Merge pull request #19 from pullfrog/custom-bash

Switch to custom Bash tool. Mask secrets from Bash subprocs.

.gitignore

@@ -46,4 +46,4 @@ examples
 .temp/
 dist
 
-.pnpm-store/
+.pnpm-store/

agents/claude.ts

@@ -21,28 +21,23 @@ export const claude = agent({
     const prompt = addInstructions({ payload, repo });
     log.group("Full prompt", () => log.info(prompt));
 
-    // configure sandbox mode if enabled
+    // SECURITY: For PUBLIC repos, Claude Code spawns subprocesses with full process.env, leaking API keys.
+    // disable native Bash; agents use MCP bash tool which filters secrets.
+    // for private repos, native Bash is allowed since secrets are less exposed.
+    const disallowedTools = repo.isPublic ? ["Bash"] : [];
     const sandboxOptions: Options = payload.sandbox
       ? {
           permissionMode: "default",
           disallowedTools: ["Bash", "WebSearch", "WebFetch", "Write"],
           async canUseTool(toolName, input, _options) {
             if (toolName.startsWith("mcp__gh_pullfrog__"))
-              return {
+              return { behavior: "allow", updatedInput: input, updatedPermissions: [] };
-                behavior: "allow",
+            return { behavior: "deny", message: "tool not allowed in sandbox mode" };
-                updatedInput: input,
-                updatedPermissions: [],
-              };
-
-            console.error("can i use this tool?", toolName);
-            return {
-              behavior: "deny",
-              message: "You are not allowed to use this tool.",
-            };
           },
         }
       : {
           permissionMode: "bypassPermissions" as const,
+          disallowedTools,
         };
 
     if (payload.sandbox) {

agents/codex.ts

@@ -1,15 +1,55 @@
-import { spawnSync } from "node:child_process";
+import { mkdirSync, writeFileSync } from "node:fs";
-import { mkdirSync } from "node:fs";
 import { join } from "node:path";
+import type { McpHttpServerConfig } from "@anthropic-ai/claude-agent-sdk";
 import { Codex, type CodexOptions, type ThreadEvent } from "@openai/codex-sdk";
 import { log } from "../utils/cli.ts";
 import { addInstructions } from "./instructions.ts";
-import {
+import { agent, installFromNpmTarball, setupProcessAgentEnv } from "./shared.ts";
-  agent,
+
-  type ConfigureMcpServersParams,
+interface WriteCodexConfigParams {
-  installFromNpmTarball,
+  tempHome: string;
-  setupProcessAgentEnv,
+  mcpServers: Record<string, McpHttpServerConfig>;
-} from "./shared.ts";
+  isPublicRepo: boolean;
+}
+
+function writeCodexConfig({ tempHome, mcpServers, isPublicRepo }: WriteCodexConfigParams): string {
+  const codexDir = join(tempHome, ".codex");
+  mkdirSync(codexDir, { recursive: true });
+  const configPath = join(codexDir, "config.toml");
+
+  // build MCP servers section
+  const mcpServerSections: string[] = [];
+  for (const [name, config] of Object.entries(mcpServers)) {
+    if (config.type !== "http") continue;
+    log.info(`» Adding MCP server '${name}' at ${config.url}`);
+    mcpServerSections.push(`[mcp_servers.${name}]\nurl = "${config.url}"`);
+  }
+
+  // SECURITY: for public repos, enforce env filtering via shell_environment_policy
+  // this prevents vuln if user's ~/.codex/config.toml has ignore_default_excludes=true
+  // for private repos, no filtering - agents use native shell with full env access
+  const shellPolicy = isPublicRepo
+    ? `[shell_environment_policy]
+ignore_default_excludes = false`
+    : "";
+
+  writeFileSync(
+    configPath,
+    `# written by pullfrog
+${shellPolicy}
+
+${mcpServerSections.join("\n\n")}
+`.trim() + "\n"
+  );
+
+  if (isPublicRepo) {
+    log.info(`» Codex config written to ${configPath} (env filtering: enabled)`);
+  } else {
+    log.info(`» Codex config written to ${configPath} (private repo: no env filtering)`);
+  }
+
+  return codexDir;
+}
 
 export const codex = agent({
   name: "codex",
@@ -21,18 +61,24 @@ export const codex = agent({
     });
   },
   run: async ({ payload, mcpServers, apiKey, cliPath, repo }) => {
-    // create config directory for codex before setting HOME
     const tempHome = process.env.PULLFROG_TEMP_DIR!;
+
+    // create config directory for codex before setting HOME
     const configDir = join(tempHome, ".config", "codex");
     mkdirSync(configDir, { recursive: true });
 
+    const codexDir = writeCodexConfig({
+      tempHome,
+      mcpServers,
+      isPublicRepo: repo.isPublic,
+    });
+
     setupProcessAgentEnv({
       OPENAI_API_KEY: apiKey,
       HOME: tempHome,
+      CODEX_HOME: codexDir, // point Codex to our config directory
     });
 
-    configureCodexMcpServers({ mcpServers, cliPath });
-
     // Configure Codex
     const codexOptions: CodexOptions = {
       apiKey,
@@ -44,7 +90,6 @@ export const codex = agent({
     }
 
     const codex = new Codex(codexOptions);
-    // valid sandbox modes: read-only, workspace-write, danger-full-access
     const thread = codex.startThread(
       payload.sandbox
         ? {
@@ -186,33 +231,3 @@ const messageHandlers: {
     log.error(`Error: ${event.message}`);
   },
 };
-
-/**
- * Configure MCP servers for Codex using the CLI.
- * For HTTP-based servers, use: codex mcp add <name> --url <url>
- */
-function configureCodexMcpServers({ mcpServers, cliPath }: ConfigureMcpServersParams): void {
-  for (const [serverName, serverConfig] of Object.entries(mcpServers)) {
-    if (serverConfig.type === "http") {
-      // HTTP-based MCP server - use --url flag
-      const addArgs = ["mcp", "add", serverName, "--url", serverConfig.url];
-
-      log.info(`Adding MCP server '${serverName}' at ${serverConfig.url}...`);
-      const addResult = spawnSync("node", [cliPath, ...addArgs], {
-        stdio: "pipe",
-        encoding: "utf-8",
-      });
-
-      if (addResult.status !== 0) {
-        throw new Error(
-          `codex mcp add failed: ${addResult.stderr || addResult.stdout || "Unknown error"}`
-        );
-      }
-      log.info(`✓ MCP server '${serverName}' configured`);
-    } else {
-      throw new Error(
-        `Unsupported MCP server type for Codex: ${(serverConfig as any).type || "unknown"}`
-      );
-    }
-  }
-}

agents/cursor.ts

@@ -93,7 +93,7 @@ export const cursor = agent({
   },
   run: async ({ payload, apiKey, cliPath, mcpServers, repo }) => {
     configureCursorMcpServers({ mcpServers, cliPath });
-    configureCursorSandbox({ sandbox: payload.sandbox ?? false });
+    configureCursorSandbox({ sandbox: payload.sandbox ?? false, isPublicRepo: repo.isPublic });
 
