Skip to content

Building AI agents, atomically

Note MCP error signaling and the tokenizer cache issue for 2.11.0

Eigenwise20fabca38a46
diff --git a/.release/2.11.0.md b/.release/2.11.0.md
index 77fabae..7e25389 100644
--- a/.release/2.11.0.md
+++ b/.release/2.11.0.md
@@ -8,6 +8,7 @@ Draft release notes. This release has not been published.
- Discover every page of MCP tools, resources and prompts. Propagate template continuation errors instead of silently dropping them.
- Resolve MCP schema references to non-object definitions and decode escaped definition names.
- Parse JSON text from real MCP SDK results when structured content is absent. Preserve precedence for present structured content, including empty values.
+- Raise a RuntimeError when an MCP server returns an error result (`isError`), before output parsing. The message carries only the server's text content; structured content, metadata and binary data stay out of it. Application outputs that merely contain an `isError` field are still returned as data. Based on #295 by pei711.
- Include the incoming message in the context token budget.
- Preserve chat history when loading invalid serialized data.
- Update the YouTube transcript tool for the instance API in youtube-transcript-api 1.x.
@@ -18,3 +19,7 @@ Draft release notes. This release has not been published.
Python 3.12 or newer is required. Existing public framework interfaces and provider configuration remain compatible.
Forge tools keep their own dependencies and setup instructions. Existing users of the YouTube transcript tool should update their local copy.
+
+## Known issue
+
+A fresh install that resolves litellm 1.104.x with tiktoken 0.14.x needs network access the first time `get_context_token_count()` runs: tiktoken rejects the tokenizer files bundled by that litellm version and downloads them. Offline environments can point `CUSTOM_TIKTOKEN_CACHE_DIR` at a directory holding the `cl100k_base` and `o200k_base` files. This is upstream behavior, not a framework change.