@the-bearded-bear/claude-craft 5.6.0-next.8bfdffa → 5.7.0-next.bf058ea

package/Dev/i18n/en/Common/agents/ralph-conductor.md

@@ -171,12 +171,62 @@ Summary:
 - Metrics exported: .ralph/sessions/.../metrics-export.json
 ```
 
+## Agent Teams Coordination Mode
+
+When operating in Agent Teams mode (activated via `--use-teams` on `/common:ralph-sprint`), the conductor takes on the role of **team lead** and coordinates a dev teammate through the Claude Code Agent Teams API instead of bash process management.
+
+### Prerequisites
+
+- `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` environment variable
+- Claude Code v2.1.32+
+- Adapter library: `Tools/AgentTeams/lib/ralph-teams-adapter.sh`
+
+### Coordination via Task System
+
+In Agent Teams mode, the conductor replaces PID-based tracking with the shared task system:
+
+| Bash Mode (current) | Agent Teams Mode |
+|---------------------|-----------------|
+| `spawn_ralph_for_story()` with bash `&` | `TaskCreate` + `SendMessage` to dev teammate |
+| `kill -0 $pid` polling | `TaskList` / `TaskCompleted` hook |
+| PID-based completion detection | `TaskUpdate(status=completed)` by dev |
+| `kill -9` for stuck processes | `SendMessage(type=shutdown_request)` + watchdog fallback |
+| `yq` writes to `batch-queue.yaml` | Shared `TaskList` (built-in coordination) |
+
+### Story Processing Flow
+
+1. **Claim story**: Conductor reads `sprint-status.yaml`, claims next `ready-for-dev` story
+2. **Create task**: `TaskCreate` with story details, acceptance criteria, and TDD instructions
+3. **Assign to dev**: `SendMessage(type=message, recipient=dev-1)` with the story prompt
+4. **Monitor progress**: Poll `TaskList` for status updates from the dev teammate
+5. **Handle completion**: When dev marks task as `completed`, conductor transitions story to `review`
+6. **Handle failure**: If dev reports failure or watchdog detects a stall, conductor applies recovery strategy
+7. **Next story**: Assign next ready story or send `shutdown_request` if sprint is complete
+
+### Watchdog Integration
+
+The conductor runs periodic health checks through the adapter's `teams_watchdog()`:
+
+- **Check interval**: Every 60 seconds (configurable via `TEAMS_WATCHDOG_INTERVAL`)
+- **Timeout threshold**: 5 minutes of no activity (configurable via `TEAMS_WATCHDOG_TIMEOUT`)
+- **Stall action**: Mark teammate as stalled, trigger `teams_fallback_sequential()`, reprocess story through existing `execute_story_with_ralph()`
+
+### Keeping Bash Mode Intact
+
+All existing bash-mode orchestration remains unchanged. The Agent Teams mode is activated only when:
+1. The `--use-teams` flag is passed to `/common:ralph-sprint`
+2. The `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` env var is set
+3. The adapter library is available
+
+Without these conditions, the conductor operates exactly as before.
+
 ## Integration Points
 
 - Works with `/common:ralph-run` command
 - Integrates with Claude Code 2.1.23+ hooks
 - Compatible with `/project:sprint-dev` workflow
 - Uses `@tdd-coach` principles
+- Agent Teams mode via `/common:ralph-sprint --use-teams`
 
 ## When to Stop
 

package/Dev/i18n/en/Common/commands/full-audit.md

@@ -30,6 +30,50 @@ For each detected technology:
 1. Load rules from `.claude/rules/`
 2. Apply specific audit
 
+### Step 1.5: Prepare Isolated Output Directories
+
+When **2 or more technologies** are detected, each audit agent MUST write its results to an isolated directory to prevent file write conflicts during parallel execution:
+
+```
+.audit-output/
+  {technology}-{category}/
+    result.json          # Structured audit result
+    tool-output.log      # Raw tool output
+```
+
+For example, a Symfony + React project creates:
+```
+.audit-output/
+  symfony-architecture/result.json
+  symfony-code-quality/result.json
+  symfony-testing/result.json
+  symfony-security/result.json
+  react-architecture/result.json
+  react-code-quality/result.json
+  react-testing/result.json
+  react-security/result.json
+```
+
+Each `result.json` follows this schema:
+```json
+{
+  "tech": "symfony",
+  "category": "architecture",
+  "score": 22,
+  "max": 25,
+  "findings": [
+    {
+      "severity": "warning",
+      "file": "src/Controller/UserController.php",
+      "message": "Direct repository access from controller",
+      "rule": "clean-architecture-layer-violation"
+    }
+  ]
+}
+```
+
+**Single-technology projects**: When only 1 technology is detected, isolation is not required. Results can be written directly without the `.audit-output/` directory structure.
+
 ### Step 2: Audit by Technology
 
 For EACH detected technology, verify:
@@ -84,6 +128,28 @@ docker compose exec node npm run lint
 docker compose exec node npm run test -- --coverage
 ```
 
