This commit is contained in:
Colin McDonnell
2026-01-08 13:58:08 -08:00
parent 6260b23de7
commit fe7ce4af11
11 changed files with 321 additions and 163 deletions
+4 -2
View File
@@ -21,8 +21,10 @@ export const claude = agent({
const prompt = addInstructions({ payload, repo }); const prompt = addInstructions({ payload, repo });
log.group("Full prompt", () => log.info(prompt)); log.group("Full prompt", () => log.info(prompt));
// SECURITY: Claude Code spawns subprocesses with full process.env, leaking API keys. // 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. // 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 const sandboxOptions: Options = payload.sandbox
? { ? {
permissionMode: "default", permissionMode: "default",
@@ -35,7 +37,7 @@ export const claude = agent({
} }
: { : {
permissionMode: "bypassPermissions" as const, permissionMode: "bypassPermissions" as const,
disallowedTools: ["Bash"], disallowedTools,
}; };
if (payload.sandbox) { if (payload.sandbox) {
+11 -2
View File
@@ -43,7 +43,16 @@ export const codex = agent({
log.info("🔒 sandbox mode enabled: restricting to read-only operations"); log.info("🔒 sandbox mode enabled: restricting to read-only operations");
} }
// Codex SDK isolates subprocess env - native bash is safe, no MCP override needed // SECURITY NOTE: Codex SDK does not have an option to disable native shell commands.
// For public repos, we rely on instructions telling the agent to use MCP bash instead.
// The MCP bash tool filters sensitive environment variables.
// This is not as robust as Claude/Cursor/OpenCode which can explicitly disable native bash.
if (repo.isPublic) {
log.info(
"🔒 public repo: instructions direct to MCP bash (native shell NOT disabled in SDK)"
);
}
const codex = new Codex(codexOptions); const codex = new Codex(codexOptions);
const thread = codex.startThread( const thread = codex.startThread(
payload.sandbox payload.sandbox
@@ -61,7 +70,7 @@ export const codex = agent({
); );
try { try {
const streamedTurn = await thread.runStreamed(addInstructions({ payload, repo, useNativeBash: true })); const streamedTurn = await thread.runStreamed(addInstructions({ payload, repo }));
let finalOutput = ""; let finalOutput = "";
for await (const event of streamedTurn.events) { for await (const event of streamedTurn.events) {
+18 -6
View File
@@ -93,7 +93,7 @@ export const cursor = agent({
}, },
run: async ({ payload, apiKey, cliPath, mcpServers, repo }) => { run: async ({ payload, apiKey, cliPath, mcpServers, repo }) => {
configureCursorMcpServers({ mcpServers, cliPath }); configureCursorMcpServers({ mcpServers, cliPath });
configureCursorSandbox({ sandbox: payload.sandbox ?? false }); configureCursorSandbox({ sandbox: payload.sandbox ?? false, isPublicRepo: repo.isPublic });
// track logged model_call_ids to avoid duplicates // track logged model_call_ids to avoid duplicates
// cursor emits each assistant message twice: once without model_call_id, then again with it // cursor emits each assistant message twice: once without model_call_id, then again with it
@@ -315,33 +315,45 @@ function configureCursorMcpServers({ mcpServers }: ConfigureMcpServersParams) {
/** /**
* Configure Cursor CLI sandbox mode via cli-config.json. * Configure Cursor CLI sandbox mode via cli-config.json.
* *
* SECURITY: Cursor spawns subprocesses with full process.env, leaking API keys. * 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 * We deny native Shell via Shell(*) rule, forcing use of MCP bash tool which
* filters secrets. Note: Shell(**) does NOT work, must use Shell(*). * 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 * Config path: $XDG_CONFIG_HOME/cursor/ (not ~/.cursor/) because createAgentEnv
* sets XDG_CONFIG_HOME=$HOME/.config. See issues/cursor-perms.md. * 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 realHome = homedir();
const cursorConfigDir = join(realHome, ".config", "cursor"); const cursorConfigDir = join(realHome, ".config", "cursor");
const cliConfigPath = join(cursorConfigDir, "cli-config.json"); const cliConfigPath = join(cursorConfigDir, "cli-config.json");
mkdirSync(cursorConfigDir, { recursive: true }); mkdirSync(cursorConfigDir, { recursive: true });
// deny native shell for public repos to prevent secret leakage
const denyShell = isPublicRepo ? ["Shell(*)"] : [];
const config = sandbox const config = sandbox
? { ? {
permissions: { permissions: {
allow: ["Read(**)"], allow: ["Read(**)"],
deny: ["Write(**)", "Shell(*)"], deny: ["Write(**)", ...denyShell],
}, },
} }
: { : {
permissions: { permissions: {
allow: ["Read(**)", "Write(**)"], allow: ["Read(**)", "Write(**)"],
deny: ["Shell(*)"], deny: denyShell,
}, },
}; };
writeFileSync(cliConfigPath, JSON.stringify(config, null, 2), "utf-8"); 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})`
);
} }
+38 -17
View File
@@ -157,28 +157,37 @@ export const gemini = agent({
}); });
}, },
run: async ({ payload, apiKey, mcpServers, cliPath, repo }) => { run: async ({ payload, apiKey, mcpServers, cliPath, repo }) => {
configureGeminiMcpServers({ mcpServers, cliPath }); configureGeminiMcpServers({ mcpServers, isPublicRepo: repo.isPublic });
if (!apiKey) { if (!apiKey) {
throw new Error("google_api_key or gemini_api_key is required for gemini agent"); throw new Error("google_api_key or gemini_api_key is required for gemini agent");
} }
const sessionPrompt = addInstructions({ payload, repo, useNativeBash: true }); const sessionPrompt = addInstructions({ payload, repo });
log.group("Full prompt", () => log.info(sessionPrompt)); log.group("Full prompt", () => log.info(sessionPrompt));
// Gemini CLI isolates subprocess env - native bash is safe, no MCP override needed // build CLI args based on sandbox mode
// sandbox mode uses --allowed-tools to restrict to read-only operations // for public repos, native shell is disabled via excludeTools in settings.json
const args = payload.sandbox let args: string[];
? [ if (payload.sandbox) {
"--allowed-tools", // sandbox mode: read-only tools only
"read_file,list_directory,search_file_content,glob,save_memory,write_todos", args = [
"--allowed-mcp-server-names", "--allowed-tools",
"gh_pullfrog", "read_file,list_directory,search_file_content,glob,save_memory,write_todos",
"--output-format=stream-json", "--allowed-mcp-server-names",
"-p", "gh_pullfrog",
sessionPrompt, "--output-format=stream-json",
] "-p",
: ["--yolo", "--output-format=stream-json", "-p", sessionPrompt]; 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) { if (payload.sandbox) {
log.info("🔒 sandbox mode enabled: restricting to read-only operations"); log.info("🔒 sandbox mode enabled: restricting to read-only operations");
@@ -264,12 +273,19 @@ export const gemini = agent({
}, },
}); });
type ConfigureGeminiParams = {
mcpServers: ConfigureMcpServersParams["mcpServers"];
isPublicRepo: boolean;
};
/** /**
* Configure MCP servers for Gemini by writing to settings.json. * Configure MCP servers for Gemini by writing to settings.json.
* Gemini CLI uses `httpUrl` for HTTP/streamable transport, `url` for SSE transport. * 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 * 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 realHome = homedir();
const geminiConfigDir = join(realHome, ".gemini"); const geminiConfigDir = join(realHome, ".gemini");
const settingsPath = join(geminiConfigDir, "settings.json"); const settingsPath = join(geminiConfigDir, "settings.json");
@@ -314,11 +330,16 @@ function configureGeminiMcpServers({ mcpServers }: ConfigureMcpServersParams): v
} }
// merge with existing settings, overwriting mcpServers // 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, ...existingSettings,
mcpServers: geminiMcpServers, mcpServers: geminiMcpServers,
}; };
if (isPublicRepo) {
newSettings.excludeTools = ["run_shell_command"];
}
writeFileSync(settingsPath, JSON.stringify(newSettings, null, 2), "utf-8"); writeFileSync(settingsPath, JSON.stringify(newSettings, null, 2), "utf-8");
log.info(`» MCP config written to ${settingsPath}`); log.info(`» MCP config written to ${settingsPath}`);
} }
+5 -6
View File
@@ -8,6 +8,7 @@ interface RepoInfo {
owner: string; owner: string;
name: string; name: string;
defaultBranch: string; defaultBranch: string;
isPublic: boolean;
} }
/** /**
@@ -53,14 +54,12 @@ function buildRuntimeContext(repo: RepoInfo): string {
interface AddInstructionsParams { interface AddInstructionsParams {
payload: Payload; payload: Payload;
repo: RepoInfo; repo: RepoInfo;
useNativeBash?: boolean;
} }
export const addInstructions = ({ export const addInstructions = ({ payload, repo }: AddInstructionsParams) => {
payload, // for public repos, always use MCP bash for security (filters secrets)
repo, // for private repos, agents can use their native bash
useNativeBash = false, const useNativeBash = !repo.isPublic;
}: AddInstructionsParams) => {
let encodedEvent = ""; let encodedEvent = "";
const eventKeys = Object.keys(payload.event); const eventKeys = Object.keys(payload.event);
+24 -5
View File
@@ -28,7 +28,11 @@ export const opencode = agent({
const configDir = join(tempHome, ".config", "opencode"); const configDir = join(tempHome, ".config", "opencode");
mkdirSync(configDir, { recursive: true }); mkdirSync(configDir, { recursive: true });
configureOpenCode({ mcpServers, sandbox: payload.sandbox ?? false }); configureOpenCode({
mcpServers,
sandbox: payload.sandbox ?? false,
isPublicRepo: repo.isPublic,
});
const prompt = addInstructions({ payload, repo }); const prompt = addInstructions({ payload, repo });
log.group("Full prompt", () => log.info(prompt)); log.group("Full prompt", () => log.info(prompt));
@@ -190,13 +194,14 @@ export const opencode = agent({
interface ConfigureOpenCodeParams { interface ConfigureOpenCodeParams {
mcpServers: ConfigureMcpServersParams["mcpServers"]; mcpServers: ConfigureMcpServersParams["mcpServers"];
sandbox: boolean; sandbox: boolean;
isPublicRepo: boolean;
} }
/** /**
* Configure OpenCode via opencode.json config file. * Configure OpenCode via opencode.json config file.
* Builds complete config with MCP servers and permissions in a single write to avoid race conditions. * 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 tempHome = process.env.PULLFROG_TEMP_DIR!;
const configDir = join(tempHome, ".config", "opencode"); const configDir = join(tempHome, ".config", "opencode");
mkdirSync(configDir, { recursive: true }); mkdirSync(configDir, { recursive: true });
@@ -220,11 +225,25 @@ function configureOpenCode({ mcpServers, sandbox }: ConfigureOpenCodeParams): vo
}; };
} }
// SECURITY: OpenCode spawns subprocesses with full process.env, leaking API keys. // 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. // 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 const permission = sandbox
? { edit: "deny", bash: "deny", webfetch: "deny", doom_loop: "allow", external_directory: "allow" } ? {
: { edit: "allow", bash: "deny", webfetch: "allow", doom_loop: "allow", external_directory: "allow" }; edit: "deny",
bash: "deny",
webfetch: "deny",
doom_loop: "allow",
external_directory: "allow",
}
: {
edit: "allow",
bash: bashPermission,
webfetch: "allow",
doom_loop: "allow",
external_directory: "allow",
};
// build complete config in one object // build complete config in one object
const config = { const config = {
+1
View File
@@ -27,6 +27,7 @@ export interface RepoInfo {
owner: string; owner: string;
name: string; name: string;
defaultBranch: string; defaultBranch: string;
isPublic: boolean;
} }
/** /**
+112 -43
View File
@@ -1,6 +1,8 @@
# Bash Tool Security # Bash Tool Security
## Architecture > **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)
``` ```
┌─────────────────────────────────────────────────────────────────┐ ┌─────────────────────────────────────────────────────────────────┐
@@ -15,7 +17,7 @@
│ │ ┌─────────────────────────────────────────────────────┐ │ │ │ │ ┌─────────────────────────────────────────────────────┐ │ │
│ │ │ Agent CLI (Claude/Cursor/OpenCode/etc.) │ │ │ │ │ │ Agent CLI (Claude/Cursor/OpenCode/etc.) │ │ │
│ │ │ - receives filtered env (only API key it needs) │ │ │ │ │ │ - receives filtered env (only API key it needs) │ │ │
│ │ │ - has built-in Bash tool (DISABLED) │ │ │ │ │ │ - has built-in Bash tool (DISABLED for public) │ │ │
│ │ │ - connects to MCP server for tools │ │ │ │ │ │ - connects to MCP server for tools │ │ │
│ │ │ │ │ │ │ │ │ │ │ │
│ │ │ ┌───────────────────────────────────────────────┐ │ │ │ │ │ │ ┌───────────────────────────────────────────────┐ │ │ │
@@ -35,11 +37,24 @@
└─────────────────────────────────────────────────────────────────┘ └─────────────────────────────────────────────────────────────────┘
``` ```
**Key insight**: The Pullfrog Action process has all secrets in `process.env`. Agent CLIs have built-in Bash tools that we can't trust. We disable those and provide our own MCP Bash tool that spawns subprocesses securely. **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.
--- ---
## Threat Model ## 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. A prompt-injected agent could run malicious bash commands to exfiltrate API keys.
@@ -55,7 +70,7 @@ The first two are solved by passing filtered env to subprocess. The third requir
--- ---
## Attack: /proc/$PPID/environ ## 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: On Linux, any process can read its parent's environment via `/proc/$PPID/environ`. Even if we spawn bash with a clean environment, the bash process can:
@@ -75,7 +90,7 @@ This bypasses environment filtering because we're reading the parent process's m
--- ---
## Solution: PID Namespace Isolation ## Solution: PID Namespace Isolation (Public Repos)
We use Linux PID namespaces to hide the parent process: We use Linux PID namespaces to hide the parent process:
@@ -104,53 +119,72 @@ unshare --pid --fork --mount-proc bash -c "$CMD"
```typescript ```typescript
import { spawn } from "node:child_process"; import { spawn } from "node:child_process";
// filter sensitive env vars (defense in depth) // filter sensitive env vars (only for public repos)
function filterEnv(): Record<string, string> { function filterEnv(isPublicRepo: boolean): Record<string, string> {
const SENSITIVE = [/_KEY$/i, /_SECRET$/i, /_TOKEN$/i, /^ANTHROPIC/i, ...]; const SENSITIVE = [/_KEY$/i, /_SECRET$/i, /_TOKEN$/i, /_PASSWORD$/i, /_CREDENTIAL$/i];
const filtered: Record<string, string> = {}; const filtered: Record<string, string> = {};
for (const [key, value] of Object.entries(process.env)) { for (const [key, value] of Object.entries(process.env)) {
if (value && !SENSITIVE.some(p => p.test(key))) { if (value === undefined) continue;
filtered[key] = value; // only filter sensitive vars for public repos
} if (isPublicRepo && SENSITIVE.some(p => p.test(key))) continue;
filtered[key] = value;
} }
return filtered; return filtered;
} }
// spawn with PID namespace in GitHub Actions, plain spawn locally // spawn with PID namespace in CI for public repos, plain spawn otherwise
