Launchpad

End-to-end orchestrator that chains launchpad-spec → launchpad-plan → launchpad-run into a single guided workflow. Ensures every phase completes before the next begins, and asks the user the right questions at every step.

Commands

/launchpad <feature description>   # Start full pipeline from scratch
/launchpad status                  # Show current pipeline stage and progress
/launchpad resume                  # Resume from where you left off

Pipeline Stages

 STAGE 1: SPEC          STAGE 2: PLAN           STAGE 3: RUN
 ┌──────────────┐       ┌──────────────┐       ┌──────────────┐
 │ /lp:spec     │       │ Dependency   │       │ /lp:run      │
 │      ↓       │       │ analysis     │       │      ↓       │
 │ /lp:refine   │──────→│      ↓       │──────→│ Dispatch     │
 │      ↓       │       │ Wave assign  │       │ sub-agents   │
 │ /lp:clarify  │       │      ↓       │       │      ↓       │
 │      ↓       │       │ harness-     │       │ Validate &   │
 │ /lp:tasks    │       │ tasks.json   │       │ merge        │
 └──────────────┘       └──────────────┘       └──────────────┘
   User answers            Automatic             Autonomous
   questions here          (with review)          (with recovery)

Stage 1: Spec (uses launchpad-spec skill)

Entry

When the user triggers /launchpad <description>:

1. Announce stage: "Stage 1/3: Spec — Defining requirements clearly."
2. Invoke /lp:spec with the user's description
3. Track state: Write .launchpad.json to project root (see State File below)

Iteration Loop

After /lp:spec creates the spec with questions:

1. Present questions to the user
2. Wait for answers → invoke /lp:clarify with user's response
3. If more questions arise → repeat
4. If spec needs research → suggest and invoke /lp:refine
5. Gate check: Only proceed when "Open Questions" section is empty

Gate 1 → 2

When all questions are resolved:

1. Invoke /lp:tasks to generate the task breakdown
2. Present task list to the user
3. Ask: "Please review the task list. Let me know if anything is missing or needs changes. If it looks good, we'll proceed to the next stage."
4. If user requests changes → edit task file → re-present
5. If user approves → proceed to Stage 2

---

Stage 2: Plan (uses launchpad-plan skill)

Entry

1. Announce stage: "Stage 2/3: Plan — Analyzing task dependencies and generating a parallel execution plan."
2. Invoke /lp:plan on the task file from Stage 1

User Review Points

The plan skill produces the parallelization plan. Present to the user:

1. Wave summary: "Total N tasks organized into M waves."
2. Wave detail table:

   Wave 0 (parallel): task-001, task-002, task-004
   Wave 1 (parallel): task-003, task-005
   Wave 2 (sequential): task-006
   

3. Conflict notes: Any file overlap warnings
4. Ask: "Please review the execution plan. Let me know if you'd like to adjust the wave assignments or dependencies."

Gate 2 → 3

When user approves the plan:

1. Verify harness-tasks.json has been written
2. Verify all tasks have context (sub-agent prompt content)
3. Verify all tasks have files_hint (batch conflict detection)
4. Verify all tasks have validation.command — if any are missing, ask the user: "Task task-003 'Add OAuth providers' is missing a validation command. What command should be used? (e.g., npm test -- --testPathPattern=oauth)"
5. Ask: "Everything is ready. Shall I start execution?"

---

Stage 3: Run (uses launchpad-run skill)

Entry

1. Announce stage: "Stage 3/3: Run — Sub-agents are executing tasks in parallel."
2. Invoke /lp:run

During Execution

The run skill handles autonomous execution. The pipeline skill adds:

1. Batch progress reports: After each batch completes, summarize:

   Batch 1 complete: task-001 ✓, task-002 ✓, task-004 ✗ (TEST_FAIL)
   Next batch: task-003, task-005
   

2. Failure triage: When a task fails, present the error and ask: "task-004 failed: [error details]. Should I auto-retry, or would you like to modify the task?"
   • Retry → let run skill retry automatically
   • Modify → edit task context/validation, then re-dispatch

Completion

When all tasks are done:

1. Final summary:

   Pipeline complete!
   - Spec: specs/feature-name.md (COMPLETED)
   - Tasks: 8/8 completed, 0 failed
   - Sessions: 2, Batches: 4
   - Total sub-agent dispatches: 10 (2 retries)
   

2. Suggest next steps:
   • "Let me know if you'd like a code review."
   • "Shall I create a PR?"

---

State File: .launchpad.json

Tracks which stage the pipeline is in, for session resume:

{
  "feature": "auth-system",
  "stage": "spec",
  "spec_file": "specs/auth-system.md",
  "task_file": "specs/auth-system.tasks.md",
  "harness_file": "harness-tasks.json",
  "stage_history": [
    { "stage": "spec", "status": "completed", "timestamp": "2025-07-01T10:00:00Z" },
    { "stage": "plan", "status": "in_progress", "timestamp": "2025-07-01T10:30:00Z" }
  ]
}

---

/launchpad status Command

Read .launchpad.json and report:

Launchpad: auth-system
├─ Stage 1 (Spec):  ✓ completed — specs/auth-system.md
├─ Stage 2 (Plan):  ● in progress — 5 tasks, 3 waves
└─ Stage 3 (Run):   ○ pending

Use /launchpad resume to continue.

---

/launchpad resume Command

1. Read .launchpad.json
2. Determine current stage
3. Resume from the appropriate point:
   • Stage 1: Re-read spec, check for open questions, continue iteration
   • Stage 2: Re-read task file, re-run plan if needed
   • Stage 3: Invoke /lp:run (run skill has its own session recovery)

---

Error Handling

Situation	Action
User says /launchpad without description	Ask: "What feature would you like to build?"
Spec has unresolved questions at gate	Block: "There are still N open questions. Please answer them with /lp:clarify."
Task has no validation command	Ask user for the command before proceeding
Task has no context	Generate from spec + task title, present for user review
Task has no files_hint	Ask: "Which files will this task modify?"
Run fails completely	Present error, offer to re-run plan with adjusted plan
Context window approaching limit	Save state to .launchpad.json, suggest /launchpad resume in new session

---

Key Rules

1. Never skip a stage. Even if the user says "just run it", walk through all three stages.
2. Always get user confirmation at gates. The two gate checkpoints (spec→plan, plan→run) require explicit user approval.
3. Fill gaps by asking, not guessing. If any required field is missing (validation command, files_hint, context), ask the user rather than making assumptions.
4. One feature per pipeline. If the user wants to build multiple features, run separate pipelines (or suggest splitting into one spec with multiple tasks).
5. State survives sessions. Always read/write .launchpad.json so the pipeline can resume after context reset.