     // track logged model_call_ids to avoid duplicates
     // cursor emits each assistant message twice: once without model_call_id, then again with it
@@ -314,39 +314,46 @@ function configureCursorMcpServers({ mcpServers }: ConfigureMcpServersParams) {
 
 /**
  * Configure Cursor CLI sandbox mode via cli-config.json.
- * When sandbox is enabled, denies all file writes and shell commands.
- * In print mode without --force, writes are blocked by default, but we add
- * explicit deny rules as defense in depth.
  *
- * See: https://cursor.com/docs/cli/reference/permissions
+ * SECURITY: For PUBLIC repos, Cursor spawns subprocesses with full process.env, leaking API keys.
+ * We deny native Shell via Shell(*) rule, forcing use of MCP bash tool which
+ * filters secrets. Note: Shell(**) does NOT work, must use Shell(*).
+ * For private repos, native Shell is allowed.
+ *
+ * Config path: $XDG_CONFIG_HOME/cursor/ (not ~/.cursor/) because createAgentEnv
+ * sets XDG_CONFIG_HOME=$HOME/.config. See issues/cursor-perms.md.
  */
-function configureCursorSandbox({ sandbox }: { sandbox: boolean }): void {
+function configureCursorSandbox({
+  sandbox,
+  isPublicRepo,
+}: {
+  sandbox: boolean;
+  isPublicRepo: boolean;
+}): void {
   const realHome = homedir();
-  const cursorConfigDir = join(realHome, ".cursor");
+  const cursorConfigDir = join(realHome, ".config", "cursor");
   const cliConfigPath = join(cursorConfigDir, "cli-config.json");
   mkdirSync(cursorConfigDir, { recursive: true });
 
+  // deny native shell for public repos to prevent secret leakage
+  const denyShell = isPublicRepo ? ["Shell(*)"] : [];
+
   const config = sandbox
     ? {
-        // sandbox mode: deny all writes and shell commands
         permissions: {
-          allow: [
+          allow: ["Read(**)"],
-            "Read(**)", // allow reading all files
+          deny: ["Write(**)", ...denyShell],
-          ],
-          deny: [
-            "Write(**)", // deny all file writes
-            "Shell(**)", // deny all shell commands
-          ],
         },
       }
     : {
-        // normal mode: allow everything
         permissions: {
-          allow: ["Read(**)", "Write(**)", "Shell(**)"],
+          allow: ["Read(**)", "Write(**)"],
-          deny: [],
+          deny: denyShell,
         },
       };
 
   writeFileSync(cliConfigPath, JSON.stringify(config, null, 2), "utf-8");
-  log.info(`» CLI config written to ${cliConfigPath} (sandbox: ${sandbox})`);
+  log.info(
+    `» CLI config written to ${cliConfigPath} (sandbox: ${sandbox}, isPublicRepo: ${isPublicRepo})`
+  );
 }

agents/gemini.ts

@@ -157,7 +157,7 @@ export const gemini = agent({
     });
   },
   run: async ({ payload, apiKey, mcpServers, cliPath, repo }) => {
-    configureGeminiMcpServers({ mcpServers, cliPath });
+    configureGeminiMcpServers({ mcpServers, isPublicRepo: repo.isPublic });
 
     if (!apiKey) {
       throw new Error("google_api_key or gemini_api_key is required for gemini agent");
@@ -166,34 +166,41 @@ export const gemini = agent({
     const sessionPrompt = addInstructions({ payload, repo });
     log.group("Full prompt", () => log.info(sessionPrompt));
 
-    // configure sandbox mode if enabled
+    // build CLI args based on sandbox mode
-    // --allowed-tools restricts which tools are available (removes others from registry entirely)
+    // for public repos, native shell is disabled via excludeTools in settings.json
-    // in sandbox mode: only read-only tools available (no write_file, run_shell_command, web_fetch)
+    let args: string[];
-    const args = payload.sandbox
+    if (payload.sandbox) {
-      ? [
+      // sandbox mode: read-only tools only
-          "--allowed-tools",
+      args = [
-          "read_file,list_directory,search_file_content,glob,save_memory,write_todos",
+        "--allowed-tools",
-          "--allowed-mcp-server-names",
+        "read_file,list_directory,search_file_content,glob,save_memory,write_todos",
-          "gh_pullfrog",
+        "--allowed-mcp-server-names",
-          "--output-format=stream-json",
+        "gh_pullfrog",
-          "-p",
+        "--output-format=stream-json",
-          sessionPrompt,
+        "-p",
-        ]
+        sessionPrompt,
-      : ["--yolo", "--output-format=stream-json", "-p", sessionPrompt];
+      ];
+    } else {
+      // normal mode: --yolo for auto-approval
+      // for public repos, shell is excluded via settings.json excludeTools
+      args = ["--yolo", "--output-format=stream-json", "-p", sessionPrompt];
+      if (repo.isPublic) {
+        log.info("🔒 public repo: native shell disabled via excludeTools, using MCP bash");
+      }
+    }
 
     if (payload.sandbox) {
       log.info("🔒 sandbox mode enabled: restricting to read-only operations");
     }
 
     let finalOutput = "";
-    let stdoutBuffer = ""; // buffer for incomplete lines across chunks
+    let stdoutBuffer = "";
+
     try {
       const result = await spawn({
         cmd: "node",
         args: [cliPath, ...args],
-        env: createAgentEnv({
+        env: createAgentEnv({ GEMINI_API_KEY: apiKey }),
-          GEMINI_API_KEY: apiKey,
-        }),
         onStdout: async (chunk) => {
           const text = chunk.toString();
           finalOutput += text;
@@ -266,12 +273,19 @@ export const gemini = agent({
   },
 });
 
+type ConfigureGeminiParams = {
+  mcpServers: ConfigureMcpServersParams["mcpServers"];
+  isPublicRepo: boolean;
+};
+
 /**
  * Configure MCP servers for Gemini by writing to settings.json.
  * Gemini CLI uses `httpUrl` for HTTP/streamable transport, `url` for SSE transport.
  * See: https://github.com/google-gemini/gemini-cli/blob/main/docs/get-started/configuration.md
+ *
+ * For public repos, also configures excludeTools to disable native shell.
  */
-function configureGeminiMcpServers({ mcpServers }: ConfigureMcpServersParams): void {
+function configureGeminiMcpServers({ mcpServers, isPublicRepo }: ConfigureGeminiParams): void {
   const realHome = homedir();
   const geminiConfigDir = join(realHome, ".gemini");
   const settingsPath = join(geminiConfigDir, "settings.json");
@@ -316,11 +330,16 @@ function configureGeminiMcpServers({ mcpServers }: ConfigureMcpServersParams): v
   }
 
