verifiers 0.4.1.dev27__tar.gz → 0.4.1.dev29__tar.gz

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.4.1.dev27
+Version: 0.4.1.dev29
 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

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_agent_user.py

@@ -52,5 +52,11 @@ async def test_agent_user_harness(run, harness, harness_runtime, scripted_model)
         client=scripted_model,
     )
     (trace,) = episode.traces
+    # mini-swe-agent only acts through tool calls: it drops a tool-less reply (the
+    # scripted "OK") from its history and resends only its format error, so the
+    # trace forks by design. Recheck on version bumps:
+    # https://github.com/SWE-agent/mini-swe-agent/blob/v2.4.6/src/minisweagent/agents/default.py#L114
+    if harness.id != "mini-swe-agent":
+        assert trace.num_branches == 1
     assert trace.ok, [error.message for error in trace.errors]
     assert trace.num_turns >= 1

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_agentic_judge.py

@@ -30,6 +30,7 @@ async def test_agentic_judge_shared_verifier(run, tmp_path):
         rollout_timeout=600,
     )
     traces = episode.traces
+    assert all(trace.num_branches == 1 for trace in traces)
     assert sorted(t.agent.name for t in traces) == ["judge", "solver"]
     (solver,) = [t for t in traces if t.agent.name == "solver"]
     (judge,) = [t for t in traces if t.agent.name == "judge"]
@@ -75,6 +76,7 @@ async def test_env_id_agentic_judge(run, tmp_path):
         rollout_timeout=600,
     )
     traces = episode.traces
+    assert all(trace.num_branches == 1 for trace in traces)
     assert sorted(t.agent.name for t in traces) == ["judge", "solver"]
     (solver,) = [t for t in traces if t.agent.name == "solver"]
     (judge,) = [t for t in traces if t.agent.name == "judge"]

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_examples.py

@@ -14,6 +14,8 @@ import pytest
 from conftest import CI_DECISION_JUDGE, CI_MODEL, CI_REASONING_EFFORT
 from grid import agent_seats
 
+import verifiers.v1 as vf
+
 EVAL_TIMEOUT = 600  # 10 minutes for a capped eval (-n 1 -r 1)
 
 ENVIRONMENTS = Path(__file__).parent.parent.parent / "environments"
@@ -107,26 +109,29 @@ def test_eval(taskset: str, tmp_path: Path):
         f"eval {taskset} failed: {(proc.stderr or proc.stdout)[-2000:]}"
     )
     episodes = [
-        json.loads(line)
+        vf.WireEpisode.model_validate_json(line)
         for line in (tmp_path / "run" / "traces.jsonl").read_text().splitlines()
     ]
-    assert episodes and all(episode["ok"] for episode in episodes), [
-        error["message"][:500]
+    assert episodes and all(episode.ok for episode in episodes), [
+        error.message[:500]
         for episode in episodes
-        for holder in (episode, *episode["traces"])
-        for error in holder.get("errors") or []
+        for holder in (episode, *episode.traces)
+        for error in holder.errors or []
     ]
+    assert all(
+        trace.num_branches == 1 for episode in episodes for trace in episode.traces
+    )
     if taskset == "interception" and os.getenv("PRIME_API_KEY"):
         assert len(episodes) == 4
         judged = [
             trace
             for episode in episodes
-            for trace in episode["traces"]
-            if trace["info"].get("judge_calls")
+            for trace in episode.traces
+            if trace.info.get("judge_calls")
         ]
         assert len(judged) == 1
         assert (
-            judged[0]["info"]["judge_calls"][0]["request"]["model"]
+            judged[0].info["judge_calls"][0]["request"]["model"]
             == CI_DECISION_JUDGE["model"]
         )
-        assert judged[0]["response_rewrites"]
+        assert judged[0].response_rewrites

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_limits.py

@@ -23,6 +23,7 @@ async def test_cap(run, cap, stop_condition):
         env={"agent": cap},
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.stop_condition == stop_condition
     assert trace.is_truncated
     assert trace.num_turns == 1
@@ -50,5 +51,6 @@ async def test_max_turns(run, harness, harness_runtime):
         max_tokens=8192,
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.stop_condition == "max_turns", [e.message for e in trace.errors]
     assert trace.num_turns == 1

