Analyze & triage¶

Phase 1: Analyze¶

Both register the result as an in-memory run. All other tools read from stored runs.

Phase 2: Triage¶

Workspace hygiene tips¶

Selected MCP responses may include a non-blocking tips[] array with structured workspace guidance. The first tip checks whether the repository root .gitignore covers .codeclone/ (or the broader .cache/ tree).

Tips are advisory only — not findings, gates, or failures. MCP never edits .gitignore automatically; agents must declare scope before changing it.

Surfaces: analyze_repository, get_run_summary, get_production_triage, start_controlled_change, and the CLI after a normal interactive analysis run (suppressed in --quiet, CI, and non-TTY contexts).

Health check¶

analyze_repository(root=<abs>)
  -> get_run_summary or get_production_triage
  -> list_hotspots or check_*
  -> get_finding -> get_remediation

PR review¶

analyze_changed_paths(root=<abs>, changed_paths=[...] or git_diff_ref="HEAD~1")
  -> list_findings(sort_by="priority")
  -> get_finding -> get_remediation
  -> generate_pr_summary

Implementation context¶

After analysis and triage, ask for bounded context around the files you expect to inspect:

get_implementation_context(
  root=<abs>,
  paths=["codeclone/surfaces/mcp/service.py"],
  mode="implementation",
)

The response combines canonical structural facts with a live freshness delta. Use context_artifact_digest to identify the source context artifact and context_projection_digest when citing the exact bounded response. If freshness.status is drifted, analyze again. This step informs scope; only start_controlled_change can return edit_allowed=true.

With no explicit subject, the tool resolves current work deterministically:

1. active intent allowed_files;
2. otherwise the bounded live git-dirty set;
3. otherwise status="no_current_work".

Use changed_scope=true to request the dirty set explicitly. Do not combine it with paths or symbols.

Exact qualnames are also valid subjects:

get_implementation_context(
  root=<abs>,
  symbols=["codeclone.surfaces.mcp.service:CodeCloneMCPService"],
  mode="implementation",
)

Symbol resolution uses the analyzed Unit inventory plus public API rows. Inspect both subject.resolved_symbols and subject.unresolved_symbols; CodeClone reports unknown qualnames instead of inferring a likely match.

Structural import, importer, and test-importer roles appear as collapsed related_modules entries with explicit relations. Read each collection's summary: the global budget is shared across the response. Safety context is allocated first; safety_context_overflow means even the hard cap could not show every safety entry.

Once an intent is active, pass its intent_id with the same explicit paths. The response then shows the declared scope, review context, do-not-touch boundaries, and guards beside lane-separated memory evidence. Use mode="impact" when you need transitive dependency context and baseline-sensitive findings. The context tool mirrors authorization evidence; it does not grant or widen authorization.

Change control tool tiers¶