diff --git a/package.json b/package.json
index 5fee805ee0..8da704d49f 100644
--- a/package.json
+++ b/package.json
@@ -69,6 +69,7 @@
     "test:integration": "vitest run test/integration",
     "test:upstream-contract": "vitest run test/contract/upstream-contract.test.ts",
     "test:engine-parity": "vitest run test/contract/engine-parity.test.ts",
+    "test:driver-parity": "vitest run test/contract/coding-agent-driver-parity.test.ts",
     "test:changed": "vitest run --changed=origin/main",
     "test:workers": "vitest run --config vitest.workers.config.ts",
     "test:coverage": "vitest run --coverage --pool=forks",
@@ -76,7 +77,7 @@
     "test:smoke:observability": "node scripts/smoke-observability-traces.mjs",
     "test:smoke:browser:install": "playwright install chromium",
     "test:smoke:browser": "node scripts/smoke-ui-browser.mjs",
-    "test:ci": "git diff --check && npm run actionlint && npm run db:migrations:check && npm run db:schema-drift:check && npm run selfhost:env-reference:check && npm run selfhost:validate-observability && npm run cf-typegen:check && npm run typecheck && npm run test:coverage && npm run test:engine-parity && npm run test --workspace @jsonbored/gittensory-engine && npm run test:workers && npm run build:mcp && npm run test:mcp-pack && npm run build:miner && npm run test:miner-pack && npm run rees:test && npm run ui:openapi:check && npm run ui:openapi:settings-parity && npm run ui:version-audit && npm run docs:drift-check && npm run manifest:drift-check && npm run engine-parity:drift-check && npm run command-reference:check && npm run ui:lint && npm run ui:typecheck && npm run ui:test && npm run ui:build",
+    "test:ci": "git diff --check && npm run actionlint && npm run db:migrations:check && npm run db:schema-drift:check && npm run selfhost:env-reference:check && npm run selfhost:validate-observability && npm run cf-typegen:check && npm run typecheck && npm run test:coverage && npm run test:engine-parity && npm run test:driver-parity && npm run test --workspace @jsonbored/gittensory-engine && npm run test:workers && npm run build:mcp && npm run test:mcp-pack && npm run build:miner && npm run test:miner-pack && npm run rees:test && npm run ui:openapi:check && npm run ui:openapi:settings-parity && npm run ui:version-audit && npm run docs:drift-check && npm run manifest:drift-check && npm run engine-parity:drift-check && npm run command-reference:check && npm run ui:lint && npm run ui:typecheck && npm run ui:test && npm run ui:build",
     "test:release": "npm run test:ci && npm run changelog:check",
     "test:release:mcp": "npm run test:ci",
     "test:watch": "vitest",
diff --git a/test/contract/coding-agent-driver-parity.test.ts b/test/contract/coding-agent-driver-parity.test.ts
new file mode 100644
index 0000000000..4d4a9d217f
--- /dev/null
+++ b/test/contract/coding-agent-driver-parity.test.ts
@@ -0,0 +1,232 @@
+// Behavioral contract/parity suite for the two `CodingAgentDriver` implementations (#4296): the CLI-subprocess
+// driver (#4266, injected fake `CliSubprocessSpawnFn`) and the Agent-SDK driver (#4267, injected fake
+// `AgentSdkQueryFn`). Unlike `engine-parity.test.ts` (golden-snapshot byte-parity for deterministic re-exports),
+// a coding agent's output is NOT deterministic across implementations or runs — so this suite asserts the
+// interchangeability CONTRACT instead: given equivalent backend behavior, both drivers accept the same
+// `CodingAgentDriverTask` shape, scope execution to the task's working directory, forward the composed
+// instructions verbatim, surface failures as structured results (never throws), and return the same
+// `CodingAgentDriverResult` SHAPE — so the iterate-loop orchestrator (#2333) can swap one for the other with no
+// caller-side changes.
+//
+// Adding a THIRD driver implementation later: append one entry to `DRIVER_HARNESSES` below that builds your
+// driver around an injected fake backend able to express the three behaviors (`success`, `clean_failure`,
+// `budget_exhausted`) and to record the working directory/prompt it was handed. Every contract case in this file
+// then runs against it automatically — new drivers must not need any test-body changes, only a harness entry.
+//
+// Documented (deliberately tolerated) divergences the shared interface does NOT paper over — the contract is
+// shape-level, not value-level (#4296's "close any interface gaps ... documented, not silently patched"):
+// - `changedFiles`: the CLI driver always reports `[]` (post-hoc git-diff detection over the worktree is a
+//   sibling concern, per its own header); the SDK driver reports the deduped Edit/Write/NotebookEdit tool-use
+//   paths on success. The CONTRACT is: always a string array, and `[]` on any failed attempt.
+// - `turnsUsed`: the CLI driver never reports a turn count (a subprocess exposes none); the SDK driver reports
+//   `num_turns` when the stream's result frame carries one. The CONTRACT is: `number | undefined`.
