v2.0.8

Changelog

New Features:

• CometAPI: Added support for CometAPI as a model provider.

Improvements:

• Access dependencies on tool: dependencies are now available on custom tools as a built-in argument. See an example.
• V1 → V2 Migration Script: Updated the migration script by introducing batching. This makes it significantly more performant for large databases.
• Support multiple text_contents on add_contents() in Knowledge.
• MultiMCPTools: Added allow_partial_failure flag to MultiMCPTools toolkit

Bug Fixes:

• Dependencies in Teams: Ensure dependencies are correctly passed through to member agents to reference.
• Tool call limit on default reasoning agent: Ensures that a tool_call_limit can be passed to the default reasoning agent, avoiding potential costly tool call loops.
• Fix string encoding: Addresses issue where string encoding broke when adding Chinese string content. Encoding now reliably works via utf-8 encoding.
• Fix run_input not being populated in session runs endpoint: The /sessions/<session_id>/runs endpoint does not return run_input in case of error as it references it from step_executor_runs and its empty if the workflow hasn't run yet, so in this case we fallback to directly get it from input.
• Fix State Precedence for Agents/Teams: Fixed issues related to state from the database that would override state passed on agent.run.
• Fix early stop in workflow to return correct last output: early stop functionality in workflow was not updating the last output properly for non stream cases
• Fixed OpenAIResponses to remove duplicate messages for reasoning models:
  • The OpenAIResponses API has the (default enabled) feature that they store the conversation server-side and we send a previous_response_id with any followup questions.
  • Agno's OpenAIResponses class implements this feature by looking through the message history, finding the most recent assistant response, and using that message ID as previous_response_id. This part works. However, we were then not deleting the prior messages from the request it sends to OpenAI. As a result OpenAI concatenates its own message history with the message history sent by Agno, resulting in an exponential number of repeated messages as the conversation continues.

What's Changed

• fix: langchaindb and llamaindex by @Ayush0054 in #4687
• feat: custom middleware agent os cookbook by @kausmeows in #4689
• fix: resolve circular import between Vectordb and knowledge modules by @SamJupe in #4696
• fix: Always encode string content as UTF-8 by @RockPaperSmittie in #4697
• fix: Handle extra_data in migration scripts by @willemcdejongh in #4654
• chore: add workflow id in workflow_data by @kausmeows in #4698
• fix: system message role bug fix by @Ayush0054 in #4701
• fix: Pass dependencies correctly to member agents by @dirkbrnd in #4700
• fix: Allow for passing tool_call_limit to default reasoning agent. by @RockPaperSmittie in #4703
• feat: access dependencies on tool by @kausmeows in #4706
• fix: favor session db state for subsequent runs by @RockPaperSmittie in #4704
• [fix] Fix claude reasoning tool user by @KeraStudios in #4709
• fix: Add fix for parameters on agentic filtering by @dirkbrnd in #4702
• [Fix] async ollama client with different Host not used correctly by @Mustafa-Esoofally in #4711
• fix: run_input not being populated in session runs endpoint if workflow raises error by @kausmeows in #4730
• Revert "fix: favor session db state for subsequent runs" by @dirkbrnd in #4732
• Fix: Ensure session state precedence by @RockPaperSmittie in #4735
• [fix] Fix reasoning content aggregation and streaming by @jtalmi in #4708
• [feat] Add CometAPI as a new model provider by @TensorNull in #4682
• fix: Gemini structured output by @dirkbrnd in #4647
• [Fix] Stop team task by @Nishant-Sagar in #4727
• fix: tool call mis match fix for ollama cloud by @Ayush0054 in #4733
• feat: fix initialisation of the sentence transformer client directly in get_embedding method. by @Ayush0054 in #4738
• [fix] Add message_id handling for correct tool call ordering by @Mustafa-Esoofally in #4664
• fix: early stop in workflow to return correct last output by @kausmeows in #4743
• fix: remove duplicate messages in OpenAIResponses by @mishramonalisha76 in #4739
• chore: OpenAI responses update by @kausmeows in #4744
• feat: add allow_partial_failure option to MultiMCPTools by @yinglj in #3627
• docs: added maxim support for observability by @akmadan in #3831
• feat: Support text_contents in add_contents by @willemcdejongh in #4695
• chore: Release 2.0.8 by @ysolanky in #4710

New Contributors

• @jtalmi made their first contribution in #4708
• @TensorNull made their first contribution in #4682
• @Nishant-Sagar made their first contribution in #4727
• @yinglj made their first contribution in #3627
• @akmadan made their first contribution in #3831

Full Changelog: v2.0.7...v2.0.8