ruflo-protector

Reviewed at commit 13af5f1 New version, not yet reviewed

Project Anatole: an optional, learning watchdog for unattended agents. Deterministic OWASP-mapped rules, a per-project behavioural baseline, four modes (off…

Made by
ruvnet
What we checked

Watches every tool call, spawned agent and incoming message with fixed rules (secrets sent over the network, piped remote code, changes to Claude config or shell profiles, destructive commands) and learns a per-project baseline. By default it only learns; notify mode shows toasts and enforce mode, which you must switch on yourself, can deny matching calls. It makes no network calls and stores short category records in your project.

Review details

Static review of the source at the pinned commit; not run. Delta review of the changes since 87d8687; the rest carries over from the review at that commit. The only change is a repository field added to plugin.json; no code, hook or prompt changed. Carried over: registers session.start, turn.start, turn.complete, session.end, tool.check, agent.spawn, session.receive and a /protector command (status, list, mode, rule, run, replay, alerts, ack, unack, allow, reset-baseline). Uses $.fs.read, $.fs.write, $.fs.exists, $.session.root, $.clock.now, $.clock.after, $.ui.toast and $.command.register; no $.net, no $.process, no $.model, no eval. Tool calls are normalised into categorical tokens (tool name, short command head, registrable host domain, two directory levels, flags); secret-shaped values become 'redacted' and no argument values, file contents or prompt text are stored. Prompt text is only used in memory to tell whether an action was asked for. tool.check can return a deny only when mode is enforce and a rule marked block fires; it never approves or asks, and an internal error passes the call through. State files under .claude-flow/protector-mod/: status.json, baseline.json, alerts.jsonl, rules.json, allow.json and ring.json, all bounded. /protector run reads .claude/settings.json and .claude/settings.local.json and reports finding names only. Default mode is learn (silent); auto-graduation moves learn to notify after a mature baseline, never to enforce. Toasts go through a shared toast policy (levels, one-line text with credentials, e-mail addresses and home paths masked, dedupe, at most four a minute); it reads .claude-flow/console/toast-prefs.json and, only when the ruflo console folder .claude-flow/console already exists, rewrites a ring of at most 60 masked toast digests to .claude-flow/console/toasts/protector.jsonl. The pre-scan hits for permission, settings and reviewer-text are detection regexes, not behaviour.

What it uses
  • Reads your prompts: yes
  • Sees tool calls: yes
  • Blocks or changes tool calls: yes
  • Reads or writes files: yes
  • Runs programs on your machine: no
  • Calls the model (uses tokens): no
  • Sends data over the network: no
  • Stores data: yes
  • Shows something on screen: yes
Hooked events

Project Anatole: an optional, learning watchdog for unattended agents. Deterministic OWASP-mapped rules, a per-project behavioural baseline, four modes (off, learn, notify, enforce), alerts you can read later.

Made this mod? Add a badge to your README

modsforClaude badge for ruflo-protector

Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.

[![Listed on modsforClaude](https://modsforclaude.com/badge/ruflo-protector.svg)](https://modsforclaude.com/mods/ruflo-protector/)

← All mods