swarm

---

name: swarm description: Orchestrated multi-agent implementation using Tech Lead + Developer pattern. Use when implementing phased plans from docs/PLAN.md. Triggers: /swarm, "implement the plan", "run phased implementation", "execute PLAN.md phases". Spawns Developer agents for each phase, reviews PRs, handles retries and escalation.

Swarm - Orchestrated Multi-Agent Implementation

Execute a phased implementation plan using a Tech Lead + Developer agent pattern.

Arguments

• /swarm - Use default plan file docs/PLAN.md
• /swarm <plan-file> - Use specified plan file

Workflow

For each phase in plan:
  1. Create branch from current main
  2. Spawn Developer agent to implement
  3. Developer creates PR when done
  4. Tech Lead (this agent) reviews PR
  5. If APPROVED: merge and continue
  6. If CHANGES_REQUESTED: Developer fixes (max 3 attempts)
  7. If max attempts exceeded: ESCALATE to human

Step 1: Parse Plan

Read the plan file and extract phases. Each phase has this structure:

<!-- PHASE:N -->
## Phase N: Name

### Branch
`phase-N-name`

### Scope
...

### Files to Create/Modify
...

### Acceptance Criteria
- [ ] Criterion 1
- [ ] Criterion 2

### Tests Required
...
<!-- /PHASE:N -->

Step 2: For Each Phase

git checkout main && git pull
git checkout -b <branch-name>

Step 3: Spawn Developer Agent

Task(
  subagent_type: "general-purpose",
  description: "Implement Phase N",
  prompt: """
You are a Developer agent implementing Phase N.

## Task
Read the plan file, find Phase N (between <!-- PHASE:N --> markers), implement EVERYTHING in Scope.

## CRITICAL: Your Work Will Be Rigorously Reviewed

The Tech Lead will verify:
1. **Every file** in "Files to Create/Modify" exists and has real implementation
2. **Every acceptance criterion** is fully implemented (not stubbed)
3. **Every test** in "Tests Required" exists and passes
4. **Integration points** - router registered, config in .env.example, migration matches ORM

DO NOT:
- Create stub implementations (empty functions, pass, TODO comments)
- Skip any file from the list
- Write trivial tests that don't verify real behavior
- Leave acceptance criteria partially implemented
- Forget to register routers or add config vars

## Rules
1. Follow CLAUDE.md configuration standards (no hardcoded values, fail fast)
2. Create ALL files listed in "Files to Create/Modify" - every single one
3. Write ALL tests specified in "Tests Required" - run them to verify they pass
4. For each acceptance criterion, identify WHERE in your code it's satisfied
5. Commit with clear messages referencing the phase

## Branch
You are on branch: <branch-name>

## Before Creating PR - Self-Review Checklist

Before creating the PR, verify yourself:
- [ ] All files from "Files to Create/Modify" exist
- [ ] No TODO/FIXME/placeholder comments in new code
- [ ] All acceptance criteria have corresponding implementation
- [ ] All tests pass: `python -m pytest <module>/tests/ -v`
- [ ] Router registered in main.py (if new module)
- [ ] Config vars in .env.example (if new config)

## When Done
Create PR with acceptance criteria as checklist:

gh pr create --title "Phase N: <name>" --body "$(cat <<'EOF'
## Implementation Summary
<brief description>

## Acceptance Criteria
- [ ] Criterion 1 - implemented in `file.py:function()`
- [ ] Criterion 2 - implemented in `file.py:function()`
...

## Tests
- `pytest <module>/tests/ -v` - X tests pass

## Files Changed
<list of files created/modified>
EOF
)"

Report back with PR number.
"""
)

Step 4: Review the PR (RIGOROUS)

IMPORTANT: Previous swarm runs had phases that were incomplete or sloppy. This review MUST be thorough. Do NOT approve until ALL checks pass.

