fix: preserve rollout meta in the v0 -> v1 legacy bridge

[mikasenghaas]

Summary

The v0 → v1 legacy bridge (rollout_output_to_trace) was producing a near-empty Turn.response: it kept only the token ids and dropped almost everything else a native v1 trace carries.

Root cause: a v0 RolloutOutput nests its data as pydantic objects in-process (SystemMessage/UserMessage/AssistantMessage, Response), and records finish_reason on response.message (not the response). The bridge's helpers only handled plain dicts (if not isinstance(m, dict): continue) and read finish_reason off the response top-level — so the message lists were silently skipped and the response fields came back null.

Fix:

• _as_dict() coerces v0 pydantic objects to dicts before mapping, so the helpers read them whether they arrive as objects or dicts.
• _to_v1_response now reads finish_reason from response.message (its v0 location), keeps usage, and mirrors the token ids onto response.tokens (as the native v1 client does).
• New _to_wire_task carries the prompt's system message → system_prompt, user message(s) → instruction, and the reference answer (as a WireTask extra) onto the task.

A v0-bridged Trace now matches the native v1 schema.

Discrepancies fixed (v0-bridge vs native v1, reverse-text step 0)

Before, the bridge dropped: turn.prompt (empty), response.message.content / reasoning_content (null), finish_reason (null), usage (null), response.tokens (null), task.instruction (empty), task.system_prompt (null), task.answer (missing). is_truncated also stayed wrong because it's computed from finish_reason.

After, the only remaining schema differences are benign and not lost information:

• rewards/metrics naming: v0 names its aggregate rewards.reward + metrics.lcs_reward_func; v1 names it rewards.lcs. The value is preserved and trace.reward resolves identically — an env-level naming difference the bridge shouldn't rewrite.
• message.content str vs null|str: sampling artifact; the field is nullable in both.

Verification

Captured real v0 RolloutOutputs through the legacy path against a live vLLM server and diffed the resulting traces against native v1 reverse-text rollouts: prompt messages, response content/reasoning, finish_reason, usage, tokens (turn + response), and task.system_prompt/instruction/answer all now present and matching. finish_reason="length" correctly drives the computed is_truncated.

---

Note

Medium Risk
Changes only the legacy adapter path but affect training-visible trace shape (tokens, finish_reason, task fields); incorrect mapping could still skew truncation or RL reads if edge cases remain.

Overview
The v0→v1 legacy bridge in rollout_output_to_trace was dropping most trace metadata because mapping helpers only accepted plain dicts while in-process v0 rollouts use Pydantic models, and finish_reason was read from the wrong level.

_as_dict() normalizes v0 objects via model_dump() before _to_v1_messages, _tool_calls, and _to_v1_response run, so prompts, assistant content/reasoning/tool calls, and usage populate correctly instead of being skipped.

_to_v1_response now resolves finish_reason from response.message (v0) with a fallback to the response root, restricts values to stop / length / tool_calls, and attaches the same turn tokens on response.tokens to match native v1 clients.

_to_wire_task replaces flattening the whole prompt into instruction: system → system_prompt, user text → instruction, and optional reference answer as a WireTask extra.

Together, bridged traces align with native v1 for training-critical fields; is_truncated can compute correctly once finish_reason="length" is preserved.

^{Reviewed by Cursor Bugbot for commit fdf82e8. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix rollout metadata preservation in the v0 to v1 legacy bridge

• Adds _as_dict helper in legacy.py to coerce pydantic-like v0 objects to dicts before mapping, fixing dropped messages and tool calls.
• _to_v1_response now populates Response.tokens and sets finish_reason only when it is one of stop, length, or tool_calls; otherwise None.
• rollout_output_to_trace now sets system_prompt from the first system message, builds instruction from user messages only, and carries answer in task extras.
• Response.tokens is now populated per turn by passing computed TurnTokens through _to_v1_response.

^{Macroscope summarized fdf82e8.}

[macroscopeapp]

Approvability

Verdict: Approved

Bug fix to preserve more metadata (tokens, finish_reason, system_prompt, answer) during v0->v1 trace conversion in the legacy bridge. Changes are self-contained to the conversion layer and authored by the module's original creator.

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