skills/team-implement-plan-full/SKILL.md
Execute implementation plans with a full team — one implementer per phase running in parallel where the dependency graph allows, plus a shared cross-phase Reviewer. Phases execute in waves based on dependencies. Use for large plans with 4+ phases and independent work streams. Highest token cost but fastest execution for parallelizable work. Triggers on "full team implement", "parallel implement", or when the user explicitly wants maximum parallelism.
npx skillsauth add mhylle/claude-skills-collection team-implement-plan-fullInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
This skill executes implementation plans with maximum parallelism. Each phase gets a dedicated implementer teammate. Independent phases run simultaneously within waves. A shared Reviewer provides cross-phase quality checks.
Team composition:
When to use this:
Token cost: ~100-150K per wave (scales with concurrent implementers)
References:
references/team-lifecycle.md for team lifecycle patternreferences/quality-pipeline-distribution.md for pipeline distributionWhen invoked with a plan path:
"I'll set up a full implementation team with parallel execution. Let me read the plan, analyze the dependency graph, and present the execution waves before starting."
Step 1a: Read and validate the plan
Read($0) # Plan path from argument
Validate:
Step 1b: Check existing progress
TaskList # Check for existing tasks
If tasks exist with some completed, resume from current state (skip completed phases, adjust waves).
Step 1c: Build dependency graph
Parse phase dependencies to build execution waves:
Algorithm:
1. Read all tasks and their blockedBy relationships
2. Wave 1 = phases with no dependencies (blockedBy is empty)
3. Wave 2 = phases whose ALL dependencies are in Wave 1
4. Wave N = phases whose ALL dependencies are in Waves 1..N-1
5. Apply file conflict detection (Step 1d)
Step 1d: Detect file conflicts
For phases in the same wave, check if they touch overlapping files:
For each pair of phases in the same wave:
- Extract file scope from plan tasks (files mentioned)
- If scopes overlap: add synthetic dependency, move one phase to next wave
- Priority: keep the phase with more downstream dependents in earlier wave
Shared files that multiple phases need to modify (index files, module registrations, route configs) are handled by the Lead after a wave completes, not by individual implementers.
Step 1e: Build file ownership map
file_ownership = {}
For each wave:
For each phase in wave:
For each file in phase scope:
file_ownership[file] = phase_id
shared_files = files appearing in multiple phases across waves
Shared files are excluded from all implementer scopes. Lead handles them at wave boundaries.
Step 1f: Present wave plan to user
## Execution Plan
### Wave 1 (parallel)
- Phase 1: [Name] — files: [scope]
- Phase 3: [Name] — files: [scope]
### Wave 2 (after Wave 1 completes)
- Phase 2: [Name] — depends on Phase 1 — files: [scope]
### Wave 3 (after Wave 2 completes)
- Phase 4: [Name] — depends on Phase 2 — files: [scope]
- Phase 5: [Name] — depends on Phase 2 — files: [scope]
### Shared files (handled by Lead at wave boundaries)
- src/app.module.ts (modified by Phase 1, 2, 4)
- src/routes/index.ts (modified by Phase 3, 5)
Estimated team size: [max wave size + 1 reviewer]
Estimated token cost: ~[cost estimate]
Proceed?
Wait for user confirmation before starting.
Step 2a: Create team
TeamCreate(team_name="impl-full-{plan-slug}")
Step 2b: Spawn Reviewer
The Reviewer is spawned once and persists across all waves.
Task(subagent_type="general-purpose",
team_name="impl-full-{plan-slug}",
name="reviewer",
prompt="You are the cross-phase Reviewer on a full implementation team.
PLAN: {full plan content}
YOUR ROLE: Quality gate for ALL phases. You review each phase as implementers complete them. You are the only reviewer — process requests FIFO.
REVIEW PROTOCOL:
When an implementer messages you 'Phase N ready for review':
1. READ all changed files completely
2. RUN exit condition commands independently:
- Build: {build commands from plan}
- Lint: {lint commands}
- Test: {test commands}
3. CHECK code quality:
- Does the code follow existing patterns?
- Any security issues?
- Error handling adequate?
- Tests meaningful?
- Design decision compliance?
4. CHECK cross-phase consistency:
- Do types/interfaces match across phases?
- Are shared contracts maintained?
- Any naming conflicts with other completed phases?
5. CHECK ADR compliance:
- Read docs/decisions/INDEX.md
- Verify against relevant ADRs
DECISION:
- PASS: Message team lead: 'PASS: Phase N — [quality summary]'
- NEEDS_CHANGES: Message the specific implementer: 'NEEDS_CHANGES: [issue list with file:line refs]'
Wait for re-request, then re-review
RULES:
- You are read-only — NEVER modify code. Only implementers write code
- Process reviews FIFO — first come, first served
- Be thorough but pragmatic — real issues only, not style nitpicks
- Always include file:line references
- If you notice cross-phase conflicts, message the team lead immediately")
For each wave:
Step 3a: Spawn wave implementers
For each phase in the current wave, spawn a dedicated implementer:
Task(subagent_type="general-purpose",
team_name="impl-full-{plan-slug}",
name="phase-{N}-impl",
prompt="You are the implementer for Phase {N} on a full implementation team.
PLAN CONTEXT: {relevant plan sections}
YOUR PHASE: {phase N details — objective, tasks, exit conditions}
YOUR FILE SCOPE — you may ONLY modify these files:
{file list from ownership map}
DO NOT MODIFY these shared files (the Lead handles them):
{shared file list}
IMPLEMENTATION PROTOCOL:
1. Read all files in your scope and relevant context files
2. Implement phase tasks IN ORDER (tests first, then implementation)
3. Run all exit condition commands:
- Build: {build commands}
- Lint: {lint commands}
- Test: {test commands}
4. Fix any failures — iterate until all exit conditions pass
5. When ALL exit conditions pass, message 'reviewer':
'Phase {N} ready for review. Files changed: [list]. Exit conditions passing.'
6. If reviewer sends NEEDS_CHANGES, fix the issues and re-request review
7. Do NOT start any other phase — your scope is Phase {N} only
RULES:
- Stay within your file scope — do NOT modify files outside it
- If you need to change a shared file, message the team lead with what change is needed
- If you need information from another phase's files, READ them but do not WRITE
- If you hit a blocker, message the team lead
- Follow existing codebase patterns documented in the plan")
Step 3b: Monitor wave progress
While wave executes:
Step 3c: Handle shared file requests
When implementers message that they need a shared file modified:
Step 3d: Wave review tracking
Track review status per phase:
Wave 1 Review Status:
Phase 1: ✅ PASS
Phase 3: 🔄 In review (attempt 2)
Phase 5: ⏳ Awaiting review
Step 3e: Handle fix loops
If a fix loop exceeds 3 iterations for any phase:
Step 3f: Wave completion
When ALL phases in the wave have passed review:
Step 3g: User confirmation between waves
Wave {W} complete.
Phases completed: {list with summaries}
Files changed: {aggregated list}
Issues caught by Reviewer: {summary}
Integration status: {pass/issues}
Next wave: {phase list}
Continue? (or /clear and resume later — progress is saved)
Wait for user confirmation before spawning next wave.
Step 3h: Reviewer health check
After each wave:
After all waves complete:
Step 4a: Final integration check
Run the full test suite one final time. All exit conditions from all phases must still pass.
Step 4b: Completion report
## Implementation Complete
**Plan**: {plan name}
**Method**: Full team (parallel wave execution)
**Waves executed**: {count}
**Total phases**: {count}
**Phases per wave**: {breakdown}
### Quality Summary
- Issues caught by Reviewer: {count}
- Fix loops required: {count}
- Cross-phase conflicts resolved: {count}
- Shared file changes: {count}
### Per-Phase Summary
| Phase | Wave | Implementer | Review Attempts | Key Changes |
|-------|------|-------------|-----------------|-------------|
| 1 | 1 | phase-1-impl | 1 | {summary} |
| 2 | 2 | phase-2-impl | 2 | {summary} |
...
### Files Changed (all phases)
{aggregated file list}
Step 4c: Shutdown and cleanup
/e2e-testing to validate the full implementation"If the session ends mid-wave:
/team-implement-plan-full [plan-path]What persists: Task status, committed code, plan file What's lost: Teammate context, uncommitted changes, review history
Mitigation: Lead encourages committing after each wave completes.
Common shared files:
app.module.ts, main.py)routes/index.ts)types/index.ts)config/*.ts)package.json — for new dependencies)Protocol:
| Scenario | Max Concurrent Implementers | Rationale | |----------|---------------------------|-----------| | Default | 4 | Balance between parallelism and coordination overhead | | Simple phases (< 3 tasks each) | 5 | Lower per-phase complexity allows more concurrency | | Complex phases (5+ tasks each) | 3 | Higher per-phase complexity needs more lead attention | | User override | As specified | User can request specific wave sizes |
The Reviewer is always 1 (shared across wave). Total team size = wave implementers + 1 reviewer + lead.
| Pipeline Step | Owner | Notes | |---|---|---| | 1. Implementation | phase-N-impl | Direct code writing within file scope | | 2. Verification-loop | phase-N-impl | Build/lint/test scoped to phase | | 3. Integration testing | phase-N-impl | Tests within phase scope | | 4. Code review | reviewer | Cross-phase consistency + per-phase quality | | 5. ADR compliance | reviewer | Centralized ADR knowledge | | 6. Plan sync | Lead | Single writer to task status | | 7. Prompt archival | Lead | Single writer to prompt files | | 8. Completion report | Lead | Aggregates per wave and final |
Before completing each wave:
Before completing the plan:
testing
One-command issue-to-merge pipeline orchestrator. Drives a GitHub issue through nine stages (preflight, plan, implement, review, ci, cloud_review, deploy, e2e, logs) with two human gates, persisting all run state to files so a crashed or interrupted run resumes losslessly. Triggers on "/ship-issue" with an issue number or URL. User-invoked only.
tools
--- name: tt-workflow-build description: Tasktracker-native trigger for a PARALLEL build via the Claude Code Workflow tool. Thin by design — it does two things, then drives to done: (1) ensure a tasktracker project exists (use the existing one, or create one), then (2) start a dynamic `Workflow` that builds it, tracking the work in tasktracker and using the build + verify skills. It does NOT analyze parallelism up front, ask the user to choose a mode, hand back, or fall back to a sequential skil
tools
--- name: grumpy-reviewer description: A single grumpy, nitpicky structural code reviewer that runs as an isolated subagent and treats the code as third-party work submitted by a junior programmer for validation. It cares about exactly one thing — maintainability — judged through separation of concerns, service-oriented design, helper-method extraction, small files, and the rule of 7 (as any grouping nears 7 members, it pushes for sub-groupings). It is deliberately kept OUT of the implementation
development
--- name: tt-workflow-run description: Tasktracker-native autonomous build-loop orchestrator. Drives a first-class `workflow_run` end-to-end — create the run (Gate 1 lifecycle completeness + Gate 2 zero-defects-in), then loop while `getNextReadyTask(projectId)` returns a slice — `setActiveTask` → record a pre-slice `scanArchitectureDrift` baseline → delegate the slice to `/tt-implement-phase` (which does the code work, registers the architecture delta in-slice, and auto-logs defects/learnings/fr