why
+- **Error handling approach**: how errors flow through this code/content
+- **Data flow**: how data enters, transforms, and exits
+- **Assumptions**: explicit and implicit assumptions identified
+
+Cross-reference each pattern against existing `coding-conventions.md` entries:
+- Already documented → note as "confirmed convention"
+- Not documented → flag as "undocumented pattern" (candidate for `spec-add`)
+
+### Stage 6: Produce Understanding Map
+Build a structured summary document:
+
+```markdown
+# Follow-Along: {target name}
+
+## Key Concepts
+- {concept}: {one-line explanation}
+
+## Patterns Identified
+| Pattern | Location | Convention Status |
+|---------|----------|-------------------|
+| {name} | {file:line} | documented / undocumented |
+
+## Assumptions
+- {assumption}: {what depends on it}
+
+## Open Questions
+- {question}: {why it matters}
+
+## Connections
+- Links to: {forward link entries}
+- Referenced by: {backlink entries}
+- Related lessons: {matching lessons.jsonl entries}
+```
+
+### Stage 7: Persist
+1. Write `.workflow/learning/follow-{slug}-{date}.md` with the understanding map
+2. Append each new pattern/technique as an insight to `lessons.jsonl`:
+   - `source: "follow"`, `category: "pattern"` or `"technique"`
+   - Tags: `["follow", "{target-slug}"]`
+   - Stable INS-id from `hash("follow" + target + pattern_name)`
+3. Update `learning-index.json`
+4. If `--save-wiki`: run `maestro wiki create --type note --slug follow-{slug} --title "Follow-Along: {target}" --body-file .workflow/learning/follow-{slug}-{date}.md`
+5. Display summary with key findings and next steps
+
+**Next-step routing:**
+- Deep dive into a discovered pattern → `/learn-decompose <path>`
+- Add undocumented pattern to specs → `/spec-add coding <description>`
+- Get second opinion on a finding → `/learn-second-opinion <file>`
+- Browse related wiki entries → `/wiki-digest <topic>`
+</execution>
+
+<error_codes>
+| Code | Severity | Condition | Recovery |
+|------|----------|-----------|----------|
+| E001 | error | Target not resolvable (file not found, wiki ID not found, search returned 0) | Check path/ID, or rephrase topic for search |
+| E002 | error | `.workflow/` not initialized | Run `/maestro-init` first |
+| W001 | warning | Wiki graph unavailable (no .workflow/ wiki entries) — skipping context web | Proceed with code-only context (imports/exports) |
+| W002 | warning | coding-conventions.md not found — skipping convention comparison | Patterns flagged as "unknown convention status" |
+| W003 | warning | Target is very large (>1000 lines) — auto-switching to shallow depth | Use --depth deep to override |
+</error_codes>
+
+<success_criteria>
+- [ ] Target resolved to concrete content (file, wiki entry, or search result)
+- [ ] Context web loaded (forward/backlinks for wiki, imports/exports for code)
+- [ ] Reading order established (sections/files ordered logically)
+- [ ] All 4 forcing questions applied per section
+- [ ] Patterns extracted with file:line anchors
+- [ ] Convention comparison performed against coding-conventions.md
+- [ ] Understanding map produced with: concepts, patterns, assumptions, questions, connections
+- [ ] `follow-{slug}-{date}.md` written
+- [ ] `lessons.jsonl` appended with discovered patterns (stable INS-ids)
+- [ ] `learning-index.json` updated
+- [ ] If --save-wiki: wiki note entry created
+- [ ] No files modified outside `.workflow/learning/` (and optionally wiki)
+- [ ] Summary displayed with next-step routing
+</success_criteria>

package/claude/maestro-flow/commands/learn/investigate.md

@@ -0,0 +1,221 @@
+---
+name: learn-investigate
+description: Systematic question investigation with hypothesis testing, evidence logging, and 3-strike escalation
+argument-hint: "<question> [--scope <path>] [--max-hypotheses N]"
+allowed-tools:
+  - Read
+  - Write
+  - Bash
+  - Glob
+  - Grep
+  - Agent
+  - AskUserQuestion
+---
+<purpose>
+Systematic investigation workflow for understanding questions (not bug-fixing). Inspired by gstack `/investigate` with its 4-phase approach, scope lock, and 3-strike escalation rule.
