@firststep-studio/sdk 0.17.0 → 0.19.0

package/README.md

@@ -38,3 +38,31 @@ createServer(handler, {
 ## LLM Reference
 
 See `llms.txt` for full API specification.
+
+## Release notes
+
+### 0.19.0
+
+- **Cohere model support.** `llm.generate` and `llm.generateStream` accept
+  `provider: 'cohere'` with a `command-*` model (Command A, Command A+,
+  Command R, Command R+, Command R7B), routed through Studio's `/api/llm/*`
+  proxy the same way Gemini requests are. Text generation, JSON output,
+  streaming, finish reasons and token usage are supported. Gemini-only
+  features (explicit prompt caches, thinking/safety settings, image parts,
+  function-calling payloads) are outside the Cohere adapter's contract.
+- **Observable SSE stream errors.** `generateStream` only ever skipped a
+  chunk it could not parse as JSON at all (stray SSE comments/heartbeats). A
+  vendor error frame is valid JSON, so it used to be yielded as an ordinary
+  stream value, a plain object with an `error` field, indistinguishable from
+  a successful chunk to anything not inspecting its shape, so a failed turn
+  could look like a normal (if malformed) reply. A chunk shaped as an error
+  now throws `LLMError` instead of being yielded as data, with `status`
+  normalized to a numeric HTTP status: Gemini's numeric `code` or Cohere's
+  numeric `status` field is used when present, otherwise `502`. This throw
+  happens on the `.next()` call that reaches the error frame, not when
+  `generateStream()` creates the generator, so the `try/catch` must wrap the
+  entire iteration (the whole `for await` loop, or every subsequent manual
+  `.next()` call) and not just the call that obtains the stream.
+- **Handler adoption required.** Handler must upgrade to this SDK version to
+  pick up both changes. Existing Gemini configurations keep their current
+  defaults.