Hooks

Run your own commands when the assistant uses a tool or finishes: block risky actions, lint edits, run tests.

Hooks are your own commands, run at fixed points in the assistant's work:

before it uses a tool, to block an action, such as committing a secret;
after it uses one, to check the result, such as linting every edit;
when it says the task is done, to send it back to work, for example while tests still fail.

They follow Claude Code's format, so hooks a team already has carry over.

Where to put them

Put them in .lamied/hooks.json at the root of the workspace. Without that file, Lamied uses the hooks in .claude/settings.json.

.lamied/hooks.json

{
  "hooks": {
    "PreToolUse": [
      { "matcher": "Bash", "hooks": [{ "type": "command", "command": "./scripts/no-secrets.sh" }] }
    ],
    "PostToolUse": [
      { "matcher": "Edit|Write", "hooks": [{ "type": "command", "command": "npx eslint --quiet \"$(jq -r .tool_input.file_path)\"", "timeout": 30 }] }
    ],
    "Stop": [
      { "hooks": [{ "type": "command", "command": "pnpm test 1>&2 || exit 2", "timeout": 300 }] }
    ]
  }
}

When they run

Event	Runs	It can
`PreToolUse`	Before a tool call the matcher names	Block the call. The assistant is told why.
`PostToolUse`	After such a call succeeds	Tell the assistant what it found, such as lint errors.
`Stop`	When the assistant has checked that the task is done	Send it back to work with a reason.

A matcher names the tools a hook is for. Several can be joined with |, and .* matches any text. An empty matcher, or *, means every tool. Lamied's tool names work, and so do Claude Code's:

Claude Code	Lamied
`Bash`	`run_terminal_command`
`Read`	`read_file`
`Write`	`write_file`
`Edit`	`edit_file`
`Glob`	`find_files`
`Grep`	`search_files`
`LS`	`list_directory`

Other tools go by their Lamied names, such as delete_file and move_file. Tools from MCP servers are named mcp__server__tool.

What a hook is given

A hook runs in the workspace folder. The variables LAMIED_PROJECT_DIR and CLAUDE_PROJECT_DIR hold its path. The event comes as JSON on standard input:

Standard input, PreToolUse

{
  "session_id": "…",
  "cwd": "/home/me/shop",
  "hook_event_name": "PreToolUse",
  "tool_name": "edit_file",
  "tool_input": { "path": "src/cart.ts", "file_path": "/home/me/shop/src/cart.ts", "old_string": "…", "new_string": "…" },
  "tool_use_id": "…"
}

A PostToolUse hook also gets tool_response, the tool's result. A Stop hook gets stop_hook_active, which is true when a Stop hook has already sent this task back to work.

How a hook answers

Exit code 0: carry on.
Exit code 2: object. What the hook printed to standard error goes to the assistant. A PreToolUse objection blocks the call; a PostToolUse one reports a problem; a Stop one sends the task back to work.
Any other exit code, or running past its timeout (60 seconds unless set; at most 600): a PreToolUse hook blocks the call, because a check that can't decide mustn't let the action through. Other hooks are ignored.

With exit code 0, a hook can also answer in JSON on standard output, as in Claude Code:

{"hookSpecificOutput": {"permissionDecision": "deny", "permissionDecisionReason": "…"}} blocks a PreToolUse call;
{"decision": "block", "reason": "…"} objects, like exit code 2;
{"hookSpecificOutput": {"additionalContext": "…"}} tells the assistant something without objecting.

You approve them first

Hooks run commands on your computer, and a repository you open may not be yours. So the first time a project's hooks would run, the chat shows their commands and asks Run this project's hooks?

An approval is remembered for that workspace, until the commands change. Then you're asked again.
If you decline, the hooks don't run, and you aren't asked again until Lamied restarts. Allowing commands in command permissions doesn't approve hooks.
A Stop hook can send a task back to work three times. After that, the task ends as incomplete, with the hook's last reason.

Differences from Claude Code

Lamied runs PreToolUse, PostToolUse and Stop hooks. Other events, such as UserPromptSubmit and SessionStart, are skipped.
A PreToolUse hook that fails or times out blocks the call. In Claude Code, the call goes ahead.
Help agents (research and worktree agents) run the same PreToolUse and PostToolUse hooks. Stop hooks run for the main task.