sdd-plan-tasks

SDD Plan And Tasks

Use after a spec is drafted/approved and the next step is design, planning, tasking, or implementation prep.

Inputs

Load only needed artifacts: spec/requirements, constitution/steering/project context, architecture docs, brownfield code patterns, prior research/contracts/quickstart.

Plan Must Include

• behavior summary; risk/evidence profile: basic, standard, strict, or regulated
• technical context, constraints, decisions, rationale, alternatives
• architecture/components; data/state model; contracts when relevant
• error handling, migration, security/privacy/performance/accessibility when relevant
• quickstart or validation scenarios
• traceability graph: requirement ID -> scenario -> decision -> task ID -> verification/evidence -> changed path
• LLM-use limits for generated code, proprietary code, credentials, or regulated data

Preserve Spec Kit layout when present: plan.md, research.md, data-model.md, contracts/, quickstart.md. Preserve Kiro design.md. Use diagrams only for real boundaries.

Tasks Must Include

• stable task IDs; requirement/story links; known file paths
• explicit dependencies; parallel markers only for independent files/work
• verification step plus expected evidence artifact/command per task or small group
• bounded implementation surface and boundary notes for multi-component/long-running work

Avoid tasks like "implement feature" or "update tests" without path, expected behavior, and validation command.

Brownfield Gap Pass

1. Search existing implementation, naming, tests, integration points.
2. Prefer extension over duplication; record options/trade-offs.
3. Mark implementation-time unknowns.
4. Save in research.md or equivalent.

Gate

python3 <plugin>/scripts/sdd_traceability_check.py <repo> --json

Resolve FAIL before implementation or record why proceeding is acceptable. Do not proceed if core artifacts are missing. For strict/regulated, create or identify evidence.md, validation.md, proof.md, or equivalent before tasks can be completed. LLM self-review/model confidence is not verification.

Done

Design satisfies the spec; every buildable requirement has tasks or deferral; tasks have dependency order, parallel boundaries, verification; traceability/evidence are audit-ready.