verifiers 0.3.2.dev32__tar.gz → 0.3.2.dev41__tar.gz

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev32
+Version: 0.3.2.dev41
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers
@@ -22,8 +22,7 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Classifier: Topic :: Software Development :: Libraries :: Python Modules
 Requires-Python: <3.14,>=3.11
 Requires-Dist: aiohttp>=3.14.1
-Requires-Dist: anthropic>=0.78.0
-Requires-Dist: datasets<6.0.0,>=3.3.0
+Requires-Dist: anthropic>=1.0.0
 Requires-Dist: gepa>=0.0.6
 Requires-Dist: httpx>=0.27.0
 Requires-Dist: loguru>=0.7.0
@@ -31,20 +30,21 @@ Requires-Dist: math-verify>=0.8.0
 Requires-Dist: mcp==2.0.0
 Requires-Dist: msgpack>=1.1.2
 Requires-Dist: numpy>=2.1.0
-Requires-Dist: openai>=2.9.0
+Requires-Dist: openai<3.0.0,>=2.54.0
 Requires-Dist: prime-pydantic-config[toml]>=0.4.3
 Requires-Dist: prime-sandboxes>=0.2.39
 Requires-Dist: prime-tunnel>=0.1.8
 Requires-Dist: pydantic>=2.12.3
 Requires-Dist: pyzmq>=27.1.0
-Requires-Dist: renderers[multimodal]>=0.1.11
+Requires-Dist: renderers[multimodal]>=0.1.12.dev2
 Requires-Dist: rich>=11.0.0
 Requires-Dist: tenacity>=8.5.0
 Requires-Dist: tomli-w>=1.0.0
 Requires-Dist: typing-extensions>=4.12.2
-Requires-Dist: uvloop>=0.21.0; sys_platform != 'win32' and sys_platform != 'cygwin' and platform_python_implementation != 'PyPy'
 Provides-Extra: harbor
 Requires-Dist: harbor==0.21.0; (python_full_version >= '3.12') and extra == 'harbor'
+Provides-Extra: lean
+Requires-Dist: datasets<6.0.0,>=3.3.0; extra == 'lean'
 Provides-Extra: modal
 Requires-Dist: modal>=1.4.0; extra == 'modal'
 Provides-Extra: openenv

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/pyproject.toml

@@ -27,11 +27,10 @@ classifiers = [
 ]
 
 dependencies = [
-    "anthropic>=0.78.0",
-    "datasets>=3.3.0,<6.0.0",
+    "anthropic>=1.0.0",
     "numpy>=2.1.0",
     "math-verify>=0.8.0",
-    "openai>=2.9.0",
+    "openai>=2.54.0,<3.0.0",
     "prime-tunnel>=0.1.8",
     "prime-sandboxes>=0.2.39",
     "pydantic>=2.12.3",
@@ -43,10 +42,9 @@ dependencies = [
     "httpx>=0.27.0",
     "aiohttp>=3.14.1",
     "prime-pydantic-config[toml]>=0.4.3",
-    "uvloop>=0.21.0; sys_platform != 'win32' and sys_platform != 'cygwin' and platform_python_implementation != 'PyPy'",
     "loguru>=0.7.0",
     "tomli-w>=1.0.0",
-    "renderers[multimodal]>=0.1.11",
+    "renderers[multimodal]>=0.1.12.dev2",
     "typing_extensions>=4.12.2",
     "mcp==2.0.0",
 ]
@@ -86,6 +84,9 @@ examples = [
 ]
 
 [project.optional-dependencies]
+lean = [
+    "datasets>=3.3.0,<6.0.0",
+]
 ta = [
     "textarena==0.7.4",
     "nltk>=3.9.2",

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/tests/v1/test_e2e.py

@@ -404,7 +404,7 @@ async def test_rubric_judge(run_v1, tmp_path):
     assert trace.ok
     assert trace.rewards["rubric"].score > 0  # the judge's verdict landed in the reward
     assert trace.metrics["rubric/always_yes"] == 1.0
-    assert trace.info["judge"]  # the call was recorded onto the trace
+    assert trace.info["judge_calls"]  # the call was recorded onto the trace
 
 
 @pytest.mark.e2e
@@ -701,11 +701,29 @@ async def test_replay_round_trip(run_v1, tmp_path):
     assert source.ok
     assert "lcs" in source.rewards
 
