chore: v2.3.21 (#5840)

# Changelog

## New Features:

- **Agent as Judge on AgentOS**: Our new [AgentAsJudge
Evals](https://docs.agno.com/basics/evals/agent-as-judge/overview) are
now completely supported on the AgentOS: you can configure and trigger
new runs, and will see existing runs listed with the rest in the Evals
page.

## Bug Fixes:

- **RunInput parsing**: Fixed an issue persisting RunInput objects when
providing it as a list of Message objects.
- **Mistral Embedder**: Fixed an issue setting timeout values for the
MistralEmbedder class.

chore: v2.3.20 (#5822)

# Changelog

## Improvements

- Run Cancellation: Add async methods to our Run Cancellation setup, and
allow users to set a custom one using `set_cancellation_manager()`.
- `reasoning_content` for LiteLLM model wrapper: Extract reasoning
content from the models which support it via lite llm

## Bug Fixes:

- **Agent Run Retries**: Fixed an issue duplicating responses when
running asynchronously with `retries` set to more than 1.
- **Complex Memories on AgentOS**: Improve handling for nested memory
content on the OS API. This also fixes an issue with some memories
resulting from the v1 → v2 database migration.
- **Session sorting on AgentOS**: Fixed an issue sorting sessions
migrated from v1 to v2, that can have empty `updated_at` values.
- **Memory Dates**: Fixed an issue handling some date formats when
initializing User Memories.

---------

Co-authored-by: Kaustubh <shuklakaustubh84@gmail.com>

chore: release 2.3.19 (#5808)

# Changelog

## Bug Fixes:

- **Vector DB imports**: Improved import paths for vector database
integrations

chore: release v2.3.18 (#5795)

# Changelog

## Improvements:

- **Google VertexAI**: Added support for a google oauth2 credentials
file for direct VertexAI authentication

## Bug Fixes:

- **Db Migration Router:** Fixed a bug causing Db migration router to
not get provisioned during resync

chore: Release 2.3.17 (#5790)

# Changelog

## New Features:

- **ChromaDB Hybrid Search:** Added support for hybrid search using
dense vector similarity search (semantic) with full-text search
(keyword/lexical) using RRF fusion for local chroma db.
- **Remote Agents:** Added the `RemoteAgent`, `RemoteTeam` and
`RemoteWorkflow` classes enabling usage of Agents, Teams and Workflows
running on a remote AgentOS. Docs coming soon.
- **Remote AgentOS client**: Added the `AgentOSClient` class enabling
usage of an AgentOS running remotely. Docs coming soon.

## Improvements:

- **SemanticChunking Embedder Support:** Use any embedder with semantic
chunking - pass a model string, any Agno embedder (AzureOpenAI, Mistral,
etc.), or custom chonkie BaseEmbeddings implementation.
- **Reconnection for Workflows via socket:** Extended the existing
websocket implementation for workflows to be reconnected via socket in
case of interruption in AgentOS client.

---------

Co-authored-by: manu <manuhortet@gmail.com>
Co-authored-by: Yash Pratap Solanky <101447028+ysolanky@users.noreply.github.com>

chore: release 2.3.16 (#5779)

# Changelog

## Bug Fixes:

- Readers: Fixes an issue with `read` and `async_read` function of some
Reader classes

chore: Release 2.3.15 (#5776)

# Changelog

## New Features:

- **OpenRouter Cost:** Now tracking `cost` on metrics for runs on
`OpenRouter`. More providers to come in future.

## Improvements:

- **RBAC & Security Key:** Improved the handling in cases where security
key and JWT authorization are both enabled, that JWT authorization is
favoured.
- **Migrate All DBs**: Added endpoint to migrate all DBs via AgentOS.

## Bug Fixes:

- **Storage for Run Requirements**: Fix an issue storing run
requirements (HITL related field) in certain streaming scenarios.
- **Async database streaming:** Fixed an issue where `_arun_stream` was
calling sync session method.

chore: v2.3.14 (#5746)

# Changelog

## New Features:

- **Reasoning Streaming**: Add support to stream reasoning chunks for
the case where `reasoning_model` is provided.
- **A2A interface endpoints:** Add new endpoints to our A2A interface.
You can now obtain the Agent Card for any of the available Agents, Teams
and Workflows. We also updated the endpoints to run your Agents, Teams
and Workflows to fit the updated protocol. See the
[docs](https://docs.agno.com/agent-os/interfaces/a2a/introduction).

