Launchpad plan

Launchpad Plan

Analyze a task breakdown and produce a parallelization-aware implementation plan.

Purpose

Bridge between the spec's task list and the runner's parallel execution. The planner determines what can run concurrently and what must be sequential by analyzing dependency declarations and file mutation overlaps.

Workflow

1. Locate the Spec

Resolve the spec name from the user's request. Read specs/<spec-name>.tasks.md (or equivalent task breakdown file).

If missing or ambiguous, list specs/ and ask.

2. Parse Task Sections

Extract all task sections. For each task, collect:

• Task ID: Identifier (e.g., task-001, T-0, etc.)
• Files: Each file path and its operation (create vs modify)
• Steps: The implementation steps
• Tests: The test/validation scenarios
• Dependencies: Explicit from the spec, plus implicit from section ordering

3. Build the Dependency Graph

For each task, determine what it depends on:

1. Explicit dependencies: From the spec's task definitions (e.g., depends_on fields).
2. File mutation conflicts: Two tasks that both modify the same file cannot safely run in parallel. If task A appears before task B and both modify src/index.ts, B depends on A.
3. Shared types/foundation dependency: If a foundational task exists (e.g., types, schema, config), all dependent tasks must wait for it.
4. Create-then-modify: If task A creates a file and task B modifies or imports from it, B depends on A.

4. Determine Task Granularity

Each task section becomes one task by default. Adjust when:

• Split a task if it contains clearly independent substeps touching different files with no data flow between them, AND splitting would enable more parallelism. Don't split just for granularity's sake.
• Merge tasks that are trivially small into a parent task or mark as no-op.

5. Assign Parallel Waves

Group tasks into execution waves — sets of tasks that can run concurrently:

• Wave 0: Tasks with no dependencies (foundational setup, independent features)
• Wave 1: Tasks whose dependencies are all in wave 0
• Wave N: Tasks whose dependencies are all in waves < N

Tasks in the same wave run in parallel. Waves execute sequentially.

6. Output Format

For launchpad-run integration (default)

When the output is intended for /lp:run, produce harness-tasks.json compatible format:

{
  "version": 3,
  "created": "<ISO timestamp>",
  "session_config": {
    "max_tasks_per_session": 20,
    "max_sessions": 50,
    "max_parallel_agents": 3
  },
  "tasks": [
    {
      "id": "task-001",
      "title": "Short task name",
      "status": "pending",
      "priority": "P0",
      "depends_on": [],
      "attempts": 0,
      "max_attempts": 3,
      "context": "Detailed description with enough context for a sub-agent to execute without reading the spec. Include: files to touch, implementation steps, constraints, expected behavior.",
      "files_hint": ["src/auth/", "src/middleware/auth.ts"],
      "validation": {
        "command": "npm test -- --testPathPattern=auth",
        "timeout_seconds": 300
      },
      "on_failure": {
        "cleanup": null
      },
      "error_log": [],
      "agent_id": null,
      "worktree_branch": null,
      "completed_at": null
    }
  ],
  "session_count": 0,
  "last_session": null
}

Field mapping from spec to runner:

• Task title → title
• Task description + steps + test scenarios → context (must be self-contained for sub-agent)
• File list → files_hint (used for batch conflict detection)
• Test command → validation.command
• Dependencies → depends_on (task IDs)
• Wave 0 tasks → priority: "P0", Wave 1 → priority: "P1", Wave 2+ → priority: "P2"

For generic use

When not targeting launchpad-run, produce implementation.json:

[
  {
    "id": "T-0",
    "name": "Short task name",
    "depends": [],
    "wave": 0,
    "kind": "code",
    "files": {
      "create": ["src/auth/service.ts"],
      "modify": ["src/index.ts"]
    },
    "description": "Detailed description with enough context for an agent to execute."
  }
]

7. Confirm with User

After writing the output, present:

• Total task count and wave count
• Which tasks run in parallel vs sequentially and why
• Any granularity decisions made (splits, merges)
• Offer to adjust groupings or wave assignments

Guidelines

• Each task's context/description must contain enough detail that an agent can execute it without reading the original spec. The orchestrator feeds these as agent prompts.
• Preserve the spec's step details and test scenarios — don't summarize away actionable information.
• Flag potential conflict risks in task descriptions (e.g., "Modifies index.ts — coordinate with task-003 if running in same wave").
• When files_hint is empty for a task, the runner will run it solo (not batched) to avoid unknown conflicts.

Integration with Launchpad

The typical workflow:

/lp:spec → /lp:refine → /lp:clarify → /lp:tasks
    ↓
/lp:plan  (reads specs/*.tasks.md → writes harness-tasks.json)
    ↓
/lp:run  (reads harness-tasks.json → dispatches sub-agents)

The planner's job is to transform human-readable task breakdowns into machine-executable task plans with parallelization awareness.