Skip to content

Implementation context

get_implementation_context is a read-only projection over one stored MCP run. It never re-analyzes, never changes edit_allowed, and never substitutes start_controlled_change. For workflow placement, see Analysis tools.

Parameters

Parameter Default Purpose
root — (required) Absolute repository root
paths null Repo-relative file or directory subjects
symbols null module:symbol qualnames (colon separator; dot notation rejected)
intent_id null Active intent — pins run and adds change_control block
changed_scope false Use bounded live git-dirty set as subject; mutually exclusive with explicit paths or symbols
mode implementation implementation, impact, or contract
include null Optional closed facet set
depth 1 Structural traversal depth (0–3)
detail_level compact compact, normal, or full
budget 50 Global evidence-entry cap (1–200)
run_id null Stored run; latest when omitted

changed_scope=true selects the dirty set explicitly. Without explicit subjects, precedence is: paths/symbols → active intent allowed_files → bounded git-dirty set. A clean tree with no subject returns no_current_work, never whole-repository context.

Modes and facets

Mode Orientation
implementation Editing context: module role, imports/importers, callees, public API, blast radius, tests, docs, memory
impact Transitive dependency context, baseline-sensitive findings; adds callers
contract Truth-map: definition_sites, version_constants, contract_tests, memory_conflicts

contract mode emits path-specific caller facets (persistence_path_callers, serialization_path_callers, deserialization_path_callers, store_api_consumers) only with a typed contract-registry, protocol, or Engineering Memory anchor. Without an anchor they report status: "not_available" rather than being guessed from names.

call_context projects callers, callees, references, and test_callers from run-bound relationship facts. Every edge is tagged relation_kind × resolution_status. Production and test-origin callers stay in separate lanes; test edges never make production code live. Unresolved calls use target_qualname: null. analysis.call_graph_status is complete, partial, or unavailable.

Import, importer, and test-importer roles collapse into structural_context.related_modules with explicit relations (imports, imported_by, tested_by).

Freshness and digests

flowchart LR
    R["Stored MCP run"] --> C["Canonical report facts"]
    R --> M["Run manifest"]
    C --> P["Bounded context projection"]
    M --> F["Live freshness delta"]
    F --> P
    P --> A["context_artifact_digest"]
    P --> E["context_projection_digest"]
  • context_artifact_digest binds the canonical run and off-report context artifact.
  • context_projection_digest binds the normalized request and exact bounded evidence returned.
  • analysis.freshness compares run manifest with live mtime+size and, when available, git DirtySnapshot delta.
  • freshness.status="drifted" means analyze again before relying on the projection.

A missing run returns needs_analysis. Invalid facets and paths outside the root raise a contract error.

Budget and safety overflow

budget is one global evidence-entry cap, not per-facet. Every bounded collection reports total, shown, truncated, and omitted. Intent do_not_touch and review-required entries consume budget first. The effective limit expands up to the server hard cap so a small requested budget cannot hide safety context.

If safety entries alone exceed that cap, the response uses status="safety_context_overflow" and reports the omitted count.

Symbol-only queries that resolve nothing return status="subject_not_found" with actionable next_steps and omit empty facet scaffolding.

Intent and memory lanes

With intent_id, the selected active intent pins the source run and adds change_control:

  • allowed_files and allowed_related from declared scope;
  • report-derived review_context;
  • explicit and built-in do_not_touch boundaries;
  • guards with authorization_source="start_controlled_change".

Engineering Memory records, test anchors, doc anchors, trajectories, and Experiences project into separate bounded lanes. Memory is evidence, not edit authority.