+    # Replay must not mix the source run's judge transcript into newly computed
+    # judge calls. Seed a saved call directly because this task scores without a
+    # judge; cleanup happens before task scoring and therefore does not inspect it.
+    import json
+
+    stream = run_dir / "traces.jsonl"
+    record = json.loads(stream.read_text())
+    record["traces"][0]["info"]["judge_calls"] = [
+        {
+            "name": "source-judge",
+            "request": {"model": "source-model", "messages": []},
+            "response": {
+                "message": {"role": "assistant", "content": "stale"},
+                "parsed": None,
+                "usage": None,
+            },
+        }
+    ]
+    stream.write_text(json.dumps(record) + "\n")
+
     async def replay(source_dir: Path, out: Path):
         # The CLI's layering, minus the argv plumbing: the saved run's config is the base
         # (`ReplayConfig` ignores its eval-only keys), the source's output_dir is dropped.
-        import json
-
         data = json.loads(saved_config_path(source_dir).read_text())
         data.pop("output_dir", None)
         config = ReplayConfig(**{**data, "rich": False})
@@ -720,6 +738,7 @@ async def test_replay_round_trip(run_v1, tmp_path):
         # and recomputed the same value.
         assert replayed.rewards.keys() == source.rewards.keys()
         assert replayed.reward == pytest.approx(source.reward)
+        assert "judge_calls" not in replayed.info
     # The wire task keeps its taskset-specific fields in the replay's own output.
     raw = (tmp_path / "replay2" / "traces.jsonl").read_text()
     assert '"answer"' in raw

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/tests/v1/test_judges.py

@@ -92,7 +92,20 @@ def fake_judge_model(monkeypatch):
         if parse is not None:
             response.parsed = parse(response)
         if trace is not None:
-            trace.record_judge(response)
+            from verifiers.v1.dialects.chat import message_to_wire
+
+            wire = (
+                [{"role": "user", "content": messages}]
+                if isinstance(messages, str)
+                else [message_to_wire(message) for message in messages]
+            )
+            kwargs = {"model": self.config.model, "messages": wire}
+            kwargs.update(sampling)
+            trace.record_judge_call(
+                name=self.reward_name,
+                request={"model": kwargs["model"], "messages": kwargs["messages"]},
+                response=response,
+            )
         return response
 
     monkeypatch.setattr(Judge, "complete", fake_complete)
@@ -225,7 +238,43 @@ async def test_reference_score(fake_judge_model):
     assert (
         "Capital of France?" in fake_judge_model[0]
     )  # the task prompt is in the judge prompt
-    assert len(trace.info["judge"]) == 1  # the call is recorded onto the trace
+    assert len(trace.info["judge_calls"]) == 1  # the call is recorded onto the trace
+    judge_record = trace.info["judge_calls"][0]
+    assert judge_record["name"] == "reference"
+    assert judge_record["request"]["model"] == "openai/gpt-5.4-nano"
+    assert "Capital of France?" in judge_record["request"]["messages"][0]["content"]
+    assert judge_record["response"]["message"] == {
+        "role": "assistant",
+        "content": "yes",
+        "reasoning_content": None,
+        "tool_calls": None,
+        "provider_state": None,
+    }
+    assert judge_record["response"]["parsed"] == 1.0
+
+    request_messages = [{"role": "user", "content": "original"}]
+    request = {
+        "model": "openai/gpt-5.4-nano",
+        "messages": request_messages,
+    }
+    trace.record_judge_call(
+        name="snapshot",
+        request=request,
+        response=JudgeResponse(text="yes"),
+    )
+    request_messages[0]["content"] = "mutated"
+    assert (
+        trace.info["judge_calls"][-1]["request"]["messages"][0]["content"] == "original"
+    )
+
+    override_trace = make_trace()
+    await vf.ReferenceJudge().complete(
+        "Judge this response.", trace=override_trace, model="openai/gpt-5.4-mini"
+    )
+    assert (
+        override_trace.info["judge_calls"][0]["request"]["model"]
+        == "openai/gpt-5.4-mini"
+    )
 
     trace = make_trace(reply="It is Rome.")
     assert await vf.ReferenceJudge().score(trace.task.data, trace) == 0.0
@@ -446,7 +495,14 @@ async def test_error_attribution(monkeypatch, tmp_path):
             return response
         finally:
             if trace is not None:
-                trace.record_judge(response)
+                trace.record_judge_call(
+                    name=self.reward_name,
+                    request={
+                        "model": self.config.model,
+                        "messages": [{"role": "user", "content": messages}],
+                    },
+                    response=response,
+                )
 
