Hacktoberfest 2026: le issue che i maintainer hanno segnato per ottobre, aperte e adatte ai principianti. Sfoglia le issue Hacktoberfest

Codex rollout reader drops raw reasoning from open-model servers

Aperta
  • Qualcosa non va?
#1,017 2 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

@jiwasawa ci sta già lavorando.

Dal 5/10/2026.

  • #1 di @jiwasawa — aperta
  • #1063 di @jiwasawa — aperta

Valutazione

Difficoltà
2/5
Tempo stimato
1-3 ore
Idoneità per principianti
45/100
Tipo di issue
Bug
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
python
Ambito
data

Direzione di ricerca

Read engine/resources/agent_rollout/codex.py, especially extract_codex_reasoning_summaries, and the rollout tests under tests/engine/resources/agent_rollout/. Use the provided JSONL reproduction to check how reasoning items with content and summary are parsed. Done means the assistant message has reasoning_content from content, with summary as fallback, and the existing rollout tests still pass.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

bug triaged
Priority Level

Medium (Annoying but has workaround)

Describe the bug

AgentRolloutSeedSource with format=codex drops the reasoning of Codex sessions that ran against an open model served by vLLM.
extract_codex_reasoning_summaries in engine/resources/agent_rollout/codex.py reads only the summary field of a reasoning response item.
vLLM's Responses API returns reasoning as content: [{"type": "reasoning_text", ...}] with an empty summary, and Codex writes that item to the rollout file unchanged.
The resulting assistant messages have no reasoning_content, so the reasoning is missing from the seed data.

Steps/Code to reproduce bug

Reproduced on main (c05a87e) with Python 3.14.3 and dependencies installed with uv sync --all-packages --frozen.
No model calls are needed.
The code calls CodexAgentRolloutFormatHandler, the handler that AgentRolloutSeedSource(format="codex") uses (registered in agent_rollout/registry.py).

import json
import tempfile
from pathlib import Path

from data_designer.engine.resources.agent_rollout.codex import CodexAgentRolloutFormatHandler

root = Path(tempfile.mkdtemp())
rows = [
    {"type": "session_meta", "payload": {"id": "sess-1"}},
    {"type": "response_item", "payload": {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "What is 2+2?"}]}},
    {"type": "response_item", "payload": {"type": "reasoning", "summary": [], "content": [{"type": "reasoning_text", "text": "2+2 is 4."}], "encrypted_content": None}},
    {"type": "response_item", "payload": {"type": "message", "role": "assistant", "content": [{"type": "output_text", "text": "4"}]}},
]
(root / "rollout-repro.jsonl").write_text("\n".join(json.dumps(r) for r in rows) + "\n")
record = CodexAgentRolloutFormatHandler().parse_file(root_path=root, relative_path="rollout-repro.jsonl")[0]
print(next(m for m in record.messages if m["role"] == "assistant").get("reasoning_content"))

Output: None

Expected behavior

The assistant message has reasoning_content == "2+2 is 4.".
When a reasoning item has text in both content and summary, the reader should prefer content (the full reasoning) and fall back to summary, OpenAI-hosted models do not return raw reasoning content, so their rollouts keep the current behavior.

Agent Diagnostic / Prior Investigation
  • Searched open and closed issues, PRs, and discussions for "codex reasoning", "codex rollout", "reasoning_content", "reasoning_text", "agent_rollout", and "AgentRolloutSeedSource": no existing report.
    #374 also concerns vLLM reasoning fields, but in model client response parsing, not in rollout ingestion.
  • Searched the docs: the agent rollout ingestion page (agent-rollout-ingestion.mdx) does not mention reasoning, so it does not say which Codex reasoning fields are read.
  • End to end: Codex CLI 0.159.2 against vLLM 0.30.0 serving nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8, two turns with tool calls, ingested with Data Designer 0.9.2.
    The rollout file had 4 reasoning items with reasoning_text content and an empty summary.
    The reader set reasoning_content on 0 of 4 corresponding assistant messages.
  • Workaround: rewrite the rollout files to copy content into summary before ingestion.
Additional context

I have a fix ready: a small change in codex.py (+11/−6 lines) that reads content first and falls back to summary, plus one unit test.
On main, tests/engine/resources/agent_rollout/ passes (49 tests) with the fix applied, and the new test fails without it.
I will open the PR once this issue is triaged.

Checklist
  • I reproduced this issue or provided a minimal example
  • I searched the docs/issues myself, or had my agent do so
  • If I used an agent, I included its diagnostics above
Lingua principale
Python
Stelle
2.4k
Fork
234
Merge medio
22h 5m
PR unite (30g)
70

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di NVIDIA-NeMo/DataDesigner

Tutte le issue di NVIDIA-NeMo/DataDesigner

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.