Guards

Sends a desktop notification when a question or a plan waits for your answer, and when a turn ends or fails.

After each commit the model makes, adds the doc lines that commit made stale (file:line references, symbol names) to the commit's result, from ripwire doc-drift, and holds each finding open until the doc holds again.

Adds a note to the fifth edit of one file in one turn, so the model re-reads the code path and states the root cause instead of trying again.

After each commit the model makes, names the env variables its added lines read that .env.example lacks, with the file and line of each.

Remembers which tests failed on which code, and tells the model when a failing test has both passed and failed on the same code, so it runs the test again instead of changing code.

Gives the model a Gemini advisor tool it calls by itself: Gemini reads the whole conversation and the model's message and answers with a second opinion. Off until /gemini-advisor on; gemini-core holds the key, tier, model and thinking level.

Has Gemini review each plan before the approval dialog, from the plan and the conversation, and sends a plan with a blocking finding back to the model, at most twice. Off until /gemini-plan-review on; gemini-core holds the key, tier, model and thinking level.

Has Gemini review every git commit the model runs, from the staged diff and the conversation, and stops a commit with a blocking finding: a bug, data loss, a security hole or a committed secret. Off until /gemini-review on; gemini-core holds the key, tier, model and thinking level.

Ships a commit skill and holds every Bash git commit to it: a commit the skill did not open, a blanket git add, a signature trailer, a secret, an ignored path, or a push or branch change nobody asked for is stopped before git runs.

After an edit that calls translation keys, names the keys the JSON, Laravel PHP, YAML or gettext locale files lack, and the languages that lack them.

After each commit the model makes, names the manifests whose dependencies it changed without their lockfile (npm, Composer, Cargo, Go, Python, Bundler, Dart, Mix).

Loads MEMORY.md into the session and saves project learnings to it after every turn through a tool-less fork in the background, without blocking the stop.

For mod development: runs a live check of plugins in a fresh Claude Code session in tmux, types the steps one by one, collects the pane and the transcript, then deletes the temp directory, its transcripts and the inline store files.

Keeps project and global memories in SQLite through a shared local daemon, hands the model the entries that match its file tool calls, prompts and subagent tasks, and saves new ones with a haiku consolidator after each main-loop turn.

For mod development: lets the model run a slash command in its own session, such as /reload-plugins after a plugin update, and read the output as the next prompt.

Shows this session's state in the sidebar: context fill, token totals, cost, model and effort, the Claude Code version, the other sessions of this machine, and the git branch and status.

Draws each PNG or JPG the model saves or reads under its tool row: pixels in kitty and Ghostty, quadrant block cells in every other terminal.

Opens one shared pane beside the transcript and draws the sections, lines and buttons every other mod writes into it through $.sidebar; each stream entry is stamped with its time and kept in a per-project log the next session takes back.

Runs slash commands joined with && one after another, waits for each command's turn, pane or dialog to end, and stops the chain at a command that failed or a turn that ended without an answer.

After an edit that builds SQL by joining or interpolating strings, names each line, so the model passes the values as query parameters.

After an edit that keeps browser data in localStorage or sessionStorage, names each line, so the model stores the data in a cookie instead.

Stops the model from repeating a tool call that just failed with the same input, until a file or command has changed something.

After a curl or wget an automated-client filter answered 403 or 429, gives the model the browser User-Agent to retry with, and the two cases where it must not.

AGENTS.md as project instructions, by one option. Under claude-md-or-agents-md a project with no instruction files of its own gets its AGENTS.md files instead: a prompt.context hook asks $.fs.ancestors for every AGENTS.md and .claude/AGENTS.md from the filesystem root down to the working directory and hands them to the engine as project instruction files, which it renders, announces and...