function spawnSandboxed(command: string, options: { env, cwd }): ChildProcess { function spawnSandboxed(command: string, options: { env, cwd, isPublicRepo }): ChildProcess {
if (process.env.GITHUB_ACTIONS === "true") { const useNamespaceIsolation = process.env.CI === "true" && options.isPublicRepo;
if (useNamespaceIsolation) {
return spawn("unshare", ["--pid", "--fork", "--mount-proc", "bash", "-c", command], options); return spawn("unshare", ["--pid", "--fork", "--mount-proc", "bash", "-c", command], options);
} }
return spawn("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:** **Defense in depth (public repos only):**
1. `filterEnv()` - prevents `env` and `echo $VAR` attacks 1. `filterEnv(true)` - prevents `env` and `echo $VAR` attacks
2. `unshare` - prevents `/proc/$PPID/environ` attack 2. `unshare` - prevents `/proc/$PPID/environ` attack
--- ---
## Disabling Native Bash Tools ## Disabling Native Bash Tools (Public Repos)
Each agent has built-in Bash/Shell tools that we can't control. We disable them and force agents to use our MCP Bash tool: For **public repos**, each agent's built-in Bash/Shell tools are disabled. Agents use our MCP Bash tool which filters secrets:
```typescript ```typescript
// Claude // Claude - conditional based on repo.isPublic
disallowedTools: ["Bash"], const disallowedTools = repo.isPublic ? ["Bash"] : [];
{ permissionMode: "bypassPermissions", disallowedTools }
// Cursor // Cursor - conditional shell denial
permissions: { deny: ["Shell(**)"] } const denyShell = isPublicRepo ? ["Shell(*)"] : [];
{ permissions: { allow: ["Read(**)", "Write(**)"], deny: denyShell } }
// OpenCode // OpenCode - conditional bash denial
permission: { bash: "deny" } const bashPermission = isPublicRepo ? "deny" : "allow";
{ permission: { edit: "allow", bash: bashPermission, ... } }
// Gemini - uses excludeTools in ~/.gemini/settings.json
newSettings.excludeTools = ["run_shell_command"];
// Codex - NO SDK mechanism to disable native shell
// Relies on instructions only (limitation)
``` ```
For **private repos**, native bash is allowed for all agents.
--- ---
## Testing ## Testing (Public Repo Scenario)
Run the vulnerability test in Docker: Run the vulnerability test in Docker to verify protection for public repos:
```bash ```bash
# from action/ directory # from action/ directory
@@ -177,7 +211,26 @@ Expected output:
--- ---
## What This Does NOT Protect Against ## 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 - **Network exfiltration**: Child has full network access
- **File access**: Child can read any file the runner can (same UID) - **File access**: Child can read any file the runner can (same UID)
@@ -185,31 +238,47 @@ Expected output:
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 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 ## Agent-Specific Notes
### Platform Environments ### Claude, Cursor, OpenCode (Public Repos)
| Environment | `CI` | Our approach | These agents have their native Bash disabled via configuration. They use our `gh_pullfrog` MCP server's `bash` tool which implements `filterEnv()` + `unshare`.
|-------------|------|--------------|
| GitHub Actions (Linux) | `"true"` | filterEnv + unshare |
| Local dev (any OS) | unset | filterEnv only |
We check `CI=true` (set automatically by GitHub) rather than platform detection. This means: For private repos, native bash is enabled for these agents.
- **In CI**: Full protection with PID namespace isolation
- **Locally**: Easier testing without Docker/unshare requirements
GitHub Actions uses Ubuntu runners where `unshare` works without root. ### Gemini (Public Repos)
### Agents Using MCP Bash (Claude, Cursor, OpenCode) Gemini CLI supports `excludeTools` in its user-level settings file (`~/.gemini/settings.json`). For public repos, we exclude the native shell tool:
These agents have their native Bash disabled. They use our `gh_pullfrog` MCP server's `bash` tool which implements `filterEnv()` + `unshare`. ```typescript
// written to ~/.gemini/settings.json
newSettings.excludeTools = ["run_shell_command"];
```
### Gemini This is a blocklist approach which explicitly excludes the shell tool while allowing all other tools.
Has built-in CI detection that filters shell env when `GITHUB_SHA` or `SURFACE=Github` is set. We set `SURFACE=Github` in our env. Double protection with our `createAgentEnv()`. Additionally, Gemini has built-in CI detection that filters shell env when `GITHUB_SHA` is set.
### Codex ### Codex (Limitation)
Uses `shell_environment_policy` in config. Needs proper configuration or MCP bash fallback. **⚠️ Codex SDK does not support disabling native shell commands.** The SDK only offers `sandboxMode` options which control filesystem access, not specific tool availability.
For public repos, we rely on:
1. **Instructions** telling the agent to use MCP bash instead of native shell
2. **MCP bash tool** being available as an alternative
This is a known limitation. Codex may still use native shell if it doesn't follow instructions.
