diff --git a/scripts/review/claude-run.ts b/scripts/review/claude-run.ts
new file mode 100644
index 0000000..e89c67a
--- /dev/null
+++ b/scripts/review/claude-run.ts
@@ -0,0 +1,51 @@
+import { isRecord, shellQuote } from "../../src/util.js";
+import { runOnce } from "../lib/subprocess.js";
+import type { ReviewConfig } from "./config.js";
+import type { ReviewRunner } from "./runner.js";
+
+function checkResult(stdout: string): void {
+  for (const line of stdout.split("\n")) {
+    const trimmed = line.trim();
+    if (!trimmed) continue;
+    let msg: unknown;
+    try {
+      msg = JSON.parse(trimmed);
+    } catch {
+      continue;
+    }
+    if (isRecord(msg) && msg.type === "result") {
+      if (msg.is_error === true) {
+        throw new Error(
+          `claude returned an error: ${typeof msg.result === "string" ? msg.result : "(no message)"}`,
+        );
+      }
+      return;
+    }
+  }
+  throw new Error(
+    `claude stream-json output contained no {"type":"result"} line`,
+  );
+}
+
+export function createClaudeReviewRunner(config: ReviewConfig): ReviewRunner {
+  return {
+    async run(workspaceDir: string, prompt: string): Promise<void> {
+      const { permissionMode, denyTools } = config.claudeCode;
+      const effectiveDenyTools = denyTools.length > 0 ? denyTools : [];
+
+      let command = `claude -p --output-format stream-json --verbose --permission-mode ${shellQuote(permissionMode)}`;
+      if (effectiveDenyTools.length > 0) {
+        command += ` --disallowedTools ${shellQuote(effectiveDenyTools.join(","))}`;
+      }
+      if (config.agent.model) {
+        command += ` --model ${shellQuote(config.agent.model)}`;
+      }
+
+      const stdout = await runOnce(command, prompt, {
+        timeoutMs: config.agent.timeoutMs,
+        cwd: workspaceDir,
+      });
+      checkResult(stdout);
+    },
+  };
+}
diff --git a/scripts/review/config.ts b/scripts/review/config.ts
index b6a8189..09353f3 100644
--- a/scripts/review/config.ts
+++ b/scripts/review/config.ts
@@ -32,6 +32,7 @@ export interface AgentConfig {
 export interface CopilotConfig {
   allowAllTools: boolean;
   allowTools: string[];
+  denyTools: string[];
 }
 
 export interface ClaudeCodeConfig {
@@ -223,6 +224,11 @@ export function parseReviewConfig(
       cp.allow_tools.every((x) => typeof x === "string")
         ? (cp.allow_tools as string[])
         : [],
+    denyTools:
+      Array.isArray(cp.deny_tools) &&
+      cp.deny_tools.every((x) => typeof x === "string")
+        ? (cp.deny_tools as string[])
+        : [],
   };
 
   const cc = section(raw, "claude_code");
diff --git a/scripts/review/copilot-run.ts b/scripts/review/copilot-run.ts
new file mode 100644
index 0000000..392a847
--- /dev/null
+++ b/scripts/review/copilot-run.ts
@@ -0,0 +1,71 @@
+import { randomUUID } from "node:crypto";
+import { shellQuote } from "../../src/util.js";
+import { runOnce } from "../lib/subprocess.js";
+import type { ReviewConfig } from "./config.js";
+import type { ReviewRunner } from "./runner.js";
+
+function checkCopilotResult(stdout: string): void {
+  for (const line of stdout.split("\n")) {
+    const trimmed = line.trim();
+    if (!trimmed) continue;
+    let msg: unknown;
+    try {
+      msg = JSON.parse(trimmed);
+    } catch {
+      continue;
+    }
+    if (
+      typeof msg === "object" &&
+      msg !== null &&
+      "type" in msg &&
+      msg.type === "result"
+    ) {
+      const exitCode =
+        "exitCode" in msg && typeof msg.exitCode === "number"
+          ? msg.exitCode
+          : 1;
+      if (exitCode !== 0) {
+        throw new Error(`copilot returned a non-zero exitCode: ${exitCode}`);
+      }
+      return;
+    }
+  }
+  throw new Error(`copilot json output contained no {"type":"result"} line`);
+}
+
+export function createCopilotReviewRunner(config: ReviewConfig): ReviewRunner {
+  return {
+    async run(workspaceDir: string, prompt: string): Promise<void> {
+      const promptBytes = Buffer.byteLength(prompt, "utf8");
+      if (promptBytes > 128 * 1024) {
+        throw new Error(
