Tools

Programs that run next to the agent. Most are wired in through hooks, so they apply without me asking.

Context savers

rtk: a CLI proxy that filters and summarises command output before it reaches the model. rtk git status or rtk grep -rn returns the useful part instead of hundreds of lines. It is wired in as a PreToolUse hook on Bash. I skip it when exact bytes matter, for example a diff I am about to apply.

codegraph: indexes a repo into a SQLite graph of symbols, call edges, and files. One codegraph_explore call returns the source of the relevant symbols plus the call paths between them, which replaces a grep-then-read loop. It runs as an MCP server, and a UserPromptSubmit hook injects graph context for the prompt. It only applies in repos that have a .codegraph/ directory. Elsewhere, the agent ignores it.

Gates

• no-mistakes: a local git proxy. Pushing through it runs code review, tests, lint, and docs checks before anything reaches the real remote.
• CodeRabbit CLI: a second, independent code review from the terminal.

Proof

proving-it-works: a Claude Code plugin that records a narrated movie proving a change works, from a browser, terminal, desktop app or stills. Its movie check gate fails a movie whose picture freezes while narration keeps going, never changes, or has silent audio: defects that per-frame checks miss. I use it for the GIF/video evidence on frontend PRs.

Review surface

lavish-axi: turns an HTML page the agent wrote into a review surface. I annotate elements in the browser, queue comments, and the agent picks them up. Good for plans and comparisons that are painful to read in a terminal.

Other agents

• Codex CLI and Cursor agent: second opinions. A different model reviewing the first one's work catches things a self-review does not.
• Ollama: local models, for work that should not leave the laptop.

Workspace glue

• herdr: a terminal workspace manager for coding agents. A SessionStart hook reports each agent's state to it.
• iTerm status hooks: every lifecycle event (Stop, Notification, PermissionRequest and the rest) calls one small script that updates the tab badge, so I can see which session is waiting on me.
• Status line: directory, branch or worktree, model, context used, rate limit used.

Config snippets

Hooks in ~/.claude/settings.json

The three hooks that do real work. Every other lifecycle event calls the iTerm status script.

{
  "hooks": {
    "SessionStart": [
      { "hooks": [{ "type": "command", "command": "lavish-axi", "timeout": 10 }] }
    ],
    "UserPromptSubmit": [
      { "hooks": [{ "type": "command", "command": "codegraph prompt-hook" }] }
    ],
    "PreToolUse": [
      { "matcher": "Bash", "hooks": [{ "type": "command", "command": "rtk hook claude" }] }
    ]
  }
}

codegraph MCP server

{
  "mcpServers": {
    "codegraph": { "type": "stdio", "command": "codegraph", "args": ["serve", "--mcp"] }
  }
}