### 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 | Instructions only (**⚠️ not enforced**) | Native bash allowed |
+83 -53
View File
@@ -100536,11 +100536,8 @@ function buildRuntimeContext(repo) {
} }
return lines.join("\n"); return lines.join("\n");
} }
var addInstructions = ({ var addInstructions = ({ payload, repo }) => {
payload, const useNativeBash = !repo.isPublic;
repo,
useNativeBash = false
}) => {
let encodedEvent = ""; let encodedEvent = "";
const eventKeys = Object.keys(payload.event); const eventKeys = Object.keys(payload.event);
if (eventKeys.length === 1 && eventKeys[0] === "trigger") { if (eventKeys.length === 1 && eventKeys[0] === "trigger") {
@@ -104754,6 +104751,7 @@ var claude = agent({
delete process.env.ANTHROPIC_API_KEY; delete process.env.ANTHROPIC_API_KEY;
const prompt = addInstructions({ payload, repo }); const prompt = addInstructions({ payload, repo });
log.group("Full prompt", () => log.info(prompt)); log.group("Full prompt", () => log.info(prompt));
const disallowedTools = repo.isPublic ? ["Bash"] : [];
const sandboxOptions = payload.sandbox ? { const sandboxOptions = payload.sandbox ? {
permissionMode: "default", permissionMode: "default",
disallowedTools: ["Bash", "WebSearch", "WebFetch", "Write"], disallowedTools: ["Bash", "WebSearch", "WebFetch", "Write"],
@@ -104764,7 +104762,7 @@ var claude = agent({
} }
} : { } : {
permissionMode: "bypassPermissions", permissionMode: "bypassPermissions",
disallowedTools: ["Bash"] disallowedTools
}; };
if (payload.sandbox) { if (payload.sandbox) {
log.info("\u{1F512} sandbox mode enabled: restricting to read-only operations"); log.info("\u{1F512} sandbox mode enabled: restricting to read-only operations");
@@ -105229,6 +105227,11 @@ var codex = agent({
if (payload.sandbox) { if (payload.sandbox) {
log.info("\u{1F512} sandbox mode enabled: restricting to read-only operations"); log.info("\u{1F512} sandbox mode enabled: restricting to read-only operations");
} }
if (repo.isPublic) {
log.info(
"\u{1F512} public repo: instructions direct to MCP bash (native shell NOT disabled in SDK)"
);
}
const codex2 = new Codex(codexOptions); const codex2 = new Codex(codexOptions);
const thread = codex2.startThread( const thread = codex2.startThread(
payload.sandbox ? { payload.sandbox ? {
@@ -105243,7 +105246,7 @@ var codex = agent({
} }
); );
try { try {
const streamedTurn = await thread.runStreamed(addInstructions({ payload, repo, useNativeBash: true })); const streamedTurn = await thread.runStreamed(addInstructions({ payload, repo }));
let finalOutput2 = ""; let finalOutput2 = "";
for await (const event of streamedTurn.events) { for await (const event of streamedTurn.events) {
const handler2 = messageHandlers2[event.type]; const handler2 = messageHandlers2[event.type];
@@ -105387,7 +105390,7 @@ var cursor = agent({
}, },
run: async ({ payload, apiKey, cliPath, mcpServers, repo }) => { run: async ({ payload, apiKey, cliPath, mcpServers, repo }) => {
configureCursorMcpServers({ mcpServers, cliPath }); configureCursorMcpServers({ mcpServers, cliPath });
configureCursorSandbox({ sandbox: payload.sandbox ?? false }); configureCursorSandbox({ sandbox: payload.sandbox ?? false, isPublicRepo: repo.isPublic });
const loggedModelCallIds = /* @__PURE__ */ new Set(); const loggedModelCallIds = /* @__PURE__ */ new Set();
const messageHandlers5 = { const messageHandlers5 = {
system: (_event) => { system: (_event) => {
@@ -105552,24 +105555,30 @@ function configureCursorMcpServers({ mcpServers }) {
writeFileSync(mcpConfigPath, JSON.stringify({ mcpServers: cursorMcpServers }, null, 2), "utf-8"); writeFileSync(mcpConfigPath, JSON.stringify({ mcpServers: cursorMcpServers }, null, 2), "utf-8");
log.info(`\xBB MCP config written to ${mcpConfigPath}`); log.info(`\xBB MCP config written to ${mcpConfigPath}`);
} }
function configureCursorSandbox({ sandbox }) { function configureCursorSandbox({
sandbox,
isPublicRepo
}) {
const realHome = homedir2(); const realHome = homedir2();
const cursorConfigDir = join7(realHome, ".config", "cursor"); const cursorConfigDir = join7(realHome, ".config", "cursor");
const cliConfigPath = join7(cursorConfigDir, "cli-config.json"); const cliConfigPath = join7(cursorConfigDir, "cli-config.json");
mkdirSync4(cursorConfigDir, { recursive: true }); mkdirSync4(cursorConfigDir, { recursive: true });
const denyShell = isPublicRepo ? ["Shell(*)"] : [];
const config4 = sandbox ? { const config4 = sandbox ? {
permissions: { permissions: {
allow: ["Read(**)"], allow: ["Read(**)"],
deny: ["Write(**)", "Shell(*)"] deny: ["Write(**)", ...denyShell]
} }
} : { } : {
permissions: { permissions: {
allow: ["Read(**)", "Write(**)"], allow: ["Read(**)", "Write(**)"],
deny: ["Shell(*)"] deny: denyShell
} }
}; };