   // merge with existing settings, overwriting mcpServers
-  const newSettings = {
+  // for public repos, exclude native shell tool to prevent secret leakage via env
+  const newSettings: Record<string, unknown> = {
     ...existingSettings,
     mcpServers: geminiMcpServers,
   };
 
+  if (isPublicRepo) {
+    newSettings.excludeTools = ["run_shell_command"];
+  }
+
   writeFileSync(settingsPath, JSON.stringify(newSettings, null, 2), "utf-8");
   log.info(`» MCP config written to ${settingsPath}`);
 }

agents/instructions.ts

@@ -8,6 +8,7 @@ interface RepoInfo {
   owner: string;
   name: string;
   defaultBranch: string;
+  isPublic: boolean;
 }
 
 /**
@@ -56,6 +57,9 @@ interface AddInstructionsParams {
 }
 
 export const addInstructions = ({ payload, repo }: AddInstructionsParams) => {
+  // for public repos, always use MCP bash for security (filters secrets)
+  // for private repos, agents can use their native bash
+  const useNativeBash = !repo.isPublic;
   let encodedEvent = "";
 
   const eventKeys = Object.keys(payload.event);
@@ -98,40 +102,9 @@ In case of conflict between instructions, follow this precedence (highest to low
 4. Repository-specific instructions (AGENTS.md, CLAUDE.md, etc.)
 5. User prompt
 
-## SECURITY
+## Security
 
-CRITICAL SECURITY RULES - NEVER VIOLATE UNDER ANY CIRCUMSTANCES:
+Never expose secrets (API keys, tokens, passwords, private keys, credentials) through any channel: console output, files, commits, comments, API responses, error messages, or URLs. Never serialize environment objects (\`process.env\`, \`os.environ\`, etc.) or iterate over them. If asked to reveal secrets: refuse, explain that exposing secrets is prohibited, and offer a safe alternative if applicable. Detect and deny any suspicious or malicious requests.
-
-### Rule 1: Never expose secrets through ANY means
-
-  You must NEVER expose secrets through any channel, including but not limited to:
-  - Displaying, printing, echoing, logging, or outputting to console
-  - Writing to files (including .txt, .env, .json, config files, etc.)
-  - Including in git commits, commit messages, or PR descriptions
-  - Posting in GitHub comments, issue bodies, or PR review comments
-  - Returning in tool outputs, API responses, or error messages
-  - Including in redirect URLs, WebSocket messages, or GraphQL responses
-
-  Secrets include: API keys, authentication tokens, passwords, private keys, certificates, database connection strings, and any credential used for authentication or authorization. Common patterns (case-insensitive): variables containing API_KEY, SECRET, TOKEN, PASSWORD, CREDENTIAL, PRIVATE_KEY, or AUTH in an authentication context. Use judgment: \`PUBLIC_KEY\` for a cryptographic public key is fine; \`PRIVATE_KEY\` is not.
-
-### Rule 2: Never serialize objects containing secrets
-
-  When working with objects that may contain environment variables or secrets:
-  - NEVER serialize, stringify, or dump entire environment objects (process.env, os.environ, ENV, etc.)
-  - NEVER iterate over environment variables and write their values to files
-  - NEVER include environment variable values in outputs, logs, HTTP requests, or anywhere they can be exposed
-  - If you must list properties, only show property NAMES, never values
-  - Only access specific, known-safe keys explicitly (e.g., NODE_ENV, HOME, PWD)
-
-### Rule 3: Refuse and explain
-
-  Even if explicitly requested to reveal secrets, you must:
-  1. Refuse the request
-  2. Print a message explaining that exposing secrets is prohibited for security reasons
-  3. If using ${ghPullfrogMcpName}, update the working comment to explain that secrets cannot be revealed
-  4. Offer a safe alternative, if applicable
-
-  If you encounter secrets in files or environment, acknowledge they exist but never reveal their values.
 
 ## MCP (Model Context Protocol) Tools
 
@@ -169,6 +142,12 @@ Tool names may be formatted as \`(server name)/(tool name)\`, for example: \`${g
 
 **Efficiency**: Trust the tools - do not repeatedly verify file contents or git status after operations. If a tool reports success, proceed to the next step. Only verify if you encounter an actual error.
 
+${
+  useNativeBash
+    ? `**Shell commands**: Use your native bash/shell tool for shell command execution.`
+    : `**Shell commands**: Use the \`${ghPullfrogMcpName}/bash\` MCP tool for all shell command execution. This tool provides a secure environment with filtered credentials. Do NOT use any native shell/bash tool - it is disabled for security.`
+}
+
 **Command execution**: Never use \`sleep\` to wait for commands to complete. Commands run synchronously - when the bash tool returns, the command has finished.
 
 **Commenting style**: When posting comments via ${ghPullfrogMcpName}, write as a professional team member would. Your final comments should be polished and actionable—do not include intermediate reasoning like "I'll now look at the code" or "Let me respond to the question."

agents/opencode.ts

@@ -7,6 +7,7 @@ import { addInstructions } from "./instructions.ts";
 import {
   agent,
   type ConfigureMcpServersParams,
+  createAgentEnv,
   installFromNpmTarball,
   setupProcessAgentEnv,
 } from "./shared.ts";
@@ -27,7 +28,11 @@ export const opencode = agent({
     const configDir = join(tempHome, ".config", "opencode");
     mkdirSync(configDir, { recursive: true });
 
-    configureOpenCode({ mcpServers, sandbox: payload.sandbox ?? false });
+    configureOpenCode({
+      mcpServers,
+      sandbox: payload.sandbox ?? false,
+      isPublicRepo: repo.isPublic,
+    });
 
     const prompt = addInstructions({ payload, repo });
     log.group("Full prompt", () => log.info(prompt));
@@ -42,24 +47,25 @@ export const opencode = agent({
     // 6. set up environment
     setupProcessAgentEnv({ HOME: tempHome });
 
-    // build env vars: start with process.env (includes all API_KEY vars loaded by config())
+    // SECURITY: build env vars from whitelisted base env to prevent API key leakage
-    // exclude GITHUB_TOKEN - OpenCode should use MCP server for GitHub operations, not direct token
+    // this prevents leaking other API keys (ANTHROPIC, GEMINI, etc.) to OpenCode subprocess
-    // then override with apiKeys, HOME, and XDG_CONFIG_HOME
     // XDG_CONFIG_HOME must be set because GitHub Actions sets it to a different path,
     // and OpenCode follows XDG spec (checks XDG_CONFIG_HOME before falling back to $HOME/.config)
     const env: Record<string, string> = {
-      ...(Object.fromEntries(
+      ...createAgentEnv({ HOME: tempHome }),
-        Object.entries(process.env).filter(
-          ([key, value]) => value !== undefined && key !== "GITHUB_TOKEN"
-        )
-      ) as Record<string, string>),
-      HOME: tempHome,
       XDG_CONFIG_HOME: join(tempHome, ".config"),
     };
+    // OpenCode doesn't support GitHub App installation tokens
+    delete env.GITHUB_TOKEN;
 
