@lousy-agents/agent-shell

A flight recorder for AI agent tool usage and npm script execution.

agent-shell records structured JSONL telemetry for both npm script execution (via script-shell shim) and AI agent tool usage (via Copilot lifecycle hooks). It provides an audit trail that doesn't depend on agent self-reports.

Quick Start

# Install globally
npm install -g @lousy-agents/agent-shell

# Initialize hooks and policy (interactive)
agent-shell init

# Add event storage to .gitignore
echo '.agent-shell/' >> .gitignore

# Verify it's working
agent-shell --version

agent-shell init detects your project, prompts for which features to enable (flight recording, policy blocking), and writes the Copilot hook configuration and policy files.

Note: agent-shell init sets up Copilot hooks and policy files, but npm script execution recording is configured separately. To record npm script runs and emit script_end events, set npm's script-shell to agent-shell (for example by adding script-shell=agent-shell to your project's .npmrc).

Why global? agent-shell is configured as npm's script-shell and as a Copilot hook command, so it must be available before npm ci or npm install runs. A local dev dependency creates a circular dependency. Installing globally keeps it on PATH independent of node_modules.

How It Works

When npm runs a script (e.g., npm test), it invokes the configured script-shell instead of /bin/sh. agent-shell wraps the real shell:

1. Captures context (actor, environment, timestamps)
2. Spawns /bin/sh -c <command> — your script runs normally
3. Records a JSONL event with execution metadata
4. Propagates the exit code unchanged

Scripts behave identically — same output, same exit codes, same signals. agent-shell adds observability without changing behavior.

Setup with agent-shell init

The init subcommand is the recommended way to configure agent-shell for your project. It detects existing configuration and guides you through enabling features.

Interactive Mode (TTY)

agent-shell init

When run in a terminal, init prompts for each feature:

• Flight recording — Records all agent tool usage via postToolUse hook
• Policy blocking — Evaluates commands against allow/deny rules via preToolUse hook

If hooks.json already exists, only missing features are prompted.

Non-Interactive Mode (CI / Flags)

# Enable specific features
agent-shell init --flight-recorder
agent-shell init --policy
agent-shell init --flight-recorder --policy

# Disable specific features
agent-shell init --no-flight-recorder
agent-shell init --no-policy

In a non-TTY environment with no explicit flags, all missing features are auto-enabled with a message to stderr.

What It Generates

Flight Recording with agent-shell record

The record subcommand is a postToolUse hook handler that records every tool invocation as a tool_use telemetry event. It captures tool usage that happens outside npm scripts — file edits, code searches, API calls, and more.

Copilot Hook Configuration

When both flight recording and policy are enabled, hooks.json looks like this:

{
  "version": 1,
  "hooks": {
    "preToolUse": [
      {
        "type": "command",
        "bash": "agent-shell policy-check",
        "timeoutSec": 30
      }
    ],
    "postToolUse": [
      {
        "type": "command",
        "bash": "agent-shell record",
        "timeoutSec": 30
      }
    ]
  }
}

The record handler reads the JSON payload from Copilot's postToolUse hook via stdin. For terminal tools (bash, zsh, ash, sh), it extracts the command string from toolArgs. For non-terminal tools, the command is recorded as an empty string.

Telemetry Schema (v1)

Each event produces one JSON line. There are four event types:

script_end Event

Recorded when an npm script completes:

{
  "v": 1,
  "session_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "event": "script_end",
  "script": "test",
  "command": "vitest run",
  "package": "my-app",
  "package_version": "1.2.0",
  "actor": "claude-code",
  "exit_code": 1,
  "signal": null,
  "duration_ms": 3420,
  "timestamp": "2026-03-08T14:32:01.000Z",
  "env": {
    "NODE_ENV": "test",
    "CI": "true"
  },
  "tags": {
    "pr": "1234"
  }
}

tool_use Event

Recorded by the postToolUse hook when an agent uses any tool:

{
  "v": 1,
  "session_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "event": "tool_use",
  "tool_name": "bash",
  "command": "npm test",
  "actor": "copilot",
  "timestamp": "2026-03-08T14:32:01.000Z",
  "env": {
    "CI": "true"
  },
  "tags": {}
}

For non-terminal tools (file edits, searches, etc.), command is an empty string and tool_name identifies the tool (e.g., file_edit, curl).

