@xxxoooxoxo/wiff

wiff

Deterministic, resumable multi-agent workflows for Codex, Claude Code, Cursor, and Kimi, exposed as a harness-agnostic MCP server. Write plain JavaScript with agent(), /goal stages, parallel(), and pipeline(); wiff runs the children on a pluggable backend (Codex app-server threads, headless claude or kimi, or the Cursor SDK — chosen per agent from the model name), journals every step, and resumes interrupted runs without re-paying for completed work.

At a Codex stage that must keep going until a condition is genuinely met, prefix the agent prompt with /goal:

phase("Verify and repair");
await agent("/goal Make the unit tests pass and verify the final run.", {
  key: "tests-green",
  sandbox: "workspace-write",
  timeoutMs: 30 * 60 * 1_000,
});

Wiff keeps that Codex thread at the current workflow stage while its goal is active. The workflow advances only when the worker marks the goal complete; blocked or limited goals fail explicitly.

Persistent preferences live in ~/.wiff/config.json, with optional project overrides at <cwd>/.wiff/config.json. They can provide user instructions, default agent options, ordered cross-backend fallbackModels, and matching rules that select models or turn phases into /goal stages. Wiff injects applicable instructions alongside the task and includes the effective preference decision in the resume-cache identity. See the workflow API reference for the schema and precedence rules.

Run the server (any MCP client):

npx @xxxoooxoxo/wiff                      # stdio MCP server: start / status / wait / cancel / models
npx -p @xxxoooxoxo/wiff wiff-viewer       # live web viewer on http://127.0.0.1:4979

Goal stages are marked directly in the workflow graph as queued, active, met, failed, or replayed.

Requires Node >= 22, git for isolation: "worktree", and the runtime of whichever backend your agents use: Codex CLI >= 0.144.6 for gpt-*/o* models (the default), the claude CLI for current claude-fable-5/claude-opus-5/claude-sonnet-5/claude-haiku-4-5 models and the moving fable/opus/sonnet/haiku aliases, or CURSOR_API_KEY plus the bundled @cursor/sdk for composer-* models, or the kimi CLI for kimi-code/* models. Set WIFF_BACKEND to route unrecognized models to a specific backend, and WIFF_DEFAULT_MODEL to change the default model. Agent calls default to medium reasoning effort and a 10-minute execution timeout; time waiting for a concurrency slot is tracked separately. Wiff-launched Codex children disable plugins, apps, and configured MCP servers only inside their app-server process so normal interactive Codex configuration is unchanged.

Full documentation, the workflow script contract, and Codex/Claude Code plugin installs: https://github.com/xxxoooxoxo/wiff