| name | backend-tasks |
| description | Split an approved backend scope + contract into a small, ordered, test-first task list with a verification step per task and risk flags. Read-only planning — never edits code. |
| allowed-tools | Read, Grep, Glob, Bash, Write |
| tags | ["sdlc","plan","backend"] |
backend-tasks
Break the approved backend scope into the smallest ordered set of verifiable tasks that
implement the contract. Small tasks keep the implement→verify→fix loop tight and every
change reviewable. Read-only — produces a task list, not code.
Inputs
Your instructions name what to read — the cross-repo contract and the backend LLD — and the
artifact path to write. Standalone? read them and write to a path you choose (and tell the user where).
Steps
- Read the contract and the backend LLD (the inputs your instructions point to); read
the target service's structure read-only.
- Slice vertically where possible — a thin end-to-end slice (schema→API→test) before
breadth, so value is demonstrable early.
- Order by dependency — types/schema → domain/service logic → persistence →
API/controller → tests → logging/metrics → docs. Sequence migrations
expand → migrate → contract; never a breaking migration in one step.
- Make each task small — ~one commit, one testable behavior, exact files.
- Attach a test + standards + risk to every task (below).
- Write the task list artifact.
Each task must specify
id, title, and the exact files it touches.test — the check that proves it (the failing test to write first).standards — which of {security, backward-compat, rate-limiting, idempotency,
validation, observability, migrations, performance} this task must honor.needs_human_gate — true if it touches DB migration, auth/permission, payment, prod
config, or dependencies.
Coverage — the plan as a whole must include tasks for
- Input validation (all fields, bounds, formats) and the error responses.
- The negative/authz paths from the contract, not just the happy path.
- Idempotency & rate limiting where the contract requires them.
- Migration + backfill of existing data (+ a rollback task) when the schema changes.
- Observability (logs/metrics/traces) and docs.
- A feature-flag task if the HLD calls for gated rollout.
Edge cases to plan for explicitly
- Tasks touching shared/legacy code with no tests (add characterization tests first).
- Migrations needing online/zero-downtime handling or large backfills.
- Tasks that require a coordinated frontend/contract change (flag the dependency).
- Concurrency-sensitive writes (add a locking/versioning task).
External skill (provision — planner)
If the writing-plans skill (from the Superpowers pack) is installed, use it to produce
bite-sized, verifiable tasks — each task must still carry files + test + standards + risk as
above. If it is not installed, plan in-pack.
Output
Author it in one shot, then refine once — compose the whole tasks.json (manifest, all
tasks[], slices[]) in a single Write; do not stub-and-grow it with successive Edits.
Validate once; if it fails, make one corrective edit and re-validate (repeat only to fix
validation errors, never to build the file up incrementally).
Write the DAG to the tasks.json path your instructions specify (the orchestrator passes it;
run standalone? use a sensible path you choose) conforming to
engine/schemas/tasks.schema.json (the same shape the design step emits): context_manifest
(batched-read files), tasks[] (id, group_id, title, depends_on intra-group only,
reads, writes, test, standards, needs_human_gate), and slices[] (one per
independent group — two tasks share a group iff one depends on the other or they write a
common file). Validate with
python3 engine/validate_tasks.py <the tasks.json path> (must print OK).
Definition of done
Every task ≤ ~1 commit, has a test and standards, and is dependency-ordered; migrations are
expand/contract with rollback; negative paths and observability are represented; risky tasks
are flagged; tasks.json validates against the schema (no cross-group edges, disjoint writes across groups).
Output contract
Return tasks_path, task_count (integer), slice_count (integer), and risky
(true if any task needs a human gate) — all short scalars.
