fix(core): filter tool calls by origin for turn-based assertions

[tedjames]

Summary

• Fix tool call assertions to only count current turn tool calls, not from entire conversation history
• Filter out tool calls with origin: 'history' from the usage events
• Include tool calls with origin: 'current' or undefined (for backward compatibility)
• Add comprehensive tests for origin filtering

Problem

When running multi-turn conversation tests, tool call assertions counted tool calls from all turns in the conversation history, not just the current turn. This caused false test failures:

steps:
  # Turn 1: Create a task
  - chat: "Create a task called 'original title'"
    tools:
      manageTasks: 1  # ✅ Passed - 1 call made

  # Turn 2: Update the task
  - chat: "Change the title to 'updated title'"
    tools:
      manageTasks: 1  # ❌ FAILED - counted 2 calls (1 from history + 1 current)

Solution

The API returns usage events with an origin field:

• origin: "history" - Tool calls from previous turns
• origin: "current" - Tool calls made in the current turn

Added filtering in http-agent.ts to only include current-turn tool calls.

Test plan

• Added test: filter out tool calls with origin: 'history'
• Added test: include tool calls with origin: 'current'
• Added test: include tool calls with no origin (backward compatibility)
• Added test: handle multi-turn conversation with mixed origins
• All existing tests pass

🤖 Generated with Claude Code