     monkeypatch.setattr(Judge, "complete", gibberish_judge)
     taskset = JudgedTaskset(
@@ -463,7 +519,7 @@ async def test_error_attribution(monkeypatch, tmp_path):
             trace, runtime=None
         )
     assert trace.rewards["reference"] is None  # seeded: expected but never scored
-    assert len(trace.info["judge"]) == 1  # the billed call is still recorded
+    assert len(trace.info["judge_calls"]) == 1  # the billed call is still recorded
 
 
 # --- rubric --------------------------------------------------------------------------------
@@ -542,7 +598,7 @@ async def test_rubric_score(tmp_path, fake_judge_model):
     trace = make_trace()
     assert await judge.score(trace.task.data, trace) == 0.75
     assert trace.metrics == {"rubric/mentions_paris": 1.0, "rubric/is_polite": 0.0}
-    assert len(trace.info["judge"]) == 1  # one call for the whole rubric
+    assert len(trace.info["judge_calls"]) == 1  # one call for the whole rubric
 
 
 async def test_rubric_verdict_mismatch_raises(tmp_path, monkeypatch):
@@ -652,7 +708,7 @@ async def test_task_score_runs_plugged_judges(tmp_path, fake_judge_model):
         trace.rewards["quality"].score == 0.75
     )  # the rubric's aggregate, under its `name`
     assert (
-        len(trace.info["judge"]) == 2
+        len(trace.info["judge_calls"]) == 2
     )  # every judge call recorded (rubric = one call)
 
 

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/verifiers/v1/__init__.py

@@ -116,7 +116,6 @@ from verifiers.v1.types import (
     ContentPart,
     ImageUrlContentPart,
     ImageUrlSource,
-    KeptTokens,
     Message,
     MessageContent,
     Messages,
@@ -124,6 +123,7 @@ from verifiers.v1.types import (
     Response,
     Sampling,
     SamplingConfig,
+    SamplingMask,
     SystemMessage,
     TextContentPart,
     Tool,
@@ -249,7 +249,7 @@ __all__ = [  # noqa: RUF022 - grouped by public API area
     "MessageNode",
     "Branch",
     "TurnTokens",
-    "KeptTokens",
+    "SamplingMask",
     "Timing",
     "TimeSpan",
     "TimeSplit",

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/verifiers/v1/cli/replay.py

@@ -156,7 +156,7 @@ async def run_replay(config: ReplayConfig, source: Path, out: Path) -> list[Trac
             else:
                 st.state = "running"
                 # Clear the recorded scores; only offline-recomputable ones come back.
-                trace.info.pop("judge", None)
+                trace.info.pop("judge_calls", None)
                 trace.rewards, trace.metrics, trace.extra_usage = {}, {}, []
                 # The declared Task for a rebuilt row, the base Task for the
                 # WireTaskData fallback.

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/verifiers/v1/clients/train.py

@@ -23,9 +23,9 @@ from verifiers.v1.graph import PendingTurn
 from verifiers.v1.types import (
     AssistantMessage,
     FinishReason,
-    KeptTokens,
     Response,
     SamplingConfig,
+    SamplingMask,
     Tool,
     ToolCall,
     TurnTokens,
@@ -161,8 +161,8 @@ def response_from_generate(
             multi_modal_data=result.get("multi_modal_data"),
             mm_token_type_id_map=mm_token_type_id_map,
             routed_experts=result.get("routed_experts"),
-            kept_tokens=KeptTokens(**kept)
-            if (kept := result.get("kept_tokens"))
+            sampling_mask=SamplingMask.from_sampling_mask(mask)
+            if (mask := result.get("sampling_mask"))
             else None,
         ),
     )

