The AI Toolchain — issue -221, January 8, 2026

Agno v2.3.24 adds proxy support for Crawl4aiTools, base-directory sandboxing for PythonTools and MLXTranscribeTools, and heading-level chunking for MarkdownChunker.

$ git clone --branch v2.3.24 https://github.com/agno-agi/agno.git
# already have the repo? check out this version:
$ git checkout v2.3.24

from agno.tools.python import PythonTools

# Production: default sandboxed behaviour (restrict_to_base_dir=True)
tools = PythonTools(base_dir="/app/workspace")

# Local dev: opt out of sandboxing
tools_open = PythonTools(base_dir="/app/workspace", restrict_to_base_dir=False)

from agno.tools.crawl4ai import Crawl4aiTools

tools = Crawl4aiTools(
    proxy_config={
        "server": "http://proxy.corp.example.com:8080",
        "username": "user",
        "password": "pass"
    }
)

from agno.document.chunking.markdown import MarkdownChunker

chunker = MarkdownChunker(split_on_headings=True)

• ›Adds proxy_config parameter to Crawl4aiTools for configuring proxy settings on the toolkit.
• ›Adds restrict_to_base_dir parameter to PythonTools and MLXTranscribeTools; by default both tools now block operations outside their contextual base directory — pass restrict_to_base_dir=False to opt out.
• ›Adds split_on_headings parameter to MarkdownChunker for fine-grained control over how chunks are separated.
• ›MongoDB connection handshake now includes Agno version metadata, improving connection identification when multiple applications share a cluster.

Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.

CrewAI 1.8.0 adds Agent-to-Agent async/streaming/push update mechanisms and Human-in-the-Loop support for Flows.

$ git clone --branch 1.8.0 https://github.com/crewAIInc/crewAI.git
# already have the repo? check out this version:
$ git checkout 1.8.0

• ›Adds native async chain support for agent-to-agent (A2A) communication.
• ›Introduces A2A update mechanisms — poll, stream, and push — with configurable handlers.
• ›Adds Human-in-the-Loop (HITL) feedback support directly within Flows via global flow configuration.
• ›Adds streaming tool call events for real-time observability of tool execution.
• ›Introduces production-ready Flows and Crews architecture.

Haystack v2.22.0 adds semantic document splitting, auto warm-up, multi-output tools, and Qwen3 reranker support.

$ git clone --branch v2.22.0 https://github.com/deepset-ai/haystack.git
# already have the repo? check out this version:
$ git checkout v2.22.0

from haystack.components.embedders import SentenceTransformersDocumentEmbedder
from haystack.components.preprocessors import EmbeddingBasedDocumentSplitter

embedder = SentenceTransformersDocumentEmbedder()
splitter = EmbeddingBasedDocumentSplitter(
    document_embedder=embedder,
    sentences_per_group=2,
    percentile=0.95,
    min_length=50,
    max_length=1000
)
result = splitter.run(documents=[doc])

from haystack.tools import Tool

tool = Tool(
    name="search",
    description="Search for documents",
    parameters={...},
    function=search_func,
    outputs_to_string={
        "formatted_docs": {"source": "documents", "handler": format_documents},
        "summary":        {"source": "metadata",  "handler": format_summary}
        # 'debug_info' is omitted and will not be stringified
    }
)

from haystack.components.rankers.sentence_transformers_similarity import SentenceTransformersSimilarityRanker

ranker = SentenceTransformersSimilarityRanker(
    model="tomaarsen/Qwen3-Reranker-0.6B-seq-cls",
    query_prefix='<|im_start|>system\nJudge whether the Document meets the requirements...\n<Query>: ',
    query_suffix="\n",
    document_prefix="<Document>: ",
    document_suffix="<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n",
)
result = ranker.run(query="Which planet is known as the Red Planet?", documents=[...])

• ›Adds EmbeddingBasedDocumentSplitter to haystack.components.preprocessors, splitting documents by semantic similarity using a pluggable embedder; constructor accepts document_embedder, sentences_per_group, percentile, min_length, and max_length parameters.
• ›Adds outputs_to_string configuration to Tool, letting a single tool expose multiple named string outputs (each with a source and handler) so the LLM receives rich, selectively stringified context without additional tool calls.
• ›Adds query_suffix and document_suffix parameters to SentenceTransformersSimilarityRanker, enabling compatibility with the Qwen3 reranker model family (e.g., tomaarsen/Qwen3-Reranker-0.6B-seq-cls).
• ›Adds enable_thinking parameter to chat generators for thinking-capable models, allowing intermediate chain-of-thought reasoning steps before final responses.
• ›Adds reasoning content support to HuggingFaceAPIChatGenerator, extracting chain-of-thought output (e.g., from DeepSeek R1) in both streaming and non-streaming modes; accessible via reply.reasoning.reasoning_text.

