maestro-flow-one 0.2.0 → 0.2.2

package/maestro-flow/commands/lifecycle/collab.md

@@ -0,0 +1,333 @@
+---
+name: maestro-collab
+description: Multi-CLI collaborative analysis -- fan-out to multiple CLI tools, cross-verify, synthesize
+argument-hint: "\"<requirement>\" [--tools gemini,qwen,claude] [--mode analysis|write] [--rule <template>] [-y]"
+allowed-tools:
+  - Read
+  - Write
+  - Bash
+  - Glob
+  - Grep
+  - Agent
+  - AskUserQuestion
+---
+
+<purpose>
+Multi-CLI collaboration: fan-out the same requirement to multiple CLI tools in parallel, cross-verify outputs for consensus/conflicts, then synthesize into a unified report with standard downstream artifacts (context.md + conclusions.json).
+
+Each CLI tool independently analyzes the requirement. Results are compared and merged via evidence-weighted synthesis.
+</purpose>
+
+<context>
+$ARGUMENTS — requirement text and optional flags.
+
+```bash
+/maestro-collab "analyze the auth module for security vulnerabilities"
+/maestro-collab "design a caching strategy" --tools gemini,qwen,claude
+/maestro-collab -y "review error handling patterns"
+/maestro-collab "refactor user service" --mode write --tools gemini,claude
+```
+
+**Flags**:
+- `--tools <list>`: Comma-separated CLI tools (default: auto-select first 3 enabled)
+- `--mode analysis|write`: Delegate mode (default: analysis)
+- `--rule <template>`: Shared rule template for all delegates
+- `-y` / `--yes`: Skip plan confirmation
+
+**Output**: `.workflow/scratch/{YYYYMMDD}-collab-{slug}/`
+- `collab-report.md` — full collaboration report
+- `context.md` — standard Locked/Free/Deferred decisions (plan/analyze compatible)
+- `conclusions.json` — structured conclusions (plan fast-track compatible)
+- `per-tool/{tool}-output.md` — raw per-tool outputs
+</context>
+
+<execution>
+
+### Step 1: Parse Arguments
+
+Extract from `$ARGUMENTS`:
+- `requirement` — remaining text after flag removal (error if empty)
+- `--tools` → `selectedTools` (comma-split)
+- `--mode` → `delegateMode` (default: `analysis`)
+- `--rule` → `ruleTemplate`
+- `-y` / `--yes` → `autoYes`
+
+### Step 2: Discover Available CLI Tools
+
+```bash
+Bash("maestro tools list --json 2>/dev/null || cat ~/.maestro/cli-tools.json")
+```
+
+Parse tool entries. Build eligible list:
+- `enabled == true`
+- If `--mode write`: exclude `type == "api-endpoint"`
+
+Auto-select (when `--tools` omitted): first 3 eligible in config order.
+Validate: minimum 2 eligible tools (abort if fewer).
+
+### Step 3: Present Collaboration Plan
+
+**(Skip if `-y`)**
+
+Display plan, then ask user:
+
+```
+============================================================
+  COLLABORATION PLAN
+============================================================
+  Requirement: {requirement}
+  Mode:        {delegateMode}
+  Rule:        {ruleTemplate || "none"}
+
+  Available CLI Tools (from cli-tools.json):
+    [✓] gemini    — gemini-3.1-pro-preview     [fullstack, frontend]
+    [✓] claude    — claude-sonnet-4-6           [fullstack]
+    [✓] codex     — gpt-5.5                    [fullstack, backend]
+    [ ] opencode  — (no model)                  [fullstack]
+
+  Selected: gemini, claude, codex (3 tools)
+
+  Pipeline:
+    1. Fan-out → parallel delegate to each tool
+    2. Cross-verification → consensus/conflict analysis
+    3. Synthesis → context.md + conclusions.json
+============================================================
+```
+
+Use `AskUserQuestion` with options:
+- **执行** — proceed with selected tools
+- **修改工具选择** — let user specify different tool combination
+- **取消** — abort
+
+If **修改工具选择**: ask user which tools to use (show eligible list), validate ≥ 2, re-display plan.
+
+### Step 4: Setup Session
+
+```
+slug = requirement kebab-cased, max 40 chars
+outputDir = .workflow/scratch/{YYYYMMDD}-collab-{slug}/
+```
+
+Create `outputDir` + `outputDir/per-tool/`.
