Mistral Vibe: full reference

This page comes from agentltl-mistral-vibe/docs/REFERENCE.md

It is kept with the code it documents. Edit it there.

The plugin is the Claude Code plugin's guard on Mistral Vibe's hooks. Both are thin layers over agentltl-coding, which holds the rule language, the library, the guard and what the hooks do; each plugin only translates its agent's payloads, tool names and files.

Setup

git clone --recurse-submodules https://github.com/AgentLTL/agentltl-mistral-vibe.git ~/.vibe/plugins/agentltl
~/.vibe/plugins/agentltl/bin/agentltl install

• Vibe has no plugin install command, and it doesn't run hooks shipped inside a plugin folder: it runs the ones in ~/.vibe/hooks.toml (and a trusted project's .vibe/hooks.toml). agentltl install adds three [[hooks]] entries there, between two marker comments, and leaves the rest of the file alone; agentltl uninstall removes them. The clone in ~/.vibe/plugins/ also gives Vibe the plugin's skills.
• The first run builds a Python environment in ~/.vibe/plugin-data/agentltl/venv-<pins> ($VIBE_HOME if set), from the commits pinned in vendor.lock. It needs Python 3.10+ and git. An update that changes the pins builds a new one; environments unused for 14 days are removed.
• Rules apply where an AGENTLTL.yaml is: the project's (found from the working directory up to the git root) and the user's, ~/.vibe/AGENTLTL.yaml. Without either, the hooks exit at once.
• agentltl install also links the command line into ~/.local/bin.
• Updating: agentltl update (a git pull of the clone, then install again).

Claude Code plugin features, in Mistral Vibe

Claude Code plugin	Mistral Vibe plugin
claude plugin install agentltl@agentltl	git clone into ~/.vibe/plugins/agentltl, then agentltl install
Hooks in the plugin's hooks.json	[[hooks]] in ~/.vibe/hooks.toml: pre_tool, post_tool, post_agent
SessionStart: the rules in force	no such event: the rules come with the result of the session's first call (additional_context), and every refusal names its rule
UserPromptSubmit: lifts a stop, starts a turn	no such event: when post_agent lets Vibe finish, the next call follows the user's reply
PreToolUse deny, with a reason	pre_tool decision: deny with a reason; Vibe shows "Denied tool …" and the model gets the reason
PreToolUse ask	no such answer: the call is denied, Vibe is told to ask the user with ask_user_question, and the exact same call then goes through once
systemMessage	system_message: shown to the user in Vibe's notices (rules in force, credential warnings, finally send-backs)
PostToolUse / PostToolUseFailure	post_tool with tool_status success (status 0), failure (status 1), cancelled (not recorded)
Stop decision: block for finally rules	post_agent decision: deny: the reason comes back to Vibe as a message and it goes on, at most finish_retries times per turn
permission_mode (auto modes)	not in the payload; AGENTLTL_AUTO=1 marks an unattended run (vibe -p): ask is refused, unparseable.auto applies
CLAUDE_PROJECT_DIR	VIBE_PROJECT_DIR, else the git root of the working directory
Subagents (Task)	agent_spawn (or task in the legacy runtime), recorded as Task; a subagent's calls are checked in its own session
Tools Bash, Write, Edit, Read, Grep, WebFetch	bash, write_file, edit, read_file, grep, web_fetch (hooks may get file_system.bash etc.); the arguments already have Claude Code's names
~/.claude/AGENTLTL.yaml	~/.vibe/AGENTLTL.yaml
Memory: CLAUDE.md, .claude/rules/, auto memory	Instructions: AGENTS.md (project and ~/.vibe/), custom agents' instructions (agents/*.toml), prompts (prompts/*.md)
Built-in rule memory-first	the same, on Vibe's instruction files
Skills /agentltl:setup, :rules, :import, :status	the same: Vibe names a plugin's skills <plugin>:<skill> (the clone in ~/.vibe/plugins/agentltl provides them)
Status line	Vibe has none: /agentltl:status and agentltl trace
Library no-claude-coauthor, subagents-on-sonnet	no-vibe-coauthor (Vibe's Co-Authored-By: Mistral Vibe <vibe@mistral.ai>); the Claude Code entries switch nothing on here

Hooks

Event	Input used	Answer
pre_tool	session_id, cwd, tool_name, tool_input	decision: deny, reason and system_message; nothing when there is no objection
post_tool	tool_status, tool_output_text, tool_error	hook_specific_output.additional_context: the rules (first call of a session), a note the rules left on the call, a credential warning; system_message for the user
post_agent	session_id	decision: deny with what a finally rule still needs

pre_tool is installed with strict = true: if the guard crashes or times out, Vibe refuses the call. An internal error in the guard denies the call with the error. The guard never answers allow: silence leaves the call to Vibe's own permissions. Vibe runs pre_tool before its own permission check, so the rules hold under --auto-approve too.

Vibe runs the tool calls of one model turn in parallel, and their hooks too. The guard's state is locked per call, so calls are recorded in the order their hooks run.

Modes

Mode	In Mistral Vibe
block	denied with the reason
warn	denied once; Vibe may repeat the exact same call to go ahead
ask	denied; Vibe asks you (ask_user_question), then the exact same call goes through once
retry	denied; after retries attempts, as ask
stop	denied, and every call is denied until you write again
log	allowed; Vibe gets the note with the call's result

Rules

The rule language is the same in every harness: see the rules reference and the library. In agentltl check, a step may name Vibe's own tools: 'write_file {"file_path": ".env", "content": "x"}'.

State

Session and project traces live in the plugin's data directory (sessions/, projects/), as in the Claude Code plugin: agentltl trace shows them, agentltl reset forgets them.

Development

git clone --recurse-submodules https://github.com/AgentLTL/agentltl-mistral-vibe.git
scripts/setup.sh --dev && .venv/bin/pytest -q

docker/Dockerfile is an image with Mistral Vibe to try the plugin end to end on any OpenAI-compatible model (see the README).