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_multi_agent.py

@@ -14,6 +14,7 @@ async def test_multi_agent_env(run):
         max_turns=2,
     )
     traces = episode.traces
+    assert all(trace.num_branches == 1 for trace in traces)
     assert len(traces) == 2  # one episode, one trace per role
     assert sorted(t.agent.name for t in traces) == ["a", "b"]
     (b,) = [t for t in traces if t.agent.name == "b"]
@@ -42,6 +43,7 @@ async def test_env_id_best_of_n(run):
         max_turns=2,
     )
     traces = episode.traces
+    assert all(trace.num_branches == 1 for trace in traces)
     assert len(traces) == 2  # one episode, two attempts
     assert all(t.agent.name == "agent" and t.ok for t in traces)
     assert any(t.metrics["best"] == 1.0 for t in traces)
@@ -62,6 +64,7 @@ async def test_kuhn_poker_self_play(run):
         rollout_timeout=300,
     )
     traces = episode.traces
+    assert all(trace.num_branches == 1 for trace in traces)
     assert sorted(t.agent.name for t in traces) == ["player0", "player1"]
     payoffs = {t.agent.name: t.rewards["payoff"].score for t in traces}
     assert payoffs["player0"] + payoffs["player1"] == 0  # zero-sum

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_multi_turn.py

@@ -25,6 +25,7 @@ async def test_multi_turn(run, harness, harness_runtime):
         max_tokens=8192,
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok, [error.message for error in trace.errors]
     assert trace.num_turns >= 1  # ran a command, then finished
     assert trace.reward == 1.0

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_multi_turn_tool.py

@@ -20,6 +20,7 @@ async def test_tool(run, harness, harness_runtime):
         taskset_overrides={"task": {"tools": {"colocated": True}}},
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok, [error.message for error in trace.errors]
     assert trace.num_turns >= 1
     # Reward 1.0 is the proof the tool ran. How a program surfaces an MCP tool to the
@@ -41,6 +42,7 @@ async def test_tool_placement(run, harness_runtime, tool_runtime):
         taskset_overrides={"task": {"tools": tool_runtime}},
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok, [error.message for error in trace.errors]
     assert trace.num_turns >= 2  # tool call + answer
     assert trace.reward == 1.0

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_multimodal_tool.py

@@ -12,6 +12,7 @@ async def test_multimodal_tool(run):
         max_turns=4,
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok
     assert trace.num_turns >= 2  # tool call + answer
     assert trace.reward == 1.0

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_network_policy.py

@@ -54,6 +54,12 @@ async def test_network_policy(run, harness, harness_runtime, notice):
     assert set(by_name) == set(NOTICES)
     for name, episode in by_name.items():
         [trace] = episode.traces
+        # mini-swe-agent only acts through tool calls: it drops a tool-less reply
+        # (the probe output as a final answer) from its history, so the trace forks
+        # by design. Recheck on version bumps:
+        # https://github.com/SWE-agent/mini-swe-agent/blob/v2.4.6/src/minisweagent/agents/default.py#L114
+        if harness.id != "mini-swe-agent":
+            assert trace.num_branches == 1
         assert trace.ok, [error.message for error in trace.errors]
         prompt = prompt_text(trace)
         if notice and NOTICES[name]:

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_resume_tool.py

@@ -19,6 +19,7 @@ async def test_resume_tool(run, harness, harness_runtime):
         rollout_timeout=600,
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok, [error.message for error in trace.errors]
     assert trace.stop_condition == "user_closed"
     assert trace.rewards["resumed"].score == 1.0

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_rubric_judge.py

@@ -31,6 +31,7 @@ async def test_rubric_judge(run, tmp_path, judge_id):
         max_turns=2,
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok
     assert trace.rewards[judge_id].score > 0  # the judge's verdict landed in the reward
     if judge_id == "rubric":

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_shared_tool.py

@@ -22,6 +22,7 @@ async def test_shared_tool(run, harness_runtime, tool_runtime):
         taskset_overrides={"tools": tool_runtime},
     )
     traces = [trace for episode in episodes for trace in episode.traces]
+    assert all(trace.num_branches == 1 for trace in traces)
     assert len(traces) == 2
     for trace in traces:
         assert trace.ok, [error.message for error in trace.errors]

