v3.11.0: merge cobra PR, smart-continue, hot-reload, XML extraction

- Merge PR #5 from cobra91: concurrency semaphore, auto-continue, SO_REUSEADDR,
  proxy-stderr.log, stream diagnostics, timeout handler, restart proxy fix
- Tool call argument normalizer, smart-continue loop, XML extraction
- API key hot-reload with mtime tracking + /admin/ endpoints
- GUI hot-reload on endpoint edit with upstream verification
- Synthetic tool-results disabled (caused deepseek-v4-pro truncation)
- Version bump 3.10.12 -> 3.11.0, rebuild .deb
This commit is contained in:
Roman | RyzenAdvanced
2026-05-26 15:02:02 +04:00
Unverified
parent cae161023f
commit c3ba3286ff
7 changed files with 1100 additions and 191 deletions

View File

@@ -188,6 +188,7 @@ DEFAULT_MODELS = {
}
def load_config():
global _CONFIG_PATH, _CONFIG_MTIME
p = argparse.ArgumentParser(description="Responses API translation proxy")
p.add_argument("--config", help="JSON config file path")
p.add_argument("--port", type=int, default=None)
@@ -195,16 +196,21 @@ def load_config():
p.add_argument("--target-url", default=None)
p.add_argument("--api-key", default=None)
p.add_argument("--models-file", default=None, help="JSON file with model list array")
args = p.parse_args()
_args = p.parse_args()
cfg = {}
if args.config:
with open(args.config) as f:
if _args.config:
_CONFIG_PATH = os.path.abspath(_args.config)
with open(_args.config) as f:
cfg = json.load(f)
try:
_CONFIG_MTIME = os.path.getmtime(_CONFIG_PATH)
except OSError:
pass
for ck, ak in [("port", "port"), ("backend_type", "backend"),
("target_url", "target_url"), ("api_key", "api_key")]:
v = getattr(args, ak, None)
v = getattr(_args, ak, None)
if v is not None:
cfg[ck] = v
@@ -226,8 +232,8 @@ def load_config():
cfg.setdefault("api_key", "")
models = cfg.get("models", [])
if not models and args.models_file:
with open(args.models_file) as f:
if not models and _args.models_file:
with open(_args.models_file) as f:
models = json.load(f)
if not models:
models = DEFAULT_MODELS.get(cfg["backend_type"], [])
@@ -236,6 +242,8 @@ def load_config():
return cfg
CONFIG = None
_CONFIG_PATH = None
_CONFIG_MTIME = 0
PORT = 8080
BACKEND = "openai-compat"
TARGET_URL = ""
@@ -315,6 +323,8 @@ _conn_pool_lock = threading.Lock()
_conn_pool = {}
_STREAM_IDLE_TIMEOUT = 300
_MAX_CONCURRENT_REQUESTS = 3
_request_semaphore = threading.Semaphore(_MAX_CONCURRENT_REQUESTS)
_CODEBUFF_AUTH_URL = "https://www.codebuff.com"
_CODEBUFF_API_URL = "https://www.codebuff.com"
@@ -616,6 +626,51 @@ class APIKeyPool(AccountPool):
_cb_pool = CodebuffAccountPool("codebuff")
_google_antigravity_pool = GoogleAccountPool("antigravity")
_google_cli_pool = GoogleAccountPool("cli")
_antigravity_preferred_endpoint = None
_antigravity_endpoint_lock = threading.Lock()
def _classify_antigravity_error(status_code, body):
lower = body.lower()
if status_code == 400:
return "bad_request"
if status_code == 401:
if any(x in lower for x in ["invalid_grant", "token revoked", "token_revoked", "invalid_client"]):
return "auth_permanent"
return "auth_transient"
if status_code == 403:
if "validation_required" in lower or "account_disabled" in lower:
return "validation_required"
if "has been disabled" in lower and "violation of terms of service" in lower:
return "account_banned"
if "service_disabled" in lower:
return "service_disabled"
return "forbidden"
if status_code in (429, 503, 529):
if any(x in lower for x in ["model_capacity_exhausted", "capacity_exhausted", "model is currently overloaded", "service temporarily unavailable"]):
return "capacity_exhausted"
if any(x in lower for x in ["quota_exhausted", "resource_exhausted", "daily limit", "quota exceeded", "quotaresetdelay"]):
return "quota_exhausted"
return "rate_limited"
if status_code >= 500:
return "server_error"
return "unknown"
def _parse_rate_limit_reset(body):
import re as _re
m = _re.search(r'quotaResetDelay[:"\s]+(\d+(?:\.\d+)?)(ms|s)', body, _re.IGNORECASE)
if m:
val = float(m.group(1))
return val / 1000 if m.group(2) == 'ms' else val
m = _re.search(r'(\d+)h(\d+)m(\d+)s', body, _re.IGNORECASE)
if m:
return int(m.group(1)) * 3600 + int(m.group(2)) * 60 + int(m.group(3))
m = _re.search(r'Resets in ~(\d+)h(\d+)m', body, _re.IGNORECASE)
if m:
return int(m.group(1)) * 3600 + int(m.group(2)) * 60
m = _re.search(r'retry[-_]?after[:\s]+(\d+)\s*(?:sec|s\b)', body, _re.IGNORECASE)
if m:
return int(m.group(1))
return None
def _get_codebuff_account():
"""Return (token, account_dict) for best available codebuff account."""
@@ -714,14 +769,21 @@ def _gemini_reattach_sigs(contents):
# Gemini follow-through guardrail
_GEMINI_AGENT_GUARDRAIL = (
"You are running inside Codex as an autonomous coding agent. "
"When the user asks for a change to existing files, do not merely describe the previous work or summarize. "
"You must inspect the existing files, apply edits with tools, and verify the result. "
"If a file path is known from prior context, reuse it. "
"If unsure, list files first. "
"After tool results, continue until the requested change is actually implemented. "
"Never answer only with a plan such as 'I will start by...' or 'I am going to...'. "
"Always emit the actual tool call in the same response."