+### Step 3.5: Merge Audit Results
+
+When using isolated output directories (2+ technologies), collect and merge all results before scoring:
+
+1. **Read all `result.json` files** from `.audit-output/*/result.json`
+2. **Group by technology**: Combine the 4 category results per technology
+3. **Deduplicate findings**: Remove duplicate findings that appear across categories (e.g., a file flagged in both architecture and code quality)
+4. **Resolve conflicts**: If the same file is scored differently by two categories, use the lower (more critical) score
+5. **Produce merged result** for each technology with all 4 category scores
+
+Merge can be done using the result aggregator script if available:
+```bash
+# If result-aggregator.sh is available
+Tools/AgentTeams/lib/result-aggregator.sh \
+  --input-dir .audit-output \
+  --output-file .audit-output/merged-report.json
+```
+
+Or manually by reading each `result.json` and aggregating in memory.
+
+**Single-technology projects**: Skip this step (no merge needed).
+
 ### Step 4: Calculate Scores
 
 For each technology, calculate:

package/Dev/i18n/en/Common/commands/ralph-sprint.md

@@ -1,6 +1,6 @@
 ---
 description: Run autonomous sprint conductor for overnight/unattended sprint execution
-argument-hint: <sprint-name> [--overnight|--parallel N|--supervised|--max-stories N]
+argument-hint: <sprint-name> [--overnight|--parallel N|--supervised|--max-stories N|--use-teams]
 ---
 
 # Ralph Sprint - Autonomous Sprint Conductor (ASC)
@@ -17,6 +17,7 @@ Execute an entire sprint autonomously with minimal human intervention. The Auton
 - `--supervised`: Pause before each story for confirmation
 - `--max-stories N`: Maximum stories to process (default: 10)
 - `--timeout H`: Maximum runtime in hours (default: 12)
+- `--use-teams`: Use Agent Teams mode (2-agent prototype: conductor + 1 dev)
 
 ## Key Features
 
@@ -163,6 +164,51 @@ parallel:
     mem_percent: 80      # Don't spawn if memory > 80%
 ```
 
+## Agent Teams Mode
+
+When `--use-teams` is specified, the ASC uses Claude Code Agent Teams (v2.1.32+) instead of bash-based parallel processing. This is a prototype mode limited to a 2-agent team (conductor + 1 dev).
+
+### Prerequisites
+
+- Claude Code v2.1.32 or later
+- Environment variable: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`
+- The adapter library: `Tools/AgentTeams/lib/ralph-teams-adapter.sh`
+
+### How It Works
+
+1. **Conductor as Team Lead**: The `@ralph-conductor` agent acts as the team lead, coordinating work via the shared task system
+2. **Dev Teammate**: A single `@dev` agent receives story assignments via `SendMessage` and reports completion via `TaskUpdate`
+3. **Shared Tasks**: Stories are managed through `TaskCreate`/`TaskUpdate` instead of PID-based tracking
+4. **Watchdog Recovery**: If the dev teammate is unresponsive for 5 minutes, the adapter triggers a fallback to sequential processing
+
+```
+Sprint Conductor (Team Lead)
+  |
+  +-- TaskCreate: US-001 story task
+  +-- SendMessage: assign to dev-1
+  |
+  +-- dev-1 implements US-001 (TDD: Red -> Green -> Refactor)
+  +-- dev-1 marks TaskUpdate: completed
+  |
+  +-- Conductor transitions story to review
+  +-- Conductor assigns next story (or shuts down teammate)
+```
+
+### Fallback Behavior
+
+- Without `--use-teams`: behavior is unchanged (bash-based parallel with `--parallel N` or sequential)
+- With `--use-teams` but without `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`: the adapter falls back to sequential processing automatically
+- If the dev teammate stalls (no response for 5 minutes): the watchdog triggers `teams_fallback_sequential()` and the story is reprocessed sequentially
+
+### Limitations (Prototype)
+
+| Constraint | Detail |
+|-----------|--------|
+| Team size | Fixed at 2 agents (conductor + 1 dev) |
+| Stories per run | Processes 1 story at a time through the dev teammate |
+| API stability | Agent Teams is a Research Preview feature |
+| No persistent teams | Teams are per-session only |
+
 ## Quick Start Examples
 
