The AI Toolchain — issue -255, December 5, 2025

Agno v2.3.8 adds model-level retry control via retries=n on the model object for provider rate-limit resilience.

$ git clone --branch v2.3.8 https://github.com/agno-agi/agno.git
# already have the repo? check out this version:
$ git checkout v2.3.8

• ›Adds retries=n parameter at the model execution layer so rate-limit errors from model providers trigger retries directly on the model, independently of agent-level retries which continue to handle broader agent execution loop exceptions.

PydanticAI v1.27.0 adds dynamic built-in tool config via RunContext, MCP tool/resource caching, CoT reasoning support, and VercelAIAdapter message conversion.

$ git clone --branch v1.27.0 https://github.com/pydantic/pydantic-ai.git
# already have the repo? check out this version:
$ git checkout v1.27.0

from pydantic_ai.adapters.vercel import VercelAIAdapter

vercel_messages = VercelAIAdapter.dump_messages(result.all_messages())

• ›Adds VercelAIAdapter.dump_messages() method to convert PydanticAI messages to Vercel AI message format.
• ›Supports tool and resource caching for MCP servers that emit change notifications, reducing redundant round-trips.
• ›Enables dynamic runtime configuration of built-in tools via RunContext, allowing per-run tool behavior without rebuilding agents.
• ›Supports raw Chain-of-Thought (CoT) reasoning output from LM Studio and other OpenAI Responses-compatible APIs.
• ›Uses a custom reasoning field for OpenRouter to surface model reasoning traces.

Jan v0.7.4 adds CUDA 13 support, a browser MCP tool, desktop file attachments, and a quick-start model flow.

$ git clone --branch v0.7.4 https://github.com/janhq/jan.git
# already have the repo? check out this version:
$ git checkout v0.7.4

• ›Adds an embedding flag to model metadata with automatic detection, allowing Jan to identify and handle embedding models without manual classification.
• ›Adds CUDA 13 backend support, with CUDA 13 set as a priority backend for GPU inference.
• ›Adds an official Jan browser MCP (Model Context Protocol) integration, accessible as a chat input shortcut, enabling web browsing capabilities directly from the chat interface.
• ›Moves proactive mode configuration into MCP settings for centralised control of agentic behaviour.
• ›Adds a lockfile mechanism to kill conflicting MCP server port usage, preventing port collision errors.

llama.cpp b7276 adds CUDA/HIP GPU acceleration for CUMSUM and TRI tensor operations.

$ git clone --branch b7276 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:
$ git checkout b7276

• ›Adds GPU-accelerated CUMSUM and TRI tensor operations for CUDA and HIP (ROCm) backends, enabling models that use these ops to run on-device without CPU fallback.

The open source coding agent.

OpenCode v1.0.134 adds max-steps control for agents, experimental OpenTelemetry tracing, and a new Vercel theme.

$ git clone --branch v1.0.134 https://github.com/sst/opencode.git
# already have the repo? check out this version:
$ git checkout v1.0.134

# In your opencode config
{
  "experimental": {
    "open_telemetry": true
  }
}

• ›Adds max_steps configuration for supervisor and sub-agents to cap agentic step counts.
• ›Adds experimental experimental.open_telemetry config option to emit OTEL spans for tracing.
• ›Adds Vercel theme for the TUI.
• ›Increases default scroll speed in the TUI.

AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI

Pi v0.12.12 adds fuzzy model/session search, prompt history navigation, and a /resume command for mid-conversation session switching.

$ git clone --branch v0.12.12 https://github.com/earendil-works/pi.git
# already have the repo? check out this version:
$ git checkout v0.12.12

$ /resume

• ›Adds fuzzy search for models and sessions (e.g., codexmax matches gpt-5.1-codex-max) in the interactive selector.
• ›Adds prompt history navigation: browse up to 100 previously submitted prompts using Up/Down arrow keys when the editor is empty.
• ›New /resume command lets you switch to a different session mid-conversation via an interactive selector, without restarting the agent.
• ›Footer token counts now use M suffix for millions and shortened context display format (e.g., 61.3%/200k).

$ git clone --branch v0.12.11 https://github.com/earendil-works/pi.git
# already have the repo? check out this version:
$ git checkout v0.12.11

{
  "name": "my-provider",
  "apiKey": "sk-...",
  "authHeader": true,
  "baseUrl": "https://my-provider.example.com/v1"
}

Qwen Code v0.4.0-preview.1 adds a TypeScript SDK, DeepSeek V3.2 thinking mode, session-level conversation history management, and custom tool support via SDK-controlled MCP servers.

$ git clone --branch v0.4.0-preview.1 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:
$ git checkout v0.4.0-preview.1

• ›Adds a basic TypeScript SDK (feat: basic TypeScript SDK) enabling programmatic integration with Qwen Code.
• ›Supports custom tools via SDK-controlled MCP servers, allowing SDK consumers to register and surface their own tooling to the agent.
• ›Integrates DeepSeek V3.2 thinking mode, enabling reasoning-trace-style responses from the DeepSeek V3.2 model.
• ›Adds session-level conversation history management, giving users control over per-session context across a conversation.

LanceDB Java SDK gains REST namespace support and a generic client builder.

$ git clone --branch v0.23.0-beta.1 https://github.com/lancedb/lancedb.git
# already have the repo? check out this version:
$ git checkout v0.23.0-beta.1

• ›Adds REST namespace support to the LanceDB Java SDK client.
• ›Makes the LanceDB Java SDK client builder generic, enabling typed client construction.