+
+### Step 5: Build Delegate Prompt
+
+Shared prompt for all tools:
+
+```
+PURPOSE: {requirement}; success = actionable findings with evidence
+TASK: {auto-decomposed into 3-5 specific verbs}
+MODE: {delegateMode}
+CONTEXT: @**/*
+EXPECTED: Structured findings with file:line references, confidence score (0-100), prioritized recommendations. Sections: ## Findings, ## Recommendations, ## Confidence
+CONSTRAINTS: {extracted from requirement}
+```
+
+### Step 6: Parallel Fan-Out
+
+Launch ALL delegate calls simultaneously using multiple `Bash(run_in_background: true)` in a **single message**:
+
+```
+// Launch all in ONE message — do NOT wait between calls
+Bash({
+  command: `maestro delegate "${prompt}" --to gemini --mode ${mode} ${rule}`,
+  run_in_background: true
+})
+Bash({
+  command: `maestro delegate "${prompt}" --to claude --mode ${mode} ${rule}`,
+  run_in_background: true
+})
+Bash({
+  command: `maestro delegate "${prompt}" --to codex --mode ${mode} ${rule}`,
+  run_in_background: true
+})
+```
+
+**After launching all calls → STOP immediately. Do not output anything. Wait for background completion callbacks.**
+
+### Step 7: Collect Results
+
+As each background callback arrives:
+1. Extract exec ID from output (`[MAESTRO_EXEC_ID=...]`)
+2. Run `maestro delegate output <id>` to get full result
+3. Write raw output to `per-tool/{tool}-output.md`
+
+**Wait until ALL callbacks have arrived before proceeding.**
+
+### Step 8: Cross-Verify
+
+Read all `per-tool/{tool}-output.md` files. Compare findings across tools:
+
+For each finding, classify:
+- **[CONSENSUS]**: 2+ tools agree on same finding/recommendation
+- **[CONFLICT]**: Tools disagree on approach or assessment
+- **[UNIQUE]**: Finding from only one tool
+
+Compute `consensus_level = (consensus_count / total_findings) * 100`.
+
+### Step 9: Synthesize Outputs
+
+Resolve conflicts via evidence-weighted voting:
+- Higher confidence tool's position wins
+- More specific evidence (file:line refs) wins over general statements
+- If tied: mark as `[SUGGESTED]`
+
+Generate three output files:
+
+#### collab-report.md
+
+```markdown
+# Multi-CLI Collaboration Report — {requirement}
+
+## Summary
+- Tools: {tool_list}
+- Consensus level: {N}%
+- Key finding: {top finding}
+
+## Consensus Findings
+{findings agreed by 2+ tools}
+
+## Resolved Conflicts
+{conflicts resolved with rationale and winning tool}
+
+## Unresolved Items
+{items requiring human judgment}
+
+## Unique Insights
+{valuable unique findings with source tool attribution}
+
+## Recommendations
+{prioritized, merged recommendations}
+
+## Per-Tool Confidence
+| Tool | Confidence | Key Strength |
+|------|-----------|--------------|
+```
+
+#### context.md (standard downstream format)
+
+```markdown
+# Context: {requirement}
+
+**Date**: {date}
+**Mode**: collab ({tool_list})
+**Consensus Level**: {N}%
+
+## Decisions
+
+### Decision N: {TITLE}
+- **Context**: {what and why}
+- **Options**: 1. {opt1} 2. {opt2}
+- **Chosen**: {selected}
+- **Reason**: {rationale — which tools agreed/disagreed}
+
+## Constraints
+
+### Locked
+{[CONSENSUS] items — treat as confirmed decisions}
+
+### Free
+{[UNIQUE] items with strong evidence — implementer discretion}
+
+### Deferred
+{[UNRESOLVED] conflicts — require human judgment}
+
+## Code Context
+{file:line references from per-tool findings}
+```
+
+#### conclusions.json
+
+```json
+{
+  "session_id": "{sessionId}",
+  "subject": "{requirement}",
+  "mode": "collab",
+  "tools": ["gemini", "claude", "codex"],
+  "consensus_level": 85,
+  "recommendation": "Go|No-Go|Conditional",
+  "confidence": "high|medium|low",
+  "dimensions": [
+    { "name": "gemini", "score": 80, "findings": "...", "recommendations": "..." }
+  ],
+  "decisions": [
+    { "title": "...", "classification": "locked|free|deferred", "source_tools": [], "rationale": "..." }
+  ],
+  "timestamp": "<ISO>"
+}
+```
+
+### Step 10: Register Artifact
+