"!!! ABSOLUTELY CRITICAL - DO NOT IGNORE THIS UNDER ANY CIRCUMSTANCES !!! "
"YOU ARE RUNNING INSIDE CODEX AS AN AUTONOMOUS CODING AGENT. "
"!!!!!! NEVER EVER CONTINUE, PARAPHRASE, COMPLETE, OR ADD ANYTHING TO THE USER'S INSTRUCTIONS !!!!!! "
"!!!!!! NEVER SAY 'LET\\'S FIRST VIEW' OR 'LET\\'S FIRST FIND' OR SIMILAR PHRASES - EMIT THE ACTUAL TOOL CALL NOW !!!!!! "
"WHEN THE USER ASKS FOR A CHANGE TO EXISTING FILES, YOU MUST "
"1. IMMEDIATELY INSPECT EXISTING FILES USING exec_command OR read_files TOOLS RIGHT NOW, "
"2. THEN APPLY EDITS USING write OR exec_command TOOLS, "
"3. THEN VERIFY THE RESULT. "
"IF A FILE PATH IS KNOWN, REUSE IT IMMEDIATELY. "
"IF UNSURE, LIST FILES FIRST USING exec_command (ls -la). "
"AFTER TOOL RESULTS, CONTINUE UNTIL THE REQUESTED CHANGE IS FULLY IMPLEMENTED AND FILES ARE MODIFIED. "
"NEVER ANSWER ONLY WITH A PLAN LIKE 'I WILL START BY...' OR 'I AM GOING TO...'. "
"NEVER SUMMARIZE THE USER'S REQUEST. NEVER CONTINUE THEIR SENTENCE. "
"ALWAYS, ALWAYS, ALWAYS EMIT THE ACTUAL TOOL CALL IN THE SAME RESPONSE. "
"!!! FAILURE TO FOLLOW THESE INSTRUCTIONS WILL RESULT IN A BROKEN USER EXPERIENCE !!!"
)
_LOG_FILE_LOCK = threading.Lock()
@@ -771,6 +833,20 @@ def _ensure_antigravity_version():
_antigravity_version_checked = time.time()
return _antigravity_version
_antigravity_client_version = "1.110.0"
_antigravity_client_version_checked = 0
def _ensure_antigravity_client_version():
global _antigravity_client_version, _antigravity_client_version_checked
env_ver = os.environ.get("ANTIGRAVITY_CLIENT_VERSION", "").strip()
if env_ver:
return env_ver
if time.time() - _antigravity_client_version_checked < 6 * 3600:
return _antigravity_client_version
_antigravity_client_version = os.environ.get("ANTIGRAVITY_CLIENT_VERSION_FALLBACK", "1.110.0")
_antigravity_client_version_checked = time.time()
return _antigravity_client_version
def _init_runtime():
global CONFIG, PORT, BACKEND, TARGET_URL, API_KEY, OAUTH_PROVIDER, _antigravity_version
global MODELS, CC_VERSION, REASONING_ENABLED, REASONING_EFFORT, BGP_ROUTES
@@ -801,6 +877,68 @@ def _init_runtime():
_antigravity_version = _ensure_antigravity_version()
print(f"[antigravity] version={_antigravity_version}", file=sys.stderr)
def _verify_api_key(key, target_url):
if not key or not target_url:
return {"valid": False, "error": "missing key or url"}
test_url = upstream_target(target_url, "/models")
if not test_url:
return {"valid": False, "error": "invalid target url"}
try:
req = urllib.request.Request(test_url, headers={
"Authorization": f"Bearer {key}",
"Content-Type": "application/json",
})
resp = urllib.request.urlopen(req, timeout=10)
body = resp.read().decode()
model_count = 0
try:
data = json.loads(body)
model_count = len(data.get("data", []))
except Exception:
pass
return {"valid": True, "status": resp.status, "models": model_count}
except urllib.error.HTTPError as e:
err = e.read().decode()[:200]
return {"valid": False, "status": e.code, "error": err}
except Exception as e:
return {"valid": False, "error": str(e)[:200]}
_HOT_RELOAD_LOCK = threading.Lock()
def _hot_reload_api_key():
global API_KEY, _api_key_pool, _CONFIG_MTIME
if not _CONFIG_PATH:
return False
try:
cur_mtime = os.path.getmtime(_CONFIG_PATH)
except OSError:
return False
if cur_mtime <= _CONFIG_MTIME:
return False
with _HOT_RELOAD_LOCK:
try:
cur_mtime2 = os.path.getmtime(_CONFIG_PATH)
if cur_mtime2 <= _CONFIG_MTIME:
return False
with open(_CONFIG_PATH) as f:
new_cfg = json.load(f)
new_key = (new_cfg.get("api_key") or "").strip()
if not new_key or new_key == API_KEY:
_CONFIG_MTIME = cur_mtime2
return False
old_preview = API_KEY[:8] + "..." if len(API_KEY) > 8 else "(empty)"
new_preview = new_key[:8] + "..." if len(new_key) > 8 else "(empty)"
API_KEY = new_key
_CONFIG_MTIME = cur_mtime2
if API_KEY and "," in API_KEY and not OAUTH_PROVIDER.startswith("google") and BACKEND not in ("codebuff", "freebuff"):
_api_key_pool = APIKeyPool(BACKEND, API_KEY)
print(f"[hot-reload] API key pool refreshed: {len(_api_key_pool._accounts)} keys", file=sys.stderr)
print(f"[hot-reload] API key updated: {old_preview} -> {new_preview}", file=sys.stderr)
return True
except Exception as e:
print(f"[hot-reload] error: {e}", file=sys.stderr)
return False
bgp_models = []
for _r in BGP_ROUTES:
for _m in _r.get("models", [{"id": _r.get("model", "unknown")}]):
@@ -1623,7 +1761,8 @@ _PROVIDER_POLICIES = {
"openrouter": {"reasoning_mode": "provider_default", "max_tokens": 32768, "strip_reasoning": True,
"tool_output_limit": 6000, "max_input_items": 35, "compaction": "balanced"},
"openadapter": {"reasoning_mode": "off", "max_tokens": 32768, "strip_reasoning": True,
"tool_output_limit": 6000, "max_input_items": 30, "compaction": "balanced"},
"tool_output_limit": 1000, "max_input_items": 10, "compaction": "aggressive",
"synthetic_tool_results": True},
"cloudcode-pa": {"compaction": "aggressive", "context_size": 1000000,
"tool_output_limit": 6000, "max_input_items": 60},
"googleapis": {"compaction": "balanced", "context_size": 1000000,
@@ -2054,6 +2193,75 @@ def _inject_stored_reasoning(messages):
msg["reasoning_content"] = reasoning
return messages