+          `copilot review: prompt too large for argv (${promptBytes} bytes > 131072)`,
+        );
+      }
+
+      const sessionId = randomUUID();
+      const { allowAllTools, allowTools, denyTools } = config.copilot;
+
+      let command = `copilot -p ${shellQuote(prompt)} --output-format json --no-ask-user --log-level none --session-id ${sessionId}`;
+      if (config.agent.model) {
+        command += ` --model ${shellQuote(config.agent.model)}`;
+      }
+      if (allowAllTools) {
+        command += " --allow-all-tools";
+      } else {
+        for (const tool of allowTools) {
+          command += ` --allow-tool=${shellQuote(tool)}`;
+        }
+      }
+      for (const tool of denyTools) {
+        command += ` --deny-tool=${shellQuote(tool)}`;
+      }
+
+      const stdout = await runOnce(command, undefined, {
+        timeoutMs: config.agent.timeoutMs,
+        cwd: workspaceDir,
+      });
+      checkCopilotResult(stdout);
+    },
+  };
+}
diff --git a/scripts/review/runner.ts b/scripts/review/runner.ts
new file mode 100644
index 0000000..ef03f8f
--- /dev/null
+++ b/scripts/review/runner.ts
@@ -0,0 +1,14 @@
+import { createClaudeReviewRunner } from "./claude-run.js";
+import type { ReviewConfig } from "./config.js";
+import { createCopilotReviewRunner } from "./copilot-run.js";
+
+export interface ReviewRunner {
+  run(workspaceDir: string, prompt: string): Promise<void>;
+}
+
+export function createRunner(config: ReviewConfig): ReviewRunner {
+  if (config.agent.kind === "copilot") {
+    return createCopilotReviewRunner(config);
+  }
+  return createClaudeReviewRunner(config);
+}
diff --git a/test/review-runners.test.ts b/test/review-runners.test.ts
new file mode 100644
index 0000000..65836d4
--- /dev/null
+++ b/test/review-runners.test.ts
@@ -0,0 +1,212 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import { createClaudeReviewRunner } from "../scripts/review/claude-run.js";
+import type { ReviewConfig } from "../scripts/review/config.js";
+import { createCopilotReviewRunner } from "../scripts/review/copilot-run.js";
+
+vi.mock("../scripts/lib/subprocess.js", () => ({
+  runOnce: vi.fn(),
+}));
+
+import { runOnce } from "../scripts/lib/subprocess.js";
+
+const mockRunOnce = vi.mocked(runOnce);
+
+function makeConfig(overrides: Partial<ReviewConfig> = {}): ReviewConfig {
+  return {
+    tracker: {
+      token: "token",
+      owner: "acme",
+      ownerType: "organization",
+      projectNumber: 1,
+      statusField: "Status",
+      activeStates: ["Todo"],
+      repos: null,
+    },
+    workspace: {
+      root: "/tmp/review",
+      hookTimeoutMs: 120000,
+    },
+    hooks: {
+      afterCreate: null,
+      beforeRun: null,
+    },
+    agent: {
+      kind: "claude_code",
+      timeoutMs: 30000,
+      model: null,
+      maxConcurrent: 1,
+    },
+    copilot: {
+      allowAllTools: false,
+      allowTools: [],
+      denyTools: [],
+    },
+    claudeCode: {
+      permissionMode: "bypassPermissions",
+      denyTools: [],
+    },
+    ...overrides,
+  };
+}
+
+beforeEach(() => {
+  mockRunOnce.mockReset();
+});
+
+describe("createClaudeReviewRunner", () => {
+  it("builds command with --verbose and optional flags, then validates result line", async () => {
+    mockRunOnce.mockResolvedValue(
+      JSON.stringify({ type: "result", is_error: false, result: "ok" }),
+    );
+    const config = makeConfig({
+      agent: {
+        kind: "claude_code",
+        timeoutMs: 12345,
+        model: "claude-opus-4.7",
+        maxConcurrent: 1,
+      },
+      claudeCode: {
+        permissionMode: "acceptEdits",
+        denyTools: ["Bash(rm:*)", "Edit"],
+      },
+    });
+
+    await createClaudeReviewRunner(config).run("/tmp/ws", "review this");
+
+    const [command, stdin, opts] = mockRunOnce.mock.calls.at(0) ?? [];
+    expect(command).toContain("--verbose");
+    expect(command).toContain("--permission-mode 'acceptEdits'");
+    expect(command).toContain("--disallowedTools 'Bash(rm:*),Edit'");
+    expect(command).toContain("--model 'claude-opus-4.7'");