Your own personal AI assistant. Any OS.

OpenClaw v2026.1.8 locks down DMs by default, adds pairing-first auth, per-agent sandboxing, and a major CLI overhaul.

$ git clone --branch v2026.1.8 https://github.com/openclaw/openclaw.git
# already have the repo? check out this version:
$ git checkout v2026.1.8

$ clawdbot pairing list --provider telegram
clawdbot pairing approve --provider telegram <code>

agent.sandbox.scope="agent"   # default: one container per agent
# or for per-session isolation:
agent.sandbox.scope="session"

agent.userTimezone="America/New_York"

• ›Enables DM pairing-first security model across Telegram, WhatsApp, Signal, iMessage, Discord, and Slack — bots are no longer open to anyone by default.
• ›Adds clawdbot pairing list --provider <provider> and clawdbot pairing approve --provider <provider> <code> commands to manage inbound DM access requests.
• ›Introduces per-agent sandbox scope (agent.sandbox.scope) with "agent", "session", and "shared" isolation modes for container/workspace separation.
• ›Adds /compact slash command to manually compact session context in the agent loop.
• ›Adds agent.userTimezone config field to inform the model of the user's local time zone via the system prompt.

GitHub Copilot CLI v0.0.376 adds GraphQL ID session loading, image processing in task tool subagents, and disk-offloaded large tool outputs.

$ git clone --branch v0.0.376 https://github.com/github/copilot-cli.git
# already have the repo? check out this version:
$ git checkout v0.0.376

• ›Enables loading remote sessions by GraphQL ID or via a session picker.
• ›Task tool subagents can now process images.
• ›Large tool outputs are written to disk and models are directed toward efficient search tools to handle them.

The open source coding agent.

OpenCode v1.1.7 adds an interactive question tool, overlay sidebar for narrow screens, expandable bash output, and .claude prompt disabling.

$ git clone --branch v1.1.7 https://github.com/sst/opencode.git
# already have the repo? check out this version:
$ git checkout v1.1.7

• ›Adds an interactive question tool for gathering user preferences and clarifying instructions mid-session.
• ›Adds overlay sidebar for narrow screens to improve mobile/small-terminal experience.
• ›Adds expandable bash output for long commands to improve readability in the TUI.
• ›Supports disabling .claude prompt and skills loading via flags.
• ›Writes truncated tool outputs to files instead of silently dropping them.

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

Pi v0.39.0 adds remote SSH tool execution, --no-tools flag, Wayland clipboard, and runtime theme switching for extensions.

$ git clone --branch v0.39.0 https://github.com/earendil-works/pi.git
# already have the repo? check out this version:
$ git checkout v0.39.0

$ pi --no-tools

• ›Adds --no-tools flag to disable all built-in tools, enabling pure extension-defined tool setups.
• ›Adds pluggable operations interfaces (ReadOperations, WriteOperations, BashOperations, GrepOperations, etc.) for routing built-in tool calls over SSH or other transports.
• ›Adds user_bash event so extensions can intercept and redirect user !/!! shell commands to remote systems.
• ›Adds setActiveTools() to ExtensionAPI for dynamically managing the active tool set at runtime.
• ›Adds ctx.ui.getAllThemes(), ctx.ui.getTheme(name), and ctx.ui.setTheme(name | Theme) for extensions to list, load, and switch themes at runtime.

$ git clone --branch v0.38.0 https://github.com/earendil-works/pi.git
# already have the repo? check out this version:
$ git checkout v0.38.0

$ pi --no-extensions -e ./extensions/my-tool.ts

$ PI_SKIP_VERSION_CHECK=1 pi 'summarize the latest alerts'

{
  "thinkingBudgets": {
    "low": 1024,
    "medium": 8192,
    "high": 32768
  }
}

Qwen Code v0.7.0 adds multi-provider model configuration and changes user vs. workspace settings selection order.

