chore: timezone-stable uv.lock and trace record cleanups

docs/v1/env.md

@@ -33,7 +33,7 @@ class DebateEnv(vf.Env[DebateConfig]):
         await agents.judge.run(VerdictTask.from_traces(task, pro, con))
 
     async def finalize(self, task: vf.Task, episode: vf.Episode) -> None:
-        by_agent = {t.agent_name: t for t in episode.traces}
+        by_agent = {t.agent.name: t for t in episode.traces}
         winner = (by_agent["judge"].last_reply or "").strip().lower()
         by_agent["pro"].record_reward("won", float(winner == "pro"))
         by_agent["con"].record_reward("won", float(winner == "con"))

environments/proposer_solver_v1/proposer_solver_v1/taskset.py

@@ -150,7 +150,7 @@ async def run(self, task: vf.Task, agents: vf.Agents) -> None:
 
     @staticmethod
     def _solve_rate(traces: list[vf.Trace]) -> float:
-        solves = [t for t in traces if t.agent_name == "solver"]
+        solves = [t for t in traces if t.agent.name == "solver"]
         if not solves:
             return 0.0
         return sum(
@@ -163,7 +163,7 @@ async def finalize(self, task: vf.Task, episode: vf.Episode) -> None:
         the problem, 0 when it's trivial or impossible for them (4p(1-p))."""
         rate = self._solve_rate(episode.traces)
         for trace in episode.traces:
-            if trace.agent_name == "proposer":
+            if trace.agent.name == "proposer":
                 trace.record_metric("solve_rate", rate)
                 trace.record_reward("learnability", 4.0 * rate * (1.0 - rate))
 

examples/agent.py

@@ -20,14 +20,13 @@ async def main() -> None:
     async with solver:
         trace = await solver.run(task)
     print("stop:", trace.stop_condition)
-    print("error:", trace.error)
+    print("error:", trace.last_error)
     print("turns:", trace.num_turns)
     print("usage:", trace.usage)
     last = trace.assistant_messages[-1].content if trace.assistant_messages else None
     print("answer:", last)
-    assert trace.agent is not None
     print("agent:", trace.agent.name, trace.agent.config.model)
-    print("runtime:", trace.runtime.type if trace.runtime else None)
+    print("runtime:", trace.agent.runtime.type if trace.agent.runtime else None)
 
 
 if __name__ == "__main__":

pyproject.toml

@@ -136,8 +136,10 @@ kuhn-poker-v1 = { path = "environments/kuhn_poker_v1", editable = true }
 
 [tool.uv.exclude-newer-package]
 # Bounded cutoffs for the explicitly requested tool upgrades.
-ruff = "2026-07-27"
-ty = "2026-07-27"
+# Full UTC timestamps: bare dates resolve to local-tz midnight and churn the
+# lockfile across timezones.
+ruff = "2026-07-28T00:00:00Z"
+ty = "2026-07-28T00:00:00Z"
 # PrimeIntellect-published on PyPI (trusted publisher)
 prime-tunnel = false
 prime-sandboxes = false

skills/evaluate-environments/references/REFERENCE.md

@@ -180,7 +180,7 @@ Trainability is not a config field: it is env truth, set in place by the env's `
 
 The separate `verifiers/v1/configs/env.py` `TimeoutConfig` (the env's `--env.timeout.*`) keeps only `episode` — the bound on the whole `run()` interaction — and `finalize` — the bound on the env's `finalize()` hook.
 
-> Remote sandboxes cap any harness timeout at 24 hours (provider max lifetime).
+> Remote sandboxes cap any agent timeout at 24 hours (provider max lifetime).
 