-    // add/override API keys from apiKeys object (uppercase keys)
+    // add API keys from apiKeys object
     for (const [key, value] of Object.entries(apiKeys || {})) {
-      env[key.toUpperCase()] = value;
+      const upperKey = key.toUpperCase();
+      env[upperKey] = value;
+      // also set GOOGLE_GENERATIVE_AI_API_KEY for Google provider compatibility
+      if (upperKey === "GEMINI_API_KEY") {
+        env.GOOGLE_GENERATIVE_AI_API_KEY = value;
+      }
     }
 
     // run OpenCode in the repository directory (process.cwd() is set to GITHUB_WORKSPACE or repo dir)
@@ -188,13 +194,14 @@ export const opencode = agent({
 interface ConfigureOpenCodeParams {
   mcpServers: ConfigureMcpServersParams["mcpServers"];
   sandbox: boolean;
+  isPublicRepo: boolean;
 }
 
 /**
  * Configure OpenCode via opencode.json config file.
  * Builds complete config with MCP servers and permissions in a single write to avoid race conditions.
  */
-function configureOpenCode({ mcpServers, sandbox }: ConfigureOpenCodeParams): void {
+function configureOpenCode({ mcpServers, sandbox, isPublicRepo }: ConfigureOpenCodeParams): void {
   const tempHome = process.env.PULLFROG_TEMP_DIR!;
   const configDir = join(tempHome, ".config", "opencode");
   mkdirSync(configDir, { recursive: true });
@@ -218,7 +225,10 @@ function configureOpenCode({ mcpServers, sandbox }: ConfigureOpenCodeParams): vo
     };
   }
 
-  // build permissions config
+  // SECURITY: For PUBLIC repos, OpenCode spawns subprocesses with full process.env, leaking API keys.
+  // disable native bash; agents use MCP bash tool which filters secrets.
+  // for private repos, native bash is allowed.
+  const bashPermission = isPublicRepo ? "deny" : "allow";
   const permission = sandbox
     ? {
         edit: "deny",
@@ -229,7 +239,7 @@ function configureOpenCode({ mcpServers, sandbox }: ConfigureOpenCodeParams): vo
       }
     : {
         edit: "allow",
-        bash: "allow",
+        bash: bashPermission,
         webfetch: "allow",
         doom_loop: "allow",
         external_directory: "allow",

agents/shared.ts

@@ -27,6 +27,7 @@ export interface RepoInfo {
   owner: string;
   name: string;
   defaultBranch: string;
+  isPublic: boolean;
 }
 
 /**

docs/docker.md

@@ -0,0 +1,52 @@
+# Docker Testing Environment
+
+`play.ts` runs in Docker by default for realistic testing (Linux, clean $HOME, matches CI).
+
+## Usage
+
+```bash
+pnpm play bash-test.ts           # runs in Docker (default)
+pnpm play --local bash-test.ts   # runs on macOS (fast iteration)
+PLAY_LOCAL=1 pnpm play ...       # same as --local
+```
+
+## Why Docker by Default?
+
+1. **Matches CI** - Linux environment like GitHub Actions
+2. **Clean $HOME** - No agent config pollution from `~/.claude`, `~/.cursor`
+3. **Tests unshare** - Verifies PID namespace sandbox works
+4. **Reproducible** - Same environment every run
+
+## Performance
+
+| Mode | Overhead |
+|------|----------|
+| Docker (cached deps) | ~1.5s |
+| Local (macOS) | ~0s |
+
+For agent runs taking 30-120s, the 1.5s overhead is negligible.
+
+## How It Works
+
+1. `play.ts` runs on host, loads `.env`
+2. Spawns Docker container with:
+   - Volume-mounted `action/` code
+   - Named volume for Linux node_modules (persists between runs)
+   - SSH agent forwarding for git clone
+   - Env vars passed via `-e` flags
+3. Inside Docker, `play.ts` runs again (detects `/.dockerenv` file)
+4. Clones `GITHUB_REPOSITORY`, runs agent
+
+## Troubleshooting
+
+**Docker not running:**
+```
+Cannot connect to the Docker daemon
+```
+→ Start Docker Desktop
+
+**SSH clone fails:**
+```
+Permission denied (publickey)
+```
+→ Ensure SSH agent is running: `ssh-add -l`