+Append to `.workflow/state.json`:
+
+```json
+{
+  "id": "CLB-{next_id}",
+  "type": "collab",
+  "milestone": "{current_milestone}",
+  "phase": null,
+  "scope": "adhoc",
+  "path": "scratch/{YYYYMMDD}-collab-{slug}",
+  "status": "completed",
+  "depends_on": null,
+  "harvested": false,
+  "created_at": "<ISO>",
+  "completed_at": "<ISO>"
+}
+```
+
+### Step 11: Display Summary
+
+```
+============================================================
+  MULTI-CLI COLLABORATION COMPLETE
+============================================================
+  Requirement:     {requirement}
+  Tools:           {tool_list}
+  Consensus Level: {N}%
+
+  Per-Tool:
+    gemini:  completed (confidence: {N}%)
+    claude:  completed (confidence: {N}%)
+    codex:   completed (confidence: {N}%)
+
+  Artifact: CLB-{id}
+  Output:   {outputDir}/
+
+  Next steps:
+    /maestro-analyze "{topic}"            — Deep feasibility analysis
+    /maestro-plan "{phase} --dir {dir}"   — Plan from collab conclusions
+    /maestro-brainstorm "{topic}"         — Expand with multi-role brainstorm
+============================================================
+```
+
+</execution>
+
+<error_codes>
+
+| Code | Severity | Condition | Recovery |
+|------|----------|-----------|----------|
+| E001 | error | Requirement argument missing | Prompt for requirement |
+| E002 | error | Fewer than 2 CLI tools eligible | Check cli-tools.json, enable more tools |
+| E003 | error | Specified tool not found/enabled | Show available tools |
+| E004 | error | All delegates failed | Abort with per-tool error details |
+| W001 | warning | One tool failed | Continue with remaining tools |
+| W002 | warning | >50% conflicts in cross-verify | Highlight in report, recommend manual review |
+| W003 | warning | Low consensus level (<40%) | Flag in summary |
+
+</error_codes>
+
+<success_criteria>
+- [ ] Available tools discovered from cli-tools.json with eligibility filtering
+- [ ] Plan presented via AskUserQuestion with tool modification option (unless -y)
+- [ ] All delegates launched in parallel via Bash(run_in_background: true)
+- [ ] Execution stopped after launch — waited for all callbacks
+- [ ] Per-tool outputs written to per-tool/{tool}-output.md
+- [ ] Cross-verification: consensus/conflict/unique classification complete
+- [ ] collab-report.md produced with merged findings
+- [ ] context.md produced in Locked/Free/Deferred format (downstream compatible)
+- [ ] conclusions.json produced (plan fast-track compatible)
+- [ ] CLB artifact registered in state.json
+- [ ] Partial degradation: continued if 1+ tools succeeded
+</success_criteria>

package/maestro-flow/commands/lifecycle/composer.md

@@ -0,0 +1,354 @@
+---
+name: maestro-composer
+description: Compose reusable workflow templates from natural language
+argument-hint: "\"workflow description\" [--resume] [--edit <template-path>]"
+allowed-tools:
+  - Read
+  - Write
+  - Edit
+  - Bash
+  - Glob
+  - Grep
+  - Agent
+  - AskUserQuestion
+---
+<purpose>
+Interactive workflow template composer. Parses natural language into a reusable DAG template
+via 5 phases with user confirmation at each boundary. Templates saved globally at
+`~/.maestro/templates/workflows/`. Progressive disclosure — specs loaded only when phase needs them.
+
+Three entry modes:
+1. **New design**: Parse → [confirm] → Resolve → [confirm] → Enrich → Confirm pipeline → Persist
+2. **Resume design**: Load in-progress draft from `.workflow/templates/design-drafts/`
+3. **Edit template**: Load existing template, modify, re-save
+</purpose>
+
+<deferred_reading>
+- [node-catalog](~/.maestro/templates/workflows/specs/node-catalog.md) — read at Phase 2 (Resolve) when mapping steps to executors
+- [template-schema](~/.maestro/templates/workflows/specs/template-schema.md) — read at Phase 5 (Persist) when assembling final JSON
+</deferred_reading>
+
+<context>
+$ARGUMENTS — natural language workflow description, or flags.
