Current seven-reflex APIs, confidence semantics, lifecycle boundaries and defaults are documented in the integration guide. Examples below describe the original minimal integration.

@brainstem/pi-adapter

Wires @brainstem/reflexes into a Pi Agent — one function call, no wrapper agent, no fork of Pi's loop.

Install

npm install @brainstem/pi-adapter @brainstem/reflexes

Usage

import { Agent } from "@earendil-works/pi-agent-core";
import { createReflexes, jevJudge } from "@brainstem/reflexes";
import { attachReflexes } from "@brainstem/pi-adapter";

const agent = new Agent({ /* ...your existing Pi agent config... */ });

const reflexes = createReflexes({ judge: jevJudge({ apiKey: process.env.TYPESAFE_API_KEY }) });

attachReflexes(agent, reflexes, {
  cwd: process.cwd(),
  taskText: () => currentTaskDescription,
  onReflex: (line) => console.log(line),
});

That's it — agent now runs Gate before bash/read/write/grep/glob calls (configurable via capturedTools) and Sanitize + Verify on their results.

How it composes

attachReflexes wraps whatever beforeToolCall/afterToolCall the agent already has — it never replaces them. An existing block result is always respected (a reflex never overrides an existing block, and never spends a judge call on a call that's already refused). Calling attachReflexes twice on the same agent composes twice; that's correct, if wasteful, behavior.

Focus (optional, off by default)

attachReflexes(agent, reflexes, {
  cwd: process.cwd(),
  focusMode: "on",
  recentActivity: () => recentFindingsList,
});

With focusMode: "on", large tool results are cut down to the sections Focus judges relevant to the task — the model only sees the reduced view.

This is a real limitation, not a caveat to skim past: this package has no artifact store and no recovery tool. Content Focus omits is gone for that turn — there's no read_output/search_output the model can call to get it back, unlike brainstem's full CLI harness. Use focusMode: "on" only if losing omitted content is acceptable for your use case, or build your own recovery path on top of reflexes.focus() directly.

No approval lifecycle

A Gate ask verdict is blocked outright, with a reason explaining why — this package does not implement interactive human approval. If you need that, call reflexes.gate() yourself and build the approval flow you want, or use brainstem's full CLI harness, which has one.

License

MIT — see the repo root LICENSE.