+
+Unlike `quality-debug` which is designed for fixing bugs during execution phases, this command is for answering "how does X work?", "why does Y happen?", "what would happen if Z?" questions. It produces structured evidence trails and understanding documents that persist to the learning system.
+</purpose>
+
+<context>
+Arguments: $ARGUMENTS
+
+**Target:** First argument is the question or topic to investigate (quoted string or keywords).
+
+**Flags:**
+- `--scope <path>` — Restrict investigation to files under this directory (default: entire project)
+- `--max-hypotheses N` — Maximum hypotheses to test before escalating (default: 3)
+
+**Storage written:**
+- `.workflow/learning/investigate-{slug}/evidence.ndjson` — Structured evidence log (one JSON line per evidence)
+- `.workflow/learning/investigate-{slug}/understanding.md` — Evolving understanding document
+- `.workflow/learning/investigate-{slug}/report.md` — Final investigation report
+- `.workflow/learning/lessons.jsonl` — Investigation findings as insights (source: "investigate")
+- `.workflow/learning/learning-index.json` — Updated index
+
+**Storage read:**
+- Source files within scope
+- `maestro wiki search "<question>"` — Prior knowledge about the topic
+- `.workflow/learning/lessons.jsonl` — Prior related investigations
+- `.workflow/specs/debug-notes.md` — Known gotchas and patterns
+- `.workflow/codebase/architecture.md` — Structural context (if exists)
+</context>
+
+<execution>
+
+### Stage 1: Frame the Question
+- Parse question from arguments
+- Determine scope (--scope or full project)
+- Generate investigation slug from question keywords
+- Create `.workflow/learning/investigate-{slug}/` directory
+- Search prior knowledge:
+  - `maestro wiki search "<question>"` for related entries
+  - Grep `lessons.jsonl` for related insights
+  - Read `debug-notes.md` for known gotchas
+
+Write initial `understanding.md`:
+```markdown
+# Investigation: {question}
+## Initial Understanding
+- Prior knowledge: {summary of wiki/lessons findings}
+- Scope: {path or "full project"}
+- Started: {timestamp}
+```
+
+### Stage 2: Evidence Collection
+Systematically gather evidence related to the question:
+
+1. **Code search**: Grep for keywords from the question across the scoped files
+2. **File inspection**: Read the most relevant files identified by search
+3. **Import/dependency tracing**: Follow imports to understand the dependency chain
+4. **Git history**: `git log --oneline -10 -- <relevant-files>` for recent changes
+
+For each piece of evidence, append to `evidence.ndjson`:
+```json
+{"ts": "ISO", "type": "code|git|search|doc", "source": "file:line", "relevance": "high|medium|low", "content": "...", "note": "why this matters"}
+```
+
+### Stage 3: Pattern Matching
+Compare collected evidence against known patterns:
+- Check `debug-notes.md` entries for matching situations
+- Check `lessons.jsonl` for related technique/pattern/gotcha entries
+- Identify: does this match a documented pattern, or is it novel?
+
+Update `understanding.md` with pattern analysis section.
+
+### Stage 4: Hypothesis Formation
+From evidence and patterns, generate ranked hypotheses:
+- Each hypothesis: a specific, testable claim about "how/why"
+- Rank by plausibility (evidence strength)
+- Write hypotheses to `understanding.md`
+
+```markdown
+## Hypotheses
+1. **[HIGH]** {hypothesis 1} — Evidence: {refs}
+2. **[MEDIUM]** {hypothesis 2} — Evidence: {refs}
+3. **[LOW]** {hypothesis 3} — Evidence: {refs}
+```
+
+### Stage 4.5: CLI Supplementary Exploration (optional)
+
+**Skip if** no enabled CLI tools or hypotheses are trivially testable.
+
+```
+IF no CLI tools enabled: skip to Stage 5
+
+hypothesis_summary = hypotheses.map(h => "${h.rank}: ${h.claim}").join("\n")
+
+Bash({
+  command: 'maestro delegate "PURPOSE: Gather evidence for investigation hypotheses
+TASK: For each hypothesis, trace relevant call chains and data flows | Find corroborating or contradicting code patterns
+MODE: analysis
+CONTEXT: @${scope_path}/**/*
+EXPECTED: JSON array of { hypothesis_rank, evidence: [{ file, line, supports: bool, explanation }] }
+CONSTRAINTS: Focus on code-level evidence only | Max 5 evidence items per hypothesis
+
+Hypotheses:
+${hypothesis_summary}
+" --role explore --mode analysis',
+  run_in_background: true
+})
+```
+
+**On callback:** Parse result, append each evidence item to `evidence.ndjson` with `type: "cli-exploration"`. Pass as supplementary context to Stage 5 testing.
+
+### Stage 5: Hypothesis Testing
+For each hypothesis (in rank order):
+
+1. **Design test**: What specific evidence would confirm or disprove this?
+2. **Execute test**: Code trace, targeted search, data inspection, or experiment
+3. **Record result**: Append to `evidence.ndjson` with `type: "test"`
+4. **Update understanding**: Mark hypothesis as confirmed / disproved / inconclusive
+
+```markdown
+## Test Results
+### Hypothesis 1: {claim}
+- Test: {what was done}
+- Result: CONFIRMED / DISPROVED / INCONCLUSIVE
+- Evidence: {file:line references}
+```
+
+### Stage 6: 3-Strike Escalation
+If `--max-hypotheses` hypotheses all fail:
+
+1. **Broaden scope**: If scope was restricted, suggest expanding. AskUserQuestion:
+   ```
+   {N} hypotheses tested, none confirmed.
+   A) Broaden scope to full project
+   B) I have a new hypothesis: [user provides]
+   C) Escalate — this needs deeper investigation
+   ```
+2. **Search wiki for clues**: `maestro wiki search` with alternative keywords
+3. **If still stuck**: Mark as INCONCLUSIVE with what was learned and what remains unknown
+
+### Stage 7: Synthesize & Report
+Write final `report.md`:
+
+```markdown
+# Investigation Report: {question}
+
+## Answer
+{confirmed understanding or "INCONCLUSIVE: ..."}
+
+## Evidence Trail
+| # | Type | Source | Relevance | Finding |
+|---|------|--------|-----------|---------|
+| 1 | code | file:line | high | ... |
+
+## Hypotheses Tested
+| Hypothesis | Result | Key Evidence |
+|-----------|--------|-------------|
+| ... | confirmed/disproved | file:line |
+
+## Key Learnings
+- {learning 1}
+- {learning 2}
+
+## Open Questions
+- {what remains unknown}
+```
+
+### Stage 8: Persist
+1. Append findings to `lessons.jsonl`:
+   - Confirmed hypotheses → `category: "technique"` or `"pattern"`
+   - Disproved hypotheses → `category: "gotcha"` (what looked true but wasn't)
+   - `source: "investigate"`, tags: `["investigate", "{question-slug}"]`
+   - Stable INS-id from `hash("investigate" + question + finding_title)`
+2. Update `learning-index.json`
+3. Display summary with answer and next steps
+
+**Next-step routing:**
+- Save finding to specs → `/spec-add debug <finding>`
+- Follow-along on discovered code → `/learn-follow <path>`
+- Decompose patterns found → `/learn-decompose <module>`
+- Create wiki entry for understanding → `maestro wiki create --type note`
+</execution>
+
+<error_codes>
+| Code | Severity | Condition | Recovery |
+|------|----------|-----------|----------|
+| E001 | error | No question provided | Provide a question as the first argument |
+| E002 | error | Scope path does not exist | Check --scope path is valid |
+| W001 | warning | No prior knowledge found in wiki/lessons | Proceed with fresh investigation |
+| W002 | warning | Evidence collection found very few matches (<3) | Broaden search terms or expand scope |
+| W003 | warning | All hypotheses inconclusive — escalating | Investigation marked INCONCLUSIVE |
+</error_codes>
+
+<success_criteria>
+- [ ] Question parsed and investigation slug generated
+- [ ] Investigation directory created under `.workflow/learning/`
+- [ ] Prior knowledge loaded from wiki and lessons
+- [ ] Evidence collected and logged to `evidence.ndjson` (structured NDJSON)
+- [ ] Pattern matching performed against debug-notes and lessons
+- [ ] At least 1 hypothesis formed and tested
+- [ ] `understanding.md` tracks evolving understanding with timestamps
+- [ ] `report.md` written with answer, evidence trail, hypothesis results
+- [ ] Findings appended to `lessons.jsonl` with stable INS-ids
+- [ ] `learning-index.json` updated
+- [ ] 3-strike escalation triggered if all hypotheses fail
+- [ ] No files modified outside `.workflow/learning/`
+- [ ] Summary displayed with answer and next-step routing
+</success_criteria>

package/claude/maestro-flow/commands/learn/retro.md

@@ -0,0 +1,303 @@
+---
+name: learn-retro
+description: Unified retrospective — git activity metrics and decision quality evaluation with lens-based selection
+argument-hint: "[--lens git|decision|all] [--days N] [--author <name>] [--area <path>] [--phase N] [--tag <tag>] [--id <id>] [--compare]"
+allowed-tools:
+  - Read
+  - Write
+  - Bash
+  - Glob
+  - Grep
+  - Agent
+  - AskUserQuestion
+---
+<purpose>
+Unified retrospective that combines git activity analysis and decision quality evaluation into a single command with lens-based selection. Works on raw git history and wiki/spec data — does not require completed phase artifacts (unlike `quality-retrospective`).
+
+Two lenses, usable independently or together:
+- **git**: Commit metrics, session detection, per-author breakdown, file hotspots, trend tracking
+- **decision**: Decision tracing across wiki/specs/git, multi-perspective evaluation, lifecycle classification
+
+All insights persist to `.workflow/learning/lessons.jsonl` for cross-session queryability via `manage-learn`.
+</purpose>
+
+<context>
+Arguments: $ARGUMENTS
+
+**Lens selection:**
+- `--lens git` — Git activity retrospective only
+- `--lens decision` — Decision evaluation only
+- `--lens all` — Both lenses (default)
+
+**Git lens flags:**
+- `--days N` — Time window in days (default: 7)
+- `--author <name>` — Filter commits by author name (substring match)
+- `--area <path>` — Scope to files under a specific directory
+- `--compare` — Compare against the previous retro report if one exists
+
+**Decision lens flags:**
+- `--phase N` — Decisions from phase N's context and related specs
+- `--tag <tag>` — Decisions tagged with specific tag in wiki/specs
+- `--id <id>` — Single decision by wiki ID or lessons.jsonl INS-id
+
+**Storage written:**
+- `.workflow/learning/retro-{YYYY-MM-DD}.md` — Unified human-readable report
+- `.workflow/learning/retro-{YYYY-MM-DD}.json` — Structured metrics (machine-readable)
+- `.workflow/learning/lessons.jsonl` — Appended insights (source: "retro-git" or "retro-decision")
+- `.workflow/learning/learning-index.json` — Updated index
+
+**Storage read:**
+- `.workflow/state.json` — Current phase context (optional)
+- `.workflow/learning/retro-*.json` — Prior retro for trend comparison
+- `.workflow/learning/lessons.jsonl` — Existing insights for dedup
+- `maestro wiki list --type spec --json` — Spec entries (decision lens)
+- `.workflow/specs/architecture-constraints.md` — Documented architectural decisions (decision lens)
+- Phase context with Locked/Free/Deferred decisions (decision lens) — resolve via `state.json.artifacts[]` scratch paths
+</context>
+
+<execution>
+
+### Stage 1: Parse Arguments & Select Lenses
+- Parse `--lens` flag: `git`, `decision`, or `all` (default: `all`)
+- Extract lens-specific flags
+- Check `.workflow/learning/` exists; bootstrap if missing
+
+Display banner:
+```
+============================================================
+  LEARN RETRO
+============================================================
+  Lens:    {git | decision | all}
+  Scope:   {days/author/area for git} {phase/tag/id for decision}
+```
+
+---
+
+### Stage 2: Git Lens (skip if --lens decision)
+
+#### 2a: Gather Raw Data (parallel git commands)
+Run ALL these git commands in parallel:
+
+```bash
+# Commit stats with author, timestamp, subject, files changed
+git log --since="<start-date>T00:00:00" --format="%H|%aN|%ae|%ai|%s" --shortstat
+
+# Per-commit numstat for test vs production LOC split
+git log --since="<start-date>T00:00:00" --format="COMMIT:%H|%aN" --numstat
+
+# Timestamps for session detection (sorted)
+git log --since="<start-date>T00:00:00" --format="%at|%aN|%ai|%s" | sort -n
+
+# File hotspots (most frequently changed files)
+git log --since="<start-date>T00:00:00" --format="" --name-only | grep -v '^$' | sort | uniq -c | sort -rn | head -20
+
+# Per-author commit counts
+git shortlog --since="<start-date>T00:00:00" -sn --no-merges
+```
+
+Apply `--author` and `--area` filters if provided.
+
+#### 2b: Compute Metrics
+| Metric | Computation |
+|--------|-------------|
+| Commits | Count of non-merge commits |
+| Contributors | Unique author count |
+| Total insertions / deletions | Sum from shortstat |
+| Net LOC | insertions - deletions |
+| Test LOC (insertions) | Sum insertions for test files from numstat |
+| Test ratio | test_insertions / total_insertions x 100% |
+| Churn rate | Files changed >2 times / total unique files |
+| Active days | Distinct dates with commits |
+
+#### 2c: Detect Work Sessions
+Cluster commits by >2hr gaps in timestamps:
+- Per session: start time, end time, duration, commit count, primary focus area
+- Compute: total sessions, avg session duration, avg LOC/session-hour
+
+#### 2d: Per-Author Breakdown
+For each author:
+- Commit count, LOC added/removed, top 3 file areas
+- Test ratio (their test LOC / their total LOC)
+- Session count and patterns
+
+#### 2e: Trend Comparison (if --compare or prior report exists)
+- Find most recent `.workflow/learning/retro-*.json`
+- Compute deltas: commits, LOC, test ratio, churn rate, session count
+- Flag significant changes (>20% delta) as trend highlights
+
+#### 2f: Distill Git Insights
+- **High churn files** (changed >3 times): instability signal
+- **Low test ratio areas** (<20%): testing gap
+- **Session patterns**: scattered vs deep sessions
+- **Area drift**: commits not aligned with current roadmap phase
+
+Each insight: title, description, category (pattern/antipattern/technique), tags, confidence.
+
+---
+
+### Stage 3: Decision Lens (skip if --lens git)
+
+#### 3a: Collect Decisions (parallel)
+```bash
+maestro wiki search "decision" --json
+maestro wiki list --type spec --json
+git log --oneline --all --grep="decision\|chose\|decided\|architecture" -20
+```
+
+Also read:
+- `.workflow/specs/architecture-constraints.md` — grep for `<spec-entry category="arch"` blocks
+- Phase context files — resolve via `state.json.artifacts[]` scratch paths — scan for "Locked:", "Deferred:" sections
+- `.workflow/learning/lessons.jsonl` — filter `category == "decision"`
+
+Apply scope filter (--phase, --tag, --id).
+
+#### 3b: Build Decision Registry
+Per decision:
+```json
+{
+  "id": "source id",
+  "title": "what was decided",
+  "source": "wiki|spec|phase-context|lesson|git",
+  "date": "when decided",
+  "rationale": "why",
+  "alternatives": "what was considered",
+  "phase": "which phase",
+  "implementation_evidence": ["file paths from git"]
+}
+```
+
+#### 3c: Multi-Perspective Evaluation
+Spawn 3 Agents in a single message:
+
+**Agent 1 — Technical Soundness:**
+- Does implementation match stated intent?
+- Has technical context changed since decision was made?
+- Grade: sound / degraded / violated
+
+**Agent 2 — Cost Assessment:**
+- What complexity did this decision add?
+- Is it creating coupling or tech debt?
+- Grade: low-cost / acceptable / expensive / debt-creating
+
+**Agent 3 — Alternative Hindsight:**
+- With what we know now, was this the right call?
+- Would reversing be feasible?
