Reachability & Dead-Code Analysis

Spec 20. Cross-language. Deterministic, offline, no API key. Candidates, never deletion authority.

find_dead_code answers three graph-reachability questions grep can't (it sees text, not reach) and the model burns tokens guessing at:

• "What code is unreachable / dead?" — mark-and-sweep from roots; the unreached remainder.
• "Is anything calling X?" — reachability from every entry point.
• "What becomes dead if I delete X?" — the set reachable only through X.

Prior art is knip / ts-prune (mark-and-sweep from entry points) — but TypeScript/JavaScript-only. This is the cross-language version over the unified tree-sitter graph (15+ languages).

Read this first — honest limits

Static reachability cannot see dynamic entry points, framework magic (routes, DI, plugin registries), reflection, or public API consumed outside the repo. All of these produce false "dead" positives. So find_dead_code:

• treats tests, imported symbols, route handlers, and main as roots,
• returns confidence-tagged candidates with a reason, never a verdict,
• and never auto-deletes.

Every response carries soundness.posture: "candidates-not-authority" and explicit caveats. Treat it as a lead generator for a human/agent to verify — not a delete list.

How roots and confidence work

A node is a root (assumed live) if it is a test, is imported by name from another file, is a detected HTTP route handler, or is main-like. Reachability is a forward BFS from those roots.

Confidence is deliberately conservative — the bias is toward false-live over false-dead:

Confidence	When
high	static language · no internal caller · not imported by name · and its module is not imported anywhere
medium	reachable only from other dead code, or no dependency-graph signal available
low	dynamic language (Python/Ruby/PHP/…), or its module is imported elsewhere (namespace/default/re-export usage the named-import scan can't resolve)

That last rule matters: on a real repo it cut high-confidence candidates from ~470 to ~35 — a symbol living in a module something else imports is never flagged high, because the specific usage may be a namespace or default import this static scan doesn't resolve.

Re-export / barrel resolution raises recall (fewer false-dead)

A reachability conclusion is only as sound as the call graph is complete. The TS/JS import resolver follows re-export chains — export { x } from './impl', export * from './x' (and TypeScript's ESM .js specifiers) — through any depth of barrel to a symbol's true definition, and that resolution runs on every cross-file call edge (not just base classes). A call imported through a barrel resolves to the real target (labelled re_export) instead of stalling at the barrel and falling through to the ambiguous first-same-named-candidate (name_only). Concretely, dogfooding this repo moved 29 symbols off the false-dead / false-entry-point list — e.g. EdgeStore.open went from a reported zero callers to its real 22 — because a method/static call through an imported receiver now binds to its definition. See openspec/changes/add-call-resolution-recall/.

Tool contract

// Candidate dead-code report
{ "directory": "/abs/path", "maxResults": 100, "filePattern": "src/" }

// "What becomes dead if I delete X?"
{ "directory": "/abs/path", "ifDeleted": "parseConfig" }

Report output:

{
  "stats": { "analyzed": 1455, "roots": 399, "reachable": 790, "candidateDead": 665 },
  "rootKinds": { "tests": 0, "imported": 393, "httpHandlers": 0 },
  "byConfidence": { "high": 35, "medium": 35, "low": 595 },
  "candidateDead": [
    { "name": "isValidEmail", "file": "src/utils/validation.ts", "language": "TypeScript",
      "fanIn": 0, "confidence": "high",
      "reason": "no internal caller; not imported by name from any other file; not a test, route handler, or main entry" }
  ],
  "coverage": { "languages": ["TypeScript", "Python", "Go"], "exportSignal": "dependency-graph" },
  "soundness": { "posture": "candidates-not-authority", "caveats": ["These are CANDIDATES…", "…"] }
}

Delete-impact output:

{
  "target": "handler",
  "becomesDeadIfDeleted": [{ "name": "helper", "file": "src/app.ts", "language": "TypeScript", "fanIn": 1 }],
  "count": 1,
  "note": "These nodes are reachable only through the target. Deleting it orphans them — verify before removing."
}

How it works

Pure read over the existing graph — no schema change:

• Reachability — forward BFS over buildAdjacency's forward map from the root set; candidate-dead = code nodes not reached. External and infrastructure (IaC) nodes are excluded.
• Liveness signals — tests + HTTP route handlers + main, plus the dependency graph's imported names (symbol-level) and imported files (module-level) for the cross-language "used elsewhere" signal.
• Delete-impact — recompute reachability with the target removed from both seeds and the graph, and diff against the baseline reached set.

Implementation: reachability.ts. Tested over a two-language fixture with known live regions, a dead orphan, a dead cluster, and a delete-impact diff in reachability.test.ts.

Accuracy depends on a current analyze_codebase that includes tests and produces the dependency graph. Without test nodes as roots, test-only code is flagged; without the dependency graph, confidence is reduced and the response says so.

Index integrity. When the persisted index does not reconcile against its build-time attestation — materially smaller than the build committed (degraded) or built at a different schema (mismatched) — the response carries that verdict in confidenceBoundary.integrity and is not marked complete. A "dead" conclusion over a half-built index is the most dangerous false negative, so "looks dead to a broken index" is labeled, never asserted. Re-run analyze_codebase to rebuild.