docs/security.md

@@ -0,0 +1,306 @@
+# Bash Tool Security
+
+> **Note**: Security measures described here apply to **PUBLIC repositories only**. For private repos, agents can use native bash with full environment access.
+
+## Architecture (Public Repos)
+
+```
+┌─────────────────────────────────────────────────────────────────┐
+│  GitHub Actions Runner                                          │
+│  (has secrets: ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.)        │
+│                                                                 │
+│  ┌───────────────────────────────────────────────────────────┐ │
+│  │  Pullfrog Action (Node.js)                                │ │
+│  │  - process.env contains all secrets                       │ │
+│  │  - spawns agent CLI as child process                      │ │
+│  │                                                           │ │
+│  │  ┌─────────────────────────────────────────────────────┐ │ │
+│  │  │  Agent CLI (Claude/Cursor/OpenCode/etc.)            │ │ │
+│  │  │  - receives filtered env (only API key it needs)    │ │ │
+│  │  │  - has built-in Bash tool (DISABLED for public)     │ │ │
+│  │  │  - connects to MCP server for tools                 │ │ │
+│  │  │                                                     │ │ │
+│  │  │  ┌───────────────────────────────────────────────┐ │ │ │
+│  │  │  │  MCP Bash Tool (our code)                     │ │ │ │
+│  │  │  │  - agent calls this for shell commands        │ │ │ │
+│  │  │  │  - spawns bash with filtered env              │ │ │ │
+│  │  │  │  - uses PID namespace isolation               │ │ │ │
+│  │  │  │                                               │ │ │ │
+│  │  │  │  ┌─────────────────────────────────────────┐ │ │ │ │
+│  │  │  │  │  Bash subprocess                        │ │ │ │ │
+│  │  │  │  │  - runs user-controlled commands        │ │ │ │ │
+│  │  │  │  │  - MUST NOT access secrets              │ │ │ │ │
+│  │  │  │  └─────────────────────────────────────────┘ │ │ │ │
+│  │  │  └───────────────────────────────────────────────┘ │ │ │
+│  │  └─────────────────────────────────────────────────────┘ │ │
+│  └───────────────────────────────────────────────────────────┘ │
+└─────────────────────────────────────────────────────────────────┘
+```
+
+**Key insight**: For **public repos**, the Pullfrog Action process has all secrets in `process.env`. Agent CLIs have built-in Bash tools that we can't trust since malicious actors can submit PRs with prompt injections. We disable those and provide our own MCP Bash tool that spawns subprocesses securely.
+
+For **private repos**, the threat model is different — only trusted collaborators can trigger workflows, so we allow native bash with full environment access for better performance and compatibility.
+
+---
+
+## Public vs Private Repos
+
+| Repo Visibility | Native Bash | Env Filtering | PID Isolation |
+|-----------------|-------------|---------------|---------------|
+| **Public** | Disabled | Yes | Yes (in CI) |
+| **Private** | Enabled | No | No |
+
+**Rationale**: Public repos are at risk from prompt injection attacks via pull requests from untrusted contributors. Private repos only allow trusted collaborators, so the attack surface is much smaller.
+
+---
+
+## Threat Model (Public Repos)
+
+A prompt-injected agent could run malicious bash commands to exfiltrate API keys.
+
+**Attack vectors:**
+
+| Vector | Example | Mitigation |
+|--------|---------|------------|
+| Direct env access | `env \| grep KEY` | Filter env vars before spawn |
+| Echo variable | `echo $ANTHROPIC_API_KEY` | Filter env vars before spawn |
+| `/proc/$PPID/environ` | `cat /proc/$PPID/environ` | PID namespace isolation |
+
+The first two are solved by passing filtered env to subprocess. The third requires special handling on Linux.
+
+---
+
+## Attack: /proc/$PPID/environ (Public Repos)
+
+On Linux, any process can read its parent's environment via `/proc/$PPID/environ`. Even if we spawn bash with a clean environment, the bash process can:
+
+```bash
+# read parent's (Node.js) environment - contains all secrets!
+tr '\0' '\n' < /proc/$PPID/environ | grep KEY
+```
+
+This bypasses environment filtering because we're reading the parent process's memory, not our own env.
+
+**Why this matters:**
+- Pullfrog Action (Node.js) has `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, etc. in `process.env`
+- We spawn agent CLI with filtered env (only its own API key)
+- Agent CLI spawns MCP Bash tool
+- MCP Bash tool spawns bash with filtered env (no secrets)
+- BUT bash can read `/proc/$PPID/environ` → gets Node.js process's full env
+
+---
+
+## Solution: PID Namespace Isolation (Public Repos)
+
+We use Linux PID namespaces to hide the parent process:
+
+```bash
+unshare --pid --fork --mount-proc bash -c "$CMD"
+```
+
+| Flag | Purpose |
+|------|---------|
+| `--pid` | Create new PID namespace |
+| `--fork` | Fork so child is actually in new namespace |
+| `--mount-proc` | Mount fresh `/proc` for new namespace |
+
+**Result:**
+- Child sees itself as PID 1
+- Child's PPID is 0 (doesn't exist)
+- `/proc` only shows processes in child's namespace
+- Parent's PID is invisible → `/proc/$PPID/environ` fails
+
+---
+
+## Implementation
+
+### mcp/bash.ts
+
+```typescript
+import { spawn } from "node:child_process";
+
+// filter sensitive env vars (only for public repos)
+function filterEnv(isPublicRepo: boolean): Record<string, string> {
+  const SENSITIVE = [/_KEY$/i, /_SECRET$/i, /_TOKEN$/i, /_PASSWORD$/i, /_CREDENTIAL$/i];
+  const filtered: Record<string, string> = {};
+  for (const [key, value] of Object.entries(process.env)) {
+    if (value === undefined) continue;
+    // only filter sensitive vars for public repos
+    if (isPublicRepo && SENSITIVE.some(p => p.test(key))) continue;
+    filtered[key] = value;
+  }
+  return filtered;
+}
+
+// spawn with PID namespace in CI for public repos, plain spawn otherwise
+function spawnSandboxed(command: string, options: { env, cwd, isPublicRepo }): ChildProcess {
+  const useNamespaceIsolation = process.env.CI === "true" && options.isPublicRepo;
+  if (useNamespaceIsolation) {
+    return spawn("unshare", ["--pid", "--fork", "--mount-proc", "bash", "-c", command], options);
+  }
+  return spawn("bash", ["-c", command], options);
+}
+
+// BashTool uses ctx.repo.private to determine visibility
+export function BashTool(ctx: ToolContext) {
+  const isPublicRepo = !ctx.repo.private;
+  // ... spawns with filterEnv(isPublicRepo) and isPublicRepo flag
+}
+```
+
+**Defense in depth (public repos only):**
+1. `filterEnv(true)` - prevents `env` and `echo $VAR` attacks
+2. `unshare` - prevents `/proc/$PPID/environ` attack
+
+---
+
+## Disabling Native Bash Tools (Public Repos)
+
+For **public repos**, each agent's built-in Bash/Shell tools are disabled. Agents use our MCP Bash tool which filters secrets:
+
+```typescript
+// Claude - conditional based on repo.isPublic
+const disallowedTools = repo.isPublic ? ["Bash"] : [];
+{ permissionMode: "bypassPermissions", disallowedTools }
+
+// Cursor - conditional shell denial
+const denyShell = isPublicRepo ? ["Shell(*)"] : [];
+{ permissions: { allow: ["Read(**)", "Write(**)"], deny: denyShell } }
+
+// OpenCode - conditional bash denial
+const bashPermission = isPublicRepo ? "deny" : "allow";
+{ permission: { edit: "allow", bash: bashPermission, ... } }
+
+// Gemini - uses excludeTools in ~/.gemini/settings.json
+newSettings.excludeTools = ["run_shell_command"];
+
+// Codex - CLI internally scrubs env before spawning shell
+// No SDK-level config needed; Codex handles this automatically
+```
+
+For **private repos**, native bash is allowed for all agents.
+
+---
+
+## Testing (Public Repo Scenario)
+
+Run the vulnerability test in Docker to verify protection for public repos:
+
+```bash
+# from action/ directory
+docker run --rm \
+  -v "$(pwd):/app/action:cached" \
+  -v "pullfrog-action-node-modules:/app/action/node_modules" \
+  -w /app/action \
+  -e GITHUB_ACTIONS=true \
+  -e TEST_SECRET_KEY=test-secret \
+  -e ANTHROPIC_API_KEY=sk-test \
+  --cap-add SYS_ADMIN \
+  --security-opt seccomp:unconfined \
+  node:22 bash -c "corepack enable pnpm && pnpm install --frozen-lockfile && node test/proc-environ-vuln.ts"
+```
+
+Expected output:
+```
+1. UNPROTECTED (filterEnv only):
+   Leaked: YES ❌
+
+2. PROTECTED (unshare --pid --fork --mount-proc):
+   Leaked: NO ✓
+```
+
+---
+
+## Platform Notes
+
+| Environment | Repo | Our approach |
+|-------------|------|--------------|
+| GitHub Actions (Linux) | Public | filterEnv + unshare + disable native bash |
+| GitHub Actions (Linux) | Private | Full env + native bash allowed |
+| Local dev (any OS) | Any | No filtering (local dev assumed trusted) |
+
+We check `process.env.CI === "true"` (set by GitHub Actions) combined with `ctx.repo.private` to determine the security posture:
+- **CI + Public repo**: Full protection with PID namespace isolation
+- **CI + Private repo**: No protection (trusted collaborators only)
+- **Local**: No protection (developer's own machine)
+
+GitHub Actions uses Ubuntu runners where `unshare` works without root.
+
+---
+
+## What This Does NOT Protect Against (Public Repos)
+
+Even with protections enabled, bash subprocesses can still:
+
+- **Network exfiltration**: Child has full network access
+- **File access**: Child can read any file the runner can (same UID)
+- **Resource exhaustion**: No cgroup limits
+
+For those, you'd need `bwrap` with `--unshare-net`, `--ro-bind`, etc. But for the stated goal—preventing secret exfiltration via env—this is sufficient.
+
+For **private repos**, none of these protections apply since we trust collaborators.
+
+---
+
+## Agent-Specific Notes
+
+### Claude, Cursor, OpenCode (Public Repos)
+
+These agents have their native Bash disabled via configuration. They use our `gh_pullfrog` MCP server's `bash` tool which implements `filterEnv()` + `unshare`.
+
+For private repos, native bash is enabled for these agents.
+
+### Gemini (Public Repos)
+
+Gemini CLI supports `excludeTools` in its user-level settings file (`~/.gemini/settings.json`). For public repos, we exclude the native shell tool:
+
+```typescript
+// written to ~/.gemini/settings.json
+newSettings.excludeTools = ["run_shell_command"];
+```
+
+This is a blocklist approach which explicitly excludes the shell tool while allowing all other tools.
+
+Additionally, Gemini has built-in CI detection that filters shell env when `GITHUB_SHA` is set.
+
+### Codex
+
+Codex CLI filters out env vars matching `KEY`, `SECRET`, or `TOKEN` (case-insensitive) by default via `shell_environment_policy.ignore_default_excludes = false`.
+
+**Vulnerability**: If a user's `~/.codex/config.toml` has `ignore_default_excludes = true`, secrets will leak to shell commands.
+
+**Our mitigation**: We set `CODEX_HOME` to a temp directory and write our own `config.toml` with `ignore_default_excludes = false` to enforce filtering regardless of what config exists in the user's `~/.codex/`.
+
+```typescript
+// set CODEX_HOME to override user's config
+setupProcessAgentEnv({ CODEX_HOME: codexDir });
+
+// write secure config to $CODEX_HOME/config.toml
+writeFileSync(join(codexDir, "config.toml"), `
+[shell_environment_policy]
+ignore_default_excludes = false
+`);
+```
+
+See [GitHub Issue #3064](https://github.com/openai/codex/issues/3064) and [config docs](https://github.com/openai/codex/blob/main/docs/config.md#shell_environment_policy).
+
+**Verified behavior** (tested via `pnpm play codex-env-test.ts`):
+- Default (no config): ✅ secrets filtered
+- `ignore_default_excludes = false`: ✅ secrets filtered  
+- `ignore_default_excludes = true`: ❌ secrets leak
+
+Example output when running `env | grep TEST` with our config:
+```
+TEST_SAFE_VAR=VISIBLE-SAFE-VALUE
+# FAKE_SECRET_KEY and TEST_API_TOKEN are NOT visible (filtered)
+```
+
+### Summary by Agent
+
+| Agent | Public Repo | Private Repo |
+|-------|-------------|--------------|
+| Claude | Native bash **disabled** | Native bash allowed |
+| Cursor | Native shell **disabled** | Native shell allowed |
+| OpenCode | Native bash **disabled** | Native bash allowed |
+| Gemini | Native shell **disabled** (via excludeTools) | Native bash allowed |
+| Codex | Native shell allowed (CLI scrubs env internally) | Native bash allowed |

