Skip to content

Atomic change control

Atomic change control tools (advanced / diagnostic)

Tool Key parameters Purpose
manage_change_intent action, root, run_id, intent_id, scope, on_conflict, ttl_seconds, lease_seconds, changed_files or diff_ref Intent lifecycle: declare, get, check, clear, renew, promote, list_workspace, gc_workspace, recover, reset_workspace. Use for queue/promote/recover operations alongside workflow tools
get_blast_radius run_id, files, depth, include Pre-change risk boundary: full transitive graph, custom include filters
get_relevant_memory root, scope, intent_id, symbols, max_records, include_stale, include_drafts, detail_level Ranked engineering memory for declared edit scope. trajectories[] excludes routine run:* workflows by default (same semantics as include_routine=false on retrieval). For explicit routine control use query_engineering_memory trajectory modes with filters.include_routine. Compact by default: bounded record/trajectory subjects plus typed records, experiences, trajectories, and coverage lanes. Auto-bootstraps store when mcp_sync_policy=bootstrap_if_missing (default). See Engineering Memory
query_engineering_memory root, mode, …, optional semantic (search only), detail_level Mode router: search, get, for_path, for_symbol, stale, drafts, coverage, status, trajectory_status, trajectory_search, trajectory_get, trajectory_anomalies, trajectory_agents, trajectory_dashboard. List/search modes default compact; get, trajectory_get, or detail_level=full are explicit drill-down. filters supports types, statuses, confidences, and match_mode (any|all) for search. semantic=true blends LanceDB proximity when [tool.codeclone.memory.semantic] enabled and index built (default off). See Engineering Memory
manage_engineering_memory root, action, … Agent-side: refresh_from_run, record_candidate, promote_experience, validate_claims, propose_from_receipt, rebuild_semantic_index, rebuild_trajectories, enqueue_projection_rebuild, projection_rebuild_status, run_projection_jobs_once. promote_experience creates a human-reviewable draft; human approve/reject/archive remains VS Code/CLI only. See Engineering Memory
check_patch_contract mode, run_id, before_run_id, after_run_id, intent_id, strictness, changed_files or diff_ref Manual budget query or step-by-step verification
create_review_receipt run_id, intent_id, format, include_blast_radius, include_patch_contract Manual receipt generation
validate_review_claims text, run_id, require_citations, patch_health_delta Standalone citation-based overclaim detection; pass patch_health_delta from verify when using the atomic workflow
Blast radius: do_not_touch vs review_context

do_not_touch is limited to actionable negative context: baselines, generated CodeClone state, explicit forbidden paths. Report-only signals such as security boundary inventory and overloaded-module candidates are returned as review_context — information, not edit prohibitions. Long context sections include total, shown, and truncated summaries.

Patch contract modes

Budget reads one stored run and optional intent. Shows regression headroom per quality dimension before editing. Queued intents return edit_allowed=false. Verify compares explicit before/after stored runs, previews gates, validates scope, and reports baseline-abuse signals. When intent_id is provided but before_run_id is omitted, verify auto-resolves the before-run from the intent record. Missing runs return status="unverified". Identical before/after runs for python_structural / governance_config return reason="after_run_not_new". Non-accepted responses include a next_step hint and claim_validation_recommended flag.

Verify regressions are run-relative, not baseline-novelty-relative: a finding absent from the clean before-run and present in the after-run is a patch regression even when its fingerprint is novelty="known" against the trusted baseline.

When a change intent is active, verify mode attributes regressions and gate changes to the declared scope. Intent-scope regressions produce contract violations; external regressions are reported as informational context. Queued intents are rejected with reason="intent_not_active". See Scope-Aware Patch Contract Verification and Verify Ergonomics.