Unregister a specific global before_llm_call hook. Args: hook: The hook function to remove Returns: True if the hook was fou

Return whether provider/API should use OpenAI Responses file support.

Validate that a directory name matches the skill name. Args: skill_dir: Path to the skill directory. skill_name: The declared ski

Validate a PDF file against constraints. Args: file: The PDF file to validate. constraints: PDF constraints to validate against.

Return True when two skill versions refer to the same version. Versions arrive as ``1.2.0`` or ``v1.2.0`` depending on whether they came from

Wait for batch to be initialized. Args: timeout: Maximum time to wait in seconds (default: 2.0) Returns: Tru

Create a new Skill at a different disclosure level. Args: level: The new disclosure level. instructions: Optional ins

Support the descriptor protocol for method binding. This allows the wrapped method to be properly bound to an instance when accessed

Initialize the CustomLLM with a predefined response. Args: response: The predefined response to return from call().

Async call to OpenAI Responses API.

Add propertyOrdering to JSON schema for Gemini 2.0 compatibility. Gemini 2.0 models require an explicit propertyOrdering list to define

Add the internal loader used for request-scoped skill disclosure.

Async version of _delegate_to_a2a.

Execute a task without a timeout asynchronously. Args: task_prompt: The prompt to send to the agent. task: The task b

Async equivalent of ``_export_output`` — uses ``acall`` so the event loop is not blocked.

Cached async implementation of AgentCard fetching.

Fetch agent cards concurrently for multiple A2A agents using asyncio.

Handle async non-streaming converse API call.

Process human feedback asynchronously and refine answer. Args: formatted_answer: Initial agent result. Returns:

Async inject files into the last user message.

Invoke the agent loop asynchronously and return the result.

Execute agent loop asynchronously using ReAct text-based pattern. Returns: Final answer from the agent.

Invoke step callback (async context). Args: formatted_answer: Current agent response.

Add text tool-calling instructions after native tools are rejected.

Query the LLM for a specific field value based on its description. Args: field_name: The name of the field to extract.

Summarize multiple message chunks concurrently using asyncio. Args: chunks: List of message chunks to summarize. llm: LLM instanc

Wire an ``LLMCallCompletedEvent`` listener for the duration of one ``kickoff_async`` call.

Run the encoding pipeline in a background thread with event emission. Both started and completed events are emitted here (in the background

Emit the start event and return the invariant context for a checkpoint.

Build an isolated execution context for a single todo. Passes only final results from completed dependencies — never execution traces

Build the LLM-facing description, including the current per-call cap. Args: max_seconds: The per-call cap advertised to the model. R

Build the execution prompt, stop words, and RPM limit function. Args: raw_tools: The raw tools available to the agent. R

Build context by reading file once. Args: file: The file to build context for. Returns: FileContext with cac

Build fingerprint metadata as a config dict for tool invocation. Returns the fingerprint data in a config dict rather than injecting it

Build a multi-turn conversation that produces multiple chunks at max_tokens=200. Each non-system message is ~100-140 estimated tokens (400-560 ch

Inner builder shared by the public entry point and recursive nested-object creation. Preprocessing via ``jsonref.replace_refs`` and the sanitizat

Build a LiteAgentOutput from an executor result dict. Shared logic used by both sync and async execution paths. Args: re

Build a timestamped checkpoint file path under a branch subdirectory. Filename format: ``{ts}_{uuid8}_p-{parent_id}.json`` Args: dir

Mimic a litellm ModelResponse that contains both content and tool_calls.

Cache key for a provider's resolved model list. Keyed by the exact API key (via a short, non-reversible digest — never the key itself), so sw

Call provided synchronous handlers. Args: source: The emitting object event: The event instance handlers:

Calls the LLM with the planning prompt. Args: prompt: The prompt to send to the LLM. plan_type: The type of plan bein

Call OpenAI Responses API.

Check if an execution error should be re-raised immediately. Args: e: The exception that occurred. task: The task bei

Warn when the lockfile pins a crewai release older than 1.13.0, which is where ``_load_response_format`` was introduced.

Utility method to split text into chunks.

Utility method to split text into chunks.

Utility method to split text into chunks.

Utility method to split text into chunks.

Utility method to split text into chunks.

Utility method to split text into chunks.

Three-chunk stream mirroring LiteLLM's include_usage behavior: two content chunks where the second carries finish_reason="stop", then a final

Classify a Gemini exception as transient or permanent upload error. Checks Gemini-specific error message patterns first, then falls back to g

Classify an S3 exception as transient or permanent upload error. Args: e: The exception to classify. filename: The filename for e

Clean up batch data after successful finalization to free memory

Cleanup MCP client connections after task execution.

Drop the header's active method once it ends. Caller holds the lock. Falls back to another still-active step (methods can overlap) so the

Validate an LLM response into the configured Pydantic model. Pure post-processing — performs no I/O. Shared by ``to_pydantic`` and ``

Collect CEL root identifiers, excluding receiver macro local variables.

Append a single line with soft JSON syntax highlighting.

Combine agent and crew knowledge contexts into a single string. Args: agent: The agent with knowledge contexts. Returns: Com

Shared strict sanitization: inline refs, close objects, require all properties.

Fire-and-forget: compact the table in a daemon background thread.

Mark the final reviewed answer as the completed executor state.

Compute exponential backoff delay with jitter. Args: attempt: The current attempt number (0-indexed). Returns: Delay in seco