README.md

Hooks

[!NOTE] This document was designed for both humans and agents.

Hooks are user-defined shell scripts that run when various events happen during the agent lifecycle, allowing you to both build on top of Crush, customize its behavior, and exert deterministic control over an agent's wily behavior.

Hooks are just shell commands, and were designed to be both simple and future forward.

Hot Hook Facts

  • Hooks just shell commands
  • Hooks can be written in any language because they’re just executables: Bash, Python, Node, Rust, Haskell, whatever
  • Hooks are Claude Code-compatible
  • Crush ships with a builtin crush-hook skill write, edit, and configure hooks; just tell Crush how to configure Crush
  • Crush currently supports just one hook, PreToolUse, with plans to support the full gamut; please let us know which hooks you'd like to see next
  • Hooks run in parallel for speed, but their results compose in config order for determinism

Some things you can do with hooks:

  • Block "dangerous" commands: no more git push -f or cabal init
  • Rewrite tool input: turn node calls info deno, scrub secrets from commands, rewrite all mentions of "Haskell" into "Haskell, The Best Language", and so on
  • Inject context: add notes to the model's context whenever certain tools are called. For example: "remember to run gofumpt after editing Go files"
  • Auto-approve tools: skip the permission prompt for bash commands that you know are safe
  • Log certain tool calls

…And lots more. Show us what you're building!

Baby's First Hook

Let's just dive into it and make a simple hook. This particular hook will disallow the use of Haskell (but we love you, Simon Peyton Jones).

Config

The first thing we need to do is hook up our hook. Let's add the following to our local crush.json. You can, of course, do this globally, too.

{
  // As expected, hooks go in a "hooks" object.
  "hooks": {
    // PreToolUse is an event that fires before a tool is used.
    "PreToolUse": [
      {
        // What tool do we want to hook into? In this case, Bash, because it
        // runs the stuff we wanna block.
        "matcher": "^bash$",

        // The path to our actual hook script.
        "command": "./no-haskell.sh",
      },
    ],
  },
}

Now, let's make our no-haskell.sh hook script.

#!/usr/bin/env bash

# Disallow ghc, cabal, and stack. Pipe the bash command output
# ($CRUSH_TOOL_INPUT_COMMAND) to grep and match on a regexp.
if echo "$CRUSH_TOOL_INPUT_COMMAND" | grep -qE '(^| )((ghc|cabal|stack)(\.exe)?)( |$)'; then

  # Someone is trying to use Haskell. Let's send a message back to the model
  # and user explaining why we're blocking this. Note that we send all feedback
  # like this to stderr.
  echo "No Haskell allowed, kiddo." >&2

  # Now, block the tool call by exiting with code 2.
  exit 2
fi

That's basically it. For the full guide on how hooks work, however, read on.


Configuration

Hooks can be added to your crush.json (or .crush.json) at both the global and project-level, with project level hooks taking precedence.

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "bash", // regex tested against the tool name
        "command": "./hooks/my-hot-hook.sh", // the path to the hook
        "timeout": 10, // in seconds; default 30
      },
    ],
  },
}

Remember, hooks will run in parallel but resolve in config order. Last hook wins when rewriting input, but first deny wins when blocking.

Events

Here are the events you can hook into (spoiler: there's currently just one):

PreToolUse

This hook fires before every tool call. Use it to block dangerous commands, enforce policies, rewrite tool input, inject context the model should see, log stuff, and so on.

Matched against: the tool name (e.g. bash, edit, write, mcp_github_create_pull_request).

[!NOTE] Event names are case insensitive and snake-caseable, so PreToolUse, pretooluse, PRETOOLUSE, pre_tool_use, and PRE_TOOL_USE all work.

Scope: PreToolUse only fires on the top-level agent's tool calls. Sub-agents (the agent task tool, agentic_fetch, etc.) run without hook interception so a single delegated turn doesn't trigger your hook N times. The outer sub-agent tool call itself is hooked, so policy like "never let the agent spawn sub-agents" still works.

Hooks are keyed by event name. Only command is required, and you can omit matcher to match all tools.

Building Hooks

When a hook fires, Crush:

  1. Filters hooks whose matcher regex matches the tool name (no matcher = match all).
  2. Deduplicates by command (identical commands run once).
  3. Runs all matching hooks in parallel as subprocesses.
  4. Waits for all to finish (or time out), then aggregates results in config order: deny wins over allow, allow wins over none; updated_input patches shallow-merge in order.
  5. Applies the result before permission checks. If the aggregated decision is deny, the tool call is blocked and you never see a permission prompt for it. If it's allow, Crush treats that as affirmative pre-approval and also skips the prompt. Silence (no decision) falls through to the normal permission flow.