## Improvements:

- **JSON Schema for Structured Outputs**: `output_schema` now accepts
JSON schemas in provider-specific formats, passed directly to the model
API without transformation. This gives users full control over
structured output configuration for OpenAI, Claude, and OpenAI-like
providers.
- **Persistence for errored runs**: Improved how all kind of errors are
handled during Agent and Team runs. Now an event representing the error
will always be emitted (in streaming scenarios). And the runs containing
errors will also always be persisted, when relevant.
- **Session names on AgentOS**: Minor improvements to what is shown as
the Session name in the AgentOS, considering the case of session with
`introduction` messages.
- **Milvus Search**: Added `search_parameters` to the search and
async_search methods of the Milvus vector database class.
- **AgentOS app reloading**: Add `reload_includes` and `reload_excludes`
to the `serve` function of AgentOS. Using these you can specify which
files the AgentOS app listens to to reload.
- **Default Gemini Embedder**: The default model ID for Gemini Embedder
class has now been updated to the latest gemini-embedding-001

## Bug Fixes:

- **BaseEval Improvement:** Improved handling of agent and team
parameters inside tool functions by using lazy imports, resolving a
warning during access.
- **Milvus filtering**: Fix an issue parsing search filters under
certain conditions when using the Milvus vector database.
- **Duplicate ID validation**: Agents, Teams or Workflows with
duplicated IDs are no longer allowed in AgentOS, to avoid unexpected
behavior.

---------

Co-authored-by: Yash Pratap Solanky <101447028+ysolanky@users.noreply.github.com>

chore: v2.3.14 (#5746)

# Changelog

## New Features:

- **Reasoning Streaming**: Add support to stream reasoning chunks for
the case where `reasoning_model` is provided.
- **A2A interface endpoints:** Add new endpoints to our A2A interface.
You can now obtain the Agent Card for any of the available Agents, Teams
and Workflows. We also updated the endpoints to run your Agents, Teams
and Workflows to fit the updated protocol. See the
[docs](https://docs.agno.com/agent-os/interfaces/a2a/introduction).

## Improvements:

- **JSON Schema for Structured Outputs**: `output_schema` now accepts
JSON schemas in provider-specific formats, passed directly to the model
API without transformation. This gives users full control over
structured output configuration for OpenAI, Claude, and OpenAI-like
providers.
- **Persistence for errored runs**: Improved how all kind of errors are
handled during Agent and Team runs. Now an event representing the error
will always be emitted (in streaming scenarios). And the runs containing
errors will also always be persisted, when relevant.
- **Session names on AgentOS**: Minor improvements to what is shown as
the Session name in the AgentOS, considering the case of session with
`introduction` messages.
- **Milvus Search**: Added `search_parameters` to the search and
async_search methods of the Milvus vector database class.
- **AgentOS app reloading**: Add `reload_includes` and `reload_excludes`
to the `serve` function of AgentOS. Using these you can specify which
files the AgentOS app listens to to reload.
- **Default Gemini Embedder**: The default model ID for Gemini Embedder
class has now been updated to the latest gemini-embedding-001

## Bug Fixes:

- **BaseEval Improvement:** Improved handling of agent and team
parameters inside tool functions by using lazy imports, resolving a
warning during access.
- **Milvus filtering**: Fix an issue parsing search filters under
certain conditions when using the Milvus vector database.
- **Duplicate ID validation**: Agents, Teams or Workflows with
duplicated IDs are no longer allowed in AgentOS, to avoid unexpected
behavior.

---------

Co-authored-by: Yash Pratap Solanky <101447028+ysolanky@users.noreply.github.com>

chore: Release 2.3.13 (#5722)

# Changelog

## New Features:

- **AgentOS Role-Based Access Control (RBAC)** → Introducing RBAC for
AgentOS (see the docs). This allows you to either automatically or
manually configure JWT-based authorization checks in AgentOS with the
following benefits:
- All traffic to your AgentOS would require a signed JWT token, with the
correct user permission scopes.
- Per-endpoint authorization → Based on the scopes inside the JWT token,
access is granted or denied to each endpoint as you choose. The required
scopes are configurable as well.
- Per-agent resource control → By setting specific scopes like
`agents:my-agent:read` you can control which agents can be run (i.e.
when calling `POST /agents/my-agent/runs` ) or which agents are returned
on read (i.e. when calling `GET /agents` or `GET /agents/{id}`). This
allows for full control over which users can use and see which agents,
and this extends to AgentOS. The same also applies to `Team` and
`Workflow`.
    - AgentOS UI compatibility to be released in the coming days.
    - **Note**: On the `JWTMiddleware` class:
- `secret_key` has been deprecated (although still supported) and it is
recommended to supply `verification_keys=[...]` going forward.
- `algorithm` default value changed from `HS256` to `RS256` to make
asymmetric keys the default (better suited for production cases).