+// - budget exhaustion: the CLI driver's ceiling is wall-clock (`timedOut` -> `<command>_timeout_<ms>ms`); the
+//   SDK driver's is the turn budget (`error_max_turns` -> `agent_sdk_error_max_turns`). The CONTRACT is:
+//   `ok: false` with a non-empty machine-readable `error`, never a throw.
+
+import { describe, expect, it } from "vitest";
+import {
+  createAgentSdkCodingAgentDriver,
+  createCliSubprocessCodingAgentDriver,
+  invokeCodingAgentDriver,
+  type AttemptLogEvent,
+  type CodingAgentDriver,
+  type CodingAgentDriverResult,
+  type CodingAgentDriverTask,
+} from "../../packages/gittensory-engine/src/index";
+
+// One shared worktree path across every describe.each iteration ON PURPOSE: harness isolation comes from each
+// case calling `make()` for a fresh driver + recorder, not from unique paths — reusing the constant keeps the
+// scoping assertions comparable across drivers.
+const WORKTREE = "/tmp/worktrees/attempt-parity-1";
+
+const task: CodingAgentDriverTask = {
+  attemptId: "attempt-parity-1",
+  workingDirectory: WORKTREE,
+  acceptanceCriteriaPath: `${WORKTREE}/ACCEPTANCE-CRITERIA.md`,
+  instructions: "Implement the fix described in ACCEPTANCE-CRITERIA.md.",
+  maxTurns: 6,
+};
+
+// The #4271 edge case: a malformed/empty acceptance-criteria input must degrade to a structured result on both
+// implementations, never a throw.
+const emptyCriteriaTask: CodingAgentDriverTask = {
+  attemptId: "attempt-parity-empty",
+  workingDirectory: WORKTREE,
+  acceptanceCriteriaPath: "",
+  instructions: "",
+  maxTurns: 1,
+};
+
+type Behavior = "success" | "clean_failure" | "budget_exhausted";
+
+/** What a harness records about the backend invocation, for the scoping/input-forwarding contract cases. */
+type RecordedInvocation = {
+  /** Every working directory the backend was pointed at — the scoping contract requires exactly the task's. */
+  cwds: string[];
+  /** The instructions/prompt text the backend received — must be the task's, verbatim, un-reframed. */
+  prompts: string[];
+};
+
+type DriverHarness = {
+  name: string;
+  make: (behavior: Behavior) => { driver: CodingAgentDriver; recorded: RecordedInvocation };
+};
+
+const DRIVER_HARNESSES: DriverHarness[] = [
+  {
+    name: "cli-subprocess (#4266)",
+    make: (behavior) => {
+      const recorded: RecordedInvocation = { cwds: [], prompts: [] };
+      const driver = createCliSubprocessCodingAgentDriver({
+        command: "claude",
+        spawn: async (_cmd, args, opts) => {
+          recorded.cwds.push(opts.cwd);
+          // Capture EVERY argv element (not a positional assumption about where the instructions land): if the
+          // driver's argv-building order ever changes, the contains-assertion below fails loudly instead of
+          // silently checking the wrong argument.
+          recorded.prompts.push(...args.map(String));
+          if (behavior === "budget_exhausted") return { stdout: "partial output", code: null, timedOut: true };
+          if (behavior === "clean_failure") return { stdout: "", code: 1, stderr: "lint failed on src/a.ts" };
+          return { stdout: "edited src/a.ts and reran the suite", code: 0 };
+        },
+      });
+      return { driver, recorded };
+    },
+  },
+  {
+    name: "agent-sdk (#4267)",
+    make: (behavior) => {
+      const recorded: RecordedInvocation = { cwds: [], prompts: [] };
+      const driver = createAgentSdkCodingAgentDriver({
+        query: (input) => {
+          recorded.cwds.push(input.options.cwd);
+          recorded.prompts.push(input.prompt);
+          return (async function* (): AsyncGenerator<Record<string, unknown>> {
+            if (behavior === "budget_exhausted") {
+              yield { type: "result", subtype: "error_max_turns", is_error: true, num_turns: 6 };
+              return;
+            }
+            if (behavior === "clean_failure") {
+              yield { type: "result", subtype: "error_during_execution", is_error: true, num_turns: 2 };
+              return;
+            }
+            yield { type: "assistant", message: { content: [{ type: "tool_use", name: "Edit", input: { file_path: "src/a.ts" } }] } };
+            yield { type: "result", subtype: "success", is_error: false, num_turns: 3, result: "edited src/a.ts" };
+          })();
+        },
+      });
+      return { driver, recorded };
+    },
+  },
+];
+
+/** The #4262 result contract, field by field — identical SHAPE requirements for every implementation. */
+function expectDriverResultShape(result: CodingAgentDriverResult): void {
+  expect(typeof result.ok).toBe("boolean");
+  expect(Array.isArray(result.changedFiles)).toBe(true);