Note that you can omit matcher and match in your shell script instead, however you'll incur some additional overhead as Crush will exec each script.

Input

Each hook receives data two ways: environment variables and stdin (as JSON). Environment variables are typically easier to work with, with JSON being available when input is more complex.

Environment Variables

The available environment variables are:

Variable Description
CRUSH_EVENT The hook event name (e.g. PreToolUse).
CRUSH_TOOL_NAME The tool being called (e.g. bash).
CRUSH_SESSION_ID Current session ID.
CRUSH_CWD Working directory.
CRUSH_PROJECT_DIR Project root directory.
CRUSH_TOOL_INPUT_COMMAND For bash calls: the shell command being run.
CRUSH_TOOL_INPUT_FILE_PATH For file tools: the target file path.

JSON

Standard input provides the full context as JSON:

{
  "event": "PreToolUse", // Hook event name
  "session_id": "313909e", // Current session ID
  "cwd": "/home/user/project", // Working directory
  "tool_name": "bash", // The tool being called
  "tool_input": { "command": "rm -rf /" }, // The tool's input
}

Note that tool_input field contains the raw JSON the model sent to the tool.

To parse the stdin JSON in your hook script, read from stdin and use a tool like jq:

#!/usr/bin/env bash
read -r input
tool_name=$(echo "$input" | jq -r '.tool_name')
command=$(echo "$input" | jq -r '.tool_input.command // empty')

You can also use tools like Python:

#!/usr/bin/env python3
import json, sys

data = json.load(sys.stdin)
tool_name = data.get("tool_name", "")
command = data.get("tool_input", {}).get("command", "")

Output

Hooks communicate back to Crush via exit code and stdout/stderr. The simplest way to do this is to return an error code and print additional context to stderr. For example:

# Here, error code 2 blocks the tool, using stderr as the reason:
if some_bad_condition; then
  echo "Blocked: reason here" >&2
  exit 2
fi
Exit Code Meaning
0 Success. Stdout is parsed as JSON (see fields below).
2 Block the tool. Stderr is used as the deny reason (no JSON).
49 Halt the turn. Stderr is used as the halt reason (no JSON).
Other Non-blocking error. Logged and ignored β€” the tool call proceeds.

The difference between exit 2 and exit 49:

  • Exit 2 blocks the current tool call. The agent sees the error and can try something else.
  • Exit 49 halts the whole turn. The agent doesn't get to respond further; the user takes over. Use this when something is wrong enough that the agent shouldn't keep trying. 49 sits in an empty slice of the exit-code space β€” between the generic-error range (1-30), the BSD sysexits.h range (64-78), and the killed-by-signal range (128+) β€” so it can't be hit by accident.

That said, if you need more control, or if you need to rewrite input, you can use JSON on stdout. Exit 0 and print a JSON object to provide context, update the input, or still deny/halt with a reason:

{
  "version": 1, // Output envelope version. Optional; defaults to 1.
  "decision": "allow", // "allow", "deny", or null. Omit for no opinion.
  "halt": false, // If true, halts the turn entirely.
  "reason": "LGTM", // Shown when denying or halting.
  "context": "Scrubbed secrets", // String or array of strings. Appended to what the model sees.
  "updated_input": { "command": "…" }, // Shallow-merged into the tool's input before execution.
}

version is an optional integer at the top of the envelope. It defaults to 1 if omitted. Unknown higher versions are still parsed; the field exists so the envelope can evolve without a compatibility shim.

decision: "allow" is affirmative: it pre-approves the tool call and bypasses the permission prompt entirely. Silence (no decision, or decision: null) means "no opinion" β€” the tool still goes through the normal permission flow. Use "allow" when you want to auto-approve; omit it when you only want to inject context or rewrite input without also vouching for the call.

updated_input is a shallow-merge patch. Keys you include overwrite matching keys in tool_input; keys you don't include are preserved. If the model called bash with {"command": "npm test", "timeout": 60000} and your hook returns {"updated_input": {"command": "bun test"}}, the tool runs with {"command": "bun test", "timeout": 60000} β€” the timeout isn't dropped. The merge is shallow: nested objects are replaced wholesale, not deep-merged.

halt: true stops the turn entirely. The agent doesn't get to respond further; the user takes over. The exit-code shorthand is exit 49 with stderr as the reason.

context accepts either a string or an array of strings. Use the string form for a single observation; use the array form when a hook produces multiple distinct notes and you'd rather not concatenate them by hand. Empty strings and empty array entries are dropped.

