diff --git a/extensions/gentle-ai.ts b/extensions/gentle-ai.ts
index 39edf9a2e..faa1a6af4 100644
--- a/extensions/gentle-ai.ts
+++ b/extensions/gentle-ai.ts
@@ -5293,7 +5293,7 @@ const REVIEW_CONTROLLER_PARAMETER_FIELDS = {
 		},
 		input: {
 			type: "string",
-			description: "A JSON object or serialized object string for START/ASSESS only; every other operation requires a serialized object string. New native ordinary START uses {\"mode\":\"ordinary\"}, plus the agent lens selection {\"lenses\":[\"risk\",...],\"lensesReason\":\"<what you touched and how>\"} (both or neither, 1-4 of risk, resilience, readability, reliability; never with focus; providers that predate it keep the tier default); answer-consent uses exactly {\"consentBinding\":\"<opaque id>\",\"answer\":\"granted|declined\"}. Ordinary provider capture belongs only to gentle_review_capture. An explicit baseRef requires committedOnly: true and requests a committed range, while repository-local policyPath remains optional. baseRef must be HEAD, a full 40- or 64-character commit id, or a ref name; abbreviated commit ids are rejected as base-ref-unresolvable. ASSESS accepts an optional object with baseRef, committedOnly, writerModelId, writerEffort, and nativeReviewOutcome (gentle-pi#662/#668/#1175). The writer profile comes from the runtime-recorded model and effort of the pending mutations for this root (writerProfileSource runtime); writerModelId and writerEffort are only a fallback when no runtime evidence exists (caller), and with neither the profile fails closed to small (fallback). The writer profile is reported but never adds a verifier: the plan adds the independent verifier only when escalate names a Task Size high-risk item. ASSESS derives `closed` only from the native candidate.consumed fact for this exact candidate, which native records only when the approved review for that candidate is acknowledged; a caller-supplied closed is not authority and, without that fact, resolves to unknown. Every outcome yields the exact risk-gated plan ASSESS returns when RDD is off: a closed review never removes the separate verifier the tier requires (verify-always-rdd-high S1), unknown is never treated as closed, and a decline is candidate-scoped and never lowers the bar below the RDD-off path. A declined or unavailable outcome, stated explicitly or recorded by this process for this exact candidate (never a different one, and never from repository state alone), wins over closure. The returned outcome_source (explicit|derived|unknown) says which of these produced the value. Legacy controller input remains separate.",
+			description: "A JSON object or serialized object string for START/ASSESS/answer-consent; every other operation requires a serialized object string. New native ordinary START uses {\"mode\":\"ordinary\"}, plus the agent lens selection {\"lenses\":[\"risk\",...],\"lensesReason\":\"<what you touched and how>\"} (both or neither, 1-4 of risk, resilience, readability, reliability; never with focus; providers that predate it keep the tier default); answer-consent uses exactly {\"consentBinding\":\"<opaque id>\",\"answer\":\"granted|declined\"}. Ordinary provider capture belongs only to gentle_review_capture. An explicit baseRef requires committedOnly: true and requests a committed range, while repository-local policyPath remains optional. baseRef must be HEAD, a full 40- or 64-character commit id, or a ref name; abbreviated commit ids are rejected as base-ref-unresolvable. ASSESS accepts an optional object with baseRef, committedOnly, writerModelId, writerEffort, and nativeReviewOutcome (gentle-pi#662/#668/#1175). The writer profile comes from the runtime-recorded model and effort of the pending mutations for this root (writerProfileSource runtime); writerModelId and writerEffort are only a fallback when no runtime evidence exists (caller), and with neither the profile fails closed to small (fallback). The writer profile is reported but never adds a verifier: the plan adds the independent verifier only when escalate names a Task Size high-risk item. ASSESS derives `closed` only