+    expect(stdin).toBe("review this");
+    expect(opts).toEqual({ timeoutMs: 12345, cwd: "/tmp/ws" });
+  });
+
+  it("throws when stream-json output has no result line", async () => {
+    mockRunOnce.mockResolvedValue(
+      JSON.stringify({ type: "assistant", text: "partial" }),
+    );
+    await expect(
+      createClaudeReviewRunner(makeConfig()).run("/tmp/ws", "p"),
+    ).rejects.toThrow('{"type":"result"}');
+  });
+
+  it("throws when result line marks is_error=true", async () => {
+    mockRunOnce.mockResolvedValue(
+      JSON.stringify({
+        type: "result",
+        is_error: true,
+        result: "Permission denied",
+      }),
+    );
+    await expect(
+      createClaudeReviewRunner(makeConfig()).run("/tmp/ws", "p"),
+    ).rejects.toThrow("claude returned an error: Permission denied");
+  });
+});
+
+describe("createCopilotReviewRunner", () => {
+  it("uses --allow-all-tools for allowAllTools mode", async () => {
+    mockRunOnce.mockResolvedValue(
+      JSON.stringify({ type: "result", exitCode: 0 }),
+    );
+    const config = makeConfig({
+      agent: {
+        kind: "copilot",
+        timeoutMs: 22222,
+        model: "gpt-5",
+        maxConcurrent: 1,
+      },
+      copilot: { allowAllTools: true, allowTools: [], denyTools: [] },
+    });
+
+    await createCopilotReviewRunner(config).run("/tmp/ws", "run review");
+
+    const [command, stdin, opts] = mockRunOnce.mock.calls.at(0) ?? [];
+    expect(command).toContain("--allow-all-tools");
+    expect(command).not.toContain("--allow-tool=*");
+    expect(command).toContain("--model 'gpt-5'");
+    expect(stdin).toBeUndefined();
+    expect(opts).toEqual({ timeoutMs: 22222, cwd: "/tmp/ws" });
+  });
+
+  it("includes allow/deny tool flags when configured", async () => {
+    mockRunOnce.mockResolvedValue(
+      JSON.stringify({ type: "result", exitCode: 0 }),
+    );
+    const config = makeConfig({
+      agent: {
+        kind: "copilot",
+        timeoutMs: 30000,
+        model: null,
+        maxConcurrent: 1,
+      },
+      copilot: {
+        allowAllTools: false,
+        allowTools: ["Bash", "Read"],
+        denyTools: ["shell(git:*)", "Edit"],
+      },
+    });
+
+    await createCopilotReviewRunner(config).run("/tmp/ws", "run review");
+
+    const [command] = mockRunOnce.mock.calls.at(0) ?? [];
+    expect(command).toContain("--allow-tool='Bash'");
+    expect(command).toContain("--allow-tool='Read'");
+    expect(command).toContain("--deny-tool='shell(git:*)'");
+    expect(command).toContain("--deny-tool='Edit'");
+  });
+
+  it("throws when prompt exceeds argv limit", async () => {
+    const oversizedPrompt = "a".repeat(128 * 1024 + 1);
+    await expect(
+      createCopilotReviewRunner(
+        makeConfig({
+          agent: {
+            kind: "copilot",
+            timeoutMs: 30000,
+            model: null,
+            maxConcurrent: 1,
+          },
+        }),
+      ).run("/tmp/ws", oversizedPrompt),
+    ).rejects.toThrow("prompt too large");
+    expect(mockRunOnce).not.toHaveBeenCalled();
+  });
+
+  it("throws when output has no result line", async () => {
+    mockRunOnce.mockResolvedValue(
+      JSON.stringify({ type: "assistant.message", data: { content: "ok" } }),
+    );
+    await expect(
+      createCopilotReviewRunner(
+        makeConfig({
+          agent: {
+            kind: "copilot",
+            timeoutMs: 30000,
+            model: null,
+            maxConcurrent: 1,
+          },
+        }),
+      ).run("/tmp/ws", "p"),
+    ).rejects.toThrow('{"type":"result"}');
+  });
+
+  it("throws when copilot result exitCode is non-zero", async () => {
+    mockRunOnce.mockResolvedValue(
+      JSON.stringify({ type: "result", exitCode: 1 }),
+    );
+    await expect(
+      createCopilotReviewRunner(
+        makeConfig({
+          agent: {
+            kind: "copilot",
+            timeoutMs: 30000,
+            model: null,
+            maxConcurrent: 1,
+          },
+        }),
+      ).run("/tmp/ws", "p"),
+    ).rejects.toThrow("non-zero exitCode: 1");
+  });
+});