A real repository, one focused change question

This demo asks where to look before changing Zustand's createStore core. It uses existing SAGE analysis and MCP tools, not a new scanner or synthetic success fixture. Read the case study for source checks, the failed quality gate and the test-acquisition limitation.

1. Prepare the two separate roots

Acquire SAGE v1.4.0 and install its prerequisites using the Quickstart. Separately acquire the Zustand checkout at b57db4f86ef179285da216eeb291266da82c361c. Run SAGE commands from the directory containing sage.py; the Zustand checkout is the explicit target, not the installation directory.

The captured run reused already installed SAGE Python/TypeScript dependencies and therefore used --skip-deps. Do not use that option to skip missing SAGE prerequisites. It did not install or execute Zustand dependencies.

$sageDemoTarget = '<absolute path to the frozen Zustand checkout>'
python -B sage.py init --skip-deps --setup-only --target-root $sageDemoTarget --projects MAIN
python -B sage.py run --target-root $sageDemoTarget --profile daily --projects MAIN --step "AI Context Generator" --ai-context

--step requests that producer's dependency closure; the command is more than a single-file scan. Here MAIN resolves to src, with 15 indexed source files. Setup does not itself establish fresh analysis. The actual analysis completed with process exit 0 and governance verdict FAIL, not a clean verdict.

2. Ask the existing MCP tools

Connect this installation using the agent runbook. Use the same absolute target_root in every call:

The evidence capture called these canonical server functions locally. It did not retest an IDE client's MCP transport or grant target-edit authority.

search_symbols(query="createStore", project="MAIN", target_root=TARGET, format="json")
get_impact_radius(target_node="MAIN::vanilla.ts", target_root=TARGET, format="json", depth=2)
inspect_file(target="MAIN::vanilla.ts", target_root=TARGET, format="json")

TARGET is the checkout path, not a literal argument. MAIN::vanilla.ts is the indexed project-relative node; the returned public reference is MAIN::src/vanilla.ts. These are tool arguments, not shell commands.

Selected fields from the actual impact response, with machine-local paths and source snippets omitted:

{
  "target_ref": "MAIN::src/vanilla.ts",
  "radius_depth": 2,
  "scope_completeness": "bounded",
  "dependency_graph_source": "sqlite_dependencies",
  "direct_dependents_count": 3,
  "direct_dependents": ["src/index.ts", "src/react.ts", "src/traditional.ts"]
}

Symbol search also returned the createStore definition at lines 99-100, plus its recorded direct import bindings in react.ts and traditional.ts. File inspection was source-grounded and explicitly orientation-only, not permission to edit. Read the core implementation and returned consumers before proposing a change; seek the native compiler/lint/test proof appropriate to it.

3. Do not turn an empty result into proof

get_test_impact(target_file="MAIN::vanilla.ts", target_root=TARGET, format="json")

The captured result had target_grounding_status="grounded" and impacted_tests=[]. Relevant alias-bound tests exist outside the selected src scope, as the case study explains. No target tests were executed or automatically selected by this query. Missing-target control queries remained missing_or_unindexed; they did not become grounded empty-test success.

The outcome is a bounded dependency/symbol review starting point, not a complete impact set, passing suite, merge approval or runtime assurance.