codeclone-implementation-context

name	codeclone-implementation-context
description	Before broad rg/grep, non-trivial planning, unclear scope, contract/schema work, or resuming another agent — call get_implementation_context once to bound the implementation frontier. Then use targeted source reads. Read-only; never grants edit permission.

CodeClone Implementation Context

Bounded orientation from one stored analyze_repository run. Not edit permission.

Mandatory triggers

Call once before:

• repo-wide or recursive rg, grep, find, or custom search scripts;
• cross-module planning, contract/schema/version work, or hub edits;
• resuming another agent/session;
• declaring a non-obvious edit scope.

Skip only for obvious local work with a complete known file boundary.

Do not search broadly first and call this afterward merely to satisfy the workflow.

Pattern

get_implementation_context → targeted reads / narrow rg → codeclone-change-control

Do not loop this tool to avoid reading source.

Subject

Choose one subject strategy:

Input	Use for
query="name"	Structural name search: definitions, call targets, imports in the stored run. Not file-text or regex. mode ignored.
paths=[...]	Known files or modules
symbols=["pkg.mod:func"]	Exact qualnames (: not .)
intent_id=...	Alone: active intent allowed_files. With explicit paths/symbols: keeps that subject and adds change_control.
changed_scope=true	Live git-dirty paths (not intent scope as subject)

paths and symbols may be combined when both are known.

Rules:

• whole-repository context is never inferred;
• never mix changed_scope with paths or symbols;
• never mix query with any other subject input;
• never guess unresolved symbols;
• weaker match_tier hits require source verification.

Mode

For paths, symbols, intent scope, or changed scope (not query):

• implementation — sites, nearby relations, likely tests;
• impact — callers, dependents, baseline-sensitive findings;
• contract — schemas, version constants, persistence and public surfaces.

Examples

get_implementation_context(root=<abs>, query="as_sequence")

get_implementation_context(
    root=<abs>,
    paths=["codeclone/contracts/__init__.py"],
    mode="impact",
)

get_implementation_context(
    root=<abs>,
    paths=["pkg/mod.py"],
    intent_id="<id>",
    mode="implementation",
)

Use rg/grep afterward for text patterns, comments, and non-symbol strings.

Engineering Memory keyword search → codeclone-engineering-memory.

Read results

• needs_analysis → analyze once, retry.
• no_matches (query) → refine query; do not jump to repo-wide search.
• subject_not_found / no_current_work → fix subject; do not widen to repo root.
• freshness.status:"drifted" → verify live source; no re-analyze loop.
• truncated/omitted summaries, unresolved call edges → incomplete evidence, not absence.
• For compact/normal detail, context_governance.mode="partial_enforce" means large facet lanes may be omitted with exact drill-down in context_governance.omitted; mandatory subject/freshness facts remain inline.
• When analysis.context_page_retrieval is present, use get_implementation_context_page(root=<abs>, context_projection_digest=analysis.context_projection_digest, facet=...) for exact facet continuation. This is session-local to MCP run history and returns not_found rather than recomputing fresh context. Facet pages themselves are mode="observe" retrieval responses.

Hard limits

• MCP tools only. Do not fall back to CLI or local report files.
• Never authorizes edits (start_controlled_change + edit_allowed:true only).
• Do not widen scope, touch do_not_touch, or edit review_context/clone cohorts from this output.
• Memory lane here ≠ get_relevant_memory.

Done when you can name: implementation files, review/dependent files, contracts, uncertainties, proposed allowed_files, and what to read next.