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_single_turn.py

@@ -23,6 +23,7 @@ async def test_single_turn(run, harness, harness_runtime):
         max_turns=2,
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok, [error.message for error in trace.errors]
     assert trace.num_turns == 1
     assert trace.stop_condition == "agent_completed"

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_skills.py

@@ -17,5 +17,6 @@ async def test_skills(run, harness, harness_runtime):
         max_tokens=8192,
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok, [error.message for error in trace.errors]
     assert trace.reward == 1.0, trace.last_reply

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_stateful_tool.py

@@ -20,6 +20,7 @@ async def test_tool_state(run, harness_runtime, tool_runtime):
         taskset_overrides={"task": {"tools": tool_runtime}},
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok, [error.message for error in trace.errors]
     assert trace.num_turns >= 2  # at least two tool calls accumulated
     assert trace.reward == 1.0

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_user_echo.py

@@ -25,6 +25,7 @@ async def test_user(run, harness, harness_runtime):
         max_tokens=8192,
     )
     (trace,) = episode.traces
+    assert trace.num_branches == 1
     assert trace.ok, [error.message for error in trace.errors]
     assert trace.num_turns >= 2  # genuinely multi-turn
     assert trace.stop_condition == "user_closed"  # leaving the interaction ended it
@@ -67,3 +68,4 @@ async def test_interaction(live_ctx):
     assert trace is not None and trace.errors == []
     assert trace.stop_condition == "user_closed"  # closing the interaction ended it
     assert trace.num_turns == 2
+    assert trace.num_branches == 1

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/integration/test_user_sim.py

@@ -17,6 +17,7 @@ async def test_env_id_user_sim(run):
         rollout_timeout=300,
     )
     traces = episode.traces
+    assert all(trace.num_branches == 1 for trace in traces)
     assert sorted(t.agent.name for t in traces) == ["assistant", "user"]
     (assistant,) = [t for t in traces if t.agent.name == "assistant"]
     (user,) = [t for t in traces if t.agent.name == "user"]
@@ -53,6 +54,7 @@ async def test_env_id_user_sim_with_tools(run):
         rollout_timeout=300,
     )
     traces = episode.traces
+    assert all(trace.num_branches == 1 for trace in traces)
     (assistant,) = [t for t in traces if t.agent.name == "assistant"]
     (user,) = [t for t in traces if t.agent.name == "user"]
     assert assistant.ok and user.ok

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/tests/unit/test_graph.py

@@ -5,6 +5,7 @@ import pytest
 
 import verifiers.v1 as vf
 from verifiers.v1 import graph
+from verifiers.v1.harnesses.utils.replay import drop_blank_text, strip_text
 from verifiers.v1.types import TurnTokens
 
 
@@ -269,6 +270,45 @@ def test_reasoning_content_participates_in_graph_prefix_matching():
     assert len(tool_call_nodes) == 2
 
 
+@pytest.mark.parametrize(
+    "view,sampled,replayed",
+    [
+        (drop_blank_text, {"content": "\n\n"}, {"content": None}),
+        (
+            drop_blank_text,
+            {"content": "\n\n", "reasoning_content": "plan"},
+            {"reasoning_content": "plan"},
+        ),
+        (strip_text, {"content": "Calling it.\n\n"}, {"content": "Calling it."}),
+        (strip_text, {"content": None}, {"reasoning_content": " "}),
+    ],
+)
+def test_replay_view_matches_cosmetically_rewritten_turns(view, sampled, replayed):
+    def run(replay_view):
+        trace = vf.Trace(
+            agent=vf.AgentInfo(config=vf.AgentConfig()),
+            task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="go")),
+        )
+        trace._replay_view = replay_view
+        user = vf.UserMessage(content="go")
+        call = vf.ToolCall(id="call_0", name="lookup", arguments="{}")
+        graph.prepare_turn(trace, [user]).commit(
+            _response(vf.AssistantMessage(tool_calls=[call], **sampled))
+        )
+        graph.prepare_turn(
+            trace,
+            [
+                user,
+                vf.AssistantMessage(tool_calls=[call], **replayed),
+                vf.ToolMessage(content="result", tool_call_id="call_0"),
+            ],
+        ).commit(_response(vf.AssistantMessage(content="done")))
+        return trace
+
+    assert run(None).num_branches == 2
+    assert run(view).num_branches == 1
+
+
 @pytest.mark.parametrize("same_tools", [False, True])
 def test_parallel_commits_reconcile_shared_prompt_prefix(same_tools):
     """Two requests prepared from the same graph snapshot share any common prompt prefix that

