Implement Tasks (Pure Orchestrator)

This skill never writes code itself. It resolves the dependency graph, generates prompt files in tasks/, and spawns isolated tmux subagents that do the actual implementation. The orchestrator only validates, coordinates, and monitors — all code changes come from subagents.

⛔ Hard Restriction: No Direct File Edits

The orchestrator agent MUST NOT directly edit, create, or delete any project source files. This is a non-negotiable constraint. The orchestrator is a coordinator, not an implementer.

Permitted Actions

• ✅ Read files — Read tasks.json, prompt files, subagent output, source files
• ✅ Browser actions — Navigate to docs, search for solutions, read API references
• ✅ Run scripts — validate-dag.ts, generate-prompts.ts, spawn-wave.sh, status-tasks.ts
• ✅ Run shell commands — grep, find, git log, tmux management, cat
• ✅ Write orchestrator artifacts — Status files, progress reports, handoff notes

Forbidden Actions

• ❌ Edit source files — No edit, write, or delete on any project source files
• ❌ Create implementation files — No new .ts, .js, .go, .py, etc. files
• ❌ Modify configuration — No changes to package.json, tsconfig, etc.
• ❌ Run implementation commands — No npm install, go build, cargo build, etc.

If implementation work is needed, spawn a subagent. That's the whole point.

When to Use

• A tasks.json file exists (created by prd-to-tasks or manually)
• The user says: "implement the tasks," "build it," "execute the plan," "start coding"
• Multiple interdependent tasks need coordinated execution

When NOT to Use

• No tasks.json exists — run prd-to-tasks first (or create-prd → prd-to-tasks)
• User wants to manually review/pick each task — use project-files TODO-driven workflow
• Exploration or research phase — use explore skill first

Note: There is no "just do it directly" exception. Even single tasks are delegated to a tmux subagent. The orchestrator stays clean.

Prerequisites

• A valid tasks.json file in the current working directory
• tmux installed and available on PATH
• pi CLI available on PATH (the subagent runner)
• `spawn-subagents` skill — Required for spawning isolated pi subagents in tmux
• `mixture-of-experts` skill — Required for expert definitions and MoE delegation patterns
• terminal-multiplexer skill — For tmux session management
• Reusable scripts in .agents/scripts/:
• validate-dag.ts — validates tasks.json structure and DAG
• generate-prompts.ts — writes tasks/TASK-XXXX-prompt from tasks.json (auto-validates)
• spawn-wave.sh — launches tmux sessions respecting the DAG
• status-tasks.ts — shows task status (done/running/pending/blocked)

Output Convention

All task artifacts live under tasks/:

tasks/
├── tasks.json                  ← input plan (generated by prd-to-tasks)
├── T001-prompt                 ← self-contained prompt for the subagent
├── T002-prompt
├── T003-prompt
├── ...
├── T001.out                    ← captured subagent stdout/stderr
├── T002.out
├── T003.out
├── ...
├── T001.done                   ← marker file (written by subagent on completion)
├── T002.done
└── ...

Rule: The orchestrator reads from /tmp/ for DONE detection, but writes prompt files and output logs to the project's tasks/ directory so they are versionable and reviewable.

Overview

tasks.json
    │
    ▼
┌─────────────────┐
│ 1. Parse &      │
│    Validate DAG  │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ 2. Generate     │
│    prompt files  │  → tasks/T001-prompt, T002-prompt, ...
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ 3. Spawn tmux   │
│    subagents     │  → one per task, respecting DAG waves
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ 4. Monitor &    │
│    report        │  → poll DONE markers, aggregate results
└─────────────────┘

Workflow

Step 1: Read and Parse tasks.json

Read the file and extract the task graph:

# Read the file
cat tasks.json

Parse mentally or with a script:

• Tasks — The full task list with IDs, descriptions, dependencies
• Phases — The phase grouping and order
• Metadata — Project name, total estimates
• Dependency graph — Build the DAG (who depends on whom)

Step 2: Validate the Task Graph

Before implementing, check for common issues:

