chore(hardening): ask before writing files a later command executes

.changeset/hardening-execution-trigger-writes.md

@@ -0,0 +1,15 @@
+---
+'@moonshot-ai/agent-core-v2': minor
+---
+
+Ask before writing a file that a later command will execute. Writes inside the
+workspace are approved without prompting, which is right for source files — the
+change is visible in the diff before anything runs it. It is not right for
+`package.json`, `Makefile`, `.github/workflows/*`, `conftest.py`,
+`.pre-commit-config.yaml` and friends: nothing happens when they are written,
+and then the next install, test run or CI job executes what they now say. The
+write is the dangerous act, so the prompt belongs there.
+
+Reads are unaffected, ordinary source writes keep the fast path, and session
+approvals and user `allow` rules still take precedence, so a decision already
+made is not re-asked.

packages/agent-core-v2/src/agent/permissionPolicy/permissionPolicyService.ts

@@ -15,6 +15,7 @@ import { AutoModeAskUserQuestionDenyPermissionPolicyService } from '#/agent/perm
 import { DefaultToolApprovePermissionPolicyService } from '#/agent/permissionPolicy/policies/default-tool-approve';
 import { FallbackAskPermissionPolicyService } from '#/agent/permissionPolicy/policies/fallback-ask';
 import { GitControlPathAccessAskPermissionPolicyService } from '#/agent/permissionPolicy/policies/git-control-path-access-ask';
+import { ExecutionTriggerWriteAskPermissionPolicyService } from '#/agent/permissionPolicy/policies/execution-trigger-write-ask';
 import { GitCwdWriteApprovePermissionPolicyService } from '#/agent/permissionPolicy/policies/git-cwd-write-approve';
 import { SensitiveFileAccessAskPermissionPolicyService } from '#/agent/permissionPolicy/policies/sensitive-file-access-ask';
 import { SessionApprovalHistoryPermissionPolicyService } from '#/agent/permissionPolicy/policies/session-approval-history';
@@ -44,8 +45,9 @@ export class AgentPermissionPolicyService
     super();
     // Order matters: the first policy to return a result wins.
     //
-    // `AutoModeApprove` sits after the two content-sensitive asks (secrets on
-    // disk, and the .git control directory) rather than ahead of them, so
+    // `AutoModeApprove` sits after the content-sensitive asks (secrets on
+    // disk, the .git control directory, and files a later command executes)
+    // rather than ahead of them, so
     // enabling auto mode speeds up ordinary work without also silently
     // waiving the checks that exist for the highest-consequence paths.
     // Everything else keeps its previous relative order: an explicit prior
@@ -59,6 +61,7 @@ export class AgentPermissionPolicyService
       this.instantiation.createInstance(UserConfiguredAllowPermissionPolicyService),
       this.instantiation.createInstance(SensitiveFileAccessAskPermissionPolicyService),
       this.instantiation.createInstance(GitControlPathAccessAskPermissionPolicyService),
+      this.instantiation.createInstance(ExecutionTriggerWriteAskPermissionPolicyService),
       this.instantiation.createInstance(AutoModeApprovePermissionPolicyService),
       this.instantiation.createInstance(YoloModeApprovePermissionPolicyService),
       this.instantiation.createInstance(DefaultToolApprovePermissionPolicyService),

packages/agent-core-v2/src/agent/permissionPolicy/policies/execution-trigger-write-ask.ts