 ---
 

tests/v1/test_e2e.py

@@ -229,7 +229,7 @@ async def test_tool(run_v1, harness_runtime, tool_runtime, tmp_path):
     assert trace.reward == 1.0
     # The interception server captured the advertised tools onto the trace (for tool-use SFT):
     # the null harness offered the task's MCP tool as `echo_back`, schema included.
-    assert trace.tools is not None
+    assert trace.tools
     (echo_tool,) = [t for t in trace.tools if t.name == "echo_back"]
     assert "message" in echo_tool.parameters.get("properties", {})
 
@@ -355,9 +355,9 @@ async def test_multi_agent_env(run_v1, tmp_path):
         max_turns=2,
     )
     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"]
-    assert b.trainable is False
+    assert sorted(t.agent.name for t in traces) == ["a", "b"]
+    (b,) = [t for t in traces if t.agent.name == "b"]
+    assert b.agent.trainable is False
     for trace in traces:
         assert trace.ok
         assert trace.reward == 1.0  # each seat's own task reward
@@ -385,7 +385,7 @@ async def test_env_id_best_of_n(run_v1, tmp_path):
         max_turns=2,
     )
     assert len(traces) == 2  # one episode, two attempts
-    assert all(t.agent_name == "agent" and t.ok for t in traces)
+    assert all(t.agent.name == "agent" and t.ok for t in traces)
     assert any(t.metrics["best"] == 1.0 for t in traces)
     assert all(t.metrics["pass_at_n"] == 1.0 for t in traces)  # echo always passes
 
@@ -423,11 +423,11 @@ async def test_env_id_agentic_judge(run_v1, tmp_path):
         max_turns=10,
         rollout_timeout=600,
     )
-    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"]
+    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"]
     assert solver.ok and judge.ok
-    assert judge.trainable is False
+    assert judge.agent.trainable is False
     # The task's own reward keeps its raw score; the rescale lands on the weight.
     assert solver.rewards["echoed"].score == 1.0
     assert solver.rewards["echoed"].weight == 0.5
@@ -448,11 +448,11 @@ async def test_env_id_user_sim(run_v1, tmp_path):
         max_turns=6,
         rollout_timeout=300,
     )
-    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"]
+    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"]
     assert assistant.ok and user.ok
-    assert user.trainable is False
+    assert user.agent.trainable is False
     assert user.num_turns >= 1  # the modeled user actually spoke
     assert assistant.metrics["user_turns"] >= 1
     # `mask_prompt`: the scenario is hidden from the assistant's harness (the run's
@@ -465,7 +465,7 @@ async def test_env_id_user_sim(run_v1, tmp_path):
     from verifiers.v1.trace import WireTrace
 
     (record,) = read_episodes(tmp_path, WireTrace)
-    assert {t.agent_name for t in record.traces} == {"assistant", "user"}
+    assert {t.agent.name for t in record.traces} == {"assistant", "user"}
     assert record.id  # both traces are persisted under one durable episode identity
 
 
@@ -488,14 +488,14 @@ async def test_env_id_user_sim_with_tools(run_v1, tmp_path):
         max_tokens=8192,
         rollout_timeout=300,
     )
-    (assistant,) = [t for t in traces if t.agent_name == "assistant"]
-    (user,) = [t for t in traces if t.agent_name == "user"]
+    (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
     assert assistant.task.data.prompt is None  # the scenario stayed off the wire
     assert user.num_turns >= 1  # the modeled user actually drove the exchange
     assert assistant.rewards["echoed"].score == 1.0  # the tool ran, mid-conversation
     # The tool was advertised to the masked chat exactly as to any run.
-    assert assistant.tools is not None
+    assert assistant.tools
     assert any(tool.name == "echo_back" for tool in assistant.tools)
 
 
@@ -514,8 +514,8 @@ async def test_kuhn_poker_self_play(run_v1, tmp_path):
         max_tokens=8192,
         rollout_timeout=300,
     )
-    assert sorted(t.agent_name for t in traces) == ["player0", "player1"]
-    payoffs = {t.agent_name: t.rewards["payoff"].score for t 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
     assert abs(payoffs["player0"]) in (1.0, 2.0)
     for trace in traces:
@@ -542,7 +542,7 @@ async def test_multi_agent_env_server(run_v1_server, tmp_path):
         max_turns=2,
     )
     assert len(traces) == 2
-    assert sorted(t.agent_name for t in traces) == ["a", "b"]
+    assert sorted(t.agent.name for t in traces) == ["a", "b"]
     for trace in traces:
         assert trace.ok
         assert trace.metrics["duet"] == 1.0