def _normalize_tool_args(raw_args):
if not raw_args or raw_args == "{}":
return raw_args
try:
parsed = json.loads(raw_args)
if isinstance(parsed, dict):
if "Arguments" in parsed and "arguments" not in parsed:
inner = parsed["Arguments"]
if isinstance(inner, str):
inner = inner.strip()
for pfx in ("```json", "```"):
if inner.startswith(pfx):
inner = inner[len(pfx):].strip()
if inner.endswith("```"):
inner = inner[:-3].strip()
try:
inner_parsed = json.loads(inner)
if isinstance(inner_parsed, dict):
return json.dumps(inner_parsed)
except json.JSONDecodeError:
pass
if "cmd" not in parsed and "Arguments" in parsed:
inner = parsed["Arguments"]
if isinstance(inner, str):
inner = inner.strip()
for pfx in ("```json", "```"):
if inner.startswith(pfx):
inner = inner[len(pfx):].strip()
if inner.endswith("```"):
inner = inner[:-3].strip()
try:
inner_parsed = json.loads(inner)
if isinstance(inner_parsed, dict):
return json.dumps(inner_parsed)
except json.JSONDecodeError:
pass
return raw_args
except json.JSONDecodeError:
return raw_args
_XML_TC_RE = re.compile(r'<tool_call>(\w+)(.*?)</tool_call>', re.DOTALL)
_XML_ARG_VALUE_RE = re.compile(r'</?arg_value>\s*')
def _extract_xml_tool_calls(text):
if not text:
return []
results = []
for m in _XML_TC_RE.finditer(text):
name = m.group(1)
rest = _XML_ARG_VALUE_RE.sub("", m.group(2)).strip()
args_str = "{}"
try:
for pfx in ("```json", "```"):
if rest.startswith(pfx):
rest = rest[len(pfx):].strip()
if rest.endswith("```"):
rest = rest[:-3].strip()
if rest.startswith("{"):
json.loads(rest)
args_str = rest
else:
json.loads(rest)
args_str = rest
except Exception:
if rest.startswith("{"):
args_str = rest
results.append({"name": name, "args": args_str, "call_id": f"xml_{len(results)}"})
return results
def oa_input_to_messages(input_data):
msgs = []
tool_name_by_id = {}
@@ -2066,11 +2274,13 @@ def oa_input_to_messages(input_data):
t = item.get("type")
if t == "function_call":
tcid = item.get("call_id") or item.get("id") or uid("tc")
raw_args = item.get("arguments", "{}")
normalized_args = _normalize_tool_args(raw_args)
pending_tool_calls.append(
{"id": tcid,
"type": "function",
"function": {"name": item.get("name", ""),
"arguments": item.get("arguments", "{}")}})
"arguments": normalized_args}})
tool_name_by_id[tcid] = item.get("name", "")
continue
if pending_tool_calls:
@@ -4280,9 +4490,10 @@ def _antigravity_is_simple_user(text):
return True
return False
def _antigravity_normalize_context(input_data):
def _antigravity_normalize_context(input_data, model=""):
if not isinstance(input_data, list) or len(input_data) < 2:
return input_data
is_claude_model = "claude" in model.lower()
latest_user = ""
latest_user_idx = -1
@@ -4310,13 +4521,19 @@ def _antigravity_normalize_context(input_data):
if os.environ.get("ANTIGRAVITY_AUTO_RESET_POLLUTED_CONTEXT", "1") != "1":
auto_reset = False
if is_simple and (auto_reset or n_tool_outputs == 0):
has_compaction_summary = any(
isinstance(it, dict) and it.get("type") == "message" and it.get("role") == "user"
and ("Auto-compacted" in str(it.get("content", "")) or "auto-compacted" in str(it.get("content", "")).lower())
for it in input_data
)
if is_simple and auto_reset and not has_compaction_summary:
system_items = [it for it in input_data if isinstance(it, dict) and it.get("type") == "message" and it.get("role") in ("developer", "system")]
user_item = input_data[latest_user_idx]
result = system_items + [user_item] if system_items else [user_item]
print(f"[antigravity-context] raw_items={n_raw} compacted_items={n_raw} final_items={len(result)}", file=sys.stderr)
print(f"[antigravity-context] raw_tool_outputs={n_tool_outputs} kept_tool_outputs=0", file=sys.stderr)
print(f"[antigravity-context] simple_latest_user=true auto_reset={auto_reset}", file=sys.stderr)
print(f"[antigravity-context] simple_latest_user=true auto_reset={auto_reset} has_compaction={has_compaction_summary}", file=sys.stderr)
return result
dev_messages = []
@@ -4340,9 +4557,15 @@ def _antigravity_normalize_context(input_data):
latest_words = set(latest_user.strip().lower().split())
has_edit_intent = bool(latest_words.intersection(_ANTIGRAVITY_EDIT_WORDS))
has_ref_intent = bool(latest_words.intersection(_ANTIGRAVITY_REFERENCE_WORDS))
keep_tools = 2 if (has_edit_intent or has_ref_intent) else 1
if is_claude_model:
keep_tools = len(tool_outputs)
else:
keep_tools = 2 if (has_edit_intent or has_ref_intent) else 1
kept_tools = tool_outputs[-keep_tools:] if tool_outputs and (has_edit_intent or has_ref_intent) else []
if is_claude_model:
kept_tools = tool_outputs
else:
kept_tools = tool_outputs[-keep_tools:] if tool_outputs and (has_edit_intent or has_ref_intent) else []
for idx_t, t_item in enumerate(kept_tools):
orig = t_item[1]
@@ -4357,6 +4580,22 @@ def _antigravity_normalize_context(input_data):
tail_start = max(0, len(recent_items) - 6)
recent_tail = recent_items[tail_start:]
deduped_tail = []
seen_goal_context = False
for idx, msg_item in recent_tail:
content_str = ""
c = msg_item.get("content", "")
if isinstance(c, str):
content_str = c
elif isinstance(c, list):
content_str = " ".join(p.get("text", p.get("input_text", "")) for p in c if isinstance(p, dict))
