emrg: Redact LLM error logging — mask response headers + inline credentials in error body

Agent.md

@@ -93,7 +93,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.port; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (490) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (493) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (86: 22 daemon_client + 22 app-commands + 17 renderer smoke + 15 i18n + 7 integration + 3 commands) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 

README.md

@@ -276,7 +276,7 @@ EMRG doesn't just keep up — it catches up on its own.
 git clone https://github.com/argszero/emrg.git
 cd emrg
 uv sync # install deps
-uv run pytest tests/ -v # run tests (currently 490 items)
+uv run pytest tests/ -v # run tests (currently 493 items)
 uv run python -m emrg # launch TUI
 # CI includes actionlint workflow gate (#444): workflow parse errors fail PR CI
 

emrg/server/llm.py

@@ -23,6 +23,31 @@
 from emrg import __version__
 from emrg.config import LlmConfig
 
+
+# ── 错误信息脱敏（20260807-0107）────────────────────────────
+# LLM 错误日志/异常可能包含 response headers（set-cookie/auth 回显）与 body
+# （模型回显的密钥/令牌）。复用 daemon._redact_string 的内联凭据遮蔽能力，
+# 避免 emrgd.log / 会话历史泄露。
+def _redact_text(text: str) -> str:
+ """遮蔽字符串中的内联凭据（sk-/ghp_/Bearer/JWT/base64-JSON）。"""
+ try:
+ from emrg.server.daemon import _redact_string
+ return _redact_string(text)
+ except Exception:
+ return text
+
+
+def _redact_headers(headers: dict) -> dict:
+ """遮蔽 response headers 中的敏感键值（set-cookie/authorization/token 等）。"""
+ sensitive = ("cookie", "set-cookie", "authorization", "token", "api-key", "apikey", "x-api-key")
+ out = {}
+ for k, v in headers.items():
+ if any(s in k.lower() for s in sensitive):
+ out[k] = "***"
+ else:
+ out[k] = _redact_text(str(v))
+ return out
+
 logger = logging.getLogger(__name__)
 
 # HTTP status codes that warrant a retry
@@ -115,18 +140,22 @@ async def chat(
  delay = RETRY_BASE_DELAY * (2 ** attempt)
  logger.warning(
  "LLM transient error %d, retrying in %.1fs (attempt %d/%d): %s",
- resp.status_code, delay, attempt + 1, MAX_RETRIES, text[:200],
+ resp.status_code, delay, attempt + 1, MAX_RETRIES,
+ _redact_text(text[:200]),
  )
  await asyncio.sleep(delay)
  last_error = RuntimeError(
- f"LLM request failed: {resp.status_code} - {text}"
+ f"LLM request failed: {resp.status_code} - {_redact_text(text)}"
  )
  continue
 
- hdr = dict(resp.headers)
- logger.error("LLM error: %s headers=%s body=%s", resp.status_code, hdr, text)
+ # 错误日志与异常信息脱敏：response headers 可能回显 set-cookie/auth，
+ # body 可能含敏感回显；统一经脱敏 + 截断（防止 API key 等泄露到 emrgd.log / 会话）。
+ hdr = _redact_headers(dict(resp.headers))
+ text_redacted = _redact_text(text)
+ logger.error("LLM error: %s headers=%s body=%s", resp.status_code, hdr, text_redacted[:2000])
  raise RuntimeError(
- f"LLM request failed: {resp.status_code} headers={hdr} body={text}"
+ f"LLM request failed: {resp.status_code} headers={hdr} body={text_redacted[:2000]}"
  )
 
  raise last_error # type: ignore[misc]
@@ -186,18 +215,18 @@ async def chat_stream(
  "LLM stream transient error %d, retrying in %.1fs "
  "(attempt %d/%d): %s",
  resp.status_code, delay, attempt + 1, MAX_RETRIES,
- text[:200],
+_redact_text(text[:200]),
  )
  await asyncio.sleep(delay)
  last_error = RuntimeError(
- f"LLM stream request failed: {resp.status_code} - {text}"
+ f"LLM stream request failed: {resp.status_code} - {_redact_text(text[:500])}"
  )
  continue
- logger.error("LLM stream error: %s %s", resp.status_code, text[:500])
- hdr = dict(resp.headers)
+ logger.error("LLM stream error: %s %s", resp.status_code, _redact_text(text[:500]))
+ hdr = _redact_headers(dict(resp.headers))
  raise RuntimeError(
  f"LLM stream request failed: {resp.status_code} "
- f"headers={hdr} body={text[:1000]}"
+ f"headers={hdr} body={_redact_text(text[:1000])}"
  )
 
  # Capture response metadata for llm.jsonl logging

tests/test_llm.py

@@ -140,3 +140,34 @@ def test_payload_temperature_default():
  c = LlmClient(default_cfg)
  p = c._make_payload([{"role": "user", "content": "x"}])
  assert p["temperature"] == 0.7
+
+
+# ── LLM 错误信息脱敏（20260807-0107）──────────────────────────
+
+
+def test_redact_headers_masks_sensitive():
+ """response headers 敏感键（set-cookie/authorization/token）被遮蔽。"""
+ from emrg.server.llm import _redact_headers
+ h = {"set-cookie": "session=abc; HttpOnly", "content-type": "application/json",
+ "x-request-id": "req-123", "x-api-key": "sk-A1b2C3d4A1b2C3d4A1b2C3d4A1b2C3d4"}
+ r = _redact_headers(h)
+ assert r["set-cookie"] == "***"
+ assert r["x-api-key"] == "***"
+ assert r["content-type"] == "application/json"
+ assert r["x-request-id"] == "req-123"
+
+
+def test_redact_headers_masks_inline_secret_in_values():
+ """非敏感键但值内联密钥也被遮蔽（如 server 回显 x-error: invalid sk-...）。"""
+ from emrg.server.llm import _redact_headers
+ r = _redact_headers({"x-error": "invalid key sk-A1b2C3d4A1b2C3d4A1b2C3d4A1b2C3d4"})
+ assert "sk-" not in r["x-error"]
+ assert "invalid key ***" in r["x-error"]
+
+
+def test_redact_text_masks_inline_credentials():
+ """LLM 错误 body 内联凭据被遮蔽，普通文本保留。"""
+ from emrg.server.llm import _redact_text
+ assert "sk-" not in _redact_text("bad key sk-A1b2C3d4A1b2C3d4A1b2C3d4A1b2C3d4 supplied")
+ assert "ghp_" not in _redact_text("token ghp_ABCDEFGHIJKLMNOPQRSTUVWXYZ1234567890 rejected")
+ assert _redact_text("rate limit exceeded, try later") == "rate limit exceeded, try later"