~/wiki
Récent+ Ajouter

LLM PauseChain

Confiance : high
llm-pausechainexception-handlingworkflow-controltool-callinghuman-in-the-looppause-resumestate-preservationllm-librarydatasette-agentclaude-fable-integrationworkflow-orchestrationexecution-interruptionapproval-workflows

Exception mechanism in the llm-library that enables tools to cleanly pause execution chains mid-process, particularly useful for implementing human-in-the-loop approval workflows. Introduced in version 0.32a3 as part of advanced tool-calling enhancements.

Core Mechanism

Exception-Based Control:

  • Tools raise llm.PauseChain exception to halt execution
  • Exception propagates to caller with preserved context
  • No model call made with placeholder results when paused
  • Clean interruption without corrupting workflow state

State Preservation:

  • .tool_call attached to exception with current tool context
  • .tool_results contains completed sibling results
  • Full context available for resumption decisions
  • Workflow state remains consistent during pause

Integration with Tool Calling

Tool Implementation: Tools can access the current tool call context through the llm_tool_call parameter:

def approval_required_tool(data, llm_tool_call):
    # Access tool call ID and context
    tool_id = llm_tool_call.tool_call_id
    
    # Raise exception to pause for approval
    if requires_approval(data):
        raise llm.PauseChain("User approval required")

Concurrent Execution Handling:

  • Async sibling tool calls run to completion before pause propagation
  • Robust failure semantics for parallel tool execution
  • Proper handling of mixed success/pause scenarios

Use Cases

Human Approval Workflows:

  • Pausing before executing sensitive operations
  • Requesting user confirmation for data modifications
  • Validating AI-generated decisions before proceeding
  • Implementing multi-step approval processes

datasette-agent Integration: Primary driver for PauseChain development was enabling ask_user() functionality:

  • Pause data analysis workflows for user input
  • Request clarification on ambiguous queries
  • Confirm potentially destructive operations
  • Enable interactive data exploration sessions

Technical Implementation

Exception Propagation:

  1. Tool raises PauseChain exception during execution
  2. Exception bubbles up through tool call infrastructure
  3. Current state preserved with tool call context
  4. Caller receives exception with resumption information
  5. Workflow can be resumed with user input/approval

Unique Tool Call IDs:

  • Every tool call guaranteed unique identifier
  • Providers without native IDs get synthesized tc_-prefixed ULIDs
  • Enables precise tracking of paused operations
  • Supports complex workflow resumption scenarios

Workflow Patterns

Interactive Approval:

1. Tool execution begins
2. Tool determines approval needed
3. PauseChain exception raised with context
4. User presented with approval request
5. User provides approval/denial/modification
6. Workflow resumes with user input

Multi-Step Processes:

  • Complex workflows can pause at multiple points
  • Each pause preserves complete execution context
  • Resumption maintains state consistency
  • Support for branching based on user decisions

Development Impact

The PauseChain mechanism was developed entirely through claude-fable 5's autonomous programming during simon-willison's first-day evaluation. This demonstrates:

  • AI-driven identification of missing infrastructure needs
  • Autonomous implementation of sophisticated control flow mechanisms
  • Integration of human oversight into automated workflows
  • Evolution from hacks to supported library features

See also