policy_decision Event

Recorded when the preToolUse policy-check hook evaluates a command:

{
  "v": 1,
  "session_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "event": "policy_decision",
  "decision": "deny",
  "matched_rule": "rm -rf *",
  "command": "rm -rf /",
  "actor": "copilot",
  "timestamp": "2026-03-08T14:32:01.000Z",
  "env": {
    "CI": "true"
  },
  "tags": {}
}

shim_error Event

Recorded when agent-shell encounters an internal error (e.g., failed to parse npm context):

{
  "v": 1,
  "session_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "event": "shim_error",
  "command": "",
  "actor": "copilot",
  "timestamp": "2026-03-08T14:32:01.000Z",
  "env": {},
  "tags": {}
}

Field Reference

Fields script, package, and package_version are only present when running within an npm context.

Actor Detection

agent-shell classifies who initiated each script execution:

Querying Events

Use the log subcommand to query execution history. Both script_end and tool_use events are included in results:

# Show events from the most recent session
agent-shell log

# Show events from the last 2 hours
agent-shell log --last 2h

# Show only failures (script_end events with non-zero exit)
agent-shell log --failures

# Filter by actor
agent-shell log --actor claude-code

# Filter by script name (script_end events only)
agent-shell log --script test

# Combine filters
agent-shell log --actor claude-code --failures --last 2h

# Output as JSON (for scripting)
agent-shell log --json

# List all sessions
agent-shell log --list-sessions

Duration Formats

Policy-Based Command Blocking

agent-shell includes a policy-check subcommand that evaluates shell commands against allow/deny rules before execution. This enables repository maintainers to define which commands AI agents are permitted to run via a pre-tool-use hook.

How It Works

The policy-check command is designed to be used as a GitHub Copilot pre-tool-use hook. It reads a JSON request from stdin containing the tool name and arguments, evaluates the command against a policy file, and writes a permission decision to stdout.

Only terminal tools (bash, zsh, ash, sh) have their commands evaluated against policy rules when the policy file loads successfully; non-terminal tools are then allowed without policy evaluation. If the default policy file is missing, all tools are allowed, but if the policy file is present and malformed or invalid, all tools are denied.

Policy File

By default, the policy file is located at .github/hooks/agent-shell/policy.json relative to the repository root. Override the location with the AGENTSHELL_POLICY_PATH environment variable.

{
  "allow": ["npm test", "npm run lint*", "git status"],
  "deny": ["npm publish", "rm -rf *"]
}

Evaluation order:

1. If the command matches any deny pattern → deny
2. If an allow list exists and the command does not match any pattern → deny
3. Otherwise → allow

Patterns support * wildcards for prefix, suffix, and infix matching (e.g., npm run * matches npm run test). When using the default policy path, a missing policy file results in all commands being allowed. When AGENTSHELL_POLICY_PATH is set and the referenced policy file is missing, malformed, or cannot be loaded, commands are denied (fail-closed).

Copilot Hook Configuration

Add the following to .github/hooks/agent-shell/hooks.json, or use agent-shell init to generate it automatically:

{
  "version": 1,
  "hooks": {
    "preToolUse": [
      {
        "type": "command",
        "timeoutSec": 30,
        "bash": "agent-shell policy-check"
      }
    ],
    "postToolUse": [
      {
        "type": "command",
        "timeoutSec": 30,
        "bash": "agent-shell record"
      }
    ]
  }
}

Input/Output Format

Input (JSON via stdin):

{
  "toolName": "bash",
  "toolArgs": "{\"command\": \"npm test\"}"
}

Output (JSON to stdout):

{"permissionDecision": "allow"}

Or when denied:

{
  "permissionDecision": "deny",
  "permissionDecisionReason": "Command 'npm publish' denied by policy rule: npm publish"
}

Policy decisions are recorded as telemetry events alongside regular script execution events.

Policy Initialization (policy --init)

Manually writing a policy.json allow list requires knowing which commands your project actually uses. policy --init eliminates that blank-page problem: it scans your project and generates a ready-to-commit policy and hook configuration.

agent-shell policy --init

What It Generates

If either file already exists, the command skips writing it and notifies you. Both files are safe to commit for team-wide enforcement.

