feat(v1): add Prime Agent ACP harness

[hallerite]

Summary

• add Prime Agent over the same generic typed ACP prompt lifecycle as nano-RLM
• pass --autonomous explicitly from harness configuration
• preserve MCP programs, resume state, skills, Chat reasoning lineage, and isolated per-rollout state
• trust Prime Agent's completed session/prompt boundary from Prime Agent PR #1612
• record response-boundary and terminal-quiescence telemetry for evidence without making it a second completion gate
• enforce single-branch lineage for the ordinary two-turn resume case

Stack

PR 3 of the three-PR Verifiers stack. This targets nano-RLM integration PR #2439, which targets typed ACP base PR #2438.

Prime Agent PR #1612 is merged and included in the pinned beta used here. It makes session/prompt wait for recursive terminal quiescence and returns the final autonomous stop reason.

Validation

• full Verifiers suite on the combined stack — 919 passed, 77 credential-gated skips
• Prime Agent Prime VM MCP/resume E2E — passed against 0.8.0-beta.548.1.9bc0055
• single-branch assertion enabled and passed for the resumed conversation
• Seth's full feature canary against current Prime Agent main (9bc0055) — passed
  • reward 1.0; all 13 feature checks true
  • retained-child follow-up acknowledged and child deleted
  • terminal quiescence after response boundary
  • zero outstanding subagents and autonomous continuations
  • Docker container removed
• changed-file pre-commit hooks and push-hook CI-parity type check

---

Note

Medium Risk
Introduces network-fetched agent binaries and a new ACP harness on the eval critical path; the RLM default pin change can shift behavior for existing RLM rollouts.

Overview
Adds a Prime Agent v1 harness that runs Prime Agent in native ACP mode on the shared ACPHarness path, with MCP, resume, skills, and optional autonomous continuation via config.

The harness bootstraps a pinned beta from GitHub releases (checksum-verified tarballs, global npm install under a commit-keyed prefix), wires the intercept provider to the eval endpoint, and keeps per-rollout state (models config, daemon socket, wrapper). After each ACP turn it records lifecycle telemetry (responseBoundary, terminalQuiescence, infrastructure status) under trace.info["acp_lifecycle"] for evidence without gating completion on it.

E2E gains a Prime VM row for prime-agent in the ACP resume+MCP test, asserting two turns show OK infrastructure and zero outstanding subagents at terminal quiescence. nano-RLM’s default install ref is bumped to a newer commit.

^{Reviewed by Cursor Bugbot for commit 9af8996. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This PR adds a substantial Prime Agent ACP harness, including external package provisioning, live process management, per-rollout state, MCP/resume support, and optional autonomous execution. It also updates the default nano-RLM dependency used by existing runs, creating runtime effects beyond a small isolated change.

Not approved because:

• Monthly spending limit reached (workspace setting). Approvability relies on correctness review in order to determine eligibility

No code changes detected at 9af8996. Prior analysis still applies.

^{Review your spending limits in Billing settings. You can add or adjust custom eligibility rules. 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 10fcab6. Configure here.}