 ```bash
@@ -180,6 +226,12 @@ parallel:
 
 # Limited run (5 stories, 4 hours)
 /common:ralph-sprint "Sprint 3" --max-stories 5 --timeout 4
+
+# Agent Teams mode (2-agent prototype)
+/common:ralph-sprint "Sprint 3" --use-teams
+
+# Agent Teams overnight
+/common:ralph-sprint "Sprint 3" --use-teams --overnight
 ```
 
 ## Configuration

package/Dev/i18n/en/Common/commands/team-audit.md

@@ -0,0 +1,310 @@
+---
+description: Full Audit Team - Parallel multi-technology audit using Agent Teams
+argument-hint: [--techs=auto|tech1,tech2] [--max-workers=4]
+---
+
+# Full Audit Team - Parallel Multi-Technology Audit
+
+Orchestrate a parallel full-audit across multiple technology stacks using Claude Code Agent Teams (v2.1.32+). Spawns a lead agent (opus) plus N stack-auditor workers (haiku), one per detected technology stack, up to a configurable maximum.
+
+## Arguments
+
+$ARGUMENTS
+
+- `--techs=auto`: Auto-detect technologies (default). Or specify comma-separated: `--techs=symfony,react`
+- `--max-workers=4`: Maximum parallel auditor workers (default: 4, max: 4)
+- `--output-dir=<path>`: Custom output directory for audit results
+- `--dry-run`: Show team composition and estimated cost without executing
+- `--skip-aggregation`: Output per-stack results without merging
+
+## Prerequisites
+
+- Claude Code v2.1.32+ with Agent Teams support
+- `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` environment variable set
+- Project with 2+ detected technology stacks (single-stack projects should use sequential `/common:full-audit`)
+- `Tools/AgentTeams/lib/compatibility-check.sh` available
+- `Tools/AgentTeams/lib/result-aggregator.sh` available
+- `Tools/AgentTeams/lib/cost-estimator.sh` available
+
+## When to Use (vs. Sequential Audit)
+
+| Condition | Use Team Audit | Use Sequential `/common:full-audit` |
+|-----------|---------------|--------------------------------------|
+| 1 technology stack | No | Yes |
+| 2+ technology stacks | Yes | Also valid (simpler, cheaper) |
+| Time-sensitive | Yes (2-3x speedup) | No |
+| Budget-constrained | No (+20-35% token overhead) | Yes |
+
+**Break-even**: Parallelization benefits emerge at 2+ stacks. For a single stack, coordination overhead exceeds time saved.
+
+## Process
+
+### Step 1: Technology Detection
+
+```
+Audit Leader (opus)
+  |
+  v
+Scan project root for technology markers:
+  composer.json + symfony/*      -> Symfony
+  pubspec.yaml + flutter:        -> Flutter
+  pyproject.toml / requirements  -> Python
+  package.json + react           -> React
+  package.json + react-native    -> React Native
+  package.json + @angular/core   -> Angular
+  package.json + vue             -> Vue.js
+  artisan + laravel/*            -> Laravel
+  *.csproj + dotnet              -> C#/.NET
+  composer.json (no symfony)     -> PHP
+```
+
+If `--techs=auto`, detect all. If explicit, validate specified stacks exist.
+
+**Decision gate**: If only 1 technology detected, fall back to sequential `/common:full-audit` (no team overhead needed).
+
+### Step 2: Compatibility Check
+
+Before spawning workers, validate each auditor agent against role requirements:
+
+```bash
+# For each detected stack, verify the reviewer agent has required tools
+Tools/AgentTeams/lib/compatibility-check.sh \
+  --agent Dev/i18n/en/<Tech>/agents/<tech>-reviewer.md \
+  --require-tools Read,Glob,Grep,Bash \
+  --require-model haiku
+```
+
+If any agent fails compatibility, log a warning and exclude that stack from parallel execution (fall back to leader handling it sequentially).
+
+### Step 3: Cost Estimation
+
+Before spawning the team, estimate token costs:
+
+```bash
+Tools/AgentTeams/lib/cost-estimator.sh \
+  --team-size <N+1> \
+  --lead-model opus \
+  --worker-model haiku \
+  --task-type audit \
+  --stacks <detected_count>
+```