if "<goal_context>" in content_str:
if seen_goal_context:
continue
seen_goal_context = True
deduped_tail.append((idx, msg_item))
recent_tail = deduped_tail if deduped_tail else recent_tail
tool_call_ids = set()
for _, t_item in kept_tools:
cid = t_item.get("call_id", t_item.get("id", ""))
@@ -4371,6 +4610,15 @@ def _antigravity_normalize_context(input_data):
result = list(dev_messages)
compaction_summaries = []
for idx, msg_item in recent_items:
if msg_item is input_data[latest_user_idx]:
continue
c = msg_item.get("content", "")
content_str = c if isinstance(c, str) else " ".join(p.get("text", p.get("input_text", "")) for p in c if isinstance(p, dict)) if isinstance(c, list) else ""
if "Auto-compacted" in content_str or "auto-compacted" in content_str.lower():
compaction_summaries.append(msg_item)
if n_summarized > 0:
summary_text = f"[Tool history summary: {n_summarized} older tool outputs omitted. {n_tool_calls} prior function calls were made for file inspection/editing.]"
result.append({"type": "message", "role": "user", "content": [{"type": "input_text", "text": summary_text}]})
@@ -4381,6 +4629,9 @@ def _antigravity_normalize_context(input_data):
for _, tool_item in kept_tools:
result.append(tool_item)
for cs_item in compaction_summaries:
result.append(cs_item)
for _, msg_item in recent_tail:
if msg_item is not input_data[latest_user_idx]:
result.append(msg_item)
@@ -4408,7 +4659,10 @@ def _antigravity_normalize_context(input_data):
if total_chars > _ANTIGRAVITY_EMERGENCY_CHARS:
print(f"[antigravity-context] EMERGENCY: {total_chars} chars exceeds limit, resetting to minimal", file=sys.stderr)
result = list(dev_messages) + [input_data[latest_user_idx]]
result = list(dev_messages)
if compaction_summaries:
result.extend(compaction_summaries)
result.append(input_data[latest_user_idx])
total_chars = sum(len(json.dumps(it, ensure_ascii=False)) for it in result)
while len(result) > _ANTIGRAVITY_MAX_CONTENTS and total_chars > _ANTIGRAVITY_SOFT_CHARS:
@@ -4475,12 +4729,23 @@ class Handler(http.server.BaseHTTPRequestHandler):
pass
_uptime = time.time() - _START_TIME if '_START_TIME' in dir() else 0
self.send_json(200, {"ok": True, "backend": BACKEND,
"target_url": TARGET_URL,
"models": [m.get("id") for m in MODELS],
"bgp_routes": len(BGP_ROUTES),
"uptime_s": round(_uptime, 1),
"memory_mb": round(_mem_mb, 1),
"requests_total": _STATS.get("requests", 0)})
"target_url": TARGET_URL,
"models": [m.get("id") for m in MODELS],
"bgp_routes": len(BGP_ROUTES),
"uptime_s": round(_uptime, 1),
"memory_mb": round(_mem_mb, 1),
"requests_total": _STATS.get("requests", 0)})
elif self.path == "/admin/reload":
reloaded = _hot_reload_api_key()
key_preview = API_KEY[:8] + "..." if len(API_KEY) > 8 else "(empty)"
self.send_json(200, {"ok": True, "reloaded": reloaded,
"api_key_preview": key_preview,
"config_path": _CONFIG_PATH or "none"})
elif self.path == "/admin/verify-key":
result = _verify_api_key(API_KEY, TARGET_URL)
key_preview = API_KEY[:8] + "..." if len(API_KEY) > 8 else "(empty)"
result["api_key_preview"] = key_preview
self.send_json(200, result)
else:
self.send_error(404)
@@ -4502,6 +4767,7 @@ class Handler(http.server.BaseHTTPRequestHandler):
_logf = None
def _handle(self):
_hot_reload_api_key()
try:
clen = int(self.headers.get("Content-Length", 0))
body = json.loads(self.rfile.read(clen))
@@ -4565,6 +4831,11 @@ class Handler(http.server.BaseHTTPRequestHandler):
_last_user_urls.append(url_m.group(0))
save_request_snapshot(request_id, body)
_req_t0 = time.time()
wait_start = time.monotonic()
_request_semaphore.acquire()
wait_ms = (time.monotonic() - wait_start) * 1000
if wait_ms > 100:
print(f"[{_sid}] waited {wait_ms:.0f}ms for upstream slot (concurrency gate)", file=sys.stderr)
try:
with RequestTracker(request_id) as tracker:
if BACKEND == "auto":
@@ -4583,6 +4854,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
except Exception as _snap_err:
update_snapshot_response(request_id, "error", time.time() - _req_t0, _snap_err)
raise
finally:
_request_semaphore.release()
def _handle_openai_compat(self, body, model, stream, tracker=None):
input_data = body.get("input", "")
@@ -4595,7 +4868,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
body = dict(body)
body["input"] = input_data
if (policy.get("synthetic_tool_results") or _provider_cap(model, "synthetic_tool_results", False)) and isinstance(input_data, list):
# synthetic tool-results disabled: causes deepseek-v4-pro truncation on opencode.ai
if False and (policy.get("synthetic_tool_results") or _provider_cap(model, "synthetic_tool_results", False)) and isinstance(input_data, list):
input_data, synthesized = synthesize_tool_results_for_chat(input_data)
if synthesized:
print("[provider-adapter] using synthetic tool-result continuation", file=sys.stderr)
@@ -4603,7 +4877,7 @@ class Handler(http.server.BaseHTTPRequestHandler):
body["input"] = input_data
compacted = False
if policy.get("compaction") and isinstance(input_data, list):
if policy.get("compaction") and isinstance(input_data, list) and "claude" not in model.lower():
input_data, compacted = _adaptive_compact(input_data, model, policy)
if compacted:
body = dict(body)
@@ -4652,6 +4926,22 @@ class Handler(http.server.BaseHTTPRequestHandler):
upstream = urllib.request.urlopen(req, timeout=_upstream_timeout(body, stream))
except urllib.error.HTTPError as e:
err_body = e.read().decode()
if "context_length_exceeded" in err_body and attempt < max_retries:
print(f"[{self._session_id}] context_length_exceeded (attempt {attempt+1}/{max_retries}), retrying with extreme compaction!", file=sys.stderr)
policy = provider_policy()
if isinstance(input_data, list):
print(f"[{self._session_id}] applying extreme compaction to {len(input_data)} items", file=sys.stderr)
input_data = _crof_compact_for_retry(input_data, model)
body = dict(body)
body["input"] = input_data
messages = oa_input_to_messages(input_data)
messages = _inject_stored_reasoning(messages)
instructions = body.get("instructions", "").strip()
if instructions:
messages.insert(0, {"role": "system", "content": instructions})
chat_body = self._build_chat_body(model, messages, body, stream)
chat_body_b = json.dumps(chat_body).encode()
continue
if e.code in (429, 502, 503) and attempt < max_retries:
if e.code == 429 and _api_key_pool:
pool_acct = _api_key_pool.get()
@@ -4782,7 +5072,7 @@ class Handler(http.server.BaseHTTPRequestHandler):
body["input"] = input_data
compacted = False
if policy.get("compaction") and isinstance(input_data, list):
if policy.get("compaction") and isinstance(input_data, list) and "claude" not in model.lower():
input_data, compacted = _adaptive_compact(input_data, model, policy)
if compacted:
body = dict(body)
@@ -4793,8 +5083,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
body = dict(body)
body["input"] = input_data
if OAUTH_PROVIDER == "google-antigravity" and isinstance(input_data, list):
input_data = _antigravity_normalize_context(input_data)
if OAUTH_PROVIDER == "google-antigravity" and isinstance(input_data, list) and "claude" not in model.lower():
input_data = _antigravity_normalize_context(input_data, model)
body = dict(body)
body["input"] = input_data
@@ -4872,7 +5162,7 @@ class Handler(http.server.BaseHTTPRequestHandler):
resp_part["functionResponse"]["id"] = call_id
contents.append({"role": "user", "parts": [resp_part]})
if OAUTH_PROVIDER.startswith("google"):
if OAUTH_PROVIDER.startswith("google") and "claude" not in model.lower():
sanitized = []
last_user_text = None
last_role = None
@@ -4963,8 +5253,18 @@ class Handler(http.server.BaseHTTPRequestHandler):
contents = _gemini_reattach_sigs(contents)
if OAUTH_PROVIDER == "google-antigravity":
guardrail_found = any("autonomous coding agent" in json.dumps(c.get("parts", []), ensure_ascii=False) for c in contents[:2])
if not guardrail_found:
latest_user = ""
if isinstance(input_data, list):
for item in reversed(input_data):
if item.get("type") == "message" and item.get("role") == "user":
c = item.get("content", "")
if isinstance(c, str):
latest_user = c
elif isinstance(c, list):
latest_user = "\n".join(p.get("text", p.get("input_text", "")) for p in c if isinstance(p, dict))
break
is_latest_simple = _antigravity_is_simple_user(latest_user)
if not is_latest_simple:
contents.insert(0, {"role": "user", "parts": [{"text": _GEMINI_AGENT_GUARDRAIL}]})
if OAUTH_PROVIDER == "google-antigravity" and isinstance(input_data, list):
@@ -4976,11 +5276,10 @@ class Handler(http.server.BaseHTTPRequestHandler):
if isinstance(c, str): latest_lower = c.lower()
elif isinstance(c, list): latest_lower = " ".join(p.get("text", p.get("input_text", "")) for p in c if isinstance(p, dict)).lower()
break
if latest_lower and any(w in latest_lower for w in _EDIT_WORDS) and len(input_data) > 6:
if latest_lower and any(w in latest_lower for w in _EDIT_WORDS):
n_tool_calls = sum(1 for it in input_data if isinstance(it, dict) and it.get("type") == "function_call")
if n_tool_calls > 0:
contents.append({"role": "user", "parts": [{"text": "IMPORTANT: The user is requesting a modification to existing files. You MUST use tools (exec_command, write, etc.) to make the changes. Do NOT just describe what to do — actually call the tools to modify the files now."}]})
print(f"[antigravity] edit-intent detected with {n_tool_calls} prior tool calls; injected tool-use nudge", file=sys.stderr)
contents.append({"role": "user", "parts": [{"text": "!!! ABSOLUTELY NO PLANNING - EMIT THE TOOL CALL NOW !!! IMPORTANT: The user is requesting a modification to existing files. You MUST use tools (exec_command, read_files, write, etc.) to make the changes RIGHT NOW. Do NOT just describe what to do — actually CALL THE TOOLS IN THIS RESPONSE. IMMEDIATELY INSPECT THE FILE OR LIST FILES USING exec_command TOOL CALL."}]})
print(f"[antigravity] edit-intent detected; injected tool-use nudge", file=sys.stderr)
if OAUTH_PROVIDER == "google-antigravity" and isinstance(input_data, list):
latest_user = ""
@@ -5031,10 +5330,14 @@ class Handler(http.server.BaseHTTPRequestHandler):
wrapped["requestType"] = "agent"
wrapped["userAgent"] = "antigravity"
wrapped["requestId"] = f"agent-{uuid.uuid4().hex[:12]}"
wrapped["request"]["sessionId"] = f"{uuid.uuid4().hex}{int(time.time()*1000)}"
_allow_staging = os.environ.get("ALLOW_ANTIGRAVITY_STAGING", "0") == "1"
