diff --git a/AGENTS.md b/AGENTS.md
index 9e592a8a93..00ac5f7f8d 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -12,6 +12,7 @@ Load only the procedures relevant to the task:
 | Domain vocabulary | `CONTEXT.md`, `docs/agents/domain.md` |
 | Architecture decisions | `docs/adr/README.md` |
 | Tests or gate selection | `docs/agents/testing.md` |
+| Writing, auditing, or sweeping tests; dispatching test-audit subagents | `docs/agents/test-audit.md` |
 | Selector capture, polling, or interaction fast paths | `docs/agents/selector-capture.md` |
 | Adding or changing a CLI flag | `docs/agents/cli-flags.md` |
 | Opening or reviewing a PR | `docs/agents/pull-requests.md` |
diff --git a/docs/agents/test-audit.md b/docs/agents/test-audit.md
new file mode 100644
index 0000000000..c9e9af4c29
--- /dev/null
+++ b/docs/agents/test-audit.md
@@ -0,0 +1,57 @@
+# Test Audit
+
+A test earns its cost by guarding behavior, credible regressions, or independent contracts. Optimize for confidence, not deletions. Gate selection stays in `docs/agents/testing.md`.
+
+## Authoring gate
+
+Answer all before adding a test:
+
+1. Which observable behavior, invariant, or independent contract does it guard?
+2. Which credible regression makes it fail? Name the production edit.
+3. Why doesn't existing coverage catch it? Each contract has one owner at its strongest boundary.
+4. Does it need a production seam no production caller needs? Move the test to the real boundary.
+
+A bug regression must fail on pre-fix code for the intended reason; one that never failed visibly proves the mock.
+
+## Junk patterns
+
+Authoring rejects; audits hunt:
+
+- assertion-free probes; a lone `typeof x === 'function'` above the sibling calling `x`;
+- self-comparison (`f(x)` vs `f(x)`, `map.get(k)` vs itself, fixture vs re-spelled fixture);
+- expectation drawn from the helper under test; a `switch` in the test body compared to its arms;
+- near-identical or renamed `test()` blocks; one golden replayed twice;
+- type pins posing as runtime assertions (`Equal<A,B> = true`, `X ? true : never`, `satisfies` after erasure, `toHaveLength(n)` over literals); if erasing leaves nothing executable, use `void [fixtures]`;
+- mocks asserting their own inputs; negative controls passing for the wrong reason;
+- names promising more than the input exercises;
+- dead code called only by tests; tests preserving test-only exports.
+
+## Retention criteria
+
+Refuse to delete tests enforcing: registry, gate-manifest, layering, and DI-seam completeness; golden tables and cross-language source inspection; `@ts-expect-error` pins (an unused directive fails `tsc`); identity pins deep-equal can't subsume; exact-array facets a membership scan can't see duplicates in; vocabulary pins surviving a production+fixture co-update. Static-or-slow and implementation-similarity are not deletion-worthy: prove you can't break it.
+
+## Candidate evidence
+
+Record before editing; a missing field blocks the edit:
+
+- location, test name, detectable failure (or why none);
+- non-test callers of the covered seams; the stronger remaining owner proof (both assertion sets, `file:line`);
+- skip guards in candidate and sibling (`skipIf`, `.skip`, env flags a workflow sets);
+- the runner that executes the file: a Vitest project (`vitest.config.ts`), a `node --test` script (`test:smoke`, `check:*:test`), native XCTest, or a `CHECK_CATALOG` gate id. Only a file no runner reaches is a never-run finding;
+- what deletion unlocks; risk.
+
+## Mutation discipline
+
+Plant the production edit representing the claimed bug: the pre-edit body must stay green while the post-edit body goes red. Report both counts (`8 passed` → `1 failed | 7 passed`). Revert and show `git status` clean before committing. For a deletion, prove the sibling catches the mutant. Never argue detection: measure it.
+
+## Focused validation
+
+`pnpm exec vitest run --project unit-core <path>`, or `--project apple-runner` for `packages/platform-apple/src/runner/**`; then `pnpm check:quick` and `pnpm check:affected --run` on the pushed commit. `scripts/**` tests are in no `tsconfig` include, so `pnpm typecheck` skips helpers there.
+
+## Dispatching subagents
+
+Run parallel read-only lanes with exclusive edit ownership: subsystem lanes (per platform package, `src/daemon`, core packages, `scripts/**`), one cross-cutting junk-pattern sweep, one never-run inventory (Vitest includes, `node --test` globs, gate ids). Mandate in every prompt: READ-ONLY; read the production owner and name the bug; verify skip guards and runner/gate ownership; paste the retention list; false positives cost more than misses.
+
+## Landing
+
+One owner-boundary batch per PR; delete the test-only seam with the test; prefer net-negative production LOC. Report retained false positives, then rebase to `main` and rediscover.