writeFileSync(cliConfigPath, JSON.stringify(config4, null, 2), "utf-8"); writeFileSync(cliConfigPath, JSON.stringify(config4, null, 2), "utf-8");
log.info(`\xBB CLI config written to ${cliConfigPath} (sandbox: ${sandbox})`); log.info(
`\xBB CLI config written to ${cliConfigPath} (sandbox: ${sandbox}, isPublicRepo: ${isPublicRepo})`
);
} }
// agents/gemini.ts // agents/gemini.ts
@@ -105737,21 +105746,29 @@ var gemini = agent({
}); });
}, },
run: async ({ payload, apiKey, mcpServers, cliPath, repo }) => { run: async ({ payload, apiKey, mcpServers, cliPath, repo }) => {
configureGeminiMcpServers({ mcpServers, cliPath }); configureGeminiMcpServers({ mcpServers, isPublicRepo: repo.isPublic });
if (!apiKey) { if (!apiKey) {
throw new Error("google_api_key or gemini_api_key is required for gemini agent"); throw new Error("google_api_key or gemini_api_key is required for gemini agent");
} }
const sessionPrompt = addInstructions({ payload, repo, useNativeBash: true }); const sessionPrompt = addInstructions({ payload, repo });
log.group("Full prompt", () => log.info(sessionPrompt)); log.group("Full prompt", () => log.info(sessionPrompt));
const args3 = payload.sandbox ? [ let args3;
"--allowed-tools", if (payload.sandbox) {
"read_file,list_directory,search_file_content,glob,save_memory,write_todos", args3 = [
"--allowed-mcp-server-names", "--allowed-tools",
"gh_pullfrog", "read_file,list_directory,search_file_content,glob,save_memory,write_todos",
"--output-format=stream-json", "--allowed-mcp-server-names",
"-p", "gh_pullfrog",
sessionPrompt "--output-format=stream-json",
] : ["--yolo", "--output-format=stream-json", "-p", sessionPrompt]; "-p",
sessionPrompt
];
} else {
args3 = ["--yolo", "--output-format=stream-json", "-p", sessionPrompt];
if (repo.isPublic) {
log.info("\u{1F512} public repo: native shell disabled via excludeTools, using MCP bash");
}
}
if (payload.sandbox) { if (payload.sandbox) {
log.info("\u{1F512} sandbox mode enabled: restricting to read-only operations"); log.info("\u{1F512} sandbox mode enabled: restricting to read-only operations");
} }
@@ -105818,7 +105835,7 @@ var gemini = agent({
} }
} }
}); });
function configureGeminiMcpServers({ mcpServers }) { function configureGeminiMcpServers({ mcpServers, isPublicRepo }) {
const realHome = homedir3(); const realHome = homedir3();
const geminiConfigDir = join8(realHome, ".gemini"); const geminiConfigDir = join8(realHome, ".gemini");
const settingsPath = join8(geminiConfigDir, "settings.json"); const settingsPath = join8(geminiConfigDir, "settings.json");
@@ -105847,6 +105864,9 @@ function configureGeminiMcpServers({ mcpServers }) {
...existingSettings, ...existingSettings,
mcpServers: geminiMcpServers mcpServers: geminiMcpServers
}; };
if (isPublicRepo) {
newSettings.excludeTools = ["run_shell_command"];
}
writeFileSync2(settingsPath, JSON.stringify(newSettings, null, 2), "utf-8"); writeFileSync2(settingsPath, JSON.stringify(newSettings, null, 2), "utf-8");
log.info(`\xBB MCP config written to ${settingsPath}`); log.info(`\xBB MCP config written to ${settingsPath}`);
} }
@@ -105868,7 +105888,11 @@ var opencode = agent({
const tempHome = process.env.PULLFROG_TEMP_DIR; const tempHome = process.env.PULLFROG_TEMP_DIR;
const configDir = join9(tempHome, ".config", "opencode"); const configDir = join9(tempHome, ".config", "opencode");
mkdirSync6(configDir, { recursive: true }); mkdirSync6(configDir, { recursive: true });
configureOpenCode({ mcpServers, sandbox: payload.sandbox ?? false }); configureOpenCode({
mcpServers,
sandbox: payload.sandbox ?? false,
isPublicRepo: repo.isPublic
});
const prompt = addInstructions({ payload, repo }); const prompt = addInstructions({ payload, repo });
log.group("Full prompt", () => log.info(prompt)); log.group("Full prompt", () => log.info(prompt));
const args3 = ["run", prompt, "--format", "json"]; const args3 = ["run", prompt, "--format", "json"];
@@ -105985,7 +106009,7 @@ var opencode = agent({
}; };
} }
}); });
function configureOpenCode({ mcpServers, sandbox }) { function configureOpenCode({ mcpServers, sandbox, isPublicRepo }) {
const tempHome = process.env.PULLFROG_TEMP_DIR; const tempHome = process.env.PULLFROG_TEMP_DIR;
const configDir = join9(tempHome, ".config", "opencode"); const configDir = join9(tempHome, ".config", "opencode");
mkdirSync6(configDir, { recursive: true }); mkdirSync6(configDir, { recursive: true });
@@ -106005,7 +106029,20 @@ function configureOpenCode({ mcpServers, sandbox }) {
url: serverConfig.url url: serverConfig.url
}; };
} }
const permission = sandbox ? { edit: "deny", bash: "deny", webfetch: "deny", doom_loop: "allow", external_directory: "allow" } : { edit: "allow", bash: "deny", webfetch: "allow", doom_loop: "allow", external_directory: "allow" }; const bashPermission = isPublicRepo ? "deny" : "allow";
const permission = sandbox ? {
edit: "deny",
bash: "deny",
webfetch: "deny",