+
+Display estimated cost to user. In `--dry-run` mode, stop here.
+
+### Step 4: Team Spawn (Fan-Out)
+
+```
+Audit Leader (opus) — coordinates via TaskCreate/SendMessage
+  |
+  +-- [Parallel Workers - max 4] --------+
+  |   stack-auditor-1 (haiku): Symfony     |
+  |   stack-auditor-2 (haiku): React       |
+  |   stack-auditor-3 (haiku): Python      |
+  |   stack-auditor-4 (haiku): Angular     |
+  +---------------------------------------+
+```
+
+**Team creation pattern:**
+
+1. Leader creates isolated output directories per worker (one per stack)
+2. Leader creates tasks via `TaskCreate` for each stack audit:
+   - Task subject: `Audit <TechName> stack`
+   - Task description: includes check-architecture, check-code-quality, check-testing, check-security, check-compliance instructions
+   - Each task specifies its isolated output path
+3. Workers claim tasks via `TaskUpdate` (status: in_progress)
+4. Workers write results to their isolated directory only
+
+**Worker instructions** (per stack):
+
+Each worker executes the 4 audit categories sequentially within its stack:
+
+| Category | Points | What to Check |
+|----------|--------|---------------|
+| Architecture (25pts) | Layer separation, dependency direction, folder conventions, no framework coupling |
+| Code Quality (25pts) | Naming standards, linting, type hints, documentation, complexity < 10 |
+| Testing (25pts) | Coverage >= 80%, unit tests, integration tests, E2E tests, test pyramid |
+| Security (25pts) | No secrets, input validation, OWASP, encryption, dependency CVEs |
+
+Workers run Docker-based diagnostic commands per stack:
+
+```bash
+# Symfony
+docker compose exec php php bin/console lint:container
+docker compose exec php vendor/bin/phpstan analyse
+docker compose exec php vendor/bin/phpunit --coverage-text
+docker compose exec php composer audit
+
+# React
+docker compose exec node npm run lint
+docker compose exec node npm run test -- --coverage
+docker compose exec node npm audit
+
+# Python
+docker compose exec app ruff check .
+docker compose exec app mypy .
+docker compose exec app pytest --cov
+docker compose exec app pip-audit
+
+# Flutter
+docker run --rm -v $(pwd):/app -w /app dart dart analyze
+docker run --rm -v $(pwd):/app -w /app dart flutter test --coverage
+```
+
+Each worker writes `result.json` to its isolated output directory:
+
+```json
+{
+  "tech": "symfony",
+  "score": 82,
+  "architecture": { "score": 22, "findings": [...] },
+  "code_quality": { "score": 20, "findings": [...] },
+  "testing": { "score": 18, "findings": [...] },
+  "security": { "score": 22, "findings": [...] }
+}
+```
+
+### Step 5: Sync Barrier
+
+Leader waits for all worker tasks to reach `completed` status via `TaskList` polling. If a worker exceeds its timeout (5 minutes per stack), leader marks it as failed and proceeds with partial results.
+
+### Step 6: Result Aggregation
+
+Leader runs the result aggregator:
+
+```bash
+Tools/AgentTeams/lib/result-aggregator.sh \
+  --input-dir <isolated-output-root> \
+  --output-file audit-report.json
+```
+
+The aggregator:
+- Collects all `result.json` files from isolated directories
+- Deduplicates findings (same file + same message = duplicate)
+- Resolves score conflicts via weighted average
+- Produces unified report
+
+### Step 7: Report Generation
+
+Leader generates the formatted multi-technology audit report:
+
+```
+================================================================
+MULTI-TECHNOLOGY AUDIT (Agent Teams) - Global Score: XX/100
+================================================================
+
+Detected technologies: [list]
+Team size: 1 leader + N workers
+Execution mode: Parallel
+Date: YYYY-MM-DD
+
+----------------------------------------------------------------
+SYMFONY - Score: XX/100