How It Discovers Commands

The scanner inspects three sources:

1. package.json scripts — Each npm run <script> entry becomes an exact allow rule
2. GitHub Actions workflow files — run: commands in .github/workflows/*.yml are extracted and deduplicated
3. mise.toml tasks — Each task becomes an exact mise run <task> allow rule (no wildcards, for security)

Language detection (Node.js, Python, Go, Rust, Ruby, Java) is also performed and reported in the scan summary. Language-specific build tool patterns may be added in a future version.

AI-Enhanced Analysis

When the @github/copilot-sdk package is available and authenticated, policy --init runs a second pass using the GitHub Copilot AI to refine the generated allow list. The AI session connects to @lousy-agents/mcp as a local MCP server for structured project discovery, and exposes two additional sandboxed tools for ad-hoc exploration:

The MCP server also provides discover_feedback_loops and discover_environment tools, giving the model structured access to SDLC commands and runtime/toolchain details without raw filesystem browsing.

Fallback behavior:

The final policy file is always generated. AI enhancement is best-effort and never blocks output.

Authentication Setup

The @github/copilot-sdk authenticates using the GitHub Copilot CLI. You need both the GitHub CLI and the Copilot CLI extension installed and signed in.

1. Install the GitHub CLI

Follow the official installation guide for your platform, then authenticate:

gh auth login

2. Install the GitHub Copilot CLI extension

gh extension install github/gh-copilot

3. Authenticate with Copilot

gh copilot auth

4. Install the SDK

npm install @github/copilot-sdk

Once authenticated, policy --init will automatically detect the SDK and run AI-enhanced analysis. To debug SDK interactions, set AGENT_SHELL_COPILOT_DEBUG=1.

Prerequisites

• Must be run inside a git repository (git rev-parse --show-toplevel is used to locate the repo root)
• agent-shell must be globally installed (required for the generated hook script to be on PATH)
• @github/copilot-sdk is optional — see Authentication Setup above to enable AI-enhanced analysis

Example Output

Scanning project...
Discovered: 8 npm script(s), 12 workflow command(s), 3 mise task(s), 1 language(s)
Created .github/hooks/agent-shell/policy.json
Created .github/hooks/agent-shell/hooks.json

--- Proposed Policy ---
{
  "allow": [
    "git branch --list *",
    "git branch --show-current",
    "git diff *",
    "git log *",
    "git rev-parse *",
    "git show *",
    "git status *",
    "npm run build",
    "npm run lint",
    "npm test",
    "pwd"
  ],
  "deny": [
    "rm -rf *",
    "sudo *"
  ]
}

--- Hook Configuration ---
{
  "version": 1,
  "hooks": {
    "preToolUse": [
      {
        "type": "command",
        "bash": "agent-shell policy-check",
        "timeoutSec": 30
      }
    ]
  }
}

Environment Variables

Custom Tags

Set AGENTSHELL_TAG_* environment variables to attach metadata:

AGENTSHELL_TAG_pr=1234 AGENTSHELL_TAG_task=fix-auth npm test

Produces: "tags": { "pr": "1234", "task": "fix-auth" }

Session Correlation

By default, each script invocation gets a fresh session ID. To correlate events across a workflow:

export AGENTSHELL_SESSION_ID=$(uuidgen)
npm run lint && npm test && npm run build
# All three events share the same session_id

Scope

• npm only — Configured via .npmrc script-shell. Other package managers (yarn, pnpm, bun) are not supported.
• POSIX only — macOS and Linux. Windows requires WSL or Git Bash.

Troubleshooting

npm fails with "script-shell not found"

Your .npmrc references agent-shell but it's not installed globally:

# Fix: install the package globally
npm install -g @lousy-agents/agent-shell

# Or: bypass temporarily
AGENTSHELL_PASSTHROUGH=1 npm test

Bypass all instrumentation

Set the passthrough escape hatch — zero overhead, zero recording:

AGENTSHELL_PASSTHROUGH=1 npm test

Remove instrumentation entirely

Remove the script-shell line from .npmrc:

# Edit .npmrc and delete this line:
# script-shell=agent-shell

# Or remove it with sed:
sed -i '' '/script-shell/d' .npmrc

Then optionally uninstall:

npm uninstall -g @lousy-agents/agent-shell

Verify the shim is active

npx agent-shell --version

Clean up recorded events

rm -rf .agent-shell/

.gitignore

Add .agent-shell/ to your .gitignore:

# agent-shell telemetry
.agent-shell/