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/graph.py

@@ -321,13 +321,23 @@ def _tools_hash(tools: list[Tool] | None) -> str:
     return hashlib.blake2b(data, digest_size=16).hexdigest()
 
 
+def _message_key(trace: Trace, message: Message) -> str:
+    """`message_hash` of the message as the trace's harness replays it, so a turn the
+    harness re-sends with cosmetic rewrites still matches the node it came from."""
+    view = trace._replay_view
+    return message_hash(view(message) if view is not None else message)
+
+
 def _node_key(
-    parent: int | None, message: Message, tools: list[Tool] | None = None
+    trace: Trace,
+    parent: int | None,
+    message: Message,
+    tools: list[Tool] | None = None,
 ) -> tuple[int | None, str | None, str]:
     return (
         parent,
         _tools_hash(tools) if parent is None else None,
-        message_hash(message),
+        _message_key(trace, message),
     )
 
 
@@ -335,7 +345,7 @@ def _head_index(trace: Trace) -> dict[tuple[int | None, str | None, str], int]:
     """Physical node key -> id, rebuilt lazily after deserialization."""
     if not trace._head_index and trace.nodes:
         trace._head_index = {
-            _node_key(node.parent, node.message, node.tools): nid
+            _node_key(trace, node.parent, node.message, node.tools): nid
             for nid, node in enumerate(trace.nodes)
         }
     return trace._head_index
@@ -350,12 +360,12 @@ def message_prefix_len(trace: Trace, prompt: list[Message]) -> int:
     parents: list[int | None] = [None]
     matched = 0
     for message in prompt:
-        key = message_hash(message)
+        key = _message_key(trace, message)
         parents = [
             node_id
             for parent in parents
             for node_id in children.get(parent, [])
-            if message_hash(trace.nodes[node_id].message) == key
+            if _message_key(trace, trace.nodes[node_id].message) == key
         ]
         if not parents:
             break
@@ -377,7 +387,7 @@ def _matching_node(
     prefix breaks can leave older physical variants under the same key, so a token mismatch falls
     back to a reverse scan rather than materializing a duplicate of an already-existing variant.
     """
-    key = _node_key(parent, message, tools)
+    key = _node_key(trace, parent, message, tools)
     indexed = _head_index(trace).get(key)
     if (
         indexed is not None
@@ -397,7 +407,7 @@ def _matching_node(
         if (
             (token_ids is None or node.token_ids == token_ids)
             and (renderer_token_ids is None or node.logical_ids == renderer_token_ids)
-            and _node_key(node.parent, node.message, node.tools) == key
+            and _node_key(trace, node.parent, node.message, node.tools) == key
         ):
             return node_id
     return None
@@ -422,7 +432,7 @@ def _matching_prefix_node(
     otherwise-missing message boundary. The next attributed message's start bounds the match:
     a sampled variant must never consume tokens that the renderer assigned to that message.
     """
-    key = _node_key(parent, message, tools)
+    key = _node_key(trace, parent, message, tools)
     indexed = _head_index(trace).get(key)
     candidates: list[int] = []
     if indexed is not None:
@@ -432,6 +442,7 @@ def _matching_prefix_node(
         for node_id in range(len(trace.nodes) - 1, -1, -1)
         if node_id != indexed
         and _node_key(
+            trace,
             trace.nodes[node_id].parent,
             trace.nodes[node_id].message,
             trace.nodes[node_id].tools,
@@ -562,7 +573,7 @@ class PendingTurn:
                 )
             )
             parent = len(self.trace.nodes) - 1
-            index[_node_key(previous, message, self.tools)] = parent
+            index[_node_key(self.trace, previous, message, self.tools)] = parent
         if self.tools:
             self.trace.tools = list(
                 {
@@ -603,7 +614,7 @@ def prepare_turn(
             if len(children) == 1 and trace.nodes[children[0]].message == msg:
                 existing = children[0]
         if existing is None:
-            message_key = message_hash(msg)
+            message_key = _message_key(trace, msg)
             existing = idx.get(
                 (parent, tools_hash if parent is None else None, message_key)
             )
@@ -903,7 +914,7 @@ def _commit_turn(turn: PendingTurn, response: Response) -> int:
     # Track new nodes separately so routed-expert attribution needs only node ids, not this path.
     new_node_ids: list[int] = []
     for i, msg in enumerate(prompt[num_reused:], start=num_reused):
-        key = _node_key(parent, msg, turn.tools)
+        key = _node_key(trace, parent, msg, turn.tools)
         start = path_len if cursor is None else cursor
         span = spans[i] if spans and i < len(spans) else None
         end = span[1] if span else start
@@ -956,7 +967,7 @@ def _commit_turn(turn: PendingTurn, response: Response) -> int:
         )
     )
     assistant_id = len(trace.nodes) - 1
-    idx[_node_key(parent, response.message, turn.tools)] = assistant_id
+    idx[_node_key(trace, parent, response.message, turn.tools)] = assistant_id
     new_node_ids.append(assistant_id)
 