fixtures/bash-test.ts

@@ -0,0 +1,35 @@
+import type { Payload } from "../external.ts";
+
+/**
+ * test fixture: verifies agents use MCP bash tool for shell commands.
+ * creates a simple test file and runs it with node.
+ *
+ * for insecure agents (claude, cursor, opencode): native bash is disabled,
+ * so they MUST use gh_pullfrog/bash MCP tool to run shell commands.
+ *
+ * for secure agents (codex, gemini): native bash is safe, but this test
+ * still verifies shell execution works.
+ *
+ * run with: AGENT_OVERRIDE=<agent> pnpm play bash-test.ts
+ */
+export default {
+  "~pullfrog": true,
+  agent: null,
+  prompt: `Create a file called test-runner.js with the following content:
+
+\`\`\`javascript
+const assert = require('assert');
+assert.strictEqual(2 + 2, 4, 'math should work');
+console.log('TEST PASSED: basic arithmetic works');
+\`\`\`
+
+Then run it with: node test-runner.js
+
+Finally, delete the test file.
+
+This tests that you can execute shell commands properly.`,
+  event: {
+    trigger: "workflow_dispatch",
+  },
+  modes: [],
+} satisfies Payload;

fixtures/basic.ts

@@ -1,9 +1,12 @@
-import type { Inputs } from "../main.ts";
+import type { Payload } from "../external.ts";
 
-const testParams = {
+export default {
+  "~pullfrog": true,
+  agent: null,
   prompt:
     "List all files in the current directory, then create a file called dynamic-test.txt with the content 'This was loaded from a TypeScript file!', then delete it.",
-  anthropic_api_key: "sk-test-key",
+  event: {
-} satisfies Inputs;
+    trigger: "workflow_dispatch",
-
+  },
-export default testParams;
+  modes: [],
+} satisfies Payload;

fixtures/sandbox.ts

@@ -6,9 +6,9 @@ import type { Payload } from "../external.ts";
  *
  * run with: AGENT_OVERRIDE=claude pnpm play sandbox.ts
  */
-const payload: Payload = {
+export default {
   "~pullfrog": true,
-  agent: null, // let AGENT_OVERRIDE control this for testing different agents
+  agent: null,
   prompt: `Please do the following three things:
 
