Restore .claude/ and .mcp.json from PR base branch before CLI runs (#1066)

* Restore .claude/ and .mcp.json from PR base branch before CLI runs

The CLI's non-interactive mode trusts cwd: it reads .mcp.json and
.claude/settings{,.local}.json from the working directory and acts on
them before any tool-permission gating — executing hooks, setting env
vars (NODE_OPTIONS, LD_PRELOAD), running apiKeyHelper shell commands,
and auto-approving MCP servers. When this action checks out a PR head,
these files are attacker-controlled.

Rather than enumerate dangerous keys, replace the entire .claude/ tree
and .mcp.json with the versions from the PR base branch (which a
maintainer has reviewed). Paths absent on base are deleted. Uses local
git state, so no TOCTOU against the GitHub API.

* Read PR base ref from payload for config restore in agent mode

Agent mode's branchInfo.baseBranch defaults to "main" (or env/input
override) instead of the PR's actual target branch — it doesn't query
prData.baseRefName like tag mode does. This meant a PR targeting
develop would get .claude/ restored from main.

Fix by reading pull_request.base.ref directly from the webhook payload
for pull_request, pull_request_review, and pull_request_review_comment
events. For issue_comment on a PR (no base.ref in payload), fall back
to the mode-provided value — tag mode's value is correct (from GraphQL);
agent mode on issue_comment is an edge case that at worst restores from
the wrong trusted branch, which is still secure.

The payload value passes through validateBranchName for defense-in-depth
(GitHub enforces valid branch names server-side, but we validate anyway).

* Extend restored paths to .gitmodules, .ripgreprc, .claude.json

.gitmodules defines submodule URLs and paths; path-confusion attacks
against git submodule operations can write into .git/hooks. .ripgreprc
can set --pre (arbitrary command on each file) if RIPGREP_CONFIG_PATH
points at it. .claude.json is cheap defense-in-depth.

Documented why .git/ is excluded (not trackable in commits, and
restoring it would undo the PR checkout), along with .gitconfig
(git never reads it from cwd) and shell rc files (sourced from $HOME,
not cwd — checkout cannot reach $HOME).

Author: km-anthropic

src/entrypoints/run.ts

@@ -15,12 +15,20 @@ import { setupGitHubToken, WorkflowValidationSkipError } from "../github/token";
 import { checkWritePermissions } from "../github/validation/permissions";
 import { createOctokit } from "../github/api/client";
 import type { Octokits } from "../github/api/client";
-import { parseGitHubContext, isEntityContext } from "../github/context";
+import {
+  parseGitHubContext,
+  isEntityContext,
+  isPullRequestEvent,
+  isPullRequestReviewEvent,
+  isPullRequestReviewCommentEvent,
+} from "../github/context";
 import type { GitHubContext } from "../github/context";
 import { detectMode } from "../modes/detector";
 import { prepareTagMode } from "../modes/tag";
 import { prepareAgentMode } from "../modes/agent";
 import { checkContainsTrigger } from "../github/validation/trigger";
+import { restoreConfigFromBase } from "../github/operations/restore-config";
+import { validateBranchName } from "../github/operations/branch";
 import { collectActionInputsPresence } from "./collect-inputs";
 import { updateCommentLink } from "./update-comment-link";
 import { formatTurnsFromData } from "./format-turns";
@@ -217,6 +225,30 @@ async function run() {
 
     validateEnvironmentVariables();
 
+    // On PRs, .claude/ and .mcp.json in the checkout are attacker-controlled.
+    // Restore them from the base branch before the CLI reads them.
+    //
+    // We read pull_request.base.ref from the payload directly because agent
+    // mode's branchInfo.baseBranch defaults to "main" rather than the PR's
+    // actual target (agent/index.ts). For issue_comment on a PR the payload
+    // lacks base.ref, so we fall back to the mode-provided value — tag mode
+    // fetches it from GraphQL; agent mode on issue_comment is an edge case
+    // that at worst restores from the wrong trusted branch (still secure).
+    if (isEntityContext(context) && context.isPR) {
+      let restoreBase = baseBranch;
+      if (
+        isPullRequestEvent(context) ||
+        isPullRequestReviewEvent(context) ||
+        isPullRequestReviewCommentEvent(context)
+      ) {
+        restoreBase = context.payload.pull_request.base.ref;
+        validateBranchName(restoreBase);
+      }
+      if (restoreBase) {
+        restoreConfigFromBase(restoreBase);
+      }
+    }
+
     await setupClaudeCodeSettings(process.env.INPUT_SETTINGS);
 