     # Attribute this turn's expert-routing array onto the nodes created this turn (new input

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/harness.py

@@ -18,7 +18,7 @@ from verifiers.v1.errors import (
 )
 from verifiers.v1.runtimes import ProgramResult, Runtime
 from verifiers.v1.task import TaskData
-from verifiers.v1.types import Messages
+from verifiers.v1.types import Message, Messages
 from verifiers.v1.utils.decorators import (
     discover_decorated,
     invoke_all,
@@ -67,6 +67,13 @@ class Harness(ABC, Generic[ConfigT]):
     def __init__(self, config: ConfigT) -> None:
         self.config = config
 
+    def replay_view(self, message: Message) -> Message:
+        """`message` as this harness re-sends it on later requests. Override only for
+        deterministic cosmetic rewrites (e.g. dropped whitespace-only text), never for
+        content changes, so a replayed turn matches the turn it came from in the trace
+        graph instead of forking it."""
+        return message
+
     def resolve_prompt(
         self, task: TaskData
     ) -> tuple[str | None, str | Messages | None]:

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/harnesses/bash/harness.py

@@ -79,9 +79,20 @@ class BashHarness(Harness[BashHarnessConfig]):
             fragments.append(EDIT_SYSTEM_PROMPT)
         if self.config.search:
             fragments.append(SEARCH_PROMPT)
-        system_prompt = "\n\n".join(
-            p for p in (" ".join(fragments), system_prompt) if p
-        )
+        harness_prompt = " ".join(fragments)
+        system_prompt = "\n\n".join(p for p in (harness_prompt, system_prompt) if p)
+        # Default resume replays the transcript. If there was no task system
+        # prompt, that transcript already contains this harness prompt.
+        if (
+            data.system_prompt is None
+            and prompt is not None
+            and not isinstance(prompt, str)
+            and any(
+                message.role == "system" and message.content == harness_prompt
+                for message in prompt
+            )
+        ):
+            system_prompt = None
         env = {**self.config.resolved_env}
         args = ["--bash"]
         if tool_interception_url:

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/harnesses/claude_code/harness.py

@@ -7,9 +7,11 @@ from verifiers.v1.clients import ModelContext
 from verifiers.v1.configs.harness import HarnessConfig, PinnedVersion
 from verifiers.v1.harnesses.node import NODE_BIN_DIR, ensure_node
 from verifiers.v1.harnesses.utils.install import ensure_installed, remove_dir
+from verifiers.v1.harnesses.utils.replay import drop_blank_text
 from verifiers.v1.runtimes import Runtime
 from verifiers.v1.task import TaskData
 from verifiers.v1.trace import Trace
+from verifiers.v1.types import Message
 
