Agent edit cycle¶

Same sequence as MCP change control workflow.

Before first edit: start_controlled_change must return status: active and edit_allowed: true. Queued intents require manage_change_intent(action=promote).

Evidence: finish requires exactly one of changed_files or diff_ref listing every in-scope dirty path.

After-run: required when verification profile is python_structural or governance_config. Pass a new after_run_id — identical before/after runs return after_run_not_new.

Claims: only claims_text goes to Claim Guard; review_text is a human note.

Memory: call get_relevant_memory after start; optional finish(..., propose_memory=true) for draft candidates (human approve in VS Code Memory view).

Implementation context: after memory, call get_implementation_context with scoped paths when editing Python — bounded structural, call-graph, and contract evidence from one stored run.

Contract tables: Verification profiles, finish_controlled_change.