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

[mikasenghaas]

Summary

Companion: PrimeIntellect-ai/prime-rl#3157 (merge together).

• Pin [tool.uv.exclude-newer-package] cutoffs as full RFC 3339 UTC timestamps: bare dates resolve to midnight local time, so teammates in different timezones kept relocking uv.lock and failing the --locked pre-commit hooks.
• Treat an expired rollout deadline as an agent failure (HarnessError, ok=False, no scoring) instead of a clean harness_timeout stop that scored the partial trajectory; rename the concept harness timeout → agent timeout throughout.
• Tighten the Trace schema (TRACE_VERSION reset to 1 — a clean break, the old counter is retired): agent, verifiers, and RunInfo.id are required (verifiers auto-stamps the current build; the v0 bridge and validate/debug CLIs synthesize their seats), tools defaults to []. run and stop_condition keep their meaningful None.
• Slim the Trace surface: one verb for mutators (record_*), one reader per fact — drop the agent_name/trainable/runtime passthroughs and the duplicate error property.
• Align tool_messages with assistant_messages (nodes-based, branch-independent).
• Drop the override warnings in record_metric/record_reward: overriding is defined behavior, last write wins.
• Trim trace docstrings to the constraints they add; hoist TRACE_VERSION; rename _NODE_DUMP_EXCLUDE → EXCLUDE_FIELDS.

Breaking

• TRACE_VERSION resets to 1 on the tightened schema. No backwards compatibility: pre-existing trace records do not load, and old eval runs cannot be resumed or replayed.
• Timed-out rollouts no longer score: stop_condition="harness_timeout" + ok=True → recorded HarnessError + ok=False + stop_condition="error"; harness_timeout is gone from the stop-condition vocabulary and Trace.is_truncated (the v0 bridge maps timeout_reached through the generic truncation fallback).
• TaskTimeout.harness → TaskTimeout.agent; cap_remote_harness_timeout → cap_remote_agent_timeout.
• Trace.agent and Trace.verifiers required: constructors must pass agent=AgentInfo(config=AgentConfig()); verifiers fills itself.
• Trace.tools: list[Tool] | None → list[Tool] (default []).
• EvalRunInfo.id / TrainRunInfo.id: str | None → required str.
• Trace.stamp(run, **info) → Trace.record_run(run, **info); Trace.capture_error(e) → Trace.record_error(e).
• Trace.error → Trace.last_error; Trace.agent_name / Trace.trainable / Trace.runtime removed — read trace.agent.name / .trainable / .runtime.
• Trace.stop() requires an explicit condition (the unused "done" default is gone).
• Trace.tool_messages: last branch only → every tool result across the node graph (identical for linear rollouts).
• _NODE_DUMP_EXCLUDE → EXCLUDE_FIELDS.

Verification

• uv lock --check green under TZ=Europe/Berlin / America/Los_Angeles / Asia/Tokyo (previously failed cross-TZ with exclude newer timestamp changed).
• uv run pytest tests/v1 green; ruff + ty hooks green.
• Live 1-rollout smoke of every entrypoint on the new schema: eval (rich dashboard + hub push), validate, debug, replay (strict re-read of a v5 record), serve (boot), legacy bridge via live test_legacy e2e.

🤖 Generated with Claude Code

---

Note

High Risk
Breaking trace schema and API (TRACE_VERSION reset, no old record compatibility) plus changed timeout semantics affect eval replay, trainers, and any code reading removed Trace properties.

Overview
Resets TRACE_VERSION to 1 with a tighter trace record: agent and auto-stamped verifiers are required, tools defaults to [], and top-level shims (agent_name, trainable, runtime, error) are removed in favor of trace.agent.* and trace.last_error. Mutators are unified as record_run / record_error (replacing stamp / capture_error).

Rollout wall-clock limits are renamed harness → agent timeout (TaskTimeout.agent, cap_remote_agent_timeout). When that budget expires, the rollout now fail()s with HarnessError (ok=False, stop_condition="error") instead of a clean harness_timeout stop that could still score a partial trajectory; harness_timeout is dropped from truncation semantics and the v0 bridge mapping.

pyproject.toml pins exclude-newer-package cutoffs as full UTC RFC 3339 timestamps so uv.lock does not churn across timezones.

Call sites (envs, dashboard, push, GEPA, tests) and docs are updated to the new trace fields; validate/debug synthesize AgentInfo on traces that never run a full agent.

^{Reviewed by Cursor Bugbot for commit de8d521. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Rename harness timeout to agent timeout and migrate Trace API to agent-scoped fields

• Renames harness_timeout to agent_timeout across the stack: TaskTimeout.harness → TaskTimeout.agent, RolloutRun constructor, cap_remote_harness_timeout → cap_remote_agent_timeout, and Harbor taskset parsing.
• Restructures Trace so agent (an AgentInfo object) is required; removes top-level convenience properties agent_name, trainable, runtime, and error in favor of trace.agent.name, trace.agent.trainable, trace.agent.runtime, and trace.last_error.
• Renames Trace.stamp → Trace.record_run, Trace.capture_error → Trace.record_error, and updates all call sites across CLI, eval runner, rollout, retries, push, and environment code.
• Trace.tools now defaults to an empty list instead of None; tool_messages returns results across all graph nodes rather than only the last branch.
• Updates uv.lock exclusion timestamps to full UTC format with an explanatory comment in pyproject.toml.
• Risk: EvalRunInfo.id is now required (was optional); is_truncated no longer treats harness_timeout as a truncation condition; any code reading removed Trace properties will break.

Changes since #2172 opened

• Reworded inline comments in verifiers.v1.cli.validate._run_gold and verifiers.v1.cli.validate._run_setup functions to clarify that no agent runs occur and the info only records the runtime policy [de8d521]
• Added class docstrings to verifiers.v1.trace.TimeSpan and verifiers.v1.trace.TimeSplit classes documenting their timestamp and duration semantics [de8d521]
• Updated comments in verifiers.v1.legacy.rollout_output_to_trace function to clarify that v0 rollouts contain no agent config rather than seat config [de8d521]

^{Macroscope summarized f0723ca.}

[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 58e3a6a. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR makes significant API/schema changes to the Trace model: making agent required, renaming properties and methods, and changing timeout handling from a clean stop condition to a failure. These are substantive contract changes beyond a typical 'chore' refactor and warrant human review.

^{You can customize Macroscope's approvability policy. Learn more.}

[hallerite]

few comments, almost lgtm