ripple

Reviewed at commit b081051 New version, not yet reviewed

Caution: Lets the separate ClaudeRipple app answer the Claude app's permission prompts, including always-allow rules, and type messages into your session from its phone app.

Sends data to TypeSafe Jev (api.typesafe.ai) or OpenCode Zen (opencode.ai), only when you pick the Jev judge; the model providers you set up in ClaudeRipple, for delegated tasks

Picks the reasoning effort for every prompt and the worker for every delegated task, can keep the prompt cache warm, and shows where ClaudeRipple sent each…

Made by
ClaudeRipple
What we checked

Shows a band above the prompt with effort, context, cache timer and plan limits, and by default has Haiku on your account read each prompt with the last turns to set the reasoning effort. With the separately installed ClaudeRipple app it sends delegated tasks to the models you set up there, shows its request log, and in the Claude app passes permission prompts and phone messages through it.

Review details

Static review of the source at the pinned commit; not run. This is the status mod of the ClaudeRipple app (a local proxy installed separately and not reviewed here; the app enables the folder with `clauderipple mod on`); without the app it still draws the band and judges effort, and its router calls fail with a 'router down' line. Hooks: session.start, session.end, classic.SessionStart, classic.Stop, classic.PermissionRequest, session.receive, session.append, prompt.submit, turn.step, turn.start, turn.complete, agent.spawn, tool.call, session.compact, command.run (/ripple, /ripple-log, /ripple-bar) and ui.render/ui.close/ui.scroll for the AbovePrompt band, SessionMode footer and four panes (request log, agents, context, limits). prompt.submit sends the last user message (400 chars), the last assistant text (2000 chars) and the new prompt to a judge: by default Haiku through $.model.complete (tokens on every judged prompt; effort auto is on by default), or, when you pick Jev, a POST to api.typesafe.ai or opencode.ai with a key from TYPESAFE_API_KEY, ~/.config/jev/.env (a pasted key is written there and chmod 600) or the OpenCode Zen key in ~/.clauderipple/config.json; turn.step then rewrites the step's effort. agent.spawn registers its own 'auto' agent ('Delegate tasks here'), rewrites ripple:auto spawns to worker agents ClaudeRipple generates in ~/.claude/agents (default ladder DeepSeek, GPT, Opus, which the router sends to those providers), rewrites model and title of other spawns, refuses spawns whose named model lacks a reason, and refuses every nested spawn from a subagent (forks, workflows and teammates excepted). Timers poll the router on 127.0.0.1 (port from ~/.clauderipple/config.json, default 8792) every 1.5 to 15 s for settings (stored in the router), request log, limits and cost. Only in the Claude desktop app: classic.PermissionRequest posts the tool name, full tool input and suggestions to the router and, when the router holds the ask for the phone, applies the phone's allow or deny, including always-allow permission updates; a phone inbox poll submits phone messages as user prompts ($.prompt.submit asUser), aborts the current turn or runs /clear; and in a session whose permission mode is already bypassPermissions it changes other sessions' model and effort through the app's ccd_session_mgmt MCP tools. The router's admin API is called without any authentication. Other actions: $.model.fork for the Handoff note and the keep-warm ping (off by default), $.command.run compact/clear and a tidy-up prompt on button presses, $.process.run for uname, macOS defaults read AppleLocale, chmod and the system opener (cmd /c start, open, xdg-open) to open the settings page or a claude:// link; files written: ~/.clauderipple/logs/worker-picks.jsonl (task title, model, reason up to 200 chars, never the brief) and ~/.clauderipple/handoffs/*.md for long handoff notes; $.store keeps the band state, last pick and keep-warm memo. Errors go to the debug log and never block a turn, except managed spawns, which fail closed. register.tsx, approval.ts and handoff.ts were read in full; settings.ts, judge.ts, workers.ts, worker.ts and modes.ts in the relevant parts; i18n.ts and the ui/ drawing files by targeted search (they never receive the engine interface and contain no engine, network or process calls). The router, its phone relay and the phone app are not reviewed.

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: yes
  • Calls the model (uses tokens): yes
  • Sends data over the network: yes
  • Stores data: yes
  • Shows something on screen: yes
Hooked events

Picks the reasoning effort for every prompt and the worker for every delegated task, can keep the prompt cache warm, and shows where ClaudeRipple sent each request.

Made this mod? Add a badge to your README

modsforClaude badge for ripple

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/ripple.svg)](https://modsforclaude.com/mods/ripple/)

← All mods