SDD Implement

Use when tasks are ready and the user wants implementation, execution, continuation, or completion from SDD artifacts.

Preflight

PowerShell: use $env:PLUGIN_ROOT for $PLUGIN_ROOT command paths.

python3 "$PLUGIN_ROOT/scripts/sdd_surface_audit.py" <repo> --json
python3 "$PLUGIN_ROOT/scripts/sdd_traceability_check.py" <repo> --json
git status --short

Read active tasks, design/plan, spec/requirements, quickstart, evidence ledger if present, and relevant code. If core artifacts are missing, route to sdd-plan-tasks or sdd-specify. Preserve unrelated user changes.

Execution

• Work by dependency order, task/story at a time.
• Respect parallel markers only for safe independent work.
• Write tests first when tasks request TDD/contracts/regression coverage.
• Keep edits inside task boundary.
• If code disproves plan/spec, update artifact or report drift before continuing.
• Complete a task only after fresh evidence.
• Never complete from intent, compilation hope, agent summary, LLM self-review, model confidence, or context-free LLM-as-a-judge.
• For strict/regulated, keep requirement/task IDs in notes, evidence, or commits.

Evidence

Use the most relevant proof available: tests, build/typecheck, relevant lint, runtime/UI smoke, bug reproduction command, or labeled manual verification (reproducible by another human, scoped to the claim) when automation is unavailable.

Evidence must match the claim. Lint does not prove behavior; a unit test does not prove runtime boot; a checklist does not prove integration.

For strict/regulated, maintain evidence.md, validation.md, proof.md, or equivalent with command/manual path, current-run marker, outcome, covered IDs, and gaps/deferrals. If deterministic proof is unavailable for high-risk behavior, record risk acceptance instead of claiming full proof.

Drift

When implementation contradicts artifacts: pause, identify artifact and code evidence, update intended artifact or implementation, then re-run relevant check.

Completion Gate

• Completed tasks have evidence.
• No unchecked blocking task is hidden.
• Traceability is PASS or findings are disclosed.
• Tests/build/smoke commands and outcomes are reported.
• Final claim is no broader than evidence.
• LLM-generated or modified code has project-local proof; model judgment is advisory only.