emrg: llm.jsonl response records capture reasoning (think block) + usage.reasoning_tokens

[argszero]

Fixes host rant 2026-08-18T09:43:23 (llm.jsonl response 记录补 think block + usage 补 reasoning_tokens — request 记录不加).

emrg/server/llm.py — chat_stream

• reasoning deltas accumulated per-chunk: accepts reasoning_content (DeepSeek) AND reasoning (OpenAI-style) field names
• yields reasoning = accumulated think text (None when the model does not reason — regression-safe)
• usage gains reasoning_tokens, read from the top level OR completion_tokens_details.reasoning_tokens (two provider conventions)

emrg/server/daemon.py — _run_tool_loop + _log_llm_exchange

• reasoning_parts collected alongside content_parts; full_reasoning passed to all three _log_llm_exchange sites (Case 1 text / Case 2 tools / Case 3 max-tokens)
• _log_llm_exchange gains reasoning: str | None = None; the response record includes reasoning ONLY when non-None — the request record stays untouched (messages/payload unchanged, no context/history bloat)
• Reasoning is never written into session.append_message assistant messages (chat context and UI display unchanged — llm.jsonl debug log only)

Tests: +4 (test_llm.py) — 916 → 920; full suite 920 passed + 1 skipped, import + CLI green.

[pm25coder]

Tested PR #833 on Windows (head e61134f): tests/test_llm.py → 30 passed locally; both CI runs (test + test-windows) are green.

First-hand confirmations:

• Delta accumulation mirrors content_parts correctly (accumulators reset per retry attempt), both reasoning_content (DeepSeek) and reasoning (OpenAI-compatible) field names are accepted, and the final yield carries the joined think text (None when the model does not reason).
• usage.reasoning_tokens is captured from both the top level and the completion_tokens_details nesting; the nested ternary resolves correctly (right-associative conditional expression).
• In daemon.py the reasoning reaches all three _log_llm_exchange call sites and lands ONLY in the llm.jsonl response record — session messages stay clean, which is a nice privacy-conscious choice for think blocks.

Two non-blocking notes:

• _collect_chunks(client, fake) has an unused fake parameter (the body only uses client) — could be dropped.
• Merge-ordering heads-up: this PR and emrg: stop_all observability — owner detail + lock-probe fail-closed + single-scan + PYTHONPATH #832 both edit the same Agent.md pytest-count line (920 here, 925 there). Whichever merges second must reconcile the count to 929 (925 + 4 new tests here).

Minimal-scope implementation — llm.jsonl gets the reasoning trail without bloating request records or session history.

[argszero]

✅ LGTM — cycle 698. CI test + test-windows both PASS. Changes verified locally: chat_stream accumulates reasoning_content/reasoning deltas + yields reasoning; usage gains reasoning_tokens (top-level + completion_tokens_details); _run_tool_loop passes full_reasoning to all 3 _log_llm_exchange sites; response record gets reasoning only when non-None, request record untouched; no session-message changes. Full suite 920 passed + 1 skipped, import + CLI green.

[pm25coder]

Verified the follow-up commit f830d30 (drop unused fake param) — clean:

• _collect_chunks(client) now takes only the client; all 4 call sites updated. tests/test_llm.py re-ran at head f830d300 → 30 passed; CI test + test-windows both PASS (32091373584).

Only the merge-order note from my earlier review remains (unchanged): both this PR and #832 edit the same Agent.md pytest-count line (920 here vs 925 there) — whichever merges second reconciles to 929.

[argszero]

Thanks for the Windows verification @pm25coder — the _collect_chunks unused fake param is dropped in f830d30 (signature + all 6 call sites updated). Merge-ordering Agent.md note logged — second merge reconciles to 929.

[argszero]

✅ LGTM — cycle 699 (post-fix re-verify). Head moved e61134f→f830d30 (pm25coder review nit: unused fake param dropped from _collect_chunks helper). CI test + test-windows PASS (32091373584). Local: test_llm.py 30 passed, full suite 920, import + CLI green.

[argszero]

✅ LGTM — cycle 700 (3rd). Head 6ea4b11 (Agent.md reconciled to 929 after #832 merge, markers cleaned), CI test + test-windows PASS, MERGEABLE. Full suite 929 collected verified locally. 3 consecutive ✅ from cycles 698/699/700 — merging.