Handle late ACP session updates

[hallerite]

Summary

• wait briefly for a late ACP session update when session/prompt returns without visible output
• notify the prompt waiter when message or tool-call output changes
• preserve the existing empty-tool-turn handling behind allow_empty_tool_reply

Root cause

agent-client-protocol 0.11 dispatches notifications through background tasks but resolves request responses directly in the receive loop. If an agent sends its final session/update immediately before its session/prompt response, the prompt coroutine can resume before the update handler records the visible reply.

This adds a narrowly scoped one-second compatibility grace period only when the prompt would otherwise fail for having no visible output. It can be removed after the upstream SDK provides ordered notification handling or a completion barrier.

Impact

ACP agents such as RLM no longer fail spuriously with ACP agent produced no visible reply when their final update races the prompt response.

Validation

• uv run pytest tests/v1/ -m 'not e2e' — 69 passed, 70 deselected
• full non-e2e suite — 913 passed, 70 deselected
• uv run pre-commit run --all-files — passed
• synthetic ACP 0.11 notification/response race reproduction — passed

Live Docker/Prime ACP end-to-end tests were unavailable locally because Docker and PRIME_API_KEY were not configured.

---

Note

Low Risk
Narrow compatibility wait in the ACP runner with a 1s cap; no auth or data-path changes, and existing empty-tool-turn handling is preserved.

Overview
Fixes spurious ACP agent produced no visible reply failures when session/prompt completes before the final session/update is applied (ACP 0.11 dispatches notifications in background tasks).

VerifiersACPClient now signals an asyncio.Condition on message and tool-call updates. After connection.prompt returns, if there is still no visible text, the runner waits up to 1 second for a late text chunk (tool-only updates do not count as a reply). allow_empty_tool_reply behavior is unchanged.

The ACP resume e2e test no longer requires trace.tools; it documents that native MCP tools may not appear on intercepted model requests and that the ACP transcript is authoritative.

^{Reviewed by Cursor Bugbot for commit b31c0e1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Handle late ACP session updates by waiting up to 1 second for visible text

• Adds a LATE_UPDATE_GRACE_SECONDS (1.0s) bounded wait in the prompt function in runner.py after receiving the initial response, allowing late-arriving visible text dispatched in background tasks to be collected before deciding the agent produced no reply.
• Introduces an asyncio.Condition (output_changed) on VerifiersACPClient so session_update can notify waiters when new visible text or tool calls arrive.
• Relaxes the test_acp_resume_with_tool assertion in test_e2e.py to no longer require trace.tools to be non-empty, relying on the ACP transcript as the source of truth instead.
• Behavioral Change: prompt now waits up to 1 second longer per call when no visible reply has been received yet, reducing premature RuntimeError for valid late replies.

📊 Macroscope summarized b31c0e1. 1 file reviewed, 0 issues evaluated, 0 issues filtered, 0 comments posted

🗂️ Filtered Issues

No issues evaluated.

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 93d26d7. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs human review

2 blocking correctness issues found. This PR introduces new async synchronization logic and timing behavior. Unresolved review comments identify a high-severity potential race condition that could cause cross-turn data leakage, plus an unnecessary 1-second delay for tool-only turns. These substantive issues warrant human review.

^{You can customize Macroscope's approvability policy. Learn more.}