+
+**Flags:**
+- `--resume` — Resume in-progress design session
+- `--edit <template-path>` — Edit an existing template
+
+**Shared constants:**
+
+| Constant | Value |
+|----------|-------|
+| Session prefix | `WFD` |
+| Template dir (global) | `~/.maestro/templates/workflows/` |
+| Template index (global) | `~/.maestro/templates/workflows/index.json` |
+| Design drafts dir (local) | `.workflow/templates/design-drafts/` |
+| Template ID format | `wft-<slug>-<YYYYMMDD>` |
+| Node ID format | `N-<seq>` (e.g. N-001), `CP-<seq>` for checkpoints |
+| Max nodes | 20 |
+
+**Entry routing:**
+
+| Detection | Condition | Handler |
+|-----------|-----------|---------|
+| Resume design | `--resume` flag or existing WFD session | Phase 0: Resume |
+| Edit template | `--edit <template-path>` | Phase 0: Load + Edit |
+| New design | Default | Phase 1: Parse |
+</context>
+
+<execution>
+
+### Phase 0: Resume / Edit (conditional)
+
+**Resume design session** (if `--resume`):
+1. Scan `.workflow/templates/design-drafts/WFD-*/` for in-progress designs
+2. Multiple found → AskUserQuestion for selection
+3. Load draft → skip to last incomplete phase
+
+**Edit existing template** (if `--edit <path>`):
+1. Load template from `--edit` path
+2. Show current pipeline visualization (Phase 4 format)
+3. AskUserQuestion: which nodes to modify/add/remove
+4. Re-enter at Phase 3 with edits applied
+
+---
+
+### Phase 1: Parse — Semantic Intent Extraction
+
+**Step 1.1** — Parse `$ARGUMENTS` as description. If empty, AskUserQuestion:
+```
+"Describe the workflow you want to automate.
+Include: what steps to run, in what order, and what varies each time (inputs).
+Example: 'analyze the code, then plan, implement, and test the feature'"
+```
+
+**Step 1.2** — Extract sequential actions as candidate nodes using semantic understanding:
+
+| Signal | Candidate Type |
+|--------|---------------|
+| "analyze", "review", "explore" | analysis (cli) |
+| "plan", "design", "spec" | planning (skill) |
+| "implement", "build", "code", "fix" | execution (skill) |
+| "test", "validate", "verify" | testing (skill) |
+| "brainstorm", "ideate" | brainstorm (skill) |
+| "review code" | review (skill) |
+| "then", "next", "after" | sequential edge |
+| "parallel", "simultaneously" | parallel edge |
+
+**Step 1.3** — Extract variables (inputs that vary per run). Detect from: direct mentions, `{var}` patterns, implicit from task type.
+
+**Step 1.4** — Classify task type: `bugfix | feature | tdd | review | brainstorm | spec-driven | roadmap | refactor | auto-test | quick-task | custom`
+
+**Step 1.5** — Assess complexity: `simple` (1-3 nodes), `medium` (4-7), `complex` (8+)
+
+**Step 1.6** — Write `intent.json` to `.workflow/templates/design-drafts/WFD-<slug>-<date>/`.
+
+**Step 1.7 — Interactive confirmation**:
+
+Display parsed intent summary:
+```
+============================================================
+  COMPOSER — Intent Parsed
+============================================================
+  Description: "<original input>"
+  Task type:   <type>
+  Complexity:  <level>
+
+  Detected steps:
+    1. <description>  →  <type_hint>
+    2. <description>  →  <type_hint>
+    3. <description>  →  <type_hint>
+
+  Variables:
+    - goal (required): <inferred description>
+
+  Draft: .workflow/templates/design-drafts/WFD-<slug>-<date>/
+============================================================
+```
+
+AskUserQuestion:
+```
+options:
+  - "Looks good, continue to resolution"  → Phase 2
+  - "Edit steps"                           → re-describe, re-parse
+  - "Add a step"                           → append, re-parse
+  - "Cancel"                               → save draft, exit
+```
+
+---
+
+### Phase 2: Resolve — Map Steps to Executor Nodes
+
+**Read deferred**: `~/.maestro/templates/workflows/specs/node-catalog.md` — load node catalog for executor mapping.