1. All dependency IDs exist — No dangling references
2. No circular dependencies — The graph must be a DAG (can run a topological sort)
3. All tasks have acceptance criteria — Cannot verify completion without them
4. All tasks have agent and moeExperts — Required for delegation (added by prd-to-tasks step 6)
5. Phase ordering is logical — Foundation before features, core before polish

Run validation with the shared DAG validator:

bun .agents/scripts/validate-dag.ts tasks.json --summary

This checks: valid JSON, missing dependencies, circular dependencies, phase keys, unique IDs, agent/moeExperts fields, and agent summary consistency. With --summary it also prints the topological order, waves, and hour estimates.

If validation fails, the script exits with code 1 and prints a specific error. Stop and report the issues. Do NOT proceed until fixed.

Step 3: Determine Execution Order

Use topological sort (Kahn's algorithm) to produce a valid execution order. Group tasks into waves — same-wave tasks have no mutual deps, run in parallel:

Wait for all tasks in a wave to complete before launching the next wave.

Step 4: Generate Prompt Files

Use the shared prompt generator — it auto-validates, includes PRD context, and orders tasks topologically:

bun .agents/scripts/generate-prompts.ts

For manual generation or custom logic, write one self-contained prompt file per task to tasks/TASK-XXXX-prompt. These files are the only input subagents receive — they must contain everything needed to implement the task autonomously.

Output convention: Prompts go to tasks/TASK-XXXX-prompt. Subagent output goes to tasks/TASK-XXXX.out. DONE markers go to tasks/TASK-XXXX.done. Everything is versionable.

Step 5: Spawn tmux Subagents in Detached Sessions (NEVER implement directly)

This skill is a pure orchestrator — it never writes code itself. Every task is delegated to an isolated tmux session running pi. Use the shared spawn script which handles base64 encoding, prompt files, and wave ordering:

# Spawn next ready wave (auto-detects pending tasks with met dependencies)
# Returns IMMEDIATELY — sessions run detached in background
bash .agents/scripts/spawn-wave.sh

# Spawn specific tasks
bash .agents/scripts/spawn-wave.sh T001 T003

# Preview what would spawn
bash .agents/scripts/spawn-wave.sh --dry-run

The script auto-generates prompts if missing, validates tasks.json, and skips already-completed tasks (.done markers). Each call spawns one wave and returns immediately — the tmux sessions run detached in the background.

⚠️ NEVER use `spawn-wave.sh --all` from within an agent session. The --all flag blocks with sleep loops waiting for each wave, which will hit the tool timeout (300s). It is intended for humans running from a terminal. Agents must use the async wave-by-wave pattern below.

#### Async Wave-by-Wave Pattern

# 1. Spawn first wave (returns immediately)
bash .agents/scripts/spawn-wave.sh

# 2. Check status — poll until wave completes
bun .agents/scripts/status-tasks.ts --compact

# 3. When current wave is done, spawn the next wave
bash .agents/scripts/spawn-wave.sh

# 4. Repeat steps 2-3 until all tasks are done

#### Monitoring

bun .agents/scripts/status-tasks.ts          # full status table
bun .agents/scripts/status-tasks.ts --compact # one-line summary
bun .agents/scripts/status-tasks.ts --pending # only pending/blocked

# Peek at a specific task's live output
tmux capture-pane -t task-T005 -p | tail -20

#### Checking tmux sessions directly

tmux ls                                    # list all sessions
tmux ls | grep "task-"                     # list task sessions only
tmux kill-session -t task-T005             # kill a specific task session
tmux kill-session -t task-                 # kill all task sessions

Step 6: Verify Completion

After each task, verify against acceptance criteria:

✅ T001: Set up project structure
  - ✓ Directory structure matches conventions
  - ✓ Type definitions compile
  - ✓ Configuration loads correctly

✅ T002: JWT token generation
  - ✓ Tokens generated with correct claims
  - ✓ Expired tokens rejected
  - ✓ Tampered tokens rejected
  - ✓ Validation under 5ms

If any criterion fails, fix before marking the task complete.

