decompose-task

decompose-task

Wrapper for the decompose tool on the local NeuroDock task-fractionator server (neurodock-task-fractionator). Takes a goal and returns an ordered list of atomic tasks the user can pick up cold.

Design rationale: ADR 0003 — task-fractionator tool design ("goblin-style" decomposition; acceptance bar is I can start this right now without thinking).

When to use

• The user states a goal that is bigger than a single task ("ship the founding-scope RFC by Friday").
• A planner skill needs structured sub-tasks to schedule.
• The user is staring at a vague item and cannot start.

Do NOT use for:

• Tasks already small enough (a single email reply).
• Strategic planning over weeks/months — this tool caps total estimated_minutes at the decompose-level budget, not a roadmap.

What it does

Calls mcp__neurodock-task-fractionator__decompose. The server:

1. Decomposes the goal into 3–12 atomic tasks (hard cap 20).
2. Each task gets: id (UUIDv4), title, description, estimated_minutes (5–90), acceptance_criteria[] (≥1), dependencies[] (refs to sibling ids), sequence (1-indexed total order), tags[].
3. Returns the tasks array plus a rationale paragraph explaining the split.
4. Does NOT persist anything — the caller is responsible for writing the result to the cognitive graph (see the record-fact skill) if persistence is desired.

How to invoke

Minimal:

{
  "goal": "ship the founding-scope RFC by Friday"
}

With a time budget (ISO 8601 duration — natural language is rejected):

{
  "goal": "fix the chronometric idle-status consent bug",
  "time_budget": "PT3H"
}

Common time_budget values:

• PT30M — 30 minutes
• PT3H — 3 hours
• PT1H30M — 1 hour 30 minutes
• P1D — 1 day
• P3D — 3 days

If the budget cannot fit a credible task list, the server returns BUDGET_INFEASIBLE rather than silently truncating.

Output shape (excerpt)

{
  "tasks": [
    {
      "id": "c1e4d2a8-7b3f-4f9a-9c1e-2d3a4b5c6d7e",
      "title": "Draft the manifesto section",
      "description": "Write the manifesto section covering ...",
      "estimated_minutes": 45,
      "acceptance_criteria": [
        "File exists at the named path",
        "Section headings match the plan"
      ],
      "dependencies": [],
      "sequence": 1,
      "tags": ["writing"]
    }
  ],
  "rationale": "Split into draft, review, publish because ..."
}

The "I can start now" acceptance bar

Per ADR 0003: every returned task must be small enough that an ADHD-prone user can pick it up cold without further planning. If a task's description uses words like "consider", "think about", or "explore", that is the server's bug — the task is not atomic. Surface the issue; do not paper over it.

Errors

• GOAL_REQUIRED — empty goal.
• GOAL_TOO_LONG — over 500 chars (itself a hyperfocus signal — split the goal first).
• TIME_BUDGET_UNPARSEABLE — time_budget was not ISO 8601.
• BUDGET_INFEASIBLE — no credible task list fits the budget.
• ACCEPTANCE_CRITERIA_REQUIRED — internal; means the model produced a task with no criteria. Retry.
• DEPENDENCY_CYCLE — internal; retry.

Limitations

• Stateless. The task list is gone the moment you discard the response. Persist via record-fact (predicate: "depends_on" for the edges) if you want it back tomorrow.
• Task estimates are LLM-generated and tend optimistic. Consider adding a buffer when the user is showing distress signals; this server does not adjust the estimates for you.
• Cross-response dependencies are not modelled — dependencies[] references siblings only.

Voice

When showing the task list, lead with sequence 1 — "start here". Do not present the rationale before the tasks; the user wants the next action, not the meta-explanation. Save the rationale for a collapsible follow-up.