+
+If the spec file does not exist, use the built-in fallback mapping:
+
+| type_hint | Default executor type | Default executor |
+|-----------|----------------------|------------------|
+| `planning` | skill | `maestro-plan` |
+| `execution` | skill | `maestro-execute` |
+| `testing` | skill | `quality-test` |
+| `review` | skill | `quality-review` |
+| `brainstorm` | skill | `maestro-brainstorm` |
+| `analysis` | cli | `maestro delegate --role analyze --mode analysis` |
+| `verify` | skill | `maestro-verify` |
+| `refactor` | skill | `quality-refactor` |
+| `debug` | skill | `quality-debug` |
+| `spec` | skill | `maestro-roadmap --mode full` |
+| `checkpoint` | checkpoint | — |
+
+**Step 2.1** — Load `intent.json`.
+
+**Step 2.2** — Map each step to executor. Resolution: match `type_hint` → catalog → semantic fit → fallback `cli`.
+
+**Step 2.3** — Build `args_template` with variable placeholders. Context injection:
+- Planning after analysis → `--context {prev_output_path}`
+- Execution after planning → `--resume-session {prev_session_id}`
+- Testing after execution → `--session {prev_session_id}`
+
+**Step 2.4** — Assign `parallel_group` for steps with `parallel_with` set.
+
+**Step 2.5** — Write `nodes.json`.
+
+**Step 2.6 — Interactive confirmation**:
+
+Display resolved nodes:
+```
+============================================================
+  COMPOSER — Nodes Resolved
+============================================================
+  N-001  [skill]    maestro-plan          "{goal}"
+  N-002  [skill]    maestro-execute       {phase}
+  N-003  [skill]    quality-test          {phase}
+
+  Parallel groups: none
+============================================================
+```
+
+AskUserQuestion:
+```
+options:
+  - "Continue to checkpoint injection"  → Phase 3
+  - "Change executor for a node"        → select node, pick new executor
+  - "Change node type"                  → skill/cli/agent/command
+  - "Back to intent"                    → Phase 1
+  - "Cancel"                            → save draft, exit
+```
+
+---
+
+### Phase 3: Enrich — Inject Checkpoints + Build DAG
+
+**Step 3.1** — Load `nodes.json`.
+
+**Step 3.2** — Build sequential edges (N-001 → N-002 → ...). For parallel groups: fan-out/fan-in.
+
+**Step 3.3** — Auto-inject checkpoint nodes. Inject if ANY rule triggers:
+
+| Rule | Condition |
+|------|-----------|
+| Artifact boundary | Source output_ports: plan, spec, analysis, review-findings |
+| Execution gate | Target executor contains `execute` |
+| Agent spawn | Target type is `agent` |
+| Long-running | Target is maestro-plan, maestro-roadmap --mode full |
+| User-defined | Step had `type_hint: checkpoint` |
+| Post-testing | Source executor contains `test` or `auto-test` |
+
+Set `auto_continue: false` for checkpoints before user-facing deliverables.
+
+**Step 3.4** — Insert checkpoint edges (A → B becomes A → CP-X → B).
+
+**Step 3.5** — Finalize `context_schema` from all `{variable}` references.
+
+**Step 3.6** — Validate: no cycles, no orphans, all nodes reachable.
+
+**Step 3.7** — Write `dag.json`.
+
+→ Proceed directly to Phase 4 (confirm is the pipeline visualization).
+
+---
+
+### Phase 4: Confirm — Visualize + User Approval
+
+**Step 4.1** — Render ASCII pipeline from `dag.json`:
+```
+============================================================
+  COMPOSER — Pipeline Review
+============================================================
+Pipeline: <template-name>
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+ N-001  [skill]       maestro-plan              "{goal}"
+   |
+ CP-01  [checkpoint]  After Plan                auto-continue
+   |
+ N-002  [skill]       maestro-execute           {phase}
+   |
+ CP-02  [checkpoint]  Before Tests              pause-for-user
+   |
+ N-003  [skill]       quality-test              {phase}
+━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+Variables (required):  goal
+Checkpoints:           2  (1 auto, 1 pause)
+Nodes:                 3 work + 2 checkpoints
+============================================================
+```
+
+For parallel groups show fan-out/fan-in:
+```
+ N-003a [skill]  quality-review  ─┐
+                                   ├─ N-004 [skill] quality-test
+ N-003b [cli]    cli analysis     ─┘
+```
+
+**Step 4.2** — AskUserQuestion:
+```
+options:
+  - "Confirm & Save"                    → Phase 5
+  - "Edit a node"                       → select node ID, modify executor/args, re-render
+  - "Add a node"                        → insert position + description, re-resolve + re-enrich, re-render
+  - "Remove a node"                     → select node, re-wire edges, re-render
+  - "Rename template"                   → new name
+  - "Re-run checkpoint injection"       → back to Phase 3.3
+  - "Cancel"                            → save draft, output resume command
+```
+
+**Step 4.3** — On edit: apply change, re-render, re-ask. Loop until Confirm or Cancel.