Step 7: Handle Failures

If a task cannot be completed:

1. Diagnose the issue — Is it a code bug, missing info, design flaw?
2. If fixable: Fix and re-verify
3. If blocked by missing info: Use grill-me to ask the user
4. If the approach is wrong: Update the task or PRD and re-plan
5. If dependency was done wrong: Fix the dependency first, then retry

# Example: Task fails because dependency T002 has a bug
echo "Task T003 blocked: Middleware test fails because T002's token validation
doesn't handle edge case X. Going back to fix T002 first."

# Re-run T002 with the specific issue
pi -p "Fix T002 (JWT token validation): handle edge case X where [details].
Acceptance criteria: [original + new]"

Step 8: Report Progress

Report progress after each phase completion:

Phase 1: Foundation — ✅ Complete (3/3 tasks, 8h estimated, 7.5h actual)

✅ T001 - Project setup (2h)
✅ T002 - JWT utils (4h)
✅ T003 - Auth middleware (1.5h — simpler than expected)

Phase 2: Core — 🔄 In Progress (1/3 tasks)

✅ T004 - Login endpoint (3h)
🔄 T005 - Registration endpoint (in progress...)
⏳ T006 - Password reset (waiting on T005)

Next: T005 → T006 → Phase 3 (Polish)

Step 9: Final Report

When all tasks are complete, produce a synthesis report:

# Implementation Complete: [Project Name]

**Completed:** YYYY-MM-DD
**Tasks:** 8/8 complete
**Estimated:** 32h | **Actual:** 30h

## Summary

[2-3 sentences about what was built]

## Phase Breakdown

### Phase 1: Foundation

- T001, T002, T003 completed

### Phase 2: Core

- T004, T005, T006 completed

### Phase 3: Polish

- T007, T008 completed

## Files Changed

- `src/auth/types.ts` — JWT type definitions
- `src/auth/utils.ts` — Token generation and validation
- `src/auth/middleware.ts` — Auth middleware
- `src/routes/auth/login.ts` — Login endpoint
- `src/routes/auth/register.ts` — Registration endpoint
- `src/routes/auth/reset.ts` — Password reset endpoint
- `tests/auth/` — Test suite (12 tests)
- `PRD-auth.md` — Updated with implementation notes

## Acceptance Criteria Status

All 24 criteria across 8 tasks verified.

## Known Issues / Follow-ups

- None

## Next Steps

- Deploy to staging for QA
- Run integration tests against staging

Example: Full Orchestration Run (Async Pattern)

#!/bin/bash
# Pure orchestrator — validates, generates prompts, spawns one wave, exits.
# The agent polls status and spawns subsequent waves independently.
# NEVER implements code directly.

set -e
TASKS_FILE="tasks.json"
SCRIPTS=".agents/scripts"

echo "🚀 Orchestrating: $(bun -e "import { loadTasks } from './$SCRIPTS/lib/tasks-lib.ts'; const d = await loadTasks('$TASKS_FILE'); console.log(d.metadata.project)")"

# Step 1: Validate DAG
bun $SCRIPTS/validate-dag.ts "$TASKS_FILE" --summary

# Step 2: Generate prompt files (auto-validates, includes PRD context)
bun $SCRIPTS/generate-prompts.ts "$TASKS_FILE"

# Step 3: Spawn the first wave (returns immediately — sessions run detached)
bash $SCRIPTS/spawn-wave.sh

# The agent now reports wave status to the user, polls for completion,
# and spawns subsequent waves as they become available.
# Each wave is spawned with: bash $SCRIPTS/spawn-wave.sh
# Status is checked with: bun $SCRIPTS/status-tasks.ts --compact
echo "📡 First wave spawned. Poll with: bun $SCRIPTS/status-tasks.ts --compact"

