/tasks

name: tasks
model: inherit
effort: medium
agents: []
description: >
  Use to break a designed feature into atomic, ≤1-day tasks with a dependency graph, a
  per-task Definition of Done, and a machine-readable tasks.json that the implement engine
  consumes. Triggers on "task breakdown for {slug}", "break down tasks for {slug}",
  "tasks for {slug}", "plan the work for {slug}", "/sdd:tasks {slug}", "розбий на задачі {slug}",
  "декомпозиція {slug}", "список задач". Reads spec.md + sad.md + Accepted ADRs (+ data-model +
  openapi if present), writes docs/features/{slug}/tasks/{_epic,tracker,<task>}.md AND
  docs/features/{slug}/tasks.json. Tracker export to any issue tracker is optional and
  tool-neutral. Hard-refuses if spec.md or sad.md or an Accepted ADR is missing.

Skill: tasks

Task-breakdown generator: atomic tasks ≤1 day, each a separately reviewable change (≤~500 LOC preferred), with a visible dependency graph and a Definition of Done per task. One task = one focused session = one PR. "Build the feature" is not a task — break it down.

A task file is **self-contained**. The rule that governs every task body: **inline the slice the task actually needs, name where it came from, and keep the link as the fallback for when the slice turns out not to be enough.** A task that only points at `spec.md §AC-N` makes the executing agent go and rebuild the context this breakdown already had — it burns its budget on rediscovery and still misses. Alongside the human-facing markdown, this skill emits **`tasks.json`**, the contract the `implement` engine reads to build its dependency DAG.

Every inlined chunk carries a one-line **provenance signature** — `<file> §<section>, <identifier>, verbatim|abridged`, e.g. «`spec.md` §5, AC-02, verbatim» — never «see the spec». The signature is what makes an inline re-checkable: an inline is a snapshot taken at breakdown time, upstream can move after it, and the source always wins.

Task prose (`title` / `dod`, the markdown bodies) follows `artifact_language` — the `tasks.json` machine fields (`id`, `layer`, `deps`, `acs`, `files_hint`, `slug`) and tracker states stay English → [`../_shared/artifact-language.md`](../_shared/artifact-language.md). An inlined chunk is quoted **verbatim in the language its source artifact is written in**; the signature is not translated.

Owner

Tech Lead.

Inputs

• `<slug>` — feature slug.
• **Gate (hard refuse):** `docs/features/<slug>/spec.md` + `docs/features/<slug>/sad.md` + ≥1 Accepted ADR in `adr/`. Missing → STOP and point at the producing skill (`specify` / `design` / `decide-adr`).
• Read directly (not via an index): spec §5 AC + §6 NFR, sad §5 module boundaries + §6 runtime + §9 ADR index, each Accepted ADR, and — if present — `data-model.md`, `contracts/openapi.yaml` and `screens.md` (the screen manifest `ui` tasks cite).
• (Expected) `sad.md` frontmatter `target_surfaces` — gates which layers appear (step 4). **Absent or empty → warn** («surfaces undeclared — re-run `design`, or proceeding as `backend-service`») **and treat as `[backend-service]`** (→ [`../_shared/surfaces.md`](../_shared/surfaces.md)); never silently emit `ui` tasks for an undeclared surface.

Protocol

1. **Prereq check (hard).** spec.md + sad.md + ≥1 Accepted ADR, else refuse with the missing one named. 2. **Read upstream directly.** Each task will carry the slices it needs, quoted from the source — so read the source text itself, never a paraphrase of it: the quote goes into the file. 3. **Read the templates from disk, then scaffold output.** **Read [`./templates/task.md`](./templates/task.md) before writing a single task file** — its frontmatter keys and its `## ` section list are the contract, and its `<!-- instruction … -->` comments are the per-section brief. Same for [`./templates/_epic.md`](./templates/_epic.md) and [`./templates/tracker.md`](./templates/tracker.md). **Never write a task from memory of what a task file looks like:** the self-contained format (the `blocks` / `context_budget` keys, the nine sections, the provenance signatures) is younger than most recollections of it, and a remembered shape silently reverts this skill to the link-only tasks it exists to replace. Output: `docs/features/<slug>/tasks/` with `_epic.md` (summary + links + the DAG `flowchart`), `tracker.md` (status table), one `<task-slug>.md` per task. **Validate the `_epic.md` `flowchart` per [`../_shared/mermaid-check.md`](../_shared/mermaid-check.md)** (render-parse with `mmdc` if available, else the structural lint; fix before committing). 4. **Identify work-items by layer.** Generic, stack-agnostic layers: `migration` (DB) · `domain` (entities/invariants) · `infra` (repo/persistence) · `app` (service/use-case) · `ports` (handler/API) · `ui` (UI components / screens / view-state — only when a UI surface is declared) · `tests` · `wiring` (composition/DI) · `docs`. **`sad.md` frontmatter `target_surfaces` gates which layers appear** (→ [`../_shared/surfaces.md`](../_shared/surfaces.md)): a `web-frontend` / `mobile-app` / `desktop-app` surface adds `ui` tasks; a backend-only feature emits domain/infra/app/ports (no `ui`); a `cli` feature app/ports; a `worker` domain/infra. Each `ui` task **names the existing components / tokens / styling it reuses** (from `architecture-map.md` §Frontend and the `docs/design-system.md` inventory) — a *new* component is listed only when no existing primitive fits — and, **when `docs/features/<slug>/screens.md` exists, cites the `SCR-NN` id(s) + the states it builds** (the manifest is the task's screen contract; `implement` builds to those states). List 8–20 items by size (see [`../_shared/size-matrix.md`](../_shared/size-matrix.md)). 5. **Atomic check — work size and context size.** Each task ≤1 working day. More → split. **And atomic in context:** count the inlined lines the finished body will carry (the non-empty lines from `## Why (user story)` through `## Acceptance criteria`) against the `context_budget`