@reinsjs/lowering-pi

The lowering layer for reins built on pi-ai: timeline events → provider messages → Anthropic Messages or OpenAI Responses wire protocol, and streaming responses → event drafts. Every event type has a declared landing per API (LOSS_MATRIX: exact, lossy, dropped); nothing is dropped silently.

pnpm add @reinsjs/lowering-pi

import { anthropic, openai } from "@reinsjs/lowering-pi"
import { createAgent } from "@reinsjs/agent"

const claude = anthropic("claude-opus-5", { apiKey: process.env.ANTHROPIC_API_KEY! })
const gpt = openai("gpt-5.5", { apiKey: process.env.OPENAI_API_KEY!, baseUrl: "https://gateway.example/v1" })

export const agent = createAgent({ model: claude, store, tools })

anthropic() / openai() return a BoundModel = { model, lowering } — the only thing createAgent needs. The apiKey is a required option, never read from the environment, so the same code runs where there is no process.env.

Options

| option | meaning | | --- | --- | | apiKey | required | | baseUrl | gateway or proxy; both factories accept any endpoint speaking the respective protocol (DeepSeek's Anthropic-compatible endpoint is tested) | | requestOptions | passed to pi-ai as a whole, not merged — if you set it on openai(), include reasoningEffort yourself or reasoning turns off | | contextWindow, maxOutputTokens, reasoning, images, midConversationSystem | override the capability table (capabilitiesOf) for models the table does not know | | trustMarkers | default on: tool output is wrapped in <untrusted source="tool:…"> in the request (the log keeps the original) |

How events land

• Model roles exist only here. system_note events become system blocks where the API allows mid-conversation system content, otherwise a framed user message (declared lossy).
• A user message that arrives while tool results are still pending is moved after them on the wire (Anthropic requires tool_result right after tool_use) and recorded as lossy(user); the log order is untouched.
• Thinking blocks are replayed only with a signature from the same provider and model.
• Anthropic cache breakpoints: at most four block-level breakpoints are set; when they are exhausted the top-level breakpoint is silently skipped rather than producing a 400.
• pi-ai 0.85.1 has no system role for injected content; the layer rewrites the payload in pi-ai's onPayload hook. The LoweredRequest.payload you can inspect shows the pre-rewrite shape — remember that when debugging a 400.

Runtime footprint

pi-ai declares ten dependencies (≈65 MB installed, about 29 MB of which reins never imports). Nothing Node-specific is on the paths used here; the layer is verified on Cloudflare workerd without nodejs_compat, and on Bun, Deno and Vercel's edge-runtime (spikes/runtime-matrix). Two runtime caveats inherited from upstream: pi-ai builds its user-agent from os.release(), so Deno needs --allow-sys=osRelease; the openai SDK's edge branch reads process.version unguarded, so an Edge runtime must expose a process object (Vercel's deployed Edge runtime does — verified live; the bare edge-runtime emulator does not). If install size matters more to you than getting the provider table through pi-ai, use @reinsjs/lowering-fetch — the same Lowering interface with no dependency but @reinsjs/core, three wire protocols (it adds OpenAI Chat Completions), and a request body that is sent exactly as you inspect it. Its README has a side-by-side comparison; timelines written by one layer replay through the other.

Documentation

docs/技术方案.md §11 (loss matrix and landings), docs/模块盘点/lowering-pi.md, spikes/ (upstream behaviour probes) — in Chinese, at the repository root.

MIT © 2026 NewRate Limited.