 CLAUDE_ACP_DIR = "/var/tmp/vf-claude-agent-acp-{version}-{acp_version}"
 PACKAGES_DIR = f"{CLAUDE_ACP_DIR}/packages"
@@ -40,6 +42,11 @@ class ClaudeCodeHarness(ACPHarness[ClaudeCodeHarnessConfig]):
     SUPPORTS_SKILLS = True
     SUPPORTS_TOOL_INTERCEPTION = True
 
+    # Claude Code drops whitespace-only text blocks on replay. Closed source: observed
+    # with 2.1.278, recheck on version bumps.
+    def replay_view(self, message: Message) -> Message:
+        return drop_blank_text(message)
+
     async def setup(self, runtime: Runtime) -> None:
         await ensure_node(runtime)
         versions = {"version": self.config.version, "acp_version": ACP_VERSION}

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/harnesses/hermes_agent/harness.py

@@ -7,10 +7,12 @@ from verifiers.v1.acp import ACPConfig, ACPHarness
 from verifiers.v1.clients import ModelContext
 from verifiers.v1.configs.harness import HarnessConfig, PinnedVersion
 from verifiers.v1.harnesses.utils.install import ensure_installed, remove_dir
+from verifiers.v1.harnesses.utils.replay import strip_text
 from verifiers.v1.runtimes import Runtime
 from verifiers.v1.runtimes.base import _ENSURE_UV
 from verifiers.v1.task import TaskData
 from verifiers.v1.trace import Trace
+from verifiers.v1.types import Message
 
 PROGRAM_SOURCE = (Path(__file__).resolve().parent / "program.py").read_text()
 HERMES_DIR = "/var/tmp/vf-hermes-agent-{version}"
@@ -37,6 +39,13 @@ class HermesAgentHarness(ACPHarness[HermesAgentHarnessConfig]):
     SUPPORTS_MCP = True
     SUPPORTS_SKILLS = True
 
+    # Hermes stores assistant text stripped and pads missing reasoning with " " on
+    # replay. Recheck on version bumps:
+    # https://github.com/NousResearch/hermes-agent/blob/v2026.9.14/agent/chat_completion_helpers.py#L1444
+    # https://github.com/NousResearch/hermes-agent/blob/v2026.9.14/agent/message_sanitization.py#L545
+    def replay_view(self, message: Message) -> Message:
+        return strip_text(message)
+
     async def setup(self, runtime: Runtime) -> None:
         # Hermes needs its source-tree assets and supports editable installs only.
         directory = HERMES_DIR.format(version=self.config.version)

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/harnesses/pi/harness.py

@@ -11,9 +11,11 @@ from verifiers.v1.clients import ModelContext
 from verifiers.v1.configs.harness import HarnessConfig, PinnedVersion, skill_destination
 from verifiers.v1.harnesses.node import NODE_BIN_DIR, ensure_node
 from verifiers.v1.harnesses.utils.install import ensure_installed, remove_dir
+from verifiers.v1.harnesses.utils.replay import drop_blank_text
 from verifiers.v1.runtimes import Runtime
 from verifiers.v1.task import TaskData
 from verifiers.v1.trace import Trace
+from verifiers.v1.types import Message
 
 logger = logging.getLogger(__name__)
 
@@ -65,6 +67,11 @@ class PiHarness(ACPHarness[PiHarnessConfig]):
     # so the installed skills are passed explicitly via `--skill` at launch.
     SUPPORTS_SKILLS = True
 
+    # pi-ai drops whitespace-only assistant text on replay. Recheck on version bumps:
+    # https://github.com/earendil-works/pi/blob/v0.86.1/packages/ai/src/api/openai-completions.ts#L1290
+    def replay_view(self, message: Message) -> Message:
+        return drop_blank_text(message)
+
     async def setup(self, runtime: Runtime) -> None:
         await ensure_node(runtime)
         logger.info(

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/harnesses/prime_agent/harness.py