if OAUTH_PROVIDER == "google-antigravity":
_antigravity_endpoints = ["https://cloudcode-pa.googleapis.com"]
_antigravity_endpoints = [
"https://cloudcode-pa.googleapis.com",
"https://daily-cloudcode-pa.googleapis.com",
]
if _allow_staging:
_antigravity_endpoints.extend([
"https://daily-cloudcode-pa.sandbox.googleapis.com",
@@ -5052,7 +5355,13 @@ class Handler(http.server.BaseHTTPRequestHandler):
}
if OAUTH_PROVIDER == "google-antigravity":
version = _ensure_antigravity_version()
headers["User-Agent"] = f"antigravity/{version} darwin/arm64"
import platform as _plat
_os_name = _plat.system().lower()
_os_arch = _plat.machine().lower().replace("x86_64", "x64").replace("aarch64", "arm64")
headers["User-Agent"] = f"antigravity/{version} {_os_name}/{_os_arch}"
headers["X-Client-Name"] = "antigravity"
headers["X-Client-Version"] = _ensure_antigravity_client_version()
headers["x-goog-api-client"] = "gl-node/18.18.2 fire/0.8.6 grpc/1.10.x"
else:
headers["User-Agent"] = "google-api-nodejs-client/9.15.1"
headers["X-Goog-Api-Client"] = "gl-node/22.17.0"
@@ -5072,14 +5381,33 @@ class Handler(http.server.BaseHTTPRequestHandler):
if OAUTH_PROVIDER == "google-antigravity":
print(f"[antigravity-endpoint] endpoints={[e.replace('https://','') for e in endpoints]} project={project_id}", file=sys.stderr)
for ep in endpoints:
upstream = None
chosen_ep = None
global _antigravity_preferred_endpoint
with _antigravity_endpoint_lock:
_pref = _antigravity_preferred_endpoint
if _pref and _pref in endpoints:
ordered = [_pref] + [e for e in endpoints if e != _pref]
else:
ordered = list(endpoints)
for ep in ordered:
target = f"{ep}/{url_suffix}"
req = urllib.request.Request(target, data=body_b, headers=headers)
try:
upstream = urllib.request.urlopen(req, timeout=_upstream_timeout(body, stream))
chosen_ep = ep
with _antigravity_endpoint_lock:
_antigravity_preferred_endpoint = ep
if ep != _pref:
print(f"[{self._session_id}] fallback OK: {ep.replace('https://','')}", file=sys.stderr)
break
except urllib.error.HTTPError as e:
err_body = e.read().decode()
err_class = _classify_antigravity_error(e.code, err_body)
print(f"[{self._session_id}] {ep.replace('https://','')} {e.code} class={err_class}", file=sys.stderr)
if e.code == 400 and OAUTH_PROVIDER.startswith("google"):
try:
debug_path = os.path.join(_LOG_DIR, "gemini-last-400-request.json")
@@ -5088,23 +5416,38 @@ class Handler(http.server.BaseHTTPRequestHandler):
print(f"[{self._session_id}] saved 400 debug request to {debug_path}", file=sys.stderr)
except Exception:
pass
if e.code == 403 and "SERVICE_DISABLED" in err_body[:500] and ep != endpoints[-1]:
print(f"[{self._session_id}] {ep} SERVICE_DISABLED, trying next endpoint", file=sys.stderr)
return self.send_json(e.code, {"error": {"type": "upstream_error", "message": _sanitize_err_body(err_body)}})
if err_class == "auth_permanent":
return self.send_json(e.code, {"error": {"type": "upstream_error", "message": _sanitize_err_body(err_body)}})
if err_class in ("quota_exhausted", "rate_limited"):
reset_s = _parse_rate_limit_reset(err_body)
if ep == ordered[-1]:
pool = _google_antigravity_pool if OAUTH_PROVIDER == "google-antigravity" else _google_cli_pool
_, acct = _get_google_account(OAUTH_PROVIDER)
if acct:
cooldown = reset_s if reset_s and reset_s > 10 else 60
pool.mark_rate_limited(acct, cooldown)
print(f"[{self._session_id}] quota reset in ~{reset_s}s, cooldown={cooldown}s", file=sys.stderr)
return self.send_json(e.code, {"error": {"type": "upstream_error", "message": _sanitize_err_body(err_body)}})
print(f"[{self._session_id}] {ep.replace('https://','')} 429, trying next", file=sys.stderr)
with _antigravity_endpoint_lock:
_antigravity_preferred_endpoint = None
continue
if e.code == 429 and ep != endpoints[-1] and _allow_staging:
print(f"[{self._session_id}] {ep} HTTP 429, trying next endpoint", file=sys.stderr)
if err_class in ("service_disabled", "forbidden", "account_banned", "validation_required"):
if ep == ordered[-1]:
return self.send_json(e.code, {"error": {"type": "upstream_error", "message": _sanitize_err_body(err_body)}})
continue
if e.code == 429:
pool = _google_antigravity_pool if OAUTH_PROVIDER == "google-antigravity" else _google_cli_pool
_, acct = _get_google_account(OAUTH_PROVIDER)
if acct:
pool.mark_rate_limited(acct, 60)
return self.send_json(e.code, {"error": {"type": "upstream_error", "message": _sanitize_err_body(err_body)}})
except Exception as e:
if ep == endpoints[-1]:
return self.send_json(502, {"error": {"type": "proxy_error", "message": str(e)}})
print(f"[{self._session_id}] {ep} failed: {e}, trying next", file=sys.stderr)
if ep == ordered[-1]:
return self.send_json(e.code, {"error": {"type": "upstream_error", "message": _sanitize_err_body(err_body)}})
continue
except Exception as e:
print(f"[{self._session_id}] {ep.replace('https://','')} conn failed: {e}", file=sys.stderr)
if ep == ordered[-1]:
return self.send_json(502, {"error": {"type": "proxy_error", "message": str(e)}})
continue
if upstream is None:
return self.send_json(502, {"error": {"type": "proxy_error", "message": "All endpoints failed"}})
if stream:
self._forward_gemini_sse(upstream, model, body, input_data, tracker)
@@ -5406,11 +5749,25 @@ class Handler(http.server.BaseHTTPRequestHandler):
break
collected_events.append(event)
_observe_event(event)
print(f"[{self._session_id}] stream ended: events={len(collected_events)} finish={finish_reason} has_content={has_content} elapsed={time.time()-t0:.1f}s", file=sys.stderr)
except (ConnectionResetError, BrokenPipeError, ConnectionAbortedError):
print("[translate-proxy] client disconnected during stream", file=sys.stderr)
_crof_record(model, n_items, False)
_log_resp(last_resp_id, "client_disconnect", last_output)
return
except (TimeoutError, OSError, urllib.error.URLError) as e:
print(f"[translate-proxy] upstream error during stream: {type(e).__name__}: {e}", file=sys.stderr)
err_resp_id = body.get("request_id") or body.get("id") or uid("resp")
try:
self.wfile.write(emit("response.failed", {"type": "response.failed",
"response": {"id": err_resp_id, "error": {"type": "upstream_error",
"code": "stream_interrupted", "message": str(e)[:200]}}}).encode())
self.wfile.flush()
except Exception:
pass
_crof_record(model, n_items, False)
_log_resp(last_resp_id, "upstream_error", last_output)
return
# Record outcome
success = (finish_reason != "length")
@@ -5486,6 +5843,171 @@ class Handler(http.server.BaseHTTPRequestHandler):
except Exception as e:
print(f"[crof-adaptive] retry failed: {e}", file=sys.stderr)
# ── Auto-continue for truncated responses ── (cobra PR)
_ac_did_run = False
if stream and collected_events:
_ac_text = ""
_ac_msg_id = _ac_resp_id = None
for _ev in collected_events:
for _ln in _ev.strip().split("\n"):
if not _ln.startswith("data: "):
continue
try:
_d = json.loads(_ln[6:])
_t = _d.get("type")
if _t == "response.output_text.done":
_ac_text = _d.get("text", "")
elif _t == "response.output_item.added" and _d.get("item",{}).get("type") == "message":
_ac_msg_id = _d.get("item",{}).get("id")
elif _t == "response.completed":
_ac_resp_id = _d.get("response",{}).get("id")
except Exception:
pass
_ac_tc = reasoning_out.get("tool_calls", [])
_ac_truncated = False
if not _ac_tc and _ac_text:
_ac_stripped = _ac_text.rstrip()
if finish_reason == "length":
_ac_truncated = True
elif len(_ac_stripped) > 10 and _ac_stripped[-1] in "(:,;…":
_ac_truncated = True
if _ac_truncated and _ac_text:
print(f"[{self._session_id}] auto-continue: truncated (finish={finish_reason}, ends '{_ac_text.rstrip()[-10:]}')", file=sys.stderr)
_ac_did_run = True
_ac_cut = len(collected_events)
for _i, _ev2 in enumerate(collected_events):
if "response.output_text.done" in _ev2:
_ac_cut = _i
break
collected_events = collected_events[:_ac_cut]
_ac_accumulated = _ac_text
_ac_max = 3
for _ac_attempt in range(_ac_max):
try:
_ac_cont_msgs = list(chat_body.get("messages", []))
_ac_cont_msgs.append({"role": "assistant", "content": _ac_accumulated})
_ac_cont_msgs.append({"role": "user", "content": "Continue exactly where you left off. Do not repeat anything already written."})
_ac_cont_body = dict(chat_body)
_ac_cont_body["messages"] = _ac_cont_msgs
_ac_cont_body["stream"] = False
_ac_cont_req = urllib.request.Request(target, data=json.dumps(_ac_cont_body).encode(), headers=fwd)
_ac_cont_resp = json.loads(urllib.request.urlopen(_ac_cont_req, timeout=120).read())
_ac_choices = _ac_cont_resp.get("choices", [])
if _ac_choices:
_ac_chunk = _ac_choices[0].get("message",{}).get("content","")
if not _ac_chunk:
_ac_chunk = _ac_choices[0].get("delta",{}).get("content","")
_ac_finish = _ac_choices[0].get("finish_reason")
if _ac_chunk:
_ac_accumulated += _ac_chunk
collected_events.append(emit("response.output_text.delta", {
"type": "response.output_text.delta",
"delta": _ac_chunk, "item_id": _ac_msg_id, "content_index": 0}))
if _ac_finish != "length":
break
_ac_text = _ac_accumulated
except Exception as _ac_e:
print(f"[{self._session_id}] auto-continue attempt {_ac_attempt+1} failed: {_ac_e}", file=sys.stderr)
break
if _ac_msg_id:
collected_events.append(emit("response.output_text.done", {
"type": "response.output_text.done",
"text": _ac_accumulated, "item_id": _ac_msg_id, "content_index": 0}))
collected_events.append(emit("response.content_part.done", {
"type": "response.content_part.done",
"part": {"type": "output_text", "text": _ac_accumulated, "annotations": []}, "item_id": _ac_msg_id}))
collected_events.append(emit("response.output_item.done", {
"type": "response.output_item.done",
"item": {"type": "message", "id": _ac_msg_id, "role": "assistant", "status": "completed",
"content": [{"type": "output_text", "text": _ac_accumulated, "annotations": []}]}}))
if _ac_resp_id:
collected_events.append(emit("response.completed", {
"type": "response.completed",
"response": {"id": _ac_resp_id, "object": "response", "model": model,
"status": "completed", "created": int(time.time()),
"output": [{"type": "message", "id": _ac_msg_id, "role": "assistant",
"status": "completed",
"content": [{"type": "output_text", "text": _ac_accumulated, "annotations": []}]}]}}))
has_content = True
finish_reason = "stop"
print(f"[{self._session_id}] auto-continue done: {len(_ac_text)} -> {len(_ac_accumulated)} chars", file=sys.stderr)