Best Practices

DOs

• Validate the graph first — Don't start implementing a broken plan
• Verify each task against acceptance criteria — Don't mark done without checking
• Fix problems at the source — If T005 fails because T002 is buggy, fix T002
• Report progress clearly — Phase + task status so the user knows what's happening
• Use explicit agent and moeExperts from tasks.json — Read them directly, don't infer. They were assigned during prd-to-tasks step 6 for a reason.
• Keep the user in the loop — Especially when a task is blocked or needs clarification

DON'Ts

• Don't implement code yourself — The orchestrator delegates 100% of work to subagents
• Don't use `tmux send-keys` for prompts — Quoting breaks on multi-line content. Use prompt files + --append-system-prompt instead
• Don't embed prompts in shell scripts — Use Python to write files; avoid heredocs with quotes
• Don't skip dependencies — Tasks built on broken foundations are broken
• Don't run everything in parallel — Respect the DAG; work wave by wave
• Don't use `spawn-wave.sh --all` — It blocks with sleep loops and hits tool timeouts. Always spawn one wave at a time and poll for completion
• Don't ignore acceptance criteria — They define "done"
• Don't proceed after validation failure — Fix the tasks.json first
• Don't lose track of what's done — Keep IMPLEMENTATION_STATUS.md updated

Scaling: Large Task Sets (50+ Tasks)

For large projects:

• Process phase by phase — Complete all of Phase 1 before starting Phase 2
• Batch parallel tasks — Within a phase, run independent tasks concurrently (up to 3-5 at a time)
• Pause between phases — Report progress and let the user review before continuing
• Use tmux sessions — Long-running tasks in background tmux sessions
• Track with a status file — Write IMPLEMENTATION_STATUS.md to persist progress:

# Implementation Status: User Auth System

Last Updated: 2026-04-27 14:30

## Phase 1: Foundation

- [x] T001: Project setup
- [x] T002: JWT utils
- [x] T003: Auth middleware

## Phase 2: Core

- [x] T004: Login endpoint
- [ ] T005: Registration (in progress)
- [ ] T006: Password reset

Integration with Other Skills

create-prd    → Produces PRD
prd-to-tasks  → Produces tasks.json
implement-tasks → Executes tasks.json
    ├── Uses: spawn-subagents (delegate work to isolated pi instances)
    ├── Uses: mixture-of-experts (expert definitions, spawn/aggregate patterns)
    ├── Uses: terminal-multiplexer (for tmux session management)
    ├── Uses: grill-me (when blocked by ambiguity)
    └── Uses: project-files (for status tracking)

Edge Cases

Empty Task List

If tasks.json has zero tasks:

tasks.json has 0 tasks. Nothing to implement.
Was the PRD created? Run create-prd first.

Single Task

For a single task, skip the full orchestration and just implement it with a lighter-weight review:

Only 1 task (T001). No dependencies. Implementing directly with architect + maintainer review.

Set up a focused MoE call with just 2 experts.

Blocked by External Dependency

Task T003 depends on T000 (EXTERNAL: API credentials).
T000 is marked external/blocker. Cannot proceed past T000.

Options:
1. Wait for credentials (pause implementation)
2. Mock the external service for development (create a mock task)
3. Skip T003 and implement other independent tasks

Failed Validation

If the tasks.json has issues, report them clearly:

❌ tasks.json validation failed:

1. T007 depends on T999 (does not exist)
2. Circular dependency: T004 → T006 → T004

Fix these before implementing. Suggested fixes:
- T007: Did you mean T009?
- T004/T006: Remove one dependency direction

User Interrupts Mid-Implementation

Save progress and create a handoff:

# Write current state
cat > IMPLEMENTATION_STATUS.md << 'EOF'
# Implementation Status: Interrupted

Last Updated: 2026-04-27 15:00
Interrupted during: Phase 2, Task T005 (Registration endpoint)

## Completed
- [x] T001-T004

## In Progress
- [ ] T005 (partially done — auth utils written, endpoint stubbed)

## Not Started
- [ ] T006-T012
EOF

echo "Progress saved to IMPLEMENTATION_STATUS.md"
echo "Resume with: 'continue implementing tasks'"