@@ -10,9 +10,11 @@ from verifiers.v1.clients import ModelContext
 from verifiers.v1.configs.harness import HarnessConfig, skill_destination
 from verifiers.v1.harnesses.node import NODE_BIN_DIR, ensure_node
 from verifiers.v1.harnesses.utils.install import ensure_installed, remove_dir
+from verifiers.v1.harnesses.utils.replay import drop_blank_text
 from verifiers.v1.runtimes import Runtime
 from verifiers.v1.task import TaskData
 from verifiers.v1.trace import Trace
+from verifiers.v1.types import Message
 
 logger = logging.getLogger(__name__)
 
@@ -99,6 +101,11 @@ class PrimeAgentHarness(ACPHarness[PrimeAgentHarnessConfig]):
     SUPPORTS_MCP = True
     SUPPORTS_SKILLS = True
 
+    # Its pi-ai fork drops whitespace-only assistant text on replay. Recheck on bumps:
+    # https://github.com/PrimeIntellect-ai/prime-agent/blob/v0.9.5/packages/ai/src/providers/openai-completions.ts#L895
+    def replay_view(self, message: Message) -> Message:
+        return drop_blank_text(message)
+
     def acp_turn_result(self, trace: Trace, result: ACPTurn) -> None:
         events = [
             event

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/harnesses/rlm/harness.py

@@ -165,6 +165,9 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
             "{ apt-get update -qq && apt-get install -y -qq git; } && "
             f"rm -rf {checkout} && git clone https://{RLM_REPO} {checkout} && "
             f"git -C {checkout} checkout {shlex.quote(self.config.version)} && "
+            # The editable install needs the files, not the history; agents searching
+            # their workspace for a repository must not find the harness's own.
+            f"rm -rf {checkout}/.git && "
             f"{home_env}UV_INSTALL_DIR={directory}/bin UV_TOOL_BIN_DIR={directory}/bin "
             f"UV_TOOL_DIR={directory}/tools "
             f"RLM_CHECKOUT_PATH={checkout} bash {checkout}/install.sh && "

verifiers-0.4.1.dev29/verifiers/v1/harnesses/utils/replay.py

@@ -0,0 +1,30 @@
+"""Cosmetic rewrites agent CLIs apply when they replay an earlier assistant turn, for
+`Harness.replay_view` overrides. Each function is idempotent and only normalizes
+whitespace, so the replayed and the sampled form of a turn map to the same view."""
+
+from verifiers.v1.types import AssistantMessage, Message
+
+
+def drop_blank_text(message: Message) -> Message:
+    """Clients that filter text blocks with `text.trim()` drop whitespace-only
+    assistant text, such as the "\\n\\n" some models emit before a tool call."""
+    if (
+        isinstance(message, AssistantMessage)
+        and isinstance(message.content, str)
+        and not message.content.strip()
+    ):
+        return message.model_copy(update={"content": None})
+    return message
+
+
+def strip_text(message: Message) -> Message:
+    """Clients that store assistant text stripped, and pad a missing reasoning field
+    with a single space on replay."""
+    if not isinstance(message, AssistantMessage):
+        return message
+    update: dict = {}
+    if isinstance(message.content, str):
+        update["content"] = message.content.strip() or None
+    if message.reasoning_content is not None and not message.reasoning_content.strip():
+        update["reasoning_content"] = None
+    return message.model_copy(update=update) if update else message

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/rollout.py

@@ -145,6 +145,7 @@ class Rollout:
             # this trace can be reproduced with.
             agent=AgentInfo(config=agent_config),
         )
+        self.trace._replay_view = harness.replay_view
         if on_trace is not None:
             on_trace(self.trace)
         interceptors = [

{verifiers-0.4.1.dev27 → verifiers-0.4.1.dev29}/verifiers/v1/trace.py

@@ -433,6 +433,9 @@ class Trace(BaseModel, Generic[DataT, StateT, AgentConfigT]):
 
     _head_index: dict = PrivateAttr(default_factory=dict)
     """Physical node key -> node id for the graph builder."""
+    _replay_view: Any = PrivateAttr(default=None)
+    """The harness's `replay_view`, applied to both sides when the graph builder matches
+    a replayed message against a stored node."""
     _on_change: Any = PrivateAttr(default=None)
     _pending: dict[int, list[Message]] = PrivateAttr(default_factory=dict)
     """The messages of the request in flight that no node holds yet (the harness's tool