tests/v1/test_graph.py

@@ -40,7 +40,8 @@ def test_routed_experts_attributed_and_aligned_across_turns():
     concatenates back to a `[tokens, layers, top_k]` array aligned 1:1 with `branch.token_ids` —
     and survives the base64 wire round-trip."""
     trace = vf.Trace(
-        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="x"))
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
+        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="x")),
     )
     user = vf.UserMessage(content="u1")
     graph.prepare_turn(trace, [user]).commit(
@@ -94,7 +95,8 @@ def test_routed_experts_none_when_absent():
     """No routing captured (engine ran without `enable_return_routed_experts`) -> the branch
     reports None and the trainer simply skips replay."""
     trace = vf.Trace(
-        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="x"))
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
+        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="x")),
     )
     graph.prepare_turn(trace, [vf.UserMessage(content="u1")]).commit(
         vf.Response(
@@ -128,7 +130,10 @@ def test_tool_call_hash_matches_v0_content_and_arguments_normalization():
 
 def test_reasoning_content_participates_in_graph_prefix_matching():
     task = vf.TaskData(idx=0, prompt="use a tool")
-    trace = vf.Trace(task=vf.TraceTask(type="Task", data=task))
+    trace = vf.Trace(
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
+        task=vf.TraceTask(type="Task", data=task),
+    )
     user = vf.UserMessage(content="use a tool")
     call = vf.ToolCall(id="call_0", name="lookup", arguments="{}")
 
@@ -205,7 +210,8 @@ def second_turn(trace, prompt_ids):
 
     # Control: the prior turn re-renders to the same tokens -> stays one linear branch.
     linear = vf.Trace(
-        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="x"))
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
+        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="x")),
     )
     first_turn(linear)
     second_turn(linear, [1, 2, 3, 4, 5, 6, 7])
@@ -214,7 +220,8 @@ def second_turn(trace, prompt_ids):
 
     # Break: the assistant turn retokenizes (4 -> 99), so prompt_ids diverge at that node.
     broken = vf.Trace(
-        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="x"))
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
+        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=0, prompt="x")),
     )
     first_turn(broken)
     second_turn(broken, [1, 2, 3, 99, 5, 6, 7])
@@ -227,7 +234,10 @@ def second_turn(trace, prompt_ids):
 
 def test_prompt_supplied_assistant_messages_are_not_sampled_turns():
     task = vf.TaskData(idx=0, prompt="few-shot")
-    trace = vf.Trace(task=vf.TraceTask(type="Task", data=task))
+    trace = vf.Trace(
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
+        task=vf.TraceTask(type="Task", data=task),
+    )
     fabricated = vf.AssistantMessage(
         content=None,
         tool_calls=[vf.ToolCall(id="call_0", name="lookup", arguments="{}")],

tests/v1/test_judges.py

@@ -36,6 +36,7 @@ def make_trace(
     task_cls: type[QAData] = QAData,
 ) -> vf.Trace:
     return vf.Trace(
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
         task=vf.TraceTask(
             type="Task",
             data=task_cls(idx=0, prompt="Capital of France?", answer=answer),
@@ -244,6 +245,7 @@ async def test_reference_score_messages_prompt(fake_judge_model):
         answer="Paris",
     )
     trace = vf.Trace(
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
         task=vf.TraceTask(type="Task", data=task),
         nodes=[
             MessageNode(parent=None, message=UserMessage(content="q"), sampled=False),
@@ -269,6 +271,7 @@ class FieldTask(vf.TaskData):
         answer="Paris",
     )
     trace = vf.Trace(
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
         task=vf.TraceTask(type="Task", data=task),
         nodes=[
             MessageNode(parent=None, message=UserMessage(content="q"), sampled=False),
@@ -293,6 +296,7 @@ def full_trace_fixture() -> vf.Trace:
     from verifiers.v1.types import ToolCall, ToolMessage
 
     return vf.Trace(
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
         task=vf.TraceTask(
             type="Task", data=QAData(idx=0, prompt="Capital of France?", answer="Paris")
         ),

