feat(v1): add trainable Prime Agent harness

[hallerite]

Summary

• add a thin Prime Agent harness on top of the shared ACP harness
• install a pinned Prime Agent release and route its OpenAI-compatible model calls through Verifiers interception
• support ACP MCP servers, skills, resume, and appended system prompts
• isolate per-trace Prime Agent state and daemon sockets while retaining reusable runtime caches
• run Prime Agent evaluations in autonomous mode by default, with an explicit opt-out
• optionally require Prime Agent's correlated ACP lifecycle contract and record terminal quiescence separately from benchmark rewards
• isolate the main-agent answer from subagent, compaction, and refinement updates

Why

Prime Agent's native model API uses streaming Chat Completions. The harness itself can stay small because the generic renderer-backed streaming bridge now lives in the parent PR, #2369. This PR contains only the Prime Agent integration.

With Prime Agent #1378, ACP-provided MCP tools are translated into session-scoped Python programs by Prime Agent itself. Verifiers supplies the environment's MCP URLs while Prime Agent owns its native tool representation.

Stack and dependency

• based on Support training with streaming harnesses #2369, which enables buffered training responses for streaming harnesses
• depends on Prime Agent #1378
• currently pins prospective Prime Agent version 0.7.3; align the pin with the release containing chore(deps): bump gitpython from 3.1.47 to 3.1.50 in /packages/verifiers-rl #1378 before merge
• lifecycle enforcement is explicit opt-in (require_terminal_quiescence=false by default), preserving compatibility with 0.7.3
• lifecycle-enabled evaluations require the Prime Agent contract in Prime Agent #1494

Validation

• focused Ruff and formatting checks passed
• uv run --frozen ty check verifiers passed
• focused ACP lifecycle/scoring/config/import suite — 42 passed
• uv run pytest tests/v1 -m 'not e2e' — 70 passed
• isolated Docker MCP evaluation — reward 1.0
• real Prime VM + Prime Inference MCP evaluation — reward 1.0; generated MCP Python program executed successfully, no trace errors, sandbox teardown verified

The lifecycle commits were added append-only. Local hooks that invoke uv --locked were skipped because the existing branch lockfile is stale; the equivalent direct frozen Ruff, formatting, type, and focused test commands above passed.

Note

Add trainable Prime Agent harness with ACP lifecycle enforcement

• Adds PrimeAgentHarness and PrimeAgentHarnessConfig in harness.py, which installs Prime Agent, prepares per-trace state directories, emits models.json, and launches it in ACP mode with configurable autonomous/skills flags
• Overhauls the ACP runner in runner.py: VerifiersACPClient now tracks lifecycle metadata per turn, filters visible replies to current-turn answer chunks, and exposes stop_reason, response_boundary, and terminal_quiescence; prompt returns a structured dict instead of a plain string and enforces correlated lifecycle semantics when configured
• Adds lifecycle_meta_namespace to ACPConfig and _record_lifecycle_status in acp/init.py so lifecycle outcomes are recorded on Trace.info without affecting rewards; successful autonomous completions populate acp_answer_fallback
• Updates read_answer_file_or_last_reply in score.py to prefer the ACP-recorded fallback reply over last_reply
• Behavioral Change: prompt in runner.py now returns a dict (was a string); ACPSession.run and serve_stream propagate this structured type. In-tree callers are updated; out-of-tree consumers expecting a string return will break

Changes since #2355 opened

• Added autonomous mode budget parameters to Prime Agent CLI arguments [e494cd9]
• Implemented incomplete lifecycle status recording for cancelled and failed turns [e494cd9]
• Implemented selective process stopping based on turn lock acquisition state [e494cd9]
• Extended lifecycle status recording with terminal quiescence and phase tracking fields [e494cd9]

^{Macroscope summarized 8a869eb.}

---

Note

^{Cursor Bugbot is generating a summary for commit 8a869eb. Configure here.}

[sethkarten]

Added the generic Prime Agent ACP lifecycle consumer onto this PR at 547abf168.

• Lifecycle enforcement is explicit opt-in, so the pinned 0.7.3 default remains compatible.
• Compatible agents correlate responseBoundary and terminalQuiescence across prompt turns with monotonic event sequencing.
• Rejected prompt admissions are distinguished from accepted turns that still owe terminal settlement.
• Main-agent answer fallback excludes subagent, compaction, refinement, and foreign-turn updates.
• Lifecycle/infrastructure status is recorded independently from benchmark rewards.
• No OOLONG-specific dataset, task selection, sandbox tuning, retry, or workflow changes are included.

Paired producer contract: Prime Agent #1494 at 095ba1e5d.

Validation: frozen Ruff/format and ty pass; focused lifecycle/scoring/config/import suite passes 42/42; GitHub Ruff, Ty, and Macroscope checks are green; zero unresolved review threads.

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This PR adds a new Prime Agent harness with substantial new runtime behavior: lifecycle state management, terminal quiescence coordination, and modified answer resolution. The scope of new async logic and behavioral changes warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[mikasenghaas]

can remove this and instead add prime agent into e2e test suite

[mikasenghaas]

cant have prime agent specific code in a generic acp runner

[mikasenghaas]

feels like a lot of addition to the acp runner but dont know this part of the code super well so will let @hallerite judge

[hallerite]

yes feels excessive. I will take a deeper look.

[mikasenghaas]

what is this? does not seem like smth users should set?

[sethkarten]

autonomous mode forwards the budget given by verifiers so it is probably ok to only be true instead of user opt-out