bianhaifeng789-hue/dangling-tool-call

Dangling Tool Call Skill

Overview

A dangling tool call occurs when:

• AIMessage contains tool_calls
• No corresponding ToolMessage in history
• Causes LLM errors due to incomplete message format

This skill patches such gaps by inserting synthetic ToolMessages with error indicators.

Problem Scenario

User: "Generate a report"
AI: tool_calls=[read_file, grep, analyze]  ← interrupted!
User: "Stop, cancel"
[No ToolMessages for the tool calls]

Next LLM call: ERROR - incomplete tool call sequence

Solution

Insert placeholder ToolMessages immediately after dangling AIMessage:

ToolMessage({
  content: "[Tool call was interrupted and did not return a result.]",
  tool_call_id: tc_id,
  name: tc.get("name", "unknown"),
  status: "error"
})

Detection Algorithm

function detectDangling(messages) {
  // 1. Collect existing ToolMessage IDs
  const existingToolIds = new Set(
    messages.filter(m => m.type === 'tool').map(m => m.tool_call_id)
  );

  // 2. Find AIMessages with missing ToolMessages
  const dangling = messages.filter(m => m.type === 'ai')
    .flatMap(m => m.tool_calls || [])
    .filter(tc => tc.id && !existingToolIds.has(tc.id));

  return dangling;
}

Patching Strategy

• Insert immediately after the dangling AIMessage (not append to end)
• Use wrap_model_call instead of before_model for correct positioning
• Log warning with patch count

Integration Points

1. wrap_model_call - Intercept before LLM call
2. Build patched messages - Insert synthetic ToolMessages
3. Preserve GraphBubbleUp - Don't catch LangGraph control signals

Implementation Script

See impl/bin/dangling-tool-patcher.js for Node.js implementation.

OpenClaw Integration

This skill integrates with:

• Message history validation
• Tool call sequence integrity
• Session recovery after interruption

Status

• ✅ Skill documentation created
• 🔜 Implementation script pending