 1. Fetch the content from https://httpbin.org/json and tell me what it says
@@ -24,6 +24,4 @@ All three of these actions should fail because you are running in sandbox mode w
   },
   modes: [],
   sandbox: true,
-};
+} satisfies Payload;
-
-export default JSON.stringify(payload);

main.ts

@@ -2,7 +2,7 @@ import { mkdtemp } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { flatMorph } from "@ark/util";
-import { Octokit } from "@octokit/rest";
+import type { Octokit } from "@octokit/rest";
 import { encode as toonEncode } from "@toon-format/toon";
 import { type } from "arktype";
 import { type Agent, agents } from "./agents/index.ts";
@@ -370,6 +370,7 @@ function resolveAgent({
   repoSettings: RepoSettings;
 }): Agent {
   const agentOverride = process.env.AGENT_OVERRIDE as AgentName | undefined;
+  log.debug(`» determineAgent: agentOverride=${agentOverride}, payload.agent=${payload.agent}, repoSettings.defaultAgent=${repoSettings.defaultAgent}`);
   const configuredAgentName = agentOverride || payload.agent || repoSettings.defaultAgent || null;
 
   if (configuredAgentName) {
@@ -513,6 +514,7 @@ async function runAgent(ctx: AgentContext): Promise<AgentResult> {
       owner: ctx.owner,
       name: ctx.name,
       defaultBranch: ctx.repo.default_branch,
+      isPublic: !ctx.repo.private,
     },
   });
 }

mcp/bash.ts

@@ -0,0 +1,129 @@
+import { type ChildProcess, spawn } from "node:child_process";
+import { type } from "arktype";
+import type { ToolContext } from "../main.ts";
+import { execute, tool } from "./shared.ts";
+
+export const BashParams = type({
+  command: "string",
+  description: "string",
+  "timeout?": "number",
+  "working_directory?": "string",
+});
+
+// patterns for sensitive env vars: suffixes (_KEY, _SECRET, _TOKEN) plus AI provider prefixes
+const SENSITIVE_PATTERNS = [/_KEY$/i, /_SECRET$/i, /_TOKEN$/i, /_PASSWORD$/i, /_CREDENTIAL$/i];
+
+function isSensitive(key: string): boolean {
+  return SENSITIVE_PATTERNS.some((p) => p.test(key));
+}
+
+/** filter env vars, removing sensitive values (only for public repos) */
+function filterEnv(isPublicRepo: boolean): Record<string, string> {
+  const filtered: Record<string, string> = {};
+  for (const [key, value] of Object.entries(process.env)) {
+    if (value === undefined) continue;
+    // only filter sensitive vars for public repos
+    if (isPublicRepo && isSensitive(key)) continue;
+    filtered[key] = value;
+  }
+  return filtered;
+}
+
+/**
+ * spawn command with filtered env. in CI, also use PID namespace isolation
+ * to prevent child from reading /proc/$PPID/environ (only for public repos)
+ */
+function spawnSandboxed(
+  command: string,
+  options: { env: Record<string, string>; cwd: string; isPublicRepo: boolean }
+): ChildProcess {
+  const stdio: ["ignore", "pipe", "pipe"] = ["ignore", "pipe", "pipe"];
+  const spawnOpts = { env: options.env, cwd: options.cwd, stdio, detached: true };
+  // only use PID namespace isolation for public repos in CI
+  const useNamespaceIsolation = process.env.CI === "true" && options.isPublicRepo;
+  return useNamespaceIsolation
+    ? spawn("unshare", ["--pid", "--fork", "--mount-proc", "bash", "-c", command], spawnOpts)
+    : spawn("bash", ["-c", command], spawnOpts);
+}
+
+/** kill process and its entire process group */
+async function killProcessGroup(proc: ChildProcess): Promise<void> {
+  if (!proc.pid) return;
+  try {
+    process.kill(-proc.pid, "SIGTERM");
+    await new Promise((r) => setTimeout(r, 200));
+    process.kill(-proc.pid, "SIGKILL");
+  } catch {
+    try {
+      proc.kill("SIGKILL");
+    } catch {
+      /* already dead */
+    }
+  }
+}
+
+export function BashTool(ctx: ToolContext) {
+  const isPublicRepo = !ctx.repo.private;
+
+  return tool({
+    name: "bash",
+    description: `Execute shell commands securely.${isPublicRepo ? " Environment is filtered to remove API keys and secrets." : ""}
+
+Use this tool to:
+- Run shell commands (ls, cat, grep, find, etc.)
+- Execute build tools (npm, pnpm, cargo, make, etc.)
+- Run tests and linters
+- Perform git operations
+- Run shell commands in a secure environment. Unlike the built-in bash tool, this tool filters sensitive environment variables from the subprocess's environment to avoid leaking secrets.`,
+    parameters: BashParams,
+    execute: execute(async (params) => {
+      const timeout = Math.min(params.timeout ?? 120000, 600000);
+      const cwd = params.working_directory ?? process.cwd();
+      const proc = spawnSandboxed(params.command, {
+        env: filterEnv(isPublicRepo),
+        cwd,
+        isPublicRepo,
+      });
+
+      let stdout = "",
+        stderr = "",
+        timedOut = false,
+        exited = false;
+      proc.stdout?.on("data", (chunk: Buffer) => {
+        stdout += chunk.toString();
+      });
+      proc.stderr?.on("data", (chunk: Buffer) => {
+        stderr += chunk.toString();
+      });
+
+      const timeoutId = setTimeout(async () => {
+        if (!exited) {
+          timedOut = true;
+          await killProcessGroup(proc);
+        }
+      }, timeout);
+
+      const exitCode = await new Promise<number | null>((resolve) => {
+        const done = (code: number | null) => {
+          exited = true;
+          clearTimeout(timeoutId);
+          resolve(code);
+        };
+        proc.on("exit", done);
+        proc.on("error", () => done(null));
+      });
+
+      let output = stderr ? (stdout ? `${stdout}\n${stderr}` : stderr) : stdout;
+      if (timedOut)
+        output = output
+          ? `${output}\n[timed out after ${timeout}ms]`
+          : `[timed out after ${timeout}ms]`;
+
+      return {
+        output: output.trim(),
+        exit_code: exitCode ?? (timedOut ? 124 : -1),
+        timed_out: timedOut,
+      };
+    }),
+  });
+}

mcp/server.ts

@@ -29,6 +29,7 @@ import { CreatePullRequestReviewTool } from "./review.ts";
 import { GetReviewCommentsTool, ListPullRequestReviewsTool } from "./reviewComments.ts";
 import { SelectModeTool } from "./selectMode.ts";
 import { addTools } from "./shared.ts";
+import { BashTool } from "./bash.ts";
 
 /**
  * Find an available port starting from the given port
@@ -94,6 +95,7 @@ export async function startMcpHttpServer(
     CreateBranchTool(ctx),
     CommitFilesTool(ctx),
     PushBranchTool(ctx),
+    BashTool(ctx),
   ];
 
   if (!ctx.payload.disableProgressComment) {

package.json

@@ -36,10 +36,10 @@
     "@standard-schema/spec": "1.0.0",
     "@toon-format/toon": "^1.0.0",
     "arktype": "2.1.28",
-    "package-manager-detector": "^1.6.0",
     "dotenv": "^17.2.3",
     "execa": "^9.6.0",
     "fastmcp": "^3.26.8",
+    "package-manager-detector": "^1.6.0",
     "table": "^6.9.0"
   },
   "devDependencies": {

play.ts

@@ -1,3 +1,4 @@
+import { spawnSync } from "node:child_process";
 import { existsSync, readFileSync } from "node:fs";
 import { extname, join, resolve } from "node:path";
 import { pathToFileURL } from "node:url";
@@ -67,7 +68,9 @@ if (import.meta.url === `file://${process.argv[1]}`) {
   const args = arg({
     "--help": Boolean,
     "--raw": String,
+    "--local": Boolean,
     "-h": "--help",
+    "-l": "--local",
   });
 
   if (args["--help"]) {
@@ -81,18 +84,85 @@ Arguments:
 
