4.1.0 • Published 10h ago

@kuralle-syrinx/mastra

Licence

MIT

Version

4.1.0

Deps

Size

17 kB

Vulns

Weekly

Summary Dependency Versions

@kuralle-syrinx/mastra

The Mastra adapter for the Syrinx Reasoner seam: fromMastraAgent(agent) → Reasoner. Drop it into ReasoningBridge to drive the voice pipeline with a Mastra Agent instead of a raw AI SDK backend — no pipeline change.

Usage

import { Agent } from "@mastra/core/agent";
import { createOpenAI } from "@ai-sdk/openai";
import { ReasoningBridge } from "@kuralle-syrinx/aisdk";
import { fromMastraAgent } from "@kuralle-syrinx/mastra";

const agent = new Agent({
  name: "support",
  instructions: "You are a helpful voice assistant.",
  model: createOpenAI({ apiKey })("gpt-4.1-mini"),
});
session.registerPlugin("bridge", new ReasoningBridge(fromMastraAgent(agent)));

How it maps

fromMastraAgent awaits agent.stream(messages) and iterates output.fullStream (a ReadableStream<{type,payload}>), mapping each chunk → ReasoningPart:

Mastra chunk	→
`text-delta` (`payload.text`)	`text-delta`
`tool-call` (`payload.{toolCallId,toolName,args}`)	`tool-call`
`tool-result` (`payload.{toolCallId,toolName,result}`)	`tool-result`
`tool-call-suspended` (`payload.suspendPayload`, `output.runId`)	terminal `suspended` (Sprint 3 / suspend-resume)
`finish` (`payload.stepResult.reason`)	`finish` (stop/tool/length) or terminal `error` for abnormal reasons
`error` (`payload.error`)	terminal `error`
anything else	dropped

Verified against @mastra/core@1.41.0. A turn.resume routes to agent.resumeStream(data, {runId}) instead of stream(...). No buffering — each chunk is yielded immediately.

MastraAgentLike is the minimal structural type the adapter needs (stream + resumeStream) — the concrete @mastra/core Agent satisfies it.

Gotchas

@mastra/core is a peerDependency — the consumer instantiates and owns the Agent (and its Mastra version). The adapter only touches stream/resumeStream/fullStream/runId.
Node / nodejs_compat only. @mastra/core imports events/fs/path/crypto and won't bundle for a plain browser/edge target (--platform=browser). On Cloudflare Workers it requires nodejs_compat (see @kuralle-syrinx/server-workers-mastra). The lean AI-SDK product worker deliberately keeps Mastra out of its bundle.
History stays bridge-owned (RFC §4.5): run the agent stateless-per-turn (no Mastra memory) so the bridge's spoken-prefix barge-in remains authoritative.