worc-deco-task

---

name: worc-deco-task description: Convert a set of related steps into wastech-orchestrator's operator-authored decomposition — one root task with a subtasks: list plus per-subtask spec files, run sequentially on one branch / one PR. Use when a change is one coherent unit you already know how to split into ordered steps that should land together.

worc-deco-task

Turn a change you already know how to split into operator-authored decomposition: one root task that carries the shared context, plus an ordered set of subtask spec files. The orchestrator runs the subtasks sequentially, on one branch, into one PR — the same machinery as an agent-proposed split, but the carving is yours.

Speak in the user's language (default to the language they wrote in).

When to use

• The user has one coherent change plus a clear, ordered list of steps to get there, and wants them done together as a single PR.
• Guardrail: if the parts are independent and each should get its own PR, those are separate tasks — author each with worc-task, not this skill. Decomposition is for steps that share one branch and one review.

How to run

1. Separate root from steps. The root holds the shared context (what the whole change is and why). The subtasks are the ordered units to implement it.
2. Check the count. You need 2 ≤ n ≤ max_subtasks, where max_subtasks is the operator's agents.decomposition.max_subtasks (default 8). Fewer than 2 isn't a decomposition; more than the max is rejected — group steps together or split the work into separate tasks instead.
3. Write the root task at tasks/preparing/<root-id>.md — the staging folder the watch daemon never scans (so an in-progress batch is never picked up mid-write) — a normal task file (same front-matter and body rules as a single task: leading ---, valid id, non-empty ## Description, only allowed keys, no secrets, no flag-shaped values) plus a subtasks: list of repo-relative paths pointing into a subfolder, in execution order.
4. Write each subtask spec at tasks/preparing/subtasks/NN-<slug>.md (NN = zero-padded order):
   • Front matter: title (required); slug (optional — defaults to slugify(title); it names the NN-<slug>.md file); depends_on (optional — a list of slugs of EARLIER subtasks only).
   • No id. A spec file is a reduced spec, not a standalone task — that is what keeps it from ever running on its own.
   • Body: ## Acceptance criteria (and any per-step instructions). The body is materialized verbatim and injected into the implementation step, so write it however the step needs.
5. Promote the batch when it is complete: worc promote <root-id> moves the root and the subtask specs it references together into tasks/pending/ (subfolder preserved), atomically — so the root never reaches the queue without its specs. (worc promote --all promotes everything staged, including the whole subtasks/ subfolder.) The subtask paths are relative to the root, so they stay valid after the move; no rewriting.
6. Self-check against the hard rules below before finishing.

The decomposition contract

Root task — note the subtasks: list of ordered, repo-relative paths into a subfolder:

---
id: epic-checkout
title: "Rework checkout into a multi-step flow"
subtasks:
  - subtasks/01-cart-model.md
  - subtasks/02-payment-step.md
  - subtasks/03-confirmation.md
---

## Description

Shared context for the whole change — the once-per-task framing every subtask inherits.

## Acceptance criteria

- [ ] The whole checkout flow works end to end after all subtasks land.

Subtask spec (tasks/preparing/subtasks/02-payment-step.md, promoted to tasks/pending/subtasks/…) — a reduced spec, no id:

---
title: "Add the payment step"
slug: "payment-step" # optional; defaults to slugify(title)
depends_on: ["cart-model"] # optional; slugs of EARLIER subtasks only
---

## Acceptance criteria

- [ ] Concrete, testable unit-level criteria for this step.

Hard rules / fail-closed reasons

The whole split is validated before any branch and quarantined (to tasks/rejected/) on any violation:

• Count out of range — fewer than 2 or more than max_subtasks (subtask_count_out_of_range).
• Forward / self / unknown dependency — a subtask's depends_on may reference only the slugs of subtasks before it; the gate enforces a linear order (subtask_depends_forward).
• Missing or malformed spec file — a referenced file that doesn't exist, has no front matter, lacks a title, or has an empty body (subtask_file_missing, subtask_malformed).
• Bad path — every subtasks: entry must be repo-relative with no .., no absolute path, no traversal, and must resolve under the task directory. Spec files must live in a subfolder (e.g. tasks/preparing/subtasks/…, promoted to tasks/pending/subtasks/…); a path beside the root task is rejected (the scheduler scans only the top level, so a subfolder also keeps specs from running as standalone tasks).
• Flow can't decompose — the task's task_type selects a flow with no decomposition: block (flow_cannot_decompose). The default implementation flow supports it.

The root task itself must still obey every single-task rule (valid id, only allowed front-matter keys, no secrets, no flag-shaped values, etc.).

What not to do

• Don't give a subtask spec an id — it is a spec, not a task; an id would make it look standalone.
• Don't reference a later subtask's slug (or its own) in depends_on — dependencies are linear and backward-only.
• Don't place spec files at the top level of tasks/preparing/ or tasks/pending/ — keep them in a subtasks/ subfolder.
• Don't exceed max_subtasks (default 8) — group steps or split into separate tasks.
• Don't use decomposition for independent changes that each deserve their own PR — those are separate tasks (worc-task), optionally chained with depends_on.