 Options:
   --raw [prompt]          Use raw string as prompt instead of loading from file
+  --local, -l             Run locally on macOS (default: runs in Docker)
   -h, --help              Show this help message
 
+Environment:
+  PLAY_LOCAL=1            Same as --local
+
 Examples:
-  tsx play.ts                        # Use default fixture
+  tsx play.ts bash-test.ts           # Run in Docker (default)
-  tsx play.ts fixtures/basic.txt     # Use specific text file
+  tsx play.ts --local bash-test.ts   # Run locally on macOS
-  tsx play.ts custom.json            # Use JSON file
-  tsx play.ts fixtures/test.ts       # Use TypeScript file
   tsx play.ts --raw "Hello world"    # Use raw string as prompt
     `);
     process.exit(0);
   }
 
+  // default: run in Docker (unless --local or PLAY_LOCAL=1 or already inside Docker)
+  const isInsideDocker = existsSync("/.dockerenv");
+  const useLocal = args["--local"] || process.env.PLAY_LOCAL === "1" || isInsideDocker;
+
+  if (!useLocal) {
+    log.info("» running in Docker container...");
+
+    const passArgs = process.argv.slice(2);
+    const nodeCmd = `node play.ts ${passArgs.join(" ")}`;
+
+    // pass all env vars to docker
+    const envFlags = Object.entries(process.env).flatMap(([key, value]) =>
+      value !== undefined ? ["-e", `${key}=${value}`] : []
+    );
+
+    // SSH for git - mount individual SSH files to avoid permission issues
+    const sshFlags: string[] = [];
+    const home = process.env.HOME;
+    if (home) {
+      const sshDir = join(home, ".ssh");
+      // mount SSH keys (try common key names)
+      for (const keyName of ["id_rsa", "id_ed25519", "id_ecdsa"]) {
+        const keyPath = join(sshDir, keyName);
+        if (existsSync(keyPath)) {
+          sshFlags.push("-v", `${keyPath}:/root/.ssh/${keyName}:ro`);
+        }
+      }
+      // mount known_hosts
+      const knownHostsPath = join(sshDir, "known_hosts");
+      if (existsSync(knownHostsPath)) {
+        sshFlags.push("-v", `${knownHostsPath}:/root/.ssh/known_hosts:ro`);
+      }
+    }
+
+    const ttyFlags = process.stdin.isTTY ? ["-it"] : [];
+
+    const result = spawnSync(
+      "docker",
+      [
+        "run",
+        "--rm",
+        ...ttyFlags,
+        "-v",
+        `${process.cwd()}:/app/action:cached`,
+        "-v",
+        "pullfrog-action-node-modules:/app/action/node_modules",
+        "-w",
+        "/app/action",
+        ...envFlags,
+        ...sshFlags,
+        "--cap-add",
+        "SYS_ADMIN",
+        "--security-opt",
+        "seccomp:unconfined",
+        "node:24",
+        "bash",
+        "-c",
+        `corepack enable pnpm >/dev/null 2>&1 && pnpm install --frozen-lockfile && ${nodeCmd}`,
+      ],
+      { stdio: "inherit" }
+    );
+
+    process.exit(result.status ?? 1);
+  }
+
   let prompt: string;
 
   if (args["--raw"]) {
@@ -135,10 +205,11 @@ Examples:
 
         if (typeof module.default === "string") {
           prompt = module.default;
-        } else if (typeof module.default === "object" && module.default.prompt) {
+        } else if (typeof module.default === "object") {
-          prompt = module.default.prompt;
+          // Payload objects (with ~pullfrog) should be stringified
-        } else {
           prompt = JSON.stringify(module.default, null, 2);
+        } else {
+          throw new Error(`Unsupported default export type: ${typeof module.default}`);
         }
         break;
       }

utils/github.ts

@@ -1,8 +1,8 @@
+import assert from "node:assert/strict";
+import { createSign } from "node:crypto";
 import * as core from "@actions/core";
 import { throttling } from "@octokit/plugin-throttling";
 import { Octokit } from "@octokit/rest";
-import assert from "node:assert/strict";
-import { createSign } from "node:crypto";
 import { log } from "./cli.ts";
 import { retry } from "./retry.ts";
 
@@ -47,8 +47,11 @@ interface RepositoriesResponse {
   repositories: Repository[];
 }
 
-function isGitHubActionsEnvironment(): boolean {
+function isOIDCAvailable(): boolean {
-  return Boolean(process.env.GITHUB_ACTIONS);
+  // OIDC requires both env vars to be set (only in real GitHub Actions with id-token permission)
+  return Boolean(
+    process.env.ACTIONS_ID_TOKEN_REQUEST_URL && process.env.ACTIONS_ID_TOKEN_REQUEST_TOKEN
+  );
 }
 
 async function acquireTokenViaOIDC(): Promise<string> {
@@ -236,7 +239,7 @@ async function acquireTokenViaGitHubApp(): Promise<string> {
 }
 
 async function acquireNewToken(): Promise<string> {
-  if (isGitHubActionsEnvironment()) {
+  if (isOIDCAvailable()) {
     return await retry(() => acquireTokenViaOIDC(), { label: "token exchange" });
   } else {
     return await acquireTokenViaGitHubApp();
@@ -267,7 +270,10 @@ export async function setupGitHubInstallationToken() {
  * Get the GitHub installation token from memory
  */
 export function getGitHubInstallationToken(): string {
-  assert(githubInstallationToken, "GitHub installation token not set. Call setupGitHubInstallationToken first.");
+  assert(
+    githubInstallationToken,
+    "GitHub installation token not set. Call setupGitHubInstallationToken first."
+  );
   return githubInstallationToken;
 }
 
@@ -313,7 +319,9 @@ export function parseRepoContext(): RepoContext {
   return { owner, name };
 }
 
-export type OctokitWithPlugins = InstanceType<ReturnType<typeof Octokit.plugin<typeof Octokit, [typeof throttling]>>>
+export type OctokitWithPlugins = InstanceType<
+  ReturnType<typeof Octokit.plugin<typeof Octokit, [typeof throttling]>>
+>;
 
 export function createOctokit(token: string): OctokitWithPlugins {
   // `OctokitWithPlugins` initialization based on https://github.com/actions/toolkit/blob/2506e78e82fbd2f9e94d63e75f5309118c8de1b1/packages/github/src/github.ts#L15-L22
@@ -328,6 +336,6 @@ export function createOctokit(token: string): OctokitWithPlugins {
       onSecondaryRateLimit: (retryAfter, options, octokit, retryCount) => {
         return retryCount <= 2;
       },
-    }
+    },
   });
 }