doom_loop: "allow",
external_directory: "allow"
} : {
edit: "allow",
bash: bashPermission,
webfetch: "allow",
doom_loop: "allow",
external_directory: "allow"
};
const config4 = { const config4 = {
mcp: opencodeMcpServers, mcp: opencodeMcpServers,
permission permission
@@ -137976,38 +138013,24 @@ var BashParams = type({
"timeout?": "number", "timeout?": "number",
"working_directory?": "string" "working_directory?": "string"
}); });
var SENSITIVE_PATTERNS = [ var SENSITIVE_PATTERNS = [/_KEY$/i, /_SECRET$/i, /_TOKEN$/i, /_PASSWORD$/i, /_CREDENTIAL$/i];
/_KEY$/i,
/_SECRET$/i,
/_TOKEN$/i,
/PASSWORD/i,
/CREDENTIAL/i,
/AUTH/i,
/^ANTHROPIC/i,
/^OPENAI/i,
/^GEMINI/i,
/^GOOGLE/i,
/^CLAUDE/i,
/^CODEX/i,
/^CURSOR/i,
/^AWS/i,
/^GITHUB_TOKEN$/i,
/^GH_TOKEN$/i
];
function isSensitive(key) { function isSensitive(key) {
return SENSITIVE_PATTERNS.some((p) => p.test(key)); return SENSITIVE_PATTERNS.some((p) => p.test(key));
} }
function filterEnv() { function filterEnv(isPublicRepo) {
const filtered = {}; const filtered = {};
for (const [key, value2] of Object.entries(process.env)) { for (const [key, value2] of Object.entries(process.env)) {
if (value2 !== void 0 && !isSensitive(key)) filtered[key] = value2; if (value2 === void 0) continue;
if (isPublicRepo && isSensitive(key)) continue;
filtered[key] = value2;
} }
return filtered; return filtered;
} }
function spawnSandboxed(command, options) { function spawnSandboxed(command, options) {
const stdio = ["ignore", "pipe", "pipe"]; const stdio = ["ignore", "pipe", "pipe"];
const spawnOpts = { env: options.env, cwd: options.cwd, stdio, detached: true }; const spawnOpts = { env: options.env, cwd: options.cwd, stdio, detached: true };
return process.env.CI === "true" ? spawn5("unshare", ["--pid", "--fork", "--mount-proc", "bash", "-c", command], spawnOpts) : spawn5("bash", ["-c", command], spawnOpts); const useNamespaceIsolation = process.env.CI === "true" && options.isPublicRepo;
return useNamespaceIsolation ? spawn5("unshare", ["--pid", "--fork", "--mount-proc", "bash", "-c", command], spawnOpts) : spawn5("bash", ["-c", command], spawnOpts);
} }
async function killProcessGroup(proc) { async function killProcessGroup(proc) {
if (!proc.pid) return; if (!proc.pid) return;
@@ -138022,21 +138045,27 @@ async function killProcessGroup(proc) {
} }
} }
} }
function BashTool(_ctx) { function BashTool(ctx) {
const isPublicRepo = !ctx.repo.private;
return tool({ return tool({
name: "bash", name: "bash",
description: `Execute shell commands securely. Environment is filtered to remove API keys and secrets. description: `Execute shell commands securely.${isPublicRepo ? " Environment is filtered to remove API keys and secrets." : ""}
Use this tool to: Use this tool to:
- Run shell commands (ls, cat, grep, find, etc.) - Run shell commands (ls, cat, grep, find, etc.)
- Execute build tools (npm, pnpm, cargo, make, etc.) - Execute build tools (npm, pnpm, cargo, make, etc.)
- Run tests and linters - Run tests and linters
- Perform git operations`, - Perform git operations
${isPublicRepo ? "\nThe command runs in a bash shell with a filtered environment that excludes sensitive variables like API keys and tokens." : ""}`,
parameters: BashParams, parameters: BashParams,
execute: execute(async (params) => { execute: execute(async (params) => {
const timeout = Math.min(params.timeout ?? 12e4, 6e5); const timeout = Math.min(params.timeout ?? 12e4, 6e5);
const cwd2 = params.working_directory ?? process.cwd(); const cwd2 = params.working_directory ?? process.cwd();
const proc = spawnSandboxed(params.command, { env: filterEnv(), cwd: cwd2 }); const proc = spawnSandboxed(params.command, {
env: filterEnv(isPublicRepo),
cwd: cwd2,
isPublicRepo
});
let stdout = "", stderr = "", timedOut = false, exited = false; let stdout = "", stderr = "", timedOut = false, exited = false;
proc.stdout?.on("data", (chunk) => { proc.stdout?.on("data", (chunk) => {
stdout += chunk.toString(); stdout += chunk.toString();
@@ -138609,7 +138638,8 @@ ${encode(eventWithoutContext)}`;
repo: { repo: {
owner: ctx.owner, owner: ctx.owner,
name: ctx.name, name: ctx.name,
defaultBranch: ctx.repo.default_branch defaultBranch: ctx.repo.default_branch,
isPublic: !ctx.repo.private
} }
}); });
} }
+2 -1
View File
@@ -2,7 +2,7 @@ import { mkdtemp } from "node:fs/promises";
import { tmpdir } from "node:os"; import { tmpdir } from "node:os";
import { join } from "node:path"; import { join } from "node:path";
import { flatMorph } from "@ark/util"; import { flatMorph } from "@ark/util";
import { Octokit } from "@octokit/rest"; import type { Octokit } from "@octokit/rest";
import { encode as toonEncode } from "@toon-format/toon"; import { encode as toonEncode } from "@toon-format/toon";
import { type } from "arktype"; import { type } from "arktype";
