mohdel v0.101.0

Fixed

• Long thinking-only streams no longer fail with
  NET_ERROR / "terminated" at ~5 minutes. Node's global fetch
  (undici) closes a streaming response when no body chunk has
  arrived for bodyTimeout ms — 300 000 ms (5 min) by default.
  Reasoning models stream zero bytes during their thinking phase,
  so any non-trivial task on a thinking-capable provider tripped
  the limit mid-run and surfaced as a retryable upstream error
  with detail: "terminated". Adapters now opt out via a shared
  undici Agent with bodyTimeout: 0, so the inter-chunk idle
  timeout no longer applies to streaming inference. Cancellation
  still comes from the caller's AbortSignal, the SDK's
  request-level timeout, and provider-side stream limits.
  Headers timeout stays bounded at 60 000 ms — connect plus first
  response must still be fast.

Added

• js/session/adapters/_dispatcher.js — exports
  streamingDispatcher(), a lazy singleton undici Agent
  (bodyTimeout: 0, headersTimeout: 60_000) shared across all
  adapters that go through globalThis.fetch. One Agent per
  process keeps a single connection pool with the default
  per-origin keep-alive semantics.
• undici declared as a direct dependency (^7.24.5).
  Previously pulled in transitively; now an explicit dep because
  mohdel imports undici.Agent directly.

Changed

• All chat-completions adapters thread the dispatcher into
  fetchOptions. openai, fireworks, deepseek, mistral,
  openrouter, xai, anthropic, groq — each adds
  fetchOptions: { dispatcher: streamingDispatcher() } to its
  no-DI client construction. SDKs spread fetchOptions into the
  underlying fetch(url, opts) call (verified in
  openai/client.js:159, @anthropic-ai/sdk/client.js:74 + :446,
  and groq-sdk/client.js:82 + :388), so the dispatcher reaches
  the wire without replacing the SDK's fetch or touching its
  internals. Tests passing an explicit deps.client are
  unaffected.

Scope

• cerebras adapter unaffected. @cerebras/cerebras_cloud_sdk
  uses node-fetch@^2, a separate HTTP stack with no inter-chunk
  body timeout — the bug was undici-specific.
• gemini adapter not yet patched. @google/genai ships both
  node-fetch@^3 and undici@^7 and selects at runtime; its
  injection surface is httpOptions, not fetchOptions. Deferred
  to a follow-up release.

Tests

• New test/unit/dispatcher.test.js — singleton identity and
  Agent instance check.
• test/integration/provider.test.js — switched the tool-use smoke
  test from toolChoice: 'required' to 'auto'. DeepSeek's
  reasoner-backed models reject 'required' ('deepseek-reasoner does not support this tool_choice') even when the spec says
  supportsTools; the prompt itself forces the tool call, so the
  assertions still verify a real tool invocation under 'auto'
  without fighting one provider's API surface.