$ git clone --branch v0.7.0-nightly.20260108.f776075a https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:
$ git checkout v0.7.0-nightly.20260108.f776075a

• ›Adds multi-provider models config support, allowing configuration of multiple model providers within a single Qwen Code setup.
• ›Changes the selection order between user settings and workspace settings, altering which takes precedence during resolution.

llama.cpp b7678 adds initial FlashAttention implementation for the WebGPU backend.

$ git clone --branch b7678 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:
$ git checkout b7678

• ›Adds initial FlashAttention implementation to the WebGPU (ggml-webgpu) backend, enabling attention computation on GPU via the WebGPU API.
• ›Adds fast matrix and matrix-vector multiplication kernels to the WebGPU backend, improving throughput for inference on WebGPU-capable devices.
• ›Adds Q4_0 quantized matrix multiplication support to the WebGPU backend.
• ›Adds subgroup matrix (cooperative matrix) shader support to the WebGPU backend for improved GPU utilization where the hardware capability is present.
• ›Adds F16 accumulation with shared-memory staging to WebGPU attention and matmul kernels.

$ git clone --branch b7672 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:
$ git checkout b7672

$ git clone --branch b7668 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:
$ git checkout b7668

oobabooga textgen v3.23 improves table and separator styling in chat messages.

$ git clone --branch v3.23 https://github.com/oobabooga/textgen.git
# already have the repo? check out this version:
$ git checkout v3.23

• ›Improves the visual style of tables and horizontal separators rendered inside chat messages.

Phoenix Evals 2.8.0 adds a correctness evaluator, tool-selection correctness metric, and sync/async LLM client kwargs support.

$ git clone --branch arize-phoenix-evals-v2.8.0 https://github.com/Arize-ai/phoenix.git
# already have the repo? check out this version:
$ git checkout arize-phoenix-evals-v2.8.0

from phoenix.evals import OpenAIModel

model = OpenAIModel(
    model="gpt-4o",
    sync_client_kwargs={"timeout": 30},
    async_client_kwargs={"timeout": 60},
)

• ›Adds sync_client_kwargs and async_client_kwargs support to the LLM constructor, enabling per-client configuration for synchronous and asynchronous calls.
• ›Adds a built-in tool selection correctness metric to evaluate whether an LLM agent chose the right tool.
• ›Adds a new correctness evaluator for assessing the accuracy of LLM outputs.

Adds CLICKHOUSE_ASYNC_INSERT_BUSY_TIMEOUT_MIN_MS setting and high-cardinality measure support in the v2 metrics API.

$ git clone --branch v3.146.0 https://github.com/langfuse/langfuse.git
# already have the repo? check out this version:
$ git checkout v3.146.0

$ CLICKHOUSE_ASYNC_INSERT_BUSY_TIMEOUT_MIN_MS=500

• ›Adds CLICKHOUSE_ASYNC_INSERT_BUSY_TIMEOUT_MIN_MS environment variable to tune ClickHouse async insert busy-timeout behavior.
• ›Allows high-cardinality measures in v2/metrics API endpoint when using topN queries.

Weave v0.52.23 adds logfire/pydantic-ai parsing, annotation queue APIs, and broader LangChain autopatching.

$ git clone --branch v0.52.23 https://github.com/wandb/weave.git
# already have the repo? check out this version:
$ git checkout v0.52.23

• ›Adds new APIs for annotation queue management, including queue stats (Annotation Queues Stats API) and querying queue items, backed by a new database migration for a queue-based call annotation system.
• ›Adds input and output parsing for logfire pydantic-ai instrumentation, enabling structured trace data from pydantic-ai spans.
• ›Improves autopatching for common LangChain imports, broadening automatic tracing coverage.
• ›Adds GPT-5.2 and gpt-image-1.5 models to the playground.
• ›Allows the client to enforce a minimum trace server version for compatibility checks.

$ git clone --branch v0.52.22 https://github.com/wandb/weave.git
# already have the repo? check out this version:
$ git checkout v0.52.22

Milvus 2.5.25 adds configurable metadata batch processing and variable-length field size estimation options.

$ git clone --branch v2.5.25 https://github.com/milvus-io/milvus.git
# already have the repo? check out this version:
$ git checkout v2.5.25

• ›Improves reliability of object storage operations under high load with automatic retry on rate-limit errors.