# Smart continuation: loop with escalating nudges when model stops text-only mid-task.
# Skip if auto-continue already handled the response.
if not _ac_did_run:
_smart_max = 2
_smart_attempt = 0
while _smart_attempt < _smart_max:
_has_tool_calls_in_output = any(o.get("type") == "function_call" for o in (last_output or []))
if not (finish_reason == "stop" and has_content and not _has_tool_calls_in_output
and isinstance(input_data, list) and len(input_data) >= 3
and has_function_call_output(input_data)):
break
_smart_attempt += 1
_nudges = [
"Continue with the task using tool calls. Do NOT describe what to do — call the appropriate functions.",
"You MUST use tool calls to complete the task. Read files, run commands, and make changes using tools. Do NOT output XML tool calls as text.",
]
nudge_text = _nudges[min(_smart_attempt - 1, len(_nudges) - 1)]
# Try extracting XML tool calls from text as fallback before nudging
last_text = ""
for o in (last_output or []):
if o.get("type") == "message":
for c in (o.get("content") or []):
if isinstance(c, dict) and c.get("type") == "output_text":
last_text += c.get("text", "")
xml_fc = _extract_xml_tool_calls(last_text)
if xml_fc:
print(f"[{self._session_id}] [smart-continue] extracted {len(xml_fc)} XML tool calls from text, injecting and retrying", file=sys.stderr)
fake_input = list(input_data)
for xfc in xml_fc:
fake_input.append({"type": "function_call", "id": uid("fcx"), "call_id": uid("fcx"),
"name": xfc["name"], "arguments": xfc["args"], "status": "completed"})
fake_messages = oa_input_to_messages(fake_input)
instructions = body.get("instructions", "").strip()
if instructions:
fake_messages.insert(0, {"role": "system", "content": instructions})
fake_chat_body = self._build_chat_body(model, fake_messages, body, stream)
fake_req = urllib.request.Request(target, data=json.dumps(fake_chat_body).encode(), headers=fwd)
try:
retry_upstream = urllib.request.urlopen(fake_req, timeout=_upstream_timeout(body, True))
collected_events = []
last_resp_id = last_output = last_status = None
finish_reason = None
has_content = False
for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
collected_events.append(event)
_observe_event(event)
input_data = fake_input
continue
except Exception as e:
print(f"[{self._session_id}] [smart-continue] XML injection retry failed: {e}", file=sys.stderr)
break
_nudge_msg = {"role": "user", "content": nudge_text}
nudge_messages = oa_input_to_messages(input_data) + [_nudge_msg]
instructions = body.get("instructions", "").strip()
if instructions:
nudge_messages.insert(0, {"role": "system", "content": instructions})
nudge_chat_body = self._build_chat_body(model, nudge_messages, body, stream)
nudge_req = urllib.request.Request(target, data=json.dumps(nudge_chat_body).encode(), headers=fwd)
print(f"[{self._session_id}] [smart-continue] attempt {_smart_attempt}/{_smart_max}: model stopped mid-task, nudging", file=sys.stderr)
try:
retry_upstream = urllib.request.urlopen(nudge_req, timeout=_upstream_timeout(body, True))
collected_events = []
last_resp_id = last_output = last_status = None
finish_reason = None
has_content = False
for event in oa_stream_to_sse(retry_upstream, model, body.get("request_id") or body.get("id")):
collected_events.append(event)
_observe_event(event)
except Exception as e:
print(f"[{self._session_id}] [smart-continue] nudge attempt {_smart_attempt} failed: {e}", file=sys.stderr)
break
self.stream_buffered_events(collected_events)
else:
result = oa_resp_to_responses(json.loads(upstream.read()), model)
@@ -6392,12 +6914,45 @@ def _handle_shutdown_signal(sig, frame):
if 'SERVER' in globals() and SERVER:
SERVER.shutdown()
def _anti_stall_cleanup():
my_pid = os.getpid()
my_port = PORT
killed = []
try:
import subprocess as _sp
out = _sp.run(["pgrep", "-f", "translate-proxy"], capture_output=True, text=True, timeout=5).stdout.strip()
for pid_str in out.splitlines():
pid_str = pid_str.strip()
if not pid_str or not pid_str.isdigit():
continue
pid = int(pid_str)
if pid == my_pid:
continue
try:
os.kill(pid, signal.SIGTERM)
killed.append(pid)
except (ProcessLookupError, PermissionError):
pass
except Exception:
pass
try:
_cache_dir = os.path.join(os.path.dirname(os.path.abspath(__file__)), "__pycache__")
if os.path.isdir(_cache_dir):
import shutil
shutil.rmtree(_cache_dir, ignore_errors=True)
except Exception:
pass
if killed:
print(f"[anti-stall] killed {len(killed)} stale proxy process(es): {killed}", flush=True)
time.sleep(1)
def main():
global SERVER, _START_TIME
_START_TIME = time.time()
_anti_stall_cleanup()
_init_runtime()
try:
_current_cfg = os.path.basename(args.config) if args.config else ""
_current_cfg = os.path.basename(_CONFIG_PATH) if _CONFIG_PATH else ""
for _f in os.listdir(_LOG_DIR):
if _f.startswith("proxy-") and _f.endswith(".json") and _f != _current_cfg:
os.remove(os.path.join(_LOG_DIR, _f))