@@ -0,0 +1,89 @@
+import * as pathe from 'pathe';
+
+import type { ResolvedToolExecutionHookContext } from '#/agent/toolExecutor/toolHooks';
+import { IHostEnvironment } from '#/os/interface/hostEnvironment';
+import type { IHostEnvironment as HostEnvironment } from '#/os/interface/hostEnvironment';
+import { ISessionWorkspaceContext } from '#/session/workspaceContext/workspaceContext';
+import type { ISessionWorkspaceContext as WorkspaceContext } from '#/session/workspaceContext/workspaceContext';
+import type {
+  PermissionPolicy,
+  PermissionPolicyResult,
+} from '#/agent/permissionPolicy/types';
+
+import { writeFileAccesses } from './path-utils';
+
+/**
+ * Files whose contents a routine follow-up command executes.
+ *
+ * Writes inside the workspace are otherwise approved without asking, which is
+ * the right default for source files: editing them is the job, and the change
+ * is visible in the diff before anything runs it. These are different. Nothing
+ * happens when they are written, and then the next `npm install`, test run, or
+ * CI job executes what they now say — so the write is the dangerous act and
+ * the prompt has to happen there, not at the point it finally runs.
+ */
+const EXECUTION_TRIGGER_BASENAMES = new Set<string>([
+  'package.json',
+  'makefile',
+  'gnumakefile',
+  'justfile',
+  'taskfile.yml',
+  'taskfile.yaml',
+  'jenkinsfile',
+  'setup.py',
+  'conftest.py',
+  '.pre-commit-config.yaml',
+  '.gitlab-ci.yml',
+  'azure-pipelines.yml',
+]);
+
+/**
+ * Directories where every file is executed by CI or a git operation.
+ * Compared against workspace-relative POSIX paths.
+ */
+const EXECUTION_TRIGGER_DIR_PREFIXES = [
+  '.github/workflows/',
+  '.github/actions/',
+  '.circleci/',
+  '.husky/',
+];
+
+export function isExecutionTriggerPath(relativePath: string): boolean {
+  const normalized = relativePath.replaceAll('\\', '/').toLowerCase();
+  if (normalized.startsWith('../')) return false;
+  if (EXECUTION_TRIGGER_BASENAMES.has(pathe.basename(normalized))) return true;
+  return EXECUTION_TRIGGER_DIR_PREFIXES.some((prefix) => normalized.startsWith(prefix));
+}
+
+/**
+ * Ask before writing a file that a later command will execute.
+ *
+ * Sits ahead of the blanket in-workspace write approval, and ahead of auto
+ * mode, for the same reason the sensitive-file check does: these are the
+ * writes where "it was inside the repo" is not a good enough reason to skip
+ * the prompt. Session history and user `allow` rules still take precedence,
+ * so an operator who has decided this is fine is not asked twice.
+ */
+export class ExecutionTriggerWriteAskPermissionPolicyService implements PermissionPolicy {
+  readonly name = 'execution-trigger-write-ask';
+
+  constructor(
+    @IHostEnvironment private readonly env: HostEnvironment,
+    @ISessionWorkspaceContext private readonly workspace: WorkspaceContext,
+  ) {}
+
+  evaluate(context: ResolvedToolExecutionHookContext): PermissionPolicyResult | undefined {
+    const cwd = this.workspace.workDir;
+    if (cwd.length === 0) return undefined;
+    const writes = writeFileAccesses(context);
+    if (writes.length === 0) return undefined;
+
+    const pathClass = this.env.pathClass;
+    const base = pathClass === 'win32' ? cwd.toLowerCase() : cwd;
+    const triggers = writes.some((access) => {
+      const target = pathClass === 'win32' ? access.path.toLowerCase() : access.path;
+      return isExecutionTriggerPath(pathe.relative(base, target));
+    });
+    return triggers ? { kind: 'ask' } : undefined;
+  }
+}

packages/agent-core-v2/test/agent/permissionPolicy/permissionPolicyService.test.ts

@@ -354,6 +354,62 @@ describe('AgentPermissionPolicyService git cwd write approval', () => {
       vi.unstubAllEnvs();
     }  });
 
+  it('asks before writing a file a later command executes', async () => {
+    // These are approved-by-default in-workspace writes today; nothing runs at
+    // write time, and then the next install/test/CI run executes them.
+    for (const rel of [
+      'package.json',
+      'Makefile',
+      '.github/workflows/ci.yml',
+      'conftest.py',
+      '.pre-commit-config.yaml',
+    ]) {
+      await expect(evaluate({
+        toolName: 'Write',
+        args: { path: rel, content: 'x' },
+        accesses: ToolAccesses.writeFile(join(workspaceDir, rel)),
+      }), rel).resolves.toMatchObject({
+        policyName: 'execution-trigger-write-ask',
+        result: { kind: 'ask' },
+      });
+    }
+  });
+
+  it('still asks for those files in auto mode', async () => {
+    mode = 'auto';
+    await expect(evaluate({
+      toolName: 'Write',
+      args: { path: 'package.json', content: 'x' },
+      accesses: ToolAccesses.writeFile(join(workspaceDir, 'package.json')),
+    })).resolves.toMatchObject({
+      policyName: 'execution-trigger-write-ask',
+      result: { kind: 'ask' },
+    });
+  });
+
+  it('leaves ordinary source writes on the fast path', async () => {
+    for (const rel of ['src/a.ts', 'README.md', 'src/package.json.ts', 'docs/Makefile.md']) {
+      await expect(evaluate({
+        toolName: 'Write',
+        args: { path: rel, content: 'x' },
+        accesses: ToolAccesses.writeFile(join(workspaceDir, rel)),
+      }), rel).resolves.toMatchObject({
+        policyName: 'git-cwd-write-approve',
+        result: { kind: 'approve' },
+      });
+    }
+  });
+
+  it('does not fire on a read of an execution-triggering file', async () => {
+    // Reading package.json is routine; only writing it is the risk.
+    const result = await evaluate({
+      toolName: 'Read',
+      args: { path: 'package.json' },
+      accesses: ToolAccesses.readFile(join(workspaceDir, 'package.json')),
+    });
+    expect(result?.policyName).not.toBe('execution-trigger-write-ask');
+  });
+
   it('does not use git-cwd approval in auto mode', async () => {
     mode = 'auto';
     await expect(evaluate({