tests/v1/test_trace.py

@@ -21,7 +21,8 @@ class MyState(vf.State):
 def test_bare_trace_round_trip():
     # The minimal trace: a base task, no nodes, no extras — dump and back into a plain Trace.
     tr = vf.Trace(
-        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=3, prompt="hello"))
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
+        task=vf.TraceTask(type="Task", data=vf.TaskData(idx=3, prompt="hello")),
     )
     rt = vf.Trace.model_validate(tr.model_dump())
     assert rt.id == tr.id
@@ -35,6 +36,7 @@ def test_custom_task_state_round_trip():
     # Custom data and state round-trip into the same parameterization. Data fields are
     # typed (not just `model_extra`); `state` is runtime-only and never crosses the wire.
     tr = vf.Trace[MyTask, MyState](
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
         task=vf.TraceTask(type="MyTask", data=MyTask(idx=0, prompt="q", answer="gold")),
         state=MyState(score=7),
         nodes=[
@@ -59,6 +61,7 @@ def test_wire_trace_round_trip():
     # Two leaves off one root → 2 branches (a compaction-shaped trace), so the round-trip has to
     # carry node `parent` links for `num_branches` to survive.
     tr = vf.Trace[MyTask, vf.State](
+        agent=vf.AgentInfo(config=vf.AgentConfig()),
         task=vf.TraceTask(type="MyTask", data=MyTask(idx=0, prompt="q", answer="a")),
         tools=[vf.Tool(name="echo", description="", parameters={"type": "object"})],
         nodes=[

verifiers/v1/agent.py

@@ -45,7 +45,7 @@
     UserMessage,
 )
 from verifiers.v1.utils.compile import (
-    cap_remote_harness_timeout,
+    cap_remote_agent_timeout,
     resolve_runtime_config,
     validate_pairing,
 )
@@ -390,7 +390,7 @@ async def run(
                 attempt + 1,
                 retry.max_retries,
                 delay,
-                trace.error.type if trace.error else "?",
+                trace.last_error.type if trace.last_error else "?",
             )
             await asyncio.sleep(delay)
         if history:
@@ -419,8 +419,8 @@ async def _run_once(
             # close() never runs — free the run's servers and owned runtime first.
             await run.abort()
             raise
-        if trace.runtime is not None:
-            trace.runtime.borrowed = runtime is not None
+        if trace.agent.runtime is not None:
+            trace.agent.runtime.borrowed = runtime is not None
         return trace
 
     @asynccontextmanager
@@ -482,8 +482,8 @@ async def interaction(
             opened = await run.open()
             if not opened:
                 trace = await run.close()
-                if trace.runtime is not None:
-                    trace.runtime.borrowed = runtime is not None
+                if trace.agent.runtime is not None:
+                    trace.agent.runtime.borrowed = runtime is not None
         if not opened:
             failure = run.failure
             if failure is None:  # `open()` returning False always captures one.
@@ -499,8 +499,8 @@ async def interaction(
             raise
         finally:
             trace = run.trace if run.closed else await interaction.close()
-            if trace.runtime is not None:
-                trace.runtime.borrowed = runtime is not None
+            if trace.agent.runtime is not None:
+                trace.agent.runtime.borrowed = runtime is not None
 
     def _rollout_params(
         self, task: Task, runtime: Runtime | None, shared_tools: dict
@@ -520,10 +520,10 @@ def _rollout_params(
             self.harness, type(task), runtime_config, shared_tools=shared_tools
         )
         # Timeout precedence: agent-level wins, else the task's, else no limit.
-        harness_timeout = (
+        agent_timeout = (
             self.timeout.rollout
             if self.timeout.rollout is not None
-            else task.data.timeout.harness
+            else task.data.timeout.agent
         )
         return {
             "agent_config": self.config,
@@ -535,8 +535,8 @@ def _rollout_params(
                 if self.timeout.setup is not None
                 else task.data.timeout.setup
             ),
-            "harness_timeout": cap_remote_harness_timeout(
-                harness_timeout, runtime_config, task
+            "agent_timeout": cap_remote_agent_timeout(
+                agent_timeout, runtime_config, task
             ),
             "finalize_timeout": (
                 self.timeout.finalize