executing-plans

Executing Plans

Execute the active plan as written. Plan artifacts are authoritative; chat is not. Inherit the primary path, assurance profile, task-wide agent budget, and verification budget from using-workflow.

Start Or Resume

1. Call workflow_status once when active state is unknown or context was recovered.
2. Read manifest.json, state.json, plan.md, and the current task's referenced context/decisions.
3. Read ui-contract.md, chunk state, or older artifacts only when the current task depends on them.
4. Inspect git status --short and preserve unrelated user changes.
5. Continue from the first non-complete task; do not replay completed work.

Use workflow_plan_update, workflow_plan_artifact_write, and workflow_plan_complete when available; manual state/artifact writes are fallback only.

Read references/execution-state.md only when updating task/chunk state, delegating a write task, or recovering a complex run.

Execution Cadence

• Update state at meaningful milestones: task start, material block, review handoff, and task completion.
• Do not write workflow state after every tool call or tiny edit.
• Execute tasks within their allowed scope and record any approved scope change before editing outside it.
• Do not perform opportunistic refactors. A new API/schema/protocol/shared abstraction/layer or a higher change class requires an approved plan update and a renewed commitment reflection before implementation continues.
• For production UI, read the UI contract's reuse receipt before markup/styles. Its exact labels are Target:, Shared primitive:, Layout precedent:, Decision:, Structural verification:. Treat named reuse as architecture acceptance; stop before a local substitute, duplicate primitive, or unapproved layout deviation.
• For a small user correction, inspect and edit only the touched surface plus directly affected tests; do not restart repository discovery.
• For chunks, root owns dependency order and integration. Never create nested chunks.