Extend LoopCode

Hooks

Run your own commands at fixed points in LoopCode's agent loop — format after edits, block dangerous commands, filter test output, add context, or keep the agent working. Compatible with Claude Code hooks.

Why hooks

A prompt asks the model to do something; a hook makes it happen. Use hooks for anything that must not be skipped: formatting after every edit, refusing a dangerous command, filtering a huge test log down to the failures before it reaches the context window, or writing an audit record.

Hooks are ordinary programs — written in whatever your team already uses — configured in settings files. The events, payloads and decisions have the same names Claude Code uses, so hooks written for it mostly work here unchanged.

Events

Event	When	Can
`PreToolUse`	Before a tool call runs	Allow, deny or ask; rewrite the tool's input
`PostToolUse`	After a tool call	Add context for the model
`UserPromptSubmit`	When you send a message	Rewrite the prompt; add context; block it
`SessionStart`	When a session starts	Add context
`Stop`	When the agent is about to finish its turn	Keep it working, with an instruction
`SubagentStop`	When a subagent finishes	Keep it working
`PreCompact`	Before the conversation is compacted	Add instructions for the summary

Configure

.loopcode/settings.json

{
  "hooks": {
    "PostToolUse": [
      { "matcher": "Edit|MultiEdit|Write", "hooks": [{ "type": "command", "command": "npx prettier --write \"$(jq -r .tool_input.path)\"" }] }
    ],
    "PreToolUse": [
      { "matcher": "Bash", "hooks": [{ "type": "command", "command": "./scripts/hooks/no-prod-db.sh", "timeout": 10 }] }
    ],
    "Stop": [
      { "hooks": [{ "type": "command", "command": "./scripts/hooks/require-changelog.sh" }] }
    ]
  }
}

Field	Meaning
`matcher`	For tool events: a tool name, a `\|` list (`Edit\|Write`), or a regular expression. Omit it, or use `*`, for every call
`hooks[].command`	The command to run, through your shell, in the project folder
`hooks[].timeout`	Seconds before it is stopped (default 60, at most 600)
`hooks[].async`	`true` runs it without waiting; its result is ignored. For notifications and logging

Hooks from all settings files accumulate. /hooks lists the hooks in force and which files they came from.

What a hook receives

The event as JSON on standard input:

json

{
  "hook_event_name": "PreToolUse",
  "session_id": "…",
  "cwd": "/Users/you/code/shop",
  "permission_mode": "auto",
  "tool_name": "Bash",
  "tool_input": { "command": "psql $PROD_DATABASE_URL -c 'drop table orders'" }
}

Field	Present for
`tool_name`, `tool_input`	Tool events
`tool_result` (`text`, `ok`)	`PostToolUse`
`prompt`	`UserPromptSubmit`
`session_start_source`	`SessionStart`
`last_assistant_message`, `stop_hook_active`	`Stop` and `SubagentStop`. Check `stop_hook_active` to avoid sending the agent back in a loop
`agent_type`	`SubagentStop`
`trigger` (`auto` or `manual`), `custom_instructions`	`PreCompact`

The environment also has LOOPCODE_PROJECT_DIR (the project folder) and LOOPCODE_HOOK_EVENT.

What a hook can decide

Exit codes. 0 is success. 2 blocks the action, with whatever the hook wrote to standard error as the reason, which the agent sees. Any other exit code is treated as a broken hook — reported, but not a veto.

JSON on standard output, at the top level or under hookSpecificOutput:

Field	Effect
`permissionDecision`	`PreToolUse`: `allow`, `deny` or `ask`, overriding the permission check for this call
`permissionDecisionReason`	Why — shown to the agent
`updatedInput`	`PreToolUse`: replace fields of the tool's input before it runs
`updatedPrompt`	`UserPromptSubmit`: replace the prompt before the model sees it
`additionalContext`	Any event: text handed to the model as extra context
`systemMessage`	A message for you, shown in the transcript
`decision: "block"` with `reason`	`Stop`: keep the agent working, with `reason` as its instruction. `PreToolUse`: deny

scripts/hooks/no-prod-db.sh

#!/bin/sh
command=$(jq -r '.tool_input.command // ""')
case "$command" in
  *PROD_DATABASE_URL*) echo "Never touch the production database from an agent." >&2; exit 2 ;;
esac
exit 0

Safety

Hooks run with your permissions, without asking. A repository's committed .loopcode/settings.json can define hooks, so read the settings of a repository you do not trust before working in it. "disableAllHooks": true turns every hook off — put it in that project's .loopcode/settings.local.json, which is read last.

Hooks

#Why hooks

#Events

#Configure

#What a hook receives

#What a hook can decide

#Safety