Here's a full shell script that produces this JSON:

#!/usr/bin/env bash
# Example: rewrite a bash command using RTK

read -r input
original_cmd=$(echo "$input" | jq -r '.tool_input.command')
rewritten=$(secret-scrubber rewrite "$original_cmd")

cat <<EOF
{
  "decision": "allow",
  "context": "Scrubbed secrets",
  "updated_input": {"command": "$rewritten"}
}
EOF

Multiple Hooks

Hooks run in parallel, but their results compose in config order. Whichever hook finishes first doesn't get to "win" by virtue of timing; composition is deterministic based on the order hooks appear in crush.json.

When multiple hooks match the same tool call:

  • If any hook denies, the tool call is blocked. reason values are concatenated in config order (newline-separated).
  • If any hook halts, the turn ends after the tool call is blocked.
  • If no hook denies or halts but at least one allows, the tool call proceeds and the permission prompt is skipped.
  • context values are concatenated in config order. Strings and arrays compose uniformly β€” each string becomes one entry, and array entries are flattened in.
  • updated_input patches shallow-merge in config order against the original tool input. Later hooks override earlier ones on colliding keys. If denied or halted, updated_input patches are ignored.

Timeouts

If a hook exceeds its timeout, the process is killed and treated as a non-blocking error and the tool call proceeds. The default timeout is 30 seconds.

Examples

Block destructive commands

Prevent the agent from running rm -rf in bash:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "^bash$",
        "command": "./hooks/no-rm-rf.sh"
      }
    ]
  }
}

hooks/no-rm-rf.sh:

#!/usr/bin/env bash
# Block rm -rf commands in the bash tool. Otherwise stay silent so the
# normal permission flow runs.

if echo "$CRUSH_TOOL_INPUT_COMMAND" | grep -qE 'rm\s+-(rf|fr)\s+/'; then
  echo "Refusing to run rm -rf against root" >&2
  exit 2
fi

exit 0

Auto-approve read-only tools

Skip the permission prompt for tools that can't change anything. The hook returns decision: "allow", which tells Crush to pre-approve the call:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "^(view|ls|grep|glob)$",
        "command": "echo '{\"decision\":\"allow\"}'",
      },
    ],
  },
}

No script file needed β€” the command is inline. Every view/ls/grep/glob call now runs without prompting. Add the bash tool to this list at your own risk; consider a more targeted allowlist instead:

#!/usr/bin/env bash
# hooks/safe-bash.sh β€” auto-approve read-only bash commands.

case "$CRUSH_TOOL_INPUT_COMMAND" in
  ls*|cat*|grep*|rg*|echo*|pwd*)
    echo '{"decision":"allow"}'
    ;;
  *)
    # Silent β€” fall through to the normal permission prompt.
    exit 0
    ;;
esac

Inject context into file writes

Add a reminder to the model whenever it writes a Go file:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "^(edit|write|multiedit)$",
        "command": "./hooks/go-context.sh"
      }
    ]
  }
}

hooks/go-context.sh:

#!/usr/bin/env bash
# Remind the model about Go formatting when editing .go files.
# Emit context only; stay silent on `decision` so the normal permission
# prompt still runs for edits/writes.

if [[ "$CRUSH_TOOL_INPUT_FILE_PATH" == *.go ]]; then
  echo '{"context": "Remember: run gofumpt after editing Go files."}'
else
  echo '{}'
fi

Block all MCP tools

The command can be inline. This one-liner matches all MCP tools and blocks them:

{ "matcher": "^mcp_", "command": "echo 'MCP tools are disabled' >&2; exit 2" }

Log every tool call

With no matcher this fires for every tool. It exits 0 with no stdout so the tool call always proceeds.

{ "command": "echo \"$(date -Iseconds) $CRUSH_TOOL_NAME\" >> ./tools.log" }

A real-world Example:

For a more practical example, see rtk-rewrite.sh, which demonstrates how to rewrite tool input using RTK to save tokens.

Using other languages

Hooks aren't limited to shell scripts: any executable works. Here's the same "block rm -rf" example in some other languages.