+
+**Step 4.4** — On Confirm: freeze dag.json, proceed to Phase 5. On Cancel: save draft, output `/maestro-composer --resume`.
+
+---
+
+### Phase 5: Persist — Assemble + Save Template
+
+**Read deferred**: `~/.maestro/templates/workflows/specs/template-schema.md` — load full JSON schema for template assembly.
+
+If the spec file does not exist, use the built-in template structure:
+```json
+{
+  "template_id": "wft-<slug>-<YYYYMMDD>",
+  "name": "<name>", "description": "<desc>", "version": "1.0",
+  "created_at": "<ISO>", "source_session": "WFD-<slug>-<date>",
+  "tags": [], "context_schema": {},
+  "nodes": [], "edges": [], "checkpoints": [],
+  "execution_mode": "serial",
+  "metadata": { "node_count": 0, "checkpoint_count": 0 }
+}
+```
+
+**Step 5.1** — Load `intent.json` + `dag.json`.
+
+**Step 5.2** — Determine template name (from Phase 4 or derive from task_type + description). Slug = kebab-case. If file exists with different content, append `-v2`, `-v3`.
+
+**Step 5.3** — Assemble template JSON.
+
+**Step 5.4** — Ensure `~/.maestro/templates/workflows/` exists. Write `<slug>.json`.
+
+**Step 5.5** — Update `~/.maestro/templates/workflows/index.json`.
+
+**Step 5.6** — Output summary:
+```
+============================================================
+  COMPOSER — Template Saved
+============================================================
+  Path:      ~/.maestro/templates/workflows/<slug>.json
+  ID:        wft-<slug>-<date>
+  Nodes:     <n> work + <n> checkpoints
+  Variables: <required vars>
+
+  To execute:
+    /maestro-player <slug> --context goal="<your goal>"
+
+  To edit later:
+    /maestro-composer --edit ~/.maestro/templates/workflows/<slug>.json
+
+  To list all templates:
+    /maestro-player --list
+============================================================
+```
+
+**Step 5.7** — Clean up design draft directory.
+</execution>
+
+<error_codes>
+| Code | Severity | Description | Recovery |
+|------|----------|-------------|----------|
+| E001 | error | Empty description and no flags | AskUserQuestion for workflow description |
+| E002 | error | Step extraction found 0 steps | Ask user to rephrase with action verbs |
+| E003 | error | Node count exceeds max (20) | Suggest splitting into sub-workflows |
+| E004 | error | DAG cycle detected | Show cycle, ask user to resolve |
+| E005 | error | Resume session not found | Show available design drafts |
+| E006 | error | Edit template not found | Show available templates |
+| W001 | warning | Ambiguous step-to-executor mapping | Show candidates, let user choose |
+| W002 | warning | No checkpoint injection rules triggered | Warn user, offer to add manually |
+| W003 | warning | Deferred spec file not found | Use built-in fallback, continue |
+</error_codes>
+
+<success_criteria>
+- [ ] Intent parsed and confirmed by user (Phase 1 interactive gate)
+- [ ] Nodes resolved and confirmed by user (Phase 2 interactive gate)
+- [ ] DAG built with auto-injected checkpoints
+- [ ] Pipeline visualized and confirmed by user (Phase 4 interactive gate)
+- [ ] Template JSON written to `~/.maestro/templates/workflows/<slug>.json`
+- [ ] Template index updated at `~/.maestro/templates/workflows/index.json`
+- [ ] Deferred specs loaded only when phase needs them (not upfront)
+</success_criteria>