Run all ACP harnesses on persistent live sessions

[xeophon]

Summary

This makes persistent LiveACP the single v1 execution model for every ACP-backed harness.

• Run Claude Code, Codex, Hermes, Kimi Code, OpenClaw, Pi, Pool, and RLM through the shared ACPHarness lifecycle.
• Keep one ACP process, connection, and native agent session alive for the complete rollout.
• Remove the one-shot ACP runner, file-backed session_path continuation, and stateless fallback.
• Make harnesses declarative: each provides its command, environment, prompt, MCP ownership, and cleanup while the shared layer owns process and session lifetime.
• Preserve exact provider replay state and require the resumed tool-use E2E to remain a single trace branch.
• Keep ACP prompts non-empty and user-only; stacked PR Add native ACP session import #2348 adds capability-negotiated native history import instead of flattening prior assistant or system turns into a user prompt.
• Isolate OpenClaw's writable skills per rollout while retaining the uploaded tree as an immutable staging cache.

Why

The harnesses previously mixed live sessions, one-shot processes, and file-backed resume. Rebuilding a conversation at those boundaries changed model-visible history and produced false branches even when the agent completed the task correctly.

With LiveACP as the common backend, every interaction segment continues on the same native session. The graph can therefore represent one append-only conversation, and harness-specific code no longer needs to emulate continuation.

Provider and agent differences remain local to their harnesses: Kimi and Pi select their transport, Hermes resolves its native catalog, OpenClaw supervises its Gateway, and Pool now uses the same live lifecycle as the other ACP agents.

Stack

• Run all ACP harnesses on persistent live sessions #2291: persistent LiveACP lifecycle and one-branch continuation
• Add native ACP session import #2348: native session import for pre-existing VF history

Validation

• 70 non-E2E v1 tests
• Prime VM ACP resume validation, including Pool, RLM, and OpenClaw
• OpenClaw per-rollout skill-isolation probe
• Ruff check and format
• ty pre-push parity check
• git diff --check

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[macroscopeapp]

Approvability

Verdict: Needs human review

Major refactor replacing the ACP composition pattern with inheritance-based ACPHarness, removing one-shot launch capability, and restructuring 8+ harness implementations. Changes to shared infrastructure abstractions warrant human review.

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

[cursor]

Cursor Bugbot has reviewed your changes and found 3 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 9c027ec. Configure here.}