[mikasenghaas]

Summary

• Rebuild streamed reasoning from raw chat completion chunks in the bundled harness.
• Support current vLLM reasoning, legacy reasoning_content, and structured reasoning_details fields.
• Join reasoning text while preserving structured scalar metadata such as id, index, format, and signature.
• Use one field-aware override path for repeated role metadata and provider reasoning fields.
• Keep the interception parser and graph hashing unchanged.

Verification

• Before: two structured reasoning chunks with id: "r1" and format: "unknown" replayed as id: "r1r1" and format: "unknownunknown". The intercepted and replayed message hashes differed.
• After: the harness replays scalar metadata once and joins text deltas. The intercepted and replayed message hashes match.
• uv run eval terminal-bench-2 --env.taskset.tasks '["fix-git"]' --env.agent.harness.id bash -n 1 -r 2 -c 2 --no-push: both rollouts solved the task with reward 1.0. They took 13 and 12 turns, with 1 branch each.
• A scripted uv run eval echo-tool-v1 ... comparison completed with reward 1.0. Base main produced 2 turns and 2 branches. This branch produced 2 turns and 1 branch.
• The regression test covers repeated role metadata plus reasoning, reasoning_content, and reasoning_details stream fields.
• uv run pytest tests/: 87 passed, 75 skipped.
• uv run pre-commit run --all-files: passed.
  \n- Current-head live rerun after the review refactor: reward 1.0, 15 turns, 1 branch, and no errors.\n\n

Note

Preserve streamed reasoning details and role in core.chat

• Adds _accumulate_streamed_message to core.py to concatenate reasoning text and merge compatible reasoning-detail fragments across streamed chunks.
• Updates core.chat to apply these accumulated fields to the final completion message, defaulting the role to assistant when missing.
• Adds a mocked SSE test in test_e2e.py to verify the chat harness preserves concatenated reasoning and details.

^{Macroscope summarized 5367fa3.}

---

Note

Medium Risk
Changes the shared harness streaming completion assembly path used by eval rollouts; incorrect merging could alter intercepted messages and downstream replay/hash behavior.

Overview
Fixes how the bundled harness chat() rebuilds assistant messages from streaming chat completion chunks so provider reasoning fields survive without SDK mis-merging.

Instead of forcing role = "assistant" on the final snapshot, the stream loop now accumulates per-choice overrides for role, reasoning, reasoning_content, and reasoning_details. String reasoning fields are concatenated across deltas; reasoning_details entries of the same type merge text/summary while keeping scalar metadata (id, index, format, signature) from the first chunk rather than duplicating them.

Adds test_chat_harness_preserves_streamed_reasoning, which mocks an SSE stream and asserts the assembled message matches joined reasoning text and a single structured reasoning.text detail.

^{Reviewed by Cursor Bugbot for commit 5367fa3. Bugbot is set up for automated code reviews on this repo. Configure here.}

[mikasenghaas]

make sure this works for any streamed reasoning (no standards since oai chat completions doesnt support)

[macroscopeapp]

Approvability

Verdict: Approved at 5367fa3

Macroscope's review found this PR approvable — This is a narrowly scoped harness bug fix that preserves streamed reasoning fields and adds a mocked SSE regression test. Runtime impact is confined to the affected streamed-completion metadata path, with no schema, infrastructure, security, billing, or authentication changes.

^{You can add or adjust custom eligibility rules. Learn more.}