Tasks
Tasks are for separation, not importance. Cut one only when a ready spec contains independently
dispatchable parallel, context, repository, platform, or change-plan wave slices. Size alone proves
nothing. The common one-worker case implements directly from the spec and records ## Execution
there.
A task narrows existing requirement IDs. If it invents one, it is a bad spec in disguise. A change plan may add wave and preservation context but never replaces the source spec.
Contract
A task must:
- name its source spec by absolute path;
- copy the scoped requirement snapshots and
Verify with:lines; - define a small write surface and a
Do not changewall; - map every scoped requirement to verification;
- record findings and the run summary;
- start only from a source spec with exactly
status: ready.
sus-task writes the packet under the agent-neutral artifact location.
Dispatch names the task and ready source spec by absolute path. suspec check <task-path> --spec <spec-path> validates source binding, shape, dependency handoff, evidence, and closed-state
blockers.
Exact shape: artifact formats.
Add ## Self-review only when it carries useful checks.
Splitting
Split when:
- waves have distinct checkpoints and rollback boundaries;
- a prerequisite must land before dependent behavior;
- parallel work has disjoint write surfaces;
- one behavior needs independent implementation and verification across repositories or platforms.
Serialize tasks that write the same file. Parallel edits to one file are serialized conflict, not parallelism. Prefix multi-repository affected areas:
example.txt
fileapi: src/checkout/**
web: app/checkout/**
Each repository keeps its own verification command.
Scope wall
Name tempting excluded areas such as schemas, public APIs, payments, authentication, generated files, and unrelated modules. A polite suggestion is not a wall. A blocked worker stops for a decision. A wall with an explicit escape path prevents more violations than prohibition alone ([IMPOSSIBLE]).
Every verification item names the command and requirement ID. Manual verification names who observed what. Findings hold surprises, risks, and adjacent issues. The run summary points to evidence without copying it.
Next: execution. Previous: inventories and change plans.
Need a starting point? Install the skills