emrg: LLM client tolerates gzip response bodies without Content-Encoding

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` (502) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (508) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (88: 22 daemon_client + 22 app-commands + 19 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)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

README.cn.md

@@ -274,7 +274,7 @@ EMRG 不只是追赶——它自己追上来。
 git clone https://github.com/argszero/emrg.git
 cd emrg
 uv sync # 安装依赖
-uv run pytest tests/ -v # 跑测试（当前 480 项）
+uv run pytest tests/ -v # 跑测试（当前 508 项）
 uv run python -m emrg # 启动 TUI
 # CI 含 actionlint workflow 门禁（#444）：workflow 解析错误在 PR 即失败
 

README.md

@@ -273,7 +273,7 @@ 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 502 items)
+uv run pytest tests/ -v # run tests (currently 508 items)
 uv run python -m emrg # launch TUI
 # CI includes actionlint workflow gate (#444): workflow parse errors fail PR CI
 

emrg/server/llm.py

@@ -14,6 +14,7 @@
 from __future__ import annotations
 
 import asyncio
+import gzip
 import json
 import logging
 from typing import AsyncIterator, Optional
@@ -56,6 +57,19 @@ def _redact_headers(headers: dict) -> dict:
 RETRY_BASE_DELAY = 1.0 # seconds, doubled each retry
 
 
+def _parse_json_body(content: bytes) -> dict:
+ """Parse an LLM response body, transparently decompressing gzip.
+
+ Some gateways/proxies return gzip-compressed bodies without a proper
+ Content-Encoding header, so httpx does not decompress them and
+ ``resp.json()`` crashes with UnicodeDecodeError on the gzip magic
+ bytes (0x1f 0x8b). Detect the magic prefix and decompress first.
+ """
+ if content[:2] == b"\x1f\x8b":
+ content = gzip.decompress(content)
+ return json.loads(content)
+
+
 class LlmClient:
  """Async LLM client with tool calling and multi-turn streaming support."""
 
@@ -131,7 +145,28 @@ async def chat(
  if resp.status_code == 200:
  self.last_response_status = resp.status_code
  self.last_response_headers = dict(resp.headers)
- data = resp.json()
+ try:
+ data = _parse_json_body(resp.content)
+ except (json.JSONDecodeError, UnicodeDecodeError, OSError, EOFError) as exc:
+ # Malformed body (e.g. truncated gzip / proxy error page).
+ # Treat as transient — retry with backoff instead of crashing
+ # (20260807: memory reflection died on gzip body without
+ # Content-Encoding, resp.json() raised UnicodeDecodeError).
+ if attempt < MAX_RETRIES:
+ delay = RETRY_BASE_DELAY * (2 ** attempt)
+ logger.warning(
+ "LLM response body unparseable (%s), retrying in %.1fs "
+ "(attempt %d/%d)",
+ type(exc).__name__, delay, attempt + 1, MAX_RETRIES,
+ )
+ await asyncio.sleep(delay)
+ last_error = RuntimeError(
+ f"LLM response body unparseable: {type(exc).__name__}"
+ )
+ continue
+ raise RuntimeError(
+ f"LLM response body unparseable: {type(exc).__name__}"
+ ) from exc
  choice = data["choices"][0]
  return choice.get("message", {})
 

tests/test_llm.py

@@ -171,3 +171,100 @@ def test_redact_text_masks_inline_credentials():
  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"
+
+
+# ── gzip body 容错（20260807-1240：memory reflection UnicodeDecodeError）──
+
+
+def test_parse_json_body_plain():
+ """Plain JSON body parses unchanged."""
+ from emrg.server.llm import _parse_json_body
+ data = _parse_json_body(b'{"choices": []}')
+ assert data == {"choices": []}
+
+
+def test_parse_json_body_gzip_without_content_encoding():
+ """Gzip body (no Content-Encoding header → httpx won't decompress) is
+ transparently decompressed via magic-byte detection."""
+ import gzip as gz
+ from emrg.server.llm import _parse_json_body
+ raw = '{"choices": [{"message": {"content": "hi"}}]}'.encode()
+ data = _parse_json_body(gz.compress(raw))
+ assert data["choices"][0]["message"]["content"] == "hi"
+
+
+def test_parse_json_body_corrupt_gzip_raises():
+ """Gzip magic bytes with corrupt payload raise OSError (BadGzipFile)."""
+ import gzip as gz
+ import pytest
+ from emrg.server.llm import _parse_json_body
+ with pytest.raises(OSError):
+ _parse_json_body(b"\x1f\x8bCORRUPTED-NOT-REAL-GZIP")
+
+
+class _FakeResponse:
+ def __init__(self, status_code: int, content: bytes, headers=None):
+ self.status_code = status_code
+ self.content = content
+ self.headers = headers or {}
+
+
+class _FakeHttpClient:
+ def __init__(self, responses):
+ self.responses = list(responses)
+ self.calls = 0
+
+ async def post(self, url, headers=None, json=None):
+ self.calls += 1
+ return self.responses.pop(0)
+
+
+def _patch_fast_sleep(monkeypatch):
+ """Make retry backoff instant in tests."""
+ import emrg.server.llm as llm_mod
+
+ async def fast_sleep(_delay):
+ pass
+
+ monkeypatch.setattr(llm_mod.asyncio, "sleep", fast_sleep)
+
+
+def test_chat_gzip_body_transparent_decompress(monkeypatch, client):
+ """chat() returns the message when the 200 body is gzip-compressed
+ without Content-Encoding (the production failure mode)."""
+ import asyncio
+ import gzip as gz
+ body = gz.compress(b'{"choices": [{"message": {"content": "ok"}}]}')
+ fake = _FakeHttpClient([_FakeResponse(200, body)])
+ client._client = fake
+ msg = asyncio.run(client.chat([{"role": "user", "content": "hi"}]))
+ assert msg == {"content": "ok"}
+ assert fake.calls == 1 # no retry needed
+
+
+def test_chat_malformed_body_retries_then_succeeds(monkeypatch, client):
+ """Unparseable 200 body retries with backoff instead of crashing
+ (previously: UnicodeDecodeError killed memory reflection outright)."""
+ import asyncio
+ _patch_fast_sleep(monkeypatch)
+ good = b'{"choices": [{"message": {"content": "recovered"}}]}'
+ fake = _FakeHttpClient([
+ _FakeResponse(200, b"\x1f\x8bCORRUPT"),
+ _FakeResponse(200, good),
+ ])
+ client._client = fake
+ msg = asyncio.run(client.chat([{"role": "user", "content": "hi"}]))
+ assert msg == {"content": "recovered"}
+ assert fake.calls == 2
+
+
+def test_chat_malformed_body_exhausts_retries(monkeypatch, client):
+ """Persistently malformed body raises RuntimeError after MAX_RETRIES."""
+ import asyncio
+ import pytest
+ _patch_fast_sleep(monkeypatch)
+ fake = _FakeHttpClient([_FakeResponse(200, b"\x1f\x8bBAD")] * 4)
+ client._client = fake
+ with pytest.raises(RuntimeError, match="unparseable"):
+ asyncio.run(client.chat([{"role": "user", "content": "hi"}]))
+ assert fake.calls == 4 # 1 initial + 3 retries