     await installPlugins(

src/github/operations/restore-config.ts

@@ -0,0 +1,79 @@
+import { execFileSync } from "child_process";
+import { rmSync } from "fs";
+
+// Paths that are both PR-controllable and read from cwd at CLI startup.
+//
+// Deliberately excluded from the CLI's broader auto-edit blocklist:
+//   .git/        — not tracked by git; PR commits cannot place files there.
+//                  Restoring it would also undo the PR checkout entirely.
+//   .gitconfig   — git reads ~/.gitconfig and .git/config, never cwd/.gitconfig.
+//   .bashrc etc. — shells source these from $HOME; checkout cannot reach $HOME.
+//   .vscode/.idea— IDE config; nothing in the CLI's startup path reads them.
+const SENSITIVE_PATHS = [
+  ".claude",
+  ".mcp.json",
+  ".claude.json",
+  ".gitmodules",
+  ".ripgreprc",
+];
+
+/**
+ * Restores security-sensitive config paths from the PR base branch.
+ *
+ * The CLI's non-interactive mode trusts cwd: it reads `.mcp.json`,
+ * `.claude/settings.json`, and `.claude/settings.local.json` from the working
+ * directory and acts on them before any tool-permission gating — executing
+ * hooks (including SessionStart), setting env vars (NODE_OPTIONS, LD_PRELOAD,
+ * PATH), running apiKeyHelper/awsAuthRefresh shell commands, and auto-approving
+ * MCP servers. When this action checks out a PR head, all of these are
+ * attacker-controlled.
+ *
+ * Rather than enumerate every dangerous key, this replaces the entire `.claude/`
+ * tree and `.mcp.json` with the versions from the PR base branch, which a
+ * maintainer has reviewed and merged. Paths absent on base are deleted.
+ *
+ * Known limitation: if a PR legitimately modifies `.claude/` and the CLI later
+ * commits with `git add -A`, the revert will be included in that commit. This
+ * is a narrow UX tradeoff for closing the RCE surface.
+ *
+ * @param baseBranch - PR base branch name. Must be pre-validated (branch.ts
+ *   calls validateBranchName on it before returning).
+ */
+export function restoreConfigFromBase(baseBranch: string): void {
+  console.log(
+    `Restoring ${SENSITIVE_PATHS.join(", ")} from origin/${baseBranch} (PR head is untrusted)`,
+  );
+
+  // Fetch base first — if this fails we haven't touched the workspace and the
+  // caller sees a clean error.
+  execFileSync("git", ["fetch", "origin", baseBranch, "--depth=1"], {
+    stdio: "inherit",
+  });
+
+  // Delete PR-controlled versions. If the restore below fails for a given path,
+  // that path stays deleted — the safe fallback (no attacker-controlled config).
+  // A bare `git checkout` alone wouldn't remove files the PR added, so nuke first.
+  for (const p of SENSITIVE_PATHS) {
+    rmSync(p, { recursive: true, force: true });
+  }
+
+  for (const p of SENSITIVE_PATHS) {
+    try {
+      execFileSync("git", ["checkout", `origin/${baseBranch}`, "--", p], {
+        stdio: "pipe",
+      });
+    } catch {
+      // Path doesn't exist on base — it stays deleted.
+    }
+  }
+
+  // `git checkout <ref> -- <path>` stages the restored files. Unstage so the
+  // revert doesn't silently leak into commits the CLI makes later.
+  try {
+    execFileSync("git", ["reset", "--", ...SENSITIVE_PATHS], {
+      stdio: "pipe",
+    });
+  } catch {
+    // Nothing was staged, or paths don't exist on HEAD — either is fine.
+  }
+}