+----------------------------------------------------------------
+
+Architecture (XX/25)
+  [PASS] Clean Architecture respected
+  [PASS] CQRS implemented correctly
+  [WARN] 2 services directly access Repository
+
+Code Quality (XX/25)
+  [PASS] PHPStan level 8 - 0 errors
+  [WARN] 5 methods > 20 lines
+
+Testing (XX/25)
+  [PASS] Coverage: 85%
+  [WARN] No Panther E2E tests
+
+Security (XX/25)
+  [PASS] No secrets in code
+  [WARN] Dependency with minor CVE
+
+----------------------------------------------------------------
+REACT - Score: XX/100
+----------------------------------------------------------------
+
+[Same structure per technology]
+
+================================================================
+GLOBAL SUMMARY
+================================================================
+
+| Technology | Architecture | Code | Tests | Security | Total |
+|------------|-------------|------|-------|----------|-------|
+| Symfony    | XX/25       | XX/25| XX/25 | XX/25    | XX/100|
+| React      | XX/25       | XX/25| XX/25 | XX/25    | XX/100|
+| AVERAGE    | XX/25       | XX/25| XX/25 | XX/25    | XX/100|
+
+================================================================
+TOP 5 PRIORITY ACTIONS
+================================================================
+
+1. [CRITICAL] Action description
+   -> Impact: +X points | Effort: Low/Medium/High
+
+2. [HIGH] Action description
+   -> Impact: +X points | Effort: Low/Medium/High
+
+================================================================
+EXECUTION METRICS
+================================================================
+
+| Metric | Value |
+|--------|-------|
+| Total time | Xs (vs ~Ys sequential) |
+| Speedup | ~X.Xx |
+| Total tokens | ~XK |
+| Token overhead vs sequential | +XX% |
+| Workers spawned | N |
+| Workers completed | N |
+| Workers failed | 0 |
+```
+
+### Step 8: Cleanup
+
+Leader sends `shutdown_request` to all workers and cleans up isolated output directories (unless `--keep-artifacts` is specified).
+
+## Scoring Rules
+
+Same as `/common:full-audit`:
+
+| Violation | Points Lost |
+|-----------|-------------|
+| Architectural pattern violated | -5 |
+| Framework/domain coupling | -3 |
+| Critical linting error | -2 |
+| Linting warning | -1 |
+| Method > 30 lines | -1 |
+| Coverage < 80% | -5 |
+| No domain unit tests | -5 |
+| Secret in code | -10 |
+| Critical CVE vulnerability | -10 |
+| High CVE vulnerability | -5 |
+
+## Performance Expectations
+
+| Stacks | Sequential Estimate | Team Estimate | Speedup | Token Overhead |
+|--------|--------------------|--------------:|---------|----------------|
+| 2 | ~4 min | ~2.5 min | ~1.6x | +20% |
+| 3 | ~6 min | ~3 min | ~2x | +25% |
+| 4 | ~8 min | ~3.5 min | ~2.3x | +30% |
+| 5+ | ~10+ min | ~4 min | ~2.5x | +35% |
+
+**Note**: These are realistic estimates accounting for coordination overhead (agent spawn ~5-10s, task assignment, result aggregation). Do not expect linear speedup.
+
+## Error Handling
+
+| Error | Recovery |
+|-------|----------|
+| Worker timeout (>5min) | Leader marks failed, proceeds with partial results |
+| Worker crash | Leader logs error, excludes stack from report |
+| Docker not available | Worker reports error, leader falls back to source-only analysis |
+| No technologies detected | Abort with clear message |
+| Single technology only | Fall back to sequential `/common:full-audit` |
+| Compatibility check fails | Exclude stack from parallel, leader handles sequentially |
+
+## Limitations
+
+- Maximum 4 parallel workers (coordination overhead dominates beyond this)
+- Token cost is ~20-35% higher than sequential due to context duplication per worker
+- Requires Agent Teams Research Preview (API may change)
+- Each worker loads project context independently (~10-20K tokens overhead each)