import { type Agent, agents } from "./agents/index.ts"; import { type Agent, agents } from "./agents/index.ts";
@@ -513,6 +513,7 @@ async function runAgent(ctx: AgentContext): Promise<AgentResult> {
owner: ctx.owner, owner: ctx.owner,
name: ctx.name, name: ctx.name,
defaultBranch: ctx.repo.default_branch, defaultBranch: ctx.repo.default_branch,
isPublic: !ctx.repo.private,
}, },
}); });
} }
+23 -28
View File
@@ -11,49 +11,37 @@ export const BashParams = type({
}); });
// patterns for sensitive env vars: suffixes (_KEY, _SECRET, _TOKEN) plus AI provider prefixes // patterns for sensitive env vars: suffixes (_KEY, _SECRET, _TOKEN) plus AI provider prefixes
const SENSITIVE_PATTERNS = [ const SENSITIVE_PATTERNS = [/_KEY$/i, /_SECRET$/i, /_TOKEN$/i, /_PASSWORD$/i, /_CREDENTIAL$/i];
/_KEY$/i,
/_SECRET$/i,
/_TOKEN$/i,
/PASSWORD/i,
/CREDENTIAL/i,
/AUTH/i,
/^ANTHROPIC/i,
/^OPENAI/i,
/^GEMINI/i,
/^GOOGLE/i,
/^CLAUDE/i,
/^CODEX/i,
/^CURSOR/i,
/^AWS/i,
/^GITHUB_TOKEN$/i,
/^GH_TOKEN$/i,
];
function isSensitive(key: string): boolean { function isSensitive(key: string): boolean {
return SENSITIVE_PATTERNS.some((p) => p.test(key)); return SENSITIVE_PATTERNS.some((p) => p.test(key));
} }
/** filter env vars, removing sensitive values */ /** filter env vars, removing sensitive values (only for public repos) */
function filterEnv(): Record<string, string> { function filterEnv(isPublicRepo: boolean): Record<string, string> {
const filtered: Record<string, string> = {}; const filtered: Record<string, string> = {};
for (const [key, value] of Object.entries(process.env)) { for (const [key, value] of Object.entries(process.env)) {
if (value !== undefined && !isSensitive(key)) filtered[key] = value; if (value === undefined) continue;
// only filter sensitive vars for public repos
if (isPublicRepo && isSensitive(key)) continue;
filtered[key] = value;
} }
return filtered; return filtered;
} }
/** /**
* spawn command with filtered env. in CI, also use PID namespace isolation * spawn command with filtered env. in CI, also use PID namespace isolation
* to prevent child from reading /proc/$PPID/environ * to prevent child from reading /proc/$PPID/environ (only for public repos)
*/ */
function spawnSandboxed( function spawnSandboxed(
command: string, command: string,
options: { env: Record<string, string>; cwd: string } options: { env: Record<string, string>; cwd: string; isPublicRepo: boolean }
): ChildProcess { ): ChildProcess {
const stdio: ["ignore", "pipe", "pipe"] = ["ignore", "pipe", "pipe"]; const stdio: ["ignore", "pipe", "pipe"] = ["ignore", "pipe", "pipe"];
const spawnOpts = { env: options.env, cwd: options.cwd, stdio, detached: true }; const spawnOpts = { env: options.env, cwd: options.cwd, stdio, detached: true };
return process.env.CI === "true" // requires linux runner or this will fail. // only use PID namespace isolation for public repos in CI
const useNamespaceIsolation = process.env.CI === "true" && options.isPublicRepo;
return useNamespaceIsolation
? spawn("unshare", ["--pid", "--fork", "--mount-proc", "bash", "-c", command], spawnOpts) ? spawn("unshare", ["--pid", "--fork", "--mount-proc", "bash", "-c", command], spawnOpts)
: spawn("bash", ["-c", command], spawnOpts); : spawn("bash", ["-c", command], spawnOpts);
} }
@@ -74,21 +62,28 @@ async function killProcessGroup(proc: ChildProcess): Promise<void> {
} }
} }
export function BashTool(_ctx: ToolContext) { export function BashTool(ctx: ToolContext) {
const isPublicRepo = !ctx.repo.private;
return tool({ return tool({
name: "bash", name: "bash",
description: `Execute shell commands securely. Environment is filtered to remove API keys and secrets. description: `Execute shell commands securely.${isPublicRepo ? " Environment is filtered to remove API keys and secrets." : ""}
Use this tool to: Use this tool to:
- Run shell commands (ls, cat, grep, find, etc.) - Run shell commands (ls, cat, grep, find, etc.)
- Execute build tools (npm, pnpm, cargo, make, etc.) - Execute build tools (npm, pnpm, cargo, make, etc.)
- Run tests and linters - Run tests and linters
- Perform git operations`, - Perform git operations
${isPublicRepo ? "\nThe command runs in a bash shell with a filtered environment that excludes sensitive variables like API keys and tokens." : ""}`,
parameters: BashParams, parameters: BashParams,
execute: execute(async (params) => { execute: execute(async (params) => {
const timeout = Math.min(params.timeout ?? 120000, 600000); const timeout = Math.min(params.timeout ?? 120000, 600000);
const cwd = params.working_directory ?? process.cwd(); const cwd = params.working_directory ?? process.cwd();
const proc = spawnSandboxed(params.command, { env: filterEnv(), cwd }); const proc = spawnSandboxed(params.command, {
env: filterEnv(isPublicRepo),
cwd,
isPublicRepo,
});
let stdout = "", let stdout = "",
stderr = "", stderr = "",