@juristechllc/fireworks-serialized-provider (P0c)

Otto-owned provider that wraps [@ai-sdk/openai-compatible] and enforces strict tool-use serialization on the LanguageModelV3 stream:

At most one committed tool-call is allowed per upstream stream. If the model emits a second concurrent tool-input-start or a distinct second tool-call, the transform suppresses the extras and emits a synthetic finish part with finishReason: "tool-calls", closing the stream so the OpenCode agent loop must execute the first tool, append the tool-result, and re-issue the completion before any further tool can fire.

Why this exists

From session ses_262cf3fb8ffeIPcZkaGvrgld2Z (Kimi K2.5 on Fireworks.ai):

• 170 / 196 (87 %) tool calls were issued while a prior call was still unresolved.
• 99 % of ai.response.started events began while a previous tool call was pending.
• 85 tool results were orphaned — they arrived after the model had already moved on to subsequent decisions.

This is a fire-and-forget control-flow bug at the model/provider interface. The Anthropic SDK enforces tool-use/tool-result ordering natively; OpenAI- compatible adapters (Fireworks) do not. This wrapper imposes the same invariant at the provider layer, uniformly for every model served through Fireworks' OpenAI-compatible endpoint — so the fix is model-agnostic.

Design constraints honored

• Does not modify upstream OpenCode. Keeps the engine updatable from the open-source project, per the Otto-OpenWork architecture rule.
• Does not touch the otto-chrome bridge, extension, cookies, or the 5-minute kill switch. Browser-control routing is unchanged.
• Does not live in the UI. Serialization runs inside the OpenCode engine process, where it belongs. The Solid UI and human interaction are untouched.
• Feature-reversible. Reverting the "npm" field in opencode.jsonc back to "@ai-sdk/openai-compatible" restores pre-P0c behavior.
• Pure and testable. The serialization transform is a pure function of the chunk sequence. Tests run against mock LanguageModelV3 streams with zero network, zero mocks of the base SDK beyond a small module loader hook.

Open-source references

• Vercel AI SDK stopWhen / stepCountIs commit vercel/ai@9315076 — documents the "stop after N tool calls" pattern at the agent loop. This package enforces it one level lower (provider stream), so it applies uniformly to every agent built on OpenCode.
• sst/opencode#2720 — "tool_use blocks found without corresponding tool_result blocks": same failure mode this wrapper prevents.
• cline/cline#9092 — the matching failure on a different agent.
• BerriAI/litellm#23507 — normalizing duplicate/concurrent tool_use emissions from Kimi K2 on non-Anthropic endpoints.
• browser-use#4651 / PR #4654 — perception-side retry of the same class of bug.

Installation

This package is published on the public npm registry. OpenCode's provider loader installs it automatically when a workspace's opencode.jsonc references it via the "npm" field:

{
  "provider": {
    "fireworks-ai": {
      "npm": "@juristechllc/fireworks-serialized-provider",
      "options": {
        "baseURL": "https://api.fireworks.ai/inference/v1"
      },
      "models": { "accounts/fireworks/models/kimi-k2p5": {} }
    }
  }
}

Otto-OpenWork's workspace-init step (apps/server/src/workspace-init.ts) performs the rewrite automatically on every workspace open — this is a follow-up PR to this package. Manual installation is not required.

For ad-hoc (non-OpenCode) use in Node:

npm install @juristechllc/fireworks-serialized-provider @ai-sdk/openai-compatible

API

import { createFireworksSerialized } from "@juristechllc/fireworks-serialized-provider";

const sdk = createFireworksSerialized({
  name: "fireworks-ai",
  baseURL: "https://api.fireworks.ai/inference/v1",
  apiKey: process.env.FIREWORKS_API_KEY,
  // Any other @ai-sdk/openai-compatible options (headers, fetch, ...) pass
  // through untouched.
});

const model = sdk.languageModel("accounts/fireworks/models/kimi-k2p5");
// `model` conforms to LanguageModelV3; doStream() output is clipped to one
// tool-call per stream.

The factory export is named createFireworksSerialized so the OpenCode provider loader's discovery heuristic (Object.keys(mod).find(k => k.startsWith("create"))) picks it up; this matches the pattern documented in sst/opencode provider.ts.

Tests

cd OpenWork/starter/apps/fireworks-serialized-provider
node --test __tests__/*.test.mjs

16 tests total: 12 cover the pure transform logic, 4 cover the factory via a mocked @ai-sdk/openai-compatible module loaded through a custom Node ESM loader hook.

Publishing

Releases are cut via the Publish fireworks-serialized-provider GitHub Actions workflow, triggered manually from the Actions tab with a version input. The workflow runs pnpm install, runs the test suite, verifies the package.json version matches the input, and publishes to public npm with NPM_TOKEN.