4.1 File Inventory Check

Read the phase's "Files to Create/Modify" list. For EACH file:

# Check file exists
ls -la <file-path>

# Read and verify content is substantial (not stub/placeholder)
Read <file-path>

FAIL if:

• Any listed file is missing
• Any file contains TODO/FIXME/placeholder comments
• Any file is a stub (empty class, pass-only functions)
• Any file has # Implementation needed or similar

4.2 Acceptance Criteria Verification

For EACH criterion in "Acceptance Criteria" section:

1. Read the criterion literally - what exactly does it require?
2. Find evidence - which file/function implements it?
3. Verify behavior - run specific test or manual check

| Criterion | Evidence | Verified |
|-----------|----------|----------|
| "Can create quota for user" | quotas/service.py:create_quota() | ✓ test passes |
| "Unique constraint prevents duplicates" | migration has UNIQUE KEY | ✓ |
| "Only managers can set quotas" | router.py uses require_role("manager") | ✓ |

FAIL if:

• Any criterion has no corresponding implementation
• Any criterion is partially implemented
• Any criterion relies on code that doesn't exist yet

4.3 Test Coverage Check

# Run the phase's tests specifically
cd apps/crm-api && python -m pytest <module>/tests/ -v

# Check all tests pass
# Check test count matches "Tests Required" list

FAIL if:

• Any test fails
• Test count is significantly lower than specified in plan
• Tests are trivial (e.g., assert True)

4.4 Integration Points Check

For new modules, verify:

# Router registered in main.py?
grep -n "<router>" apps/crm-api/app/main.py

# Config vars in .env.example?
grep -n "<new_config_var>" apps/crm-api/.env.example

# ORM model matches migration?
# Compare column names, types, nullability

FAIL if:

• Router not registered (endpoints unreachable)
• Config vars missing from .env.example
• Migration and ORM model have mismatches

4.5 Code Quality Check

1. No hardcoded values - all thresholds from config
2. No silent defaults - missing required config = startup failure
3. UUID not INT - all IDs are VARCHAR(36)
4. Follows existing patterns - check similar modules for consistency

4.6 Final Verdict

Only after ALL above checks pass:

## PR Review: Phase N

### File Inventory: ✓ PASS
- [x] All 8 files created
- [x] No stubs or placeholders

### Acceptance Criteria: ✓ PASS
| Criterion | Status |
|-----------|--------|
| Criterion 1 | ✓ Verified in service.py:45 |
| Criterion 2 | ✓ Test test_unique_constraint passes |
| ... | ... |

### Tests: ✓ PASS
- 8/8 tests pass
- Matches "Tests Required" list

### Integration: ✓ PASS
- Router registered in main.py:67
- Config vars in .env.example

### Code Quality: ✓ PASS

**VERDICT: APPROVED**

If ANY check fails:

## PR Review: Phase N

### FAILED CHECKS:

1. **Acceptance Criteria #3 not implemented**
   - "Quota attainment calculates correctly from closed-won opportunities"
   - No implementation found in service.py
   - Required: Add get_quota_attainment() method

2. **Missing test**
   - "Tests Required" lists "Fiscal period calculation" but no such test exists

**VERDICT: CHANGES_REQUESTED**

Step 5: Decision

APPROVED:

gh pr merge <pr-number> --squash --delete-branch

Continue to next phase.

CHANGES_REQUESTED:

Task(
  prompt: """
Your PR for Phase N needs changes.

## Feedback
<specific feedback>

## Required Changes
1. ...

Fix on same branch and push. Attempt: <N>/3
"""
)

Attempt >= 3:

ESCALATE: Phase N requires human intervention.
PR: <url>
Issues: <summary>

Stop and notify user.

Step 6: Progress Tracking

After each phase:

## Swarm Progress

| Phase | Status | PR | Attempts |
|-------|--------|-----|----------|
| 1     | DONE   | #12 | 1        |
| 2     | IN_PROGRESS | #13 | 1   |
| 3     | PENDING | -  | -        |

Error Handling

• Git conflict: Escalate immediately
• CI failure: Count as failed review attempt
• Agent timeout: Retry once, then escalate
• Network error: Retry with backoff