Lua

{"matcher": "^bash$", "command": "lua ./hooks/no-rm-rf.lua"}

local input = io.read("*a")
local tool_input = input:match('"command":"(.-)"') or ""

if tool_input:match("rm%s+%-[rf][rf]%s+/") then
  io.stderr:write("Refusing to run rm -rf against root\n")
  os.exit(2)
end

JavaScript

{"matcher": "^bash$", "command": "node ./hooks/no-rm-rf.js"}

let input = "";
process.stdin.on("data", (chunk) => (input += chunk));
process.stdin.on("end", () => {
  const { tool_input: toolInput } = JSON.parse(input);

  if (/rm\s+-[rf]{2}\s+\//.test(toolInput.command)) {
    process.stderr.write("Refusing to run rm -rf against root\n");
    process.exit(2);
  }
});

Claude Code compatibility

Crush hooks are broadly compatible with Claude Code hooks: the config shape, stdin payload, output envelope, and exit codes line up so most Claude Code hooks run under Crush unchanged. This document covers the Crush-specific API only β€” anything not documented here isn't guaranteed to work.

One intentional divergence: Crush treats updated_input as a shallow-merge patch against the original tool_input rather than a full replacement. Keys you omit are preserved. See Output for details.


Reference

This is the official reference of the narrative above. If prose and this section disagree, the prose should be presumed canonical for intent, while this section is canonical for shape.

Both the stdin payload and the output envelope have common fields that apply to every event and per-event fields that only some events recognize. When an event doesn't understand a field, it's ignored.

Hook config

Each entry under a hooks.<EventName> array:

{
  // string. Optional. Regex tested against the tool name. Omit to match all.
  "matcher": "^bash$",

  // string. Required. Shell command to run.
  "command": "./hooks/my-hook.sh",

  // number. Optional. Seconds before the hook is killed. Defaults to 30.
  "timeout": 10,
}

Stdin payload (common)

Present in every hook event:

{
  // string. Hook event name.
  "event": "PreToolUse",

  // string. Current session ID.
  "session_id": "313909e",

  // string. Working directory when invoked.
  "cwd": "/home/user/project",
}

Stdin payload β€” PreToolUse

Extends the common payload:

{
  // ...common fields...

  // string. The tool being called.
  "tool_name": "bash",

  // object. Raw JSON input the model sent to the tool. Shape is per-tool.
  "tool_input": {
    "command": "npm test",
  },
}

Output envelope (common)

Fields a hook may print to stdout on exit 0. All are optional and apply to every event:

{
  // number. Defaults to 1. Unknown higher values still parse; exists for
  // forward-compat.
  "version": 1,

  // boolean. If true, ends the turn entirely. User takes over.
  "halt": false,

  // string. Shown when denying (to the model) or halting (to the model and
  // user).
  "reason": "not allowed",

  // string | string[]. Appended to what the model sees. Empty entries are
  // dropped.
  "context": "Rewrote with RTK",
}

Output envelope β€” PreToolUse

Extends the common envelope:

{
  // ...common fields...

  // "allow" | "deny" | null. null/omitted = no opinion, the tool still goes
  // through the normal permission prompt. "allow" is affirmative: pre-approves
  // the tool call and bypasses the prompt. "deny" blocks the call; the model
  // sees the error and may try something else.
  "decision": "allow",

  // object. Shallow-merge patch against tool_input. Nested objects are
  // replaced wholesale, not deep-merged.
  "updated_input": {
    "command": "bun test",
  },
}

Exit codes

Code Meaning
0 Success. Stdout is parsed as the output envelope.
2 Block this tool call. Stderr becomes the deny reason. Stdout is ignored.
49 Halt the whole turn. Stderr becomes the halt reason. Stdout is ignored.
other Non-blocking error. Logged and ignored; the tool call proceeds.

Exit 2 only applies to events that can block something. On events where there's nothing to block, it's treated as a non-blocking error.

Aggregation

When multiple hooks match the same event, results compose in config order.

Universal rules:

  1. halt is sticky: if any hook halts, the turn ends.
  2. reason values concatenate with \n in config order. Halt-only hooks without a deny still contribute their reason.
  3. context values concatenate with \n in config order. String entries and array entries flatten uniformly.

PreToolUse-specific rules:

  1. decision precedence: deny > allow > null. First deny determines the outcome; subsequent allows don't override. If the final aggregated decision is allow, Crush pre-approves the tool call and skips the permission prompt. If it's null (no hook allowed), the tool goes through the normal permission flow.
  2. updated_input patches shallow-merge sequentially against the original tool_input. Later patches override earlier ones on colliding keys. Patches are ignored if the final decision is deny or halt.

Environment variables

See Environment Variables above for the full list.


Whatcha think?

We'd love to hear your thoughts on this project. Need help? We gotchu. You can find us on:


Part of Charm.

The Charm logo

Charmηƒ­ηˆ±εΌ€ζΊ β€’ Charm loves open source