langroid releases

Python 3.13 support and Task.init() memory leak fix

Python 3.13 support

PR [#1011](https://github.com/langroid/langroid/pull/1011) bumps the supported Python range to >=3.10,<3.14 and adds Python 3.13 to the CI matrix alongside 3.11, so Langroid is now tested on both versions on every push.

• pyproject.toml: requires-python = "<3.14,>=3.10"
• .github/workflows/validate.yml: lint/test job now runs against ["3.11", "3.13"]

Fix ObjectRegistry memory leak in Task.init()

PR [#1022](https://github.com/langroid/langroid/pull/1022) (originally [#1021](https://github.com/langroid/langroid/pull/1021)) fixes a memory leak in Task.init(). The method built a temporary ChatDocument solely to feed log_message(Entity.SYSTEM, ...), but ChatDocument.__init__ auto-registers every instance in the class-level ObjectRegistry, so each new Task left behind a system-message doc that was never freed. In long-running services that spawn many tasks (e.g. servers), this accumulated significantly over time.

The fix:

• explicitly removes the temporary doc via ChatDocument.delete_id(...) after logging
• renames the local to system_message_temp_doc to make its transient nature obvious
• adds a regression test (test_task_init_no_registry_leak) mirroring the PR #939 leak-check pattern, asserting that repeated Task(...).init() calls leave zero leaked ChatDocuments in the registry

Thanks to @alexagr for the catch and the original fix.

Thread-safe client cache with LRU eviction

PR #993 enhances the client cache in langroid/language_models/client_cache.py with thread-safety and LRU (Least Recently Used) eviction.

What's New

• Thread safety: All cache operations are now protected by a threading.RLock() to prevent race conditions in multi-threaded environments
• LRU tracking: Cache entries store a last-used monotonic timestamp, refreshed on each access
• prune_cache(max_age_seconds): New function to evict stale client entries older than the specified age
• Helper functions: _get_cached_client() and _store_client() encapsulate cache access with consistent locking and timestamp management

Applies to all cached client getters: get_openai_client(), get_async_openai_client(), get_groq_client(), get_async_groq_client(), get_cerebras_client(), and get_async_cerebras_client().

Configurable context overflow strategy (#967, #974)

Added a context_overflow_strategy option to ChatAgentConfig for handling message history that exceeds the model's context length. Two strategies are available:

• "truncate" (default): Truncates content of early messages while preserving all messages in the sequence. Maintains backward compatibility and the alternating message structure required by LLM APIs.
• "drop_turns": Drops complete conversation turns (a USER message and all responses until the next USER message). More aggressive but cleaner — particularly useful for voice agents with limited context models (e.g., llama-3-8b with 8192 tokens), where individual messages are already short and truncation is ineffective.

config = lr.ChatAgentConfig(
    context_overflow_strategy="drop_turns",  # or "truncate" (default)
    ...
)

Cleaner OpenAI API error logging (#975)

OpenAI API errors (authentication, bad request, rate limits, etc.) are now logged without a full Python traceback, since these errors originate server-side and a local stack trace adds no diagnostic value. Network-level errors (APIConnectionError, APITimeoutError) still include the full traceback to aid in diagnosing local issues.

Minor fixes

• Fixed defensive check for empty/missing choices in OpenAI API response parsing (#975)
• Formatting cleanup in inline reasoning tests (#977)
• Updated llms*.txt documentation files

Support inline reasoning in streaming responses (#973)

Added support for handling inline reasoning content in streaming LLM responses, for models that embed thinking/reasoning inside the main content stream (e.g., using <think>...</think> delimiters) rather than in a separate reasoning field.

Key Changes

• Added extra_content field to OpenAIToolCall: New optional field to store additional metadata from tool calls, with updated from_dict() and api_dict() methods
• Implemented inline reasoning splitting: New _split_inline_reasoning() static method that separates reasoning tokens from text tokens based on configurable delimiters, tracking state across streaming chunks
• Enhanced streaming event processing: Updated _process_stream_event() and _process_stream_event_async() to route split tokens to correct streamers (TEXT vs REASONING) with proper state tracking
• Updated Ruff: Bumped pre-commit ruff version from v0.14.14 to v0.15.0

Add reasoning parameter to LLM response callbacks (#965)

This release adds support for passing chain-of-thought reasoning from LLMs (like DeepSeek R1, Claude extended thinking) to UI callbacks.

Features

• Reasoning in callbacks: show_llm_response and finish_llm_stream callbacks now receive a reasoning parameter containing the LLM's chain-of-thought reasoning (when available)

• Chainlit integration: Reasoning is automatically displayed as a nested message with a "💭 Reasoning" label in Chainlit UIs

• Backward compatible: Existing custom callbacks without the reasoning parameter continue to work - uses signature inspection to only pass reasoning if supported

Documentation

• Added "Displaying Reasoning in UI Callbacks" section to reasoning-content.md with examples for custom callback implementations

Thanks to @alexagr for the initial PR adding the reasoning parameter!

Fixed

• Respect user-provided API parameters: Removed overly aggressive parameter filtering that silently dropped params like reasoning_effort for models added to MODEL_INFO. The library now trusts user configuration and lets the API validate parameter support. (#956 - thanks @alexagr)

What's Changed

Improvements

Callback API Enhancement: Separate content and tools_content (#952, #945)

The show_llm_response and finish_llm_stream callbacks now include separate content and tools_content parameters:

• content: Always contains the text message generated by the model
• tools_content: Contains serialized functions/tools if present, empty string otherwise

Why this matters: Previously, content mixed text messages with JSON-serialized tool calls. This caused issues for applications using callbacks for purposes like text-to-speech, where tool call JSON was incorrectly processed as regular text.

Backward compatible: The tools_content parameter defaults to an empty string, so existing callback implementations continue to work.

Full Changelog: https://github.com/langroid/langroid/compare/0.59.25...0.59.26

What's Changed

• Fix: Replace deprecated datetime.utcnow() with datetime.now(timezone.utc) to remove DeprecationWarnings in Python 3.12+ (#937)