fix(agent): honor AgentToolResult.terminate

[d8rt8v]

Summary

Honor AgentToolResult.terminate in the agent loop.

The loop now collects finalized results after the complete tool-call batch has run and skips the automatic follow-up model turn only when every result requests termination. Results modified by after_tool_call are respected, while mixed batches continue normally.

Reproduction

Before this change, a custom terminal tool could return:

return AgentToolResult(
    content=[TextContent(text="result accepted")],
    terminate=True,
)

The tool completed successfully, but the loop set has_more_tools from the presence of tool calls alone and made another provider request. With max_turns=1, the attempted follow-up also turned a successful run into Agent stopped after max_turns=1.

This is the same early-termination behavior fixed in Pi:

• pi-agent-core: Early termination hints earendil-works/pi#3525
• https://github.com/earendil-works/pi/blob/v0.80.6/packages/agent/src/agent-loop.ts#L584-L600

Validation

• uv run pytest
• uv run ruff check .
• uv run ruff format --check .
• uv run mypy

[Copilot]

Pull request overview

Updates Tau’s provider/tool agent loop to respect AgentToolResult.terminate, aligning early-termination behavior with Pi and preventing unnecessary follow-up model turns (especially under tight max_turns limits).

Changes:

• Collects finalized tool execution results (ToolExecutionEndEvent.result) across the full tool-call batch.
• Skips the automatic follow-up model turn only when all tool results explicitly request termination (terminate is True), while mixed batches continue normally.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.