repo-tasks
Plan only the repository named by your instructions. Read its AGENTS.md/CLAUDE.md,
codebase map, approved LLD, contract, and the live files needed to verify paths. Do not edit
product code.
Shape
Write the requested tasks.json in one pass against the task schema your instructions name
(standalone: describe the same structure in a file you choose). Set stack to the repository
name.
- Small, verifiable tasks — about one commit and one testable behavior each, with exact
reads and writes, a concrete test to write first, the standards it must honor, and whether
it touches a sensitive area (
needs_human_gate: migrations, auth/permission, payments, production config, dependencies) — the approved LLD already authorizes it; the flag makes the implementer build it exactly as designed and the review check it closely. - Slices — split work into independent slices with disjoint writes; dependencies may only point within a slice. Prefer a thin end-to-end slice before breadth.
- Order — types and schema → domain logic → persistence → API, UI or entry point → tests → logging and metrics → docs. Migrations go expand → migrate → contract, never breaking in one step.
Coverage
The plan as a whole includes tasks for input validation and error responses, the contract's negative and authorization paths, idempotency and rate limits where required, migration plus backfill and rollback when the schema changes, every required UI state, observability, and framework-native verification for this repository's technology. Add characterization tests before touching untested legacy code.
Gaps in the design. If the LLD needs work it does not plan — a change in another repository,
an unplanned migration, permission change or dependency — do not plan around it. Plan what the
LLD does plan, then return escalate true with the gap in escalation.
Validate
Run the validator your instructions name on the file and correct every reported error before returning. Never return an invalid task file.
Output contract
Return task_count and slice_count, literal boolean escalate, and escalation (one line
naming the gap, or None) as short scalar JSON fields on the last line.