feat(v1): run RLM over persistent ACP

[hallerite]

Keeps ACP agents alive for the duration of a v1 interaction and moves RLM onto its native ACP transport.

RLM now preserves conversation and IPython state across segments, while MCP servers remain model-facing Python skills behind ipython.

Tested with:

• uv run pytest tests/ — 909 passed, 64 skipped
• uv run pre-commit run --all-files
• DeepSeek V4 Flash in a Prime sandbox — two ACP segments, resumed MCP call, reward 1.0, finalized RLM metrics

Note

Run RLM harness over a persistent ACP sidecar per trace

• The RLM harness now launches an ACP agent (rlm --acp) backed by a long-lived Unix-socket sidecar per trace instead of invoking the CLI in one-shot mode per prompt.
• _runner.py gains a PersistentSession class and a serve_sidecar Unix-socket server that serializes prompts over a single agent process and session across segments.
• acp/__init__.py extends ACP.run with a sidecar_path parameter that starts the sidecar on first use, routes subsequent prompts through it, and appends sidecar logs to stderr on failure; ACP.close shuts the sidecar down and removes its state directory.
• Per-trace RLM_HOME is now stored under .vf-rlm/<trace.id>/home, enabling meta.json metrics (e.g. turns_since_last_compaction) to persist and be read after the sidecar is stopped.
• Risk: the sidecar process outlives individual prompts; a crash or stale socket requires the cleanup path in RLMHarness.cleanup to recover correctly.

^{Macroscope summarized 7caf113.}

---

Note

Medium Risk
New long-lived sidecar lifecycle and RLM execution path affect multi-segment rollouts; client disconnect during a prompt tears down the session and can discard in-progress work.

Overview
RLM now runs through rlm-harness on native ACP (rlm --acp) instead of a one-shot CLI invocation. MCP servers are wired through ACP; the old RLM_MCP_CONFIG path is removed. Per-trace state lives under .vf-rlm/{trace.id}/ (RLM_HOME and a private acp.sock sidecar).

The shared ACP layer gains an optional persistent sidecar: probe/start a background serve process, send each segment via request, and close/shutdown with log tailing on failures. _runner.py adds a UNIX-socket server with PersistentSession (one agent process for many prompts), plus once / serve / request / shutdown / probe commands.

E2E adds rlm to ACP resume coverage and asserts RLM compaction metrics on resumed runs. Docs pin the default rlm-harness git ref and describe the ACP + IPython skill model.

^{Reviewed by Cursor Bugbot for commit 7caf113. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR introduces significant new async/concurrency functionality for persistent ACP sessions. An unresolved high-severity review comment identifies a race condition in the disconnect handling logic where the session may close while an ACP prompt is still unwinding.

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

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

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

^{Reviewed by Cursor Bugbot for commit 7caf113. Configure here.}

[cursor]

Disconnect cancels prompt tracking early

High Severity

On client disconnect, the handler cancels prompt_task (run_prompt) before stop_session. That hits run_prompt's finally, which clears active_prompt without waiting for session.run to finish, so stop_session no longer sees the in-flight prompt and can close the session while the ACP prompt is still unwinding—the race this shutdown path was meant to prevent.

Additional Locations (1)

• verifiers/v1/acp/_runner.py#L328-L350

 

^{Reviewed by Cursor Bugbot for commit 7caf113. Configure here.}