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/verifiers/v1/graph.py

@@ -35,9 +35,9 @@ from renderers.base import MultiModalData, PlaceholderRange, RenderedTokens
 from verifiers.v1.semantic import ParentLink
 from verifiers.v1.types import (
     AssistantMessage,
-    KeptTokens,
     Message,
     Response,
+    SamplingMask,
     TextContentPart,
     Tool,
     ToolMessage,
@@ -121,6 +121,12 @@ class MessageNode(BaseModel):
     reference_logprobs: list[float] | None = None
     """Reference-model logprobs over the sampled tokens, in the same compact layout as
     `logprobs`. None means no reference model scored this node."""
+    trainer_logprobs: list[float] | None = None
+    """Trainer-recomputed logprobs over the sampled tokens, in the same compact layout as
+    `logprobs`. None means no trainer forward annotated this node."""
+    entropies: list[float] | None = None
+    """Trainer policy entropies over the sampled tokens, in the same compact layout as
+    `logprobs`. None means no trainer forward annotated this node."""
     loss_weights: dict[str, list[float]] | None = None
     """Named loss-weight streams aligned to `token_ids`, consumer-stamped."""
     multi_modal_data: SkipJsonSchema[MultiModalData | None] = None
@@ -135,12 +141,12 @@ class MessageNode(BaseModel):
     the turn's `generate` payload by `_attribute_routed_experts`; `Branch.routed_experts`
     concatenates these along the path into the trainer's router-replay input. Rides the wire as
     a raw-bytes `__nd__` dict; kept off disk by the dump-site `exclude` in prime-rl."""
-    kept_tokens: SkipJsonSchema[KeptTokens | None] = None
-    """Kept-set sampling masks for this node's sampled tokens, decoded: `ids` flat int32
-    in position order, `counts` the per-token kept-set sizes (aligned with `logprobs`;
-    0 = no mask). Assistant nodes only; consumed via `Branch.kept_tokens` for
-    sampling-replay training. Rides the wire as raw-bytes `__nd__` dicts; kept off disk
-    by the dump-site `exclude` in prime-rl."""
+    sampling_mask: SkipJsonSchema[SamplingMask | None] = None
+    """Sampling masks for this node's sampled tokens.
+
+    `ids` stores the flat token ids and `counts` stores each token's row size. Assistant
+    nodes only. The arrays serialize as raw-byte `__nd__` dictionaries.
+    """
 
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
@@ -202,27 +208,26 @@ class MessageNode(BaseModel):
             return _decode_ndarray(value)
         raise TypeError(f"cannot build ndarray field from {type(value).__name__}")
 
-    @field_serializer("kept_tokens")
-    def serialize_kept_tokens(self, kept: KeptTokens | None) -> dict | None:
-        """`KeptTokens` -> dict of raw-bytes `__nd__` entries so the arrays ride the wire."""
-        if kept is None:
+    @field_serializer("sampling_mask")
+    def serialize_sampling_mask(self, mask: SamplingMask | None) -> dict | None:
+        if mask is None:
             return None
         return {
-            "ids": _encode_ndarray(kept.ids),
-            "counts": _encode_ndarray(kept.counts),
+            "ids": _encode_ndarray(mask.ids),
+            "counts": _encode_ndarray(mask.counts),
         }
 
-    @field_validator("kept_tokens", mode="before")
+    @field_validator("sampling_mask", mode="before")
     @classmethod
-    def deserialize_kept_tokens(cls, value: Any) -> KeptTokens | None:
-        if value is None or isinstance(value, KeptTokens):
+    def deserialize_sampling_mask(cls, value: Any) -> SamplingMask | None:
+        if value is None or isinstance(value, SamplingMask):
             return value
         if isinstance(value, dict):
-            return KeptTokens(
+            return SamplingMask(
                 ids=_decode_ndarray(value["ids"]),
                 counts=_decode_ndarray(value["counts"]),
             )
-        raise TypeError(f"cannot build KeptTokens from {type(value).__name__}")
+        raise TypeError(f"cannot build SamplingMask from {type(value).__name__}")
 
 
 def _canonical_tool_arguments(arguments: str) -> str:
@@ -606,21 +611,18 @@ def _attribute_routed_experts(
         off = end
 
 
-def _attribute_kept_tokens(
-    trace: Trace, assistant_id: int, payload: KeptTokens | None
+def _attribute_sampling_mask(
+    trace: Trace, assistant_id: int, payload: SamplingMask | None
 ) -> None:
-    """Attach this turn's kept-set sampling masks to the assistant node (the payload
-    covers exactly the turn's completion tokens, so no path arithmetic). A payload
-    that doesn't line up with the node's sampled tokens is dropped, not misaligned."""
+    """Attach a completion-aligned sampling mask to the assistant node."""
     if payload is None:
         return
-    counts = np.frombuffer(binascii.a2b_base64(payload.counts), dtype=np.int32)
-    ids = np.frombuffer(binascii.a2b_base64(payload.ids), dtype=np.int32)
     node = trace.nodes[assistant_id]
-    if len(counts) != sum(node.mask) or int(counts.sum()) != len(ids):
+    if len(payload.counts) != sum(node.mask) or int(payload.counts.sum()) != len(
+        payload.ids
+    ):
         return
-    # Own the buffers — the payload views reference the turn's response bytes.
-    node.kept_tokens = KeptTokens(ids=ids.copy(), counts=counts.copy())
+    node.sampling_mask = payload
 
 
 def _commit_turn(turn: PendingTurn, response: Response) -> int:
@@ -769,9 +771,10 @@ def _commit_turn(turn: PendingTurn, response: Response) -> int:
         trace, new_node_ids, path_len, tokens.routed_experts if tokens else None
     )
 
-    # Attribute this turn's kept-set sampling masks onto the assistant node (they are
-    # completion-aligned, so only the sampled node carries them).
-    _attribute_kept_tokens(trace, assistant_id, tokens.kept_tokens if tokens else None)
+    # Sampling masks are completion-aligned, so only the sampled node carries them.
+    _attribute_sampling_mask(
+        trace, assistant_id, tokens.sampling_mask if tokens else None
+    )
 
     return assistant_id
 

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/verifiers/v1/harnesses/bash/harness.py

@@ -8,6 +8,7 @@ from verifiers.v1.configs.harness import HarnessConfig
 from verifiers.v1.harness import Harness
 from verifiers.v1.harnesses.utils.launch import (
     CHAT_PROGRAM_SOURCE,
+    MCP_CHAT_PROGRAM_SOURCE,
     launch_chat_program,
 )
 from verifiers.v1.runtimes import ProgramResult, Runtime
@@ -121,6 +122,7 @@ class BashHarness(Harness[BashHarnessConfig]):
             mcp_urls,
             system_prompt,
             prompt,
+            source_with_mcp=MCP_CHAT_PROGRAM_SOURCE,
             extra_args=args,
             env=env,
             activate=False,

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/verifiers/v1/harnesses/null/harness.py

@@ -3,6 +3,7 @@ from verifiers.v1.configs.harness import HarnessConfig
 from verifiers.v1.harness import Harness
 from verifiers.v1.harnesses.utils.launch import (
     CHAT_PROGRAM_SOURCE,
+    MCP_CHAT_PROGRAM_SOURCE,
     launch_chat_program,
 )
 from verifiers.v1.runtimes import ProgramResult, Runtime
@@ -46,4 +47,5 @@ class NullHarness(Harness[NullHarnessConfig]):
             mcp_urls,
             system_prompt,
             prompt,
+            source_with_mcp=MCP_CHAT_PROGRAM_SOURCE,
         )

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/verifiers/v1/harnesses/prime_agent/harness.py

@@ -19,10 +19,10 @@ logger = logging.getLogger(__name__)
 GITHUB_RELEASE_URL = (
     "https://github.com/PrimeIntellect-ai/prime-agent/releases/download"
 )
-PRIME_AGENT_COMMIT: Literal["514633727bf26d74f39f3119c2b0e31a5ceb2a9d"] = (
-    "514633727bf26d74f39f3119c2b0e31a5ceb2a9d"
+PRIME_AGENT_COMMIT: Literal["81ae3cb34d27d38ee37f9e205a1e73694993b344"] = (
+    "81ae3cb34d27d38ee37f9e205a1e73694993b344"
 )
-PRIME_AGENT_VERSION = "0.8.1"
+PRIME_AGENT_VERSION = "0.9.1"
 PRIME_AGENT_DIR = "/var/tmp/vf-prime-agent"
 STATE_ROOT = "/tmp/vf-prime-agent-runs"
 SKILLS_DIR = ".agents/skills"
@@ -51,10 +51,10 @@ for tarball in "$agent_tarball" "$ai_tarball" "$core_tarball" "$tui_tarball"; do
         "$release_url/$tarball" -o "$download_dir/$tarball"
 done
 printf '%s  %s\n' \
-    '46c24db1782dd31adc35d5c6cbcc75564faba6ced3bf2ccf03d836ee77134475' "$agent_tarball" \
-    'f6c3bdb6093bc24a327546fe865ef9a4a172c734fcd4c4093e30c19476f0134d' "$ai_tarball" \
-    '0cc3660953545f8ac9a7e704fcb9875f954d58c3085304080ef615c280aa5748' "$core_tarball" \
-    'bd07bccee0ca495565b1d62e9411f3fdebe49e3dfa52870564f08af5e61fde15' "$tui_tarball" \
+    '573bce0cd004fc62052e9a924089941b7f39266ab71e66a94c85a1f9d35835ba' "$agent_tarball" \
+    '11b5b4cf67b6bb2d3420a44fb69181bc9d94d81e69a2b4fde07eb9c99f5faf4f' "$ai_tarball" \
+    'fb6f3a5dcc8b69c5eeb3beff722b5e0f09885c14849db50bc1d7c0f1d064151c' "$core_tarball" \
+    '4f3eaca2814944d3993073e0b88c0bde54a641ddb5132fbe107e392b997e38ec' "$tui_tarball" \
     > "$download_dir/SHA256SUMS"
 (cd "$download_dir" && sha256sum -c SHA256SUMS)
 mkdir "$download_dir/package-root"
@@ -87,7 +87,7 @@ PRIME_AGENT_BOOTSTRAP_TOOLS_ON_INSTALL=1 npm install -g \
 
 
 class PrimeAgentHarnessConfig(HarnessConfig):
-    commit: Literal["514633727bf26d74f39f3119c2b0e31a5ceb2a9d"] = PRIME_AGENT_COMMIT
+    commit: Literal["81ae3cb34d27d38ee37f9e205a1e73694993b344"] = PRIME_AGENT_COMMIT
     """Prime Agent main commit to install."""
 
     autonomous: bool = False

{verifiers-0.3.2.dev32 → verifiers-0.3.2.dev41}/verifiers/v1/harnesses/rlm/harness.py

@@ -5,7 +5,14 @@ import logging
 import shlex
 from typing import Any, Literal
 
-from pydantic import BaseModel, ConfigDict, Field, PositiveInt, model_validator
+from pydantic import (
+    BaseModel,
+    ConfigDict,
+    Field,
+    NonNegativeInt,
+    PositiveInt,
+    model_validator,
+)
 from pydantic_config import BaseConfig
 
 from verifiers.v1.acp import ACPConfig, ACPHarness, ACPTurn, JsonObject
@@ -17,7 +24,7 @@ from verifiers.v1.trace import Trace
 
 logger = logging.getLogger(__name__)
 
-BuiltinSkill = Literal["edit", "search"]
+BuiltinSkill = Literal["bash", "edit", "fetch", "search"]
 
 RLM_REPO = "github.com/PrimeIntellect-ai/nano-rlm.git"
 RLM_CACHE_DIR = "/tmp/vf-rlm"
@@ -40,18 +47,45 @@ class CompactionConfig(BaseConfig):
     summarize_at_tokens: PositiveInt | None = None
     """Compact at this token count. When unset, compact when 16k tokens remain below the
     model context window when the provider advertises it."""
+    max_compactions: PositiveInt | None = None
+    """Compactions per session before the engine stops compacting; `None` =
+    nano-rlm's default."""
 
 
 class RLMHarnessConfig(HarnessConfig):
-    version: str = Field(default="4ef3438", min_length=1)
-    """Git ref (branch, tag, or commit) of nano-rlm to install."""
-    max_depth: int = 0
-    """Recursion depth RLM may spawn sub-harnesses to."""
+    version: str = Field(default="dd2c04f", min_length=1)
+    """Git ref (branch, tag, or commit) of nano-rlm to install. Must know every
+    field this harness puts on the wire, i.e. be at least the default ref."""
+    max_depth: NonNegativeInt | None = None
+    """Recursion depth RLM may spawn sub-agents to; `None` = nano-rlm's default (1).
+    Set 0 to disable recursion."""
     builtin_skills: list[BuiltinSkill] = Field(default_factory=list)
     """Built-in rlm skills to enable (RLM_SKILLS), e.g. `["edit"]`; empty enables none.
     The tool set is fixed (ipython); the base `skills` field takes SKILL.md paths."""
-    compaction: CompactionConfig | None = None
-    """Context compaction policy. Set an empty config to use automatic thresholds."""
+    compaction: CompactionConfig | bool | None = None
+    """Context compaction: a `[compaction]` section (or `true`) enables it with the
+    given settings, `false` disables it (an overflowing session then fails), and
+    unset defers to nano-rlm's default (on, automatic thresholds, bounded by its
+    default 1M tree-token budget)."""
+    max_concurrent_subagents: PositiveInt | None = None
+    """Sub-agents running at once per session tree; `None` = nano-rlm's default (4),
+    raised to an explicit `max_depth` when needed to keep the policy valid."""
+    max_total_turns: PositiveInt | None = None
+    """Tree-total turn budget (one turn = one work-loop model call, any engine); every
+    engine stops before its next call once spent. `None` = uncapped."""
+    max_total_tokens: PositiveInt | None = None
+    """Tree-total budget of NEW tokens (completion + uncached prompt) across the session
+    tree; once spent every engine stops and no further sub-agents spawn. `None` = unbounded."""
+    max_tool_output_bytes: PositiveInt | None = None
+    """Byte budget for a single tool result entering the conversation (middle truncation);
+    overrides rlm's built-in 20KB default in either direction."""
+    append_to_system_prompt: str | None = None
+    """Appended to the root engine's system prompt, after the taskset's system prompt."""
+    subagent_append_to_system_prompt: str | None = None
+    """Append for sub-agent engines that can still delegate (depth >= 1, below
+    `max_depth`); unset falls back to `append_to_system_prompt`."""
+    leaf_append_to_system_prompt: str | None = None
+    """Append for depth-capped leaf engines; unset falls back to the sub-agent append."""
 
     @model_validator(mode="after")
     def reject_disabled_tools(self) -> "RLMHarnessConfig":
@@ -60,6 +94,15 @@ class RLMHarnessConfig(HarnessConfig):
                 "the rlm harness has a fixed tool set (ipython) and does not support "
                 "`disabled_tools`; use `builtin_skills` to enable built-in skills instead."
             )
+        if (
+            self.max_depth is not None
+            and self.max_concurrent_subagents is not None
+            and self.max_concurrent_subagents < self.max_depth
+        ):
+            raise ValueError(
+                "`max_concurrent_subagents` must be at least `max_depth` "
+                "(nano-rlm rejects the policy otherwise)."
+            )
         return self
 
 
@@ -109,6 +152,27 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
         system_prompt: str | None,
     ) -> JsonObject:
         compaction = self.config.compaction
+        max_concurrent_subagents = self.config.max_concurrent_subagents
+        if max_concurrent_subagents is None and self.config.max_depth is not None:
+            max_concurrent_subagents = max(4, self.config.max_depth)
+        policy_knobs: dict[str, Any] = {
+            "max_depth": self.config.max_depth,
+            "max_concurrent_subagents": max_concurrent_subagents,
+            "max_total_turns": self.config.max_total_turns,
+            "max_total_tokens": self.config.max_total_tokens,
+            "max_tool_output_bytes": self.config.max_tool_output_bytes,
+        }
+        if isinstance(compaction, bool):
+            policy_knobs["compaction"] = compaction
+        elif compaction is not None:
+            policy_knobs["compaction"] = True
+            policy_knobs["summarize_at_tokens"] = compaction.summarize_at_tokens
+            policy_knobs["max_compactions"] = compaction.max_compactions
+        appends = [
+            text
+            for text in (system_prompt, self.config.append_to_system_prompt)
+            if text
+        ]
         payload = {
             "session_id": trace.id,
             "model": ctx.model,
@@ -116,16 +180,17 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
                 "base_url": endpoint,
                 "api_key": secret,
             },
+            # None = passthrough: the key stays off the wire and nano-rlm's own
+            # default applies.
             "policy": {
-                "max_depth": self.config.max_depth,
-                "compaction": compaction is not None,
-                "summarize_at_tokens": (
-                    compaction.summarize_at_tokens if compaction else None
-                ),
-                "max_concurrent_subagents": max(4, self.config.max_depth),
+                key: value for key, value in policy_knobs.items() if value is not None
             },
             "system_prompt_path": None,
-            "append_to_system_prompt": system_prompt,
+            "append_to_system_prompt": "\n\n".join(appends) or None,
+            "subagent_append_to_system_prompt": (
+                self.config.subagent_append_to_system_prompt
+            ),
+            "leaf_append_to_system_prompt": self.config.leaf_append_to_system_prompt,
             "skills": list(self.config.builtin_skills),
             "kernel_env": runtime.env,
             "search_api_key": self.config.resolved_env.get("SERPER_API_KEY"),