from the native candidate.consumed fact for this exact candidate, which native records only when the approved review for that candidate is acknowledged; a caller-supplied closed is not authority and, without that fact, resolves to unknown. Every outcome yields the exact risk-gated plan ASSESS returns when RDD is off: a closed review never removes the separate verifier the tier requires (verify-always-rdd-high S1), unknown is never treated as closed, and a decline is candidate-scoped and never lowers the bar below the RDD-off path. A declined or unavailable outcome, stated explicitly or recorded by this process for this exact candidate (never a different one, and never from repository state alone), wins over closure. The returned outcome_source (explicit|derived|unknown) says which of these produced the value. Legacy controller input remains separate.",
 		},
 		outputPath: { type: "string", description: "Retired with legacy bundle export; ignored. Export returns legacy-operation-retired." },
 		inputPath: { type: "string", description: "Repository-local JSON input file for the separate legacy controller flow (alternative to input). Legacy bundle import is retired." },
@@ -5564,7 +5564,7 @@ function parseReviewControllerParameters(value: unknown): ReviewControllerParame
 	};
 	for (const key of ["changeName", "idempotencyKey", "transition", "input", "outputPath", "inputPath", "operationId", "lineageIds", "acknowledgeUntrustedBundleSource", "workspaceRoot"] as const) {
 		const optional = value[key];
-		if (key === "input" && key in value && (value.operation === REVIEW_CONTROLLER_OPERATION.START || value.operation === REVIEW_CONTROLLER_OPERATION.ASSESS)) {
+		if (key === "input" && key in value && (value.operation === REVIEW_CONTROLLER_OPERATION.START || value.operation === REVIEW_CONTROLLER_OPERATION.ASSESS || value.operation === REVIEW_CONTROLLER_OPERATION.ANSWER_CONSENT)) {
 			parameters.input = serializeReviewJsonArgument(optional);
 			continue;
 		}
diff --git a/odd/tasks/fix-2375-gentle-review-answer-consent-input-coercion.md b/odd/tasks/fix-2375-gentle-review-answer-consent-input-coercion.md
new file mode 100644
index 000000000..a50b87c5c
--- /dev/null
+++ b/odd/tasks/fix-2375-gentle-review-answer-consent-input-coercion.md
@@ -0,0 +1,30 @@
+# Task Breakdown: Fix #2375 Gentle Review Answer Consent Input Coercion
+
+**Feature:** `fix-2375-gentle-review-answer-consent-input-coercion`  
+**Issue:** #2375 (`bug(review): gentle_review answer-consent input is unreachable from direct tool calls (object never coerced to string) (4.3.0)`)  
+**Strategy:** Strict TDD, ODD workflow, Conventional Commits  
+
+---
+
+## Tasks
+
+### 1. Test (RED): Unit test for answer-consent object input coercion
+- [x] In `tests/review-json-arguments.test.ts`:
+  - Add test asserting that `validate("gentle_review", { operation: "answer-consent", input })` with object input `{ consentBinding: "bound-1", answer: "granted" }` is admitted and coerced into a serialized string by `parseReviewControllerParameters`.
+  - Assert that execution reaches the consent registry without throwing `"Review controller input must be a string"`.
+  - Evidence: Red test failure `Review controller input must be a string` observed before fix.
+
+### 2. Implementation (GREEN): Coerce answer-consent object input
+- [x] In `extensions/gentle-ai.ts`:
+  - Include `value.operation === REVIEW_CONTROLLER_OPERATION.ANSWER_CONSENT` alongside `START` and `ASSESS` in `parseReviewControllerParameters`.
+  - Update schema description in `REVIEW_CONTROLLER_PARAMETER_FIELDS.input` to `START/ASSESS/answer-consent`.
+  - In `tests/review-json-arguments.test.ts`, exempt `answer-consent` from fail-closed string-only operation test.
+  - Evidence: All 9 tests in `tests/review-json-arguments.test.ts` pass green.
+
+### 3. Verification & Regressions
+- [x] Run test suite: `node --experimental-strip-types --test tests/review-json-arguments.test.ts tests/review-controller.test.ts`.
+- [x] Run typecheck: `npm run typecheck` (0 diagnostics, 0 regressions).
+- [x] Evidence: 0 regressions, all tests passing.