+- Grade: confirmed / questionable / should-revisit
+
+#### 3d: Classify Decision Lifecycle
+| Status | Criteria |
+|--------|---------|
+| **Validated** | Sound + Low/Acceptable cost + Confirmed |
+| **Aging** | Sound but Expensive + Confirmed |
+| **Questionable** | Degraded or Violated + Questionable |
+| **Stale** | Any + Should-revisit |
+| **Reversed** | Code contradicts the decision |
+
+#### 3e: Generate Recommendations
+- **Aging**: flag for tech debt review
+- **Questionable**: create issue for investigation
+- **Stale**: suggest decision refresh
+- **Reversed**: suggest documenting the reversal
+
+---
+
+### Stage 4: Unified Report
+
+Write `.workflow/learning/retro-{date}.md`:
+
+```markdown
+# Retrospective: {date}
+**Lenses:** {active lenses} | **Period:** {days}d
+
+## Git Activity  (if git lens active)
+### Metrics
+| Metric | Value | Trend |
+|--------|-------|-------|
+| Commits | N | +/-% |
+| ...
+
+### Work Sessions
+{session timeline}
+
+### File Hotspots
+{top 10 most-changed files}
+
+### Per-Author
+{author breakdown table}
+
+## Decision Health  (if decision lens active)
+### Dashboard
+| Status | Count | Decisions |
+|--------|-------|-----------|
+| Validated | N | {list} |
+| Aging | N | {list} |
+| ...
+
+### Per-Decision Evaluation
+{detailed evaluations}
+
+## Combined Insights
+{merged insights from both lenses, deduplicated}
+
+## Recommended Actions
+1. {action}: {reason}
+```
+
+Write `.workflow/learning/retro-{date}.json` with structured data.
+
+---
+
+### Stage 5: Persist
+1. Write report files
+2. Append insights to `lessons.jsonl`:
+   - Git insights: `source: "retro-git"`, `category` per insight type
+   - Decision insights: `source: "retro-decision"`, `category: "decision"`
+   - Stable INS-id from `hash(lens + metric_or_decision + date)`
+3. Update `learning-index.json`
+4. Display summary
+
+**Next-step routing:**
+- Browse insights → `/manage-learn list --tag retro`
+- Deep dive on high-churn file → `/learn-follow <path>`
+- Fix test gaps → `/quality-auto-test <area>`
+- Create issue for questionable decision → `/manage-issue create ...`
+- Investigate stale decision → `/learn-investigate <question>`
+- Full phase retrospective → `/quality-retrospective`
+</execution>
+
+<error_codes>
+| Code | Severity | Condition | Recovery |
+|------|----------|-----------|----------|
+| E001 | error | Not inside a git repository (git lens) | Navigate to a git repo directory |
+| E002 | error | No commits found in time window (git lens) | Increase --days or check filters |
+| E003 | error | No decisions found in any source (decision lens) | Check wiki/specs content, or provide --id |
+| E004 | error | --id not found in wiki or lessons (decision lens) | Verify the decision ID exists |
+| W001 | warning | `.workflow/learning/` not found, bootstrapping | Auto-created; proceed normally |
+| W002 | warning | No prior retro report for comparison | Skip trend section; first retro establishes baseline |
+| W003 | warning | One perspective agent failed — partial evaluation (decision lens) | Proceed with available perspectives |
+| W004 | warning | No git implementation evidence for a decision | Evaluation is theoretical only |
+| W005 | warning | Phase context files not found (decision lens) | Skip phase-context decisions |
+</error_codes>
+
+<success_criteria>
+- [ ] Lens selection parsed correctly (git / decision / all)
+- [ ] Git lens (if active):
+  - [ ] All git commands executed successfully
+  - [ ] Metrics computed: commits, LOC, test ratio, churn rate, sessions
+  - [ ] Sessions detected with >2hr gap clustering
+  - [ ] Per-author breakdown generated
+  - [ ] Trend comparison computed if prior report exists
+  - [ ] At least 1 actionable insight distilled
+- [ ] Decision lens (if active):
+  - [ ] Decisions collected from available sources
+  - [ ] Scope filter applied correctly
+  - [ ] 3 perspective agents spawned in parallel
+  - [ ] Each decision classified by lifecycle status
+  - [ ] Recommendations generated for non-Validated decisions
+- [ ] Unified report written to `retro-{date}.md`
+- [ ] Structured data written to `retro-{date}.json`
+- [ ] `lessons.jsonl` appended with insights (stable INS-ids)
+- [ ] `learning-index.json` updated
+- [ ] No files modified outside `.workflow/learning/`
+- [ ] Summary displayed with next-step routing
+</success_criteria>