Repository navigation
Read before upgrading
AgentConfig.max_context_tokenshas to be a positive number orNone. Zero and negative values now fail validation instead of trimming the whole history before raising. (#475)TokenCounter.count_text()counts the text alone. It used to wrap the text in a chat message, so every result was a few tokens too high (7 for an empty string on GPT-4). If you calibrated a budget against the old numbers, they drop by that amount. (#407)- MCP: a server argument named
tool_name,resource_nameorprompt_nameused to be overwritten by the routing field and never reached the server. It now appears asmcp_tool_name,mcp_resource_nameormcp_prompt_nameon the generated input model and is sent under its own name. (#409) - MCP: fixed-length tuple arguments are validated per position, and
minProperties/maxPropertieson dict-valued arguments are enforced, so a value the server would reject fails locally first. (#391, #463) - Tracing: for Anthropic,
gen_ai.usage.input_tokensnow includes cache read and cache write tokens, which Anthropic reports outsideinput_tokens. Totals for cached prompts go up to the real number. (#404) ChatHistory.dump()writes round-trip JSON, so computed fields are no longer stored in a snapshot. (#388)
Fixes
Agents, history and schemas
- A saved history with aliased or
Json[T]fields loads again, and streamed outputs with such fields are recorded in history. (#388) - Specializing a generic
BaseIOSchema, for exampleEnvelope[int], keeps the docstring of the generic class instead of raising. (#431) - Media on a field hidden with
exclude_ifis not sent to the model. (#408) - Whitespace before the first JSON token of a streamed response no longer breaks partial parsing. (#468)
MCP
- Discovery skips a catalog the server did not advertise instead of failing with
Method not found, so tools-only servers work with the combined fetch. (#415) - An empty text block in a prompt response stays an empty string. (#411)
Tracing
- Anthropic cache tokens are reported as
gen_ai.usage.cache_read.input_tokensandgen_ai.usage.cache_write.input_tokens. (#404) - Native Gemini responses record token usage and the response model. (#430)
- Payloads and captured messages are not serialized for spans that are not recording, such as sampled-out runs. (#460)
Forge tools
webpage_scraper: pages that declare their encoding in the HTML but not in the HTTP header are decoded correctly (#418), relative links and images are resolved against the final page URL (#412), headings and bylines inside<article><header>are kept (#442), and an emptymain-contentelement no longer wins over the real article (#458).hackernews_search: queries that fail are reported in afailureslist instead of looking like an empty result. (#400)serply_search: requests don't follow redirects, so the API key header can't be forwarded to another host. (#462)searxng_search: the social media category is sent under the name SearXNG uses. (#453)bocha_search: the per-call result limit is sent to the API. (#464)arxiv_search: queries that start withsubmittedDate:orjr:are sent as written. (#402)wikipedia_search: edition codes are validated before they go into the host name. (#379)weather: the forecast is kept when the current temperature is missing. (#397)datetime_tool: a local time that does not exist (spring-forward gap) is rejected. (#406)
Examples and the atomic TUI
rag-chatbot: chunks stay within the configured size (#382), and the Qdrant backend reports a cosine distance like the Chroma one (#392).deep-research: the call cap is checked before every agent call (#377), and the research plan is part of the shared context (#459).dspy-integration: prediction fields survive the bridge (#444), nullable and union fields are kept in signatures (#445), each benchmark review starts from a clean history (#450), and a limited training set still contains every genre (#452).- YouTube examples: the configured caption languages are used. (#447)
atomic: the file picker survives unreadable folders and duplicate names (#405), shows paths with square brackets literally (#410), and loads the tool catalogue without freezing the interface (#422).
Contributing
- The contributing guide has a new section, "What gets reviewed": at most 5 open pull requests per contributor, related fixes grouped, a bug fix names a trigger a user can hit, and features and performance changes start with an issue. (#483)
Thanks
All fixes in this release are by @Afloat16.