+
+### 4. Work-Unit Commit
+- [x] Create conventional commit: `fix(review): coerce answer-consent object input to serialized string (#2375)`.
diff --git a/tests/review-json-arguments.test.ts b/tests/review-json-arguments.test.ts
index a19e31d0f..1fe6a53e3 100644
--- a/tests/review-json-arguments.test.ts
+++ b/tests/review-json-arguments.test.ts
@@ -53,7 +53,7 @@ test("public Anthropic non-strict adapter preserves the complete review tool dec
 	assert.deepEqual(controller.required, ["operation"]);
 	assert.deepEqual(Object.keys(controller.properties).sort(), ["operation", "lineageId", "selectionBinding", "intendedUntracked", "untrackedScope", "changeName", "idempotencyKey", "transition", "input", "outputPath", "inputPath", "operationId", "lineageIds", "workspaceRoot"].sort());
 	assert.ok(controller.properties.input.anyOf.some((schema: any) => schema.type === "object"));
-	assert.match(controller.properties.input.description, /START\/ASSESS only/);
+	assert.match(controller.properties.input.description, /START\/ASSESS\/answer-consent/);
 	assert.deepEqual(controller.properties.operation.enum, tools.get("gentle_review").parameters.properties.operation.enum);
 	for (const name of ["gentle_review_capture", "gentle_review_capture_group"]) {
 		const schema = emitted.get(name);
@@ -66,7 +66,7 @@ test("public Anthropic non-strict adapter preserves the complete review tool dec
 	let nativeCalls = 0;
 	const native = new Proxy({}, { get() { return async () => { nativeCalls++; }; } }) as NativeReviewCli;
 	for (const operation of ["start", "assess", "answer-consent", "inspect", "reset", "recover"]) {
-		const inputs = operation === "start" || operation === "assess" ? [null] : [null, { consentBinding: "opaque", answer: "granted" }];
+		const inputs = operation === "start" || operation === "assess" || operation === "answer-consent" ? [null] : [null, { consentBinding: "opaque", answer: "granted" }];
 		for (const input of inputs) {
 			const args = validateToolArguments({ ...tools.get("gentle_review"), parameters: controller }, { type: "toolCall", id: "provider-shell", name: "gentle_review", arguments: { operation, input } });
 			assert.deepEqual(args.input, input, "provider-shell validation accepts and retains this input");
@@ -89,6 +89,23 @@ test("registered review schemas admit START/ASSESS objects and ordered capture o
 	assert.deepEqual(__testing.parseReviewCaptureGroupParameters(args).collectBindings, [JSON.stringify(binding), ordered[1]]);
 });
 
+test("answer-consent admits JSON object input and coerces it cleanly (#2375)", async () => {
+	const input = { consentBinding: "bound-1", answer: "granted" };
+	assert.deepEqual(validate("gentle_review", { operation: "answer-consent", input }).input, input);
+	assert.equal(
+		__testing.parseReviewControllerParameters(validate("gentle_review", { operation: "answer-consent", input })).input,
+		JSON.stringify(input),
+	);
+	const native = { targetStatus: async () => ({ targetIdentity: "t1" }) } as unknown as NativeReviewCli;
+	const outcome = await __testing.executeReviewControllerOperation(
+		validate("gentle_review", { operation: "answer-consent", input }),
+		process.cwd(),
+		native,
+	);
+	assert.equal(outcome.operation, "answer-consent");
+	assert.match(String(outcome.outcome), /consent-binding-stale/);
+});
+
 test("validated ASSESS objects reach the mocked native facade with parsed fields", async () => {
 	const requests: unknown[] = [];
 	const native = { assess: async (request: unknown) => { requests.push(request); throw new Error("mock assessment unavailable"); } } as unknown as NativeReviewCli;
@@ -132,7 +149,7 @@ test("actual validation rejects arrays, null and coerced primitives rather than
 });
 
 test("all other controller operations keep object input fail-closed at the facade boundary (#1698)", async () => {
-	const operations = (tools.get("gentle_review").parameters.properties.operation.enum as string[]).filter((operation) => operation !== "start" && operation !== "assess" && operation !== "select-intended-untracked");
+	const operations = (tools.get("gentle_review").parameters.properties.operation.enum as string[]).filter((operation) => operation !== "start" && operation !== "assess" && operation !== "answer-consent" && operation !== "select-intended-untracked");
 	let calls = 0;
 	const native = new Proxy({}, { get() { return async () => { calls++; throw new Error("unexpected native execution"); }; } }) as NativeReviewCli;
 	for (const operation of operations) {