+  for (const file of result.changedFiles) expect(typeof file).toBe("string");
+  expect(typeof result.summary).toBe("string");
+  expect(result.summary.length).toBeGreaterThan(0);
+  if (result.transcript !== undefined) expect(typeof result.transcript).toBe("string");
+  if (result.turnsUsed !== undefined) expect(typeof result.turnsUsed).toBe("number");
+  if (result.error !== undefined) {
+    expect(typeof result.error).toBe("string");
+    expect(result.error.length).toBeGreaterThan(0);
+  }
+}
+
+describe.each(DRIVER_HARNESSES)("CodingAgentDriver contract — $name", ({ make }) => {
+  it("returns the shared result shape with ok=true and no error on a successful attempt", async () => {
+    const { driver } = make("success");
+    const result = await driver.run(task);
+    expectDriverResultShape(result);
+    expect(result.ok).toBe(true);
+    expect(result.error).toBeUndefined();
+  });
+
+  it("surfaces a clean failure as a structured result (never a throw) with [] changed files", async () => {
+    const { driver } = make("clean_failure");
+    const result = await driver.run(task);
+    expectDriverResultShape(result);
+    expect(result.ok).toBe(false);
+    expect(result.changedFiles).toEqual([]);
+    expect(result.error).toBeDefined();
+  });
+
+  it("surfaces budget exhaustion (wall-clock or turn budget) the same structured way", async () => {
+    const { driver } = make("budget_exhausted");
+    const result = await driver.run(task);
+    expectDriverResultShape(result);
+    expect(result.ok).toBe(false);
+    expect(result.changedFiles).toEqual([]);
+    expect(result.error).toBeDefined();
+  });
+
+  it("scopes every backend invocation to the task's working directory and nothing else (#4269 output)", async () => {
+    const { driver, recorded } = make("success");
+    await driver.run(task);
+    expect(recorded.cwds.length).toBeGreaterThan(0);
+    for (const cwd of recorded.cwds) expect(cwd).toBe(WORKTREE);
+  });
+
+  it("forwards the composed instructions verbatim — no driver-side prompt reframing", async () => {
+    const { driver, recorded } = make("success");
+    await driver.run(task);
+    expect(recorded.prompts).toContain(task.instructions);
+  });
+
+  it("keeps failure-path summaries distinguishable from success-path summaries", async () => {
+    const success = await make("success").driver.run(task);
+    const failure = await make("clean_failure").driver.run(task);
+    expect(success.summary).not.toBe(failure.summary);
+  });
+
+  it("degrades a malformed/empty acceptance-criteria input (#4271) to a structured result, never a throw", async () => {
+    const { driver } = make("success");
+    const result = await driver.run(emptyCriteriaTask);
+    expectDriverResultShape(result);
+  });
+});
+
+describe("caller-side interchangeability (#2333's requirement)", () => {
+  it("produces the same attempt-log event sequence through invokeCodingAgentDriver for both drivers", async () => {
+    const sequences: string[][] = [];
+    for (const harness of DRIVER_HARNESSES) {
+      const events: AttemptLogEvent[] = [];
+      const { driver } = harness.make("success");
+      const result = await invokeCodingAgentDriver(driver, "live", task, { append: (event) => events.push(event) });
+      expect(result.ok).toBe(true);
+      sequences.push(events.map((event) => event.eventType));
+    }
+    expect(sequences[0]).toEqual(["attempt_started", "attempt_succeeded"]);
+    expect(sequences[1]).toEqual(sequences[0]);
+  });
+
+  it("produces the same attempt-log failure sequence for both drivers", async () => {
+    const sequences: string[][] = [];
+    for (const harness of DRIVER_HARNESSES) {
+      const events: AttemptLogEvent[] = [];
+      const { driver } = harness.make("clean_failure");
+      const result = await invokeCodingAgentDriver(driver, "live", task, { append: (event) => events.push(event) });
+      expect(result.ok).toBe(false);
+      sequences.push(events.map((event) => event.eventType));
+    }
+    expect(sequences[0]).toEqual(["attempt_started", "attempt_failed"]);
+    expect(sequences[1]).toEqual(sequences[0]);
+  });
+});
+
+describe("documented divergences, locked in explicitly", () => {
+  it("CLI driver reports no turn count (a subprocess exposes none) — turnsUsed stays undefined", async () => {
+    const { driver } = DRIVER_HARNESSES[0]!.make("success");
+    const result = await driver.run(task);
+    expect(result.turnsUsed).toBeUndefined();
+  });
+
+  it("Agent-SDK driver reports the result frame's num_turns", async () => {
+    const { driver } = DRIVER_HARNESSES[1]!.make("success");
+    const result = await driver.run(task);
+    expect(result.turnsUsed).toBe(3);
+  });
+});