emrg: add llm.jsonl rotation, enrich LLM errors with headers, make stream_options configurable

emrg/config.py

@@ -28,6 +28,9 @@ class LlmConfig:
  context_window: int = 131072
  auto_compact_threshold: float = 0.0
  models: list[dict] = field(default_factory=list) # [[llm.models]] for /model switching
+ # stream_options: None means don't send stream_options at all (for APIs like Kimi).
+ # Default is {"include_usage": False} for OpenAI/DeepSeek compatibility.
+ stream_options: Optional[dict] = field(default_factory=lambda: {"include_usage": False})
 
 
 @dataclass
@@ -59,6 +62,13 @@ def load_config() -> EmrgConfig:
  data = tomllib.loads(content)
  llm_data = data.get("llm", {})
 
+ # stream_options: None = no stream_options sent (for Kimi etc.)
+ raw_stream_opts = llm_data.get("stream_options")
+ if raw_stream_opts is None and "stream_options" not in llm_data:
+ stream_opts: Optional[dict] = {"include_usage": False} # default
+ else:
+ stream_opts = raw_stream_opts # explicit None disables it
+
  llm = LlmConfig(
  base_url=llm_data.get("base_url", "https://api.openai.com/v1"),
  api_key=llm_data.get("api_key", ""),
@@ -69,6 +79,7 @@ def load_config() -> EmrgConfig:
  context_window=llm_data.get("context_window", 131072),
  auto_compact_threshold=llm_data.get("auto_compact_threshold", 0.0),
  models=llm_data.get("models", []),
+ stream_options=stream_opts,
  )
 
  # Resolve ${ENV_VAR} placeholders in the API key

emrg/server/daemon.py

@@ -578,7 +578,7 @@ def _build_history_section(self, session: Session) -> str:
  f"- **Current history** (may be compacted): `{session.dir_path}/history.jsonl`",
  f"- **Daily full history** (never compacted): `{session.dir_path}/history_{today}.jsonl`",
  f"- Daily files are named `history_YYMMDD.jsonl`",
- f"- LLM raw log: `{session.dir_path}/llm.jsonl`",
+ f"- LLM raw log: `{session.dir_path}/llm.jsonl` (rotated at 50MB, up to 2 backups)",
  "",
  "**To read history**: use the `read` tool on `history.jsonl` for the current",
  "context, or on a specific `history_YYMMDD.jsonl` file for older messages.",

emrg/server/llm.py

@@ -58,7 +58,8 @@ def _make_payload(
  payload["tools"] = tools
  if stream:
  payload["stream"] = True
- payload["stream_options"] = {"include_usage": False}
+ if self.config.stream_options is not None:
+ payload["stream_options"] = self.config.stream_options
  return payload
 
  def _headers(self) -> dict:
@@ -110,8 +111,11 @@ async def chat(
  )
  continue
 
- logger.error("LLM error: %s %s", resp.status_code, text)
- raise RuntimeError(f"LLM request failed: {resp.status_code} - {text}")
+ hdr = dict(resp.headers)
+ logger.error("LLM error: %s headers=%s body=%s", resp.status_code, hdr, text)
+ raise RuntimeError(
+ f"LLM request failed: {resp.status_code} headers={hdr} body={text}"
+ )
 
  raise last_error # type: ignore[misc]
 
@@ -173,8 +177,10 @@ async def chat_stream(
  )
  continue
  logger.error("LLM stream error: %s %s", resp.status_code, text[:500])
+ hdr = dict(resp.headers)
  raise RuntimeError(
- f"LLM stream request failed: {resp.status_code}"
+ f"LLM stream request failed: {resp.status_code} "
+ f"headers={hdr} body={text[:1000]}"
  )
 
  async for line in resp.aiter_lines():

emrg/session.py

@@ -15,6 +15,7 @@
 
 import json
 import logging
+import os
 import secrets
 from datetime import datetime
 from pathlib import Path
@@ -23,6 +24,10 @@
 
 logger = logging.getLogger(__name__)
 
+# ── llm.jsonl rotation ────────────────────────────────────────
+_LLM_LOG_MAX_BYTES = 50 * 1024 * 1024 # 50 MB
+_LLM_LOG_BACKUP_COUNT = 2
+
 
 def generate_session_id(cwd: Path) -> str:
  """Generate a human-friendly session ID: s_YYMMDD_HHMM_xxxx."""
@@ -181,17 +186,49 @@ def append_message(self, record: dict) -> None:
  self._save_meta()
 
  def append_llm(self, record: dict) -> None:
- """Append an LLM request/response record to llm.jsonl."""
+ """Append an LLM request/response record to llm.jsonl.
+
+ Automatically rotates the file when it exceeds 50 MB, keeping
+ up to 2 backup files (llm.jsonl.1, llm.jsonl.2).
+ """
  if "timestamp" not in record:
  record["timestamp"] = datetime.now().isoformat()
 
  entry = {"timestamp": record.pop("timestamp")}
  entry.update(record)
 
  line = json.dumps(entry, ensure_ascii=False) + "\n"
+
+ # Rotate if file exceeds max size
+ self._rotate_llm_log()
+
  with open(self._llm_path, "a") as f:
  f.write(line)
 
+ def _rotate_llm_log(self) -> None:
+ """Rotate llm.jsonl if it exceeds _LLM_LOG_MAX_BYTES."""
+ if not self._llm_path.exists():
+ return
+ try:
+ size = os.path.getsize(self._llm_path)
+ except OSError:
+ return
+ if size < _LLM_LOG_MAX_BYTES:
+ return
+
+ # Shift existing backups: .2 → .3 (delete), .1 → .2, main → .1
+ for i in range(_LLM_LOG_BACKUP_COUNT, 0, -1):
+ old = Path(str(self._llm_path) + f".{i}")
+ new = Path(str(self._llm_path) + f".{i + 1}")
+ if i == _LLM_LOG_BACKUP_COUNT and new.exists():
+ new.unlink()
+ if old.exists():
+ old.rename(new)
+
+ # Rename current to .1
+ backup_path = Path(str(self._llm_path) + ".1")
+ self._llm_path.rename(backup_path)
+
  # ── History reading ───────────────────────────────────────
 
  def _read_history(self) -> list[dict]: