fix(gateway): one context-overflow verdict for the reply and the transcript skip
Moving the `if response` guard below the failed branch exposed the normalizer's loose overflow predicate (bare "token"/"exceed"/"context"/"payload", or any 400 on a long session) to failed turns that carry real text: billing, rate-limit, auth and content-policy replies were rewritten to "Session too large / /compact". Hoist run_turn's stricter classifier (compression_exhausted, multi-word phrases, 400 on history > 50) into a module-level `is_context_overflow_failure_result` and use it for both the #1630 transcript skip and the user-facing rewrite, so the two can never disagree. Populated text is only rewritten when it is the bare provider envelope (`_looks_like_gateway_provider_error`) on an overflow turn; curated agent text (compression-timeout guidance, /compress hint) survives. Replaces the sanitizer-wording test with the passthrough invariants that catch the regression.
This commit is contained in:
+8
-11
@@ -2106,7 +2106,7 @@ from gateway.slash_commands import GatewaySlashCommandsMixin
|
||||
from gateway.run_voice import GatewayVoiceMixin
|
||||
from gateway.run_adapters import GatewayAdapterLifecycleMixin
|
||||
from gateway.run_topics import GatewayTopicThreadsMixin
|
||||
from gateway.run_turn import GatewayTurnMixin
|
||||
from gateway.run_turn import GatewayTurnMixin, is_context_overflow_failure_result
|
||||
from gateway.run_shutdown import GatewayShutdownMixin, _exit_with_failure_verdict, _resolve_gateway_exit_verdict
|
||||
from gateway.run_busy import GatewayBusySessionMixin
|
||||
from gateway.run_config_loaders import GatewayConfigLoadersMixin
|
||||
@@ -3027,10 +3027,13 @@ def _normalize_empty_agent_response(
|
||||
hits a stale generation token and returns an empty result, leaving the platform with nothing to send.
|
||||
(#31884)
|
||||
|
||||
Failed context-overflow turns rewrite even when ``final_response`` already holds the raw HTTP
|
||||
400 envelope. Returning that envelope unchanged lets chat sanitizers replace it with a generic
|
||||
provider-failed reply, so the user never sees /compact (#1630 message layer).
|
||||
A failed context-overflow turn whose ``final_response`` is only the raw provider envelope
|
||||
(``HTTP 400: {...}``) is rewritten too: returned unchanged, chat sanitizers turn it into a
|
||||
generic provider-failed reply and the user never sees /compact. Curated agent text survives.
|
||||
"""
|
||||
is_overflow = is_context_overflow_failure_result(agent_result, history_len)
|
||||
if response and not (is_overflow and _looks_like_gateway_provider_error(response)):
|
||||
return response
|
||||
if agent_result.get("failed"):
|
||||
# ``error`` can be an EXPLICIT None (bypasses dict.get default) -> would render "failed: None".
|
||||
error_detail = agent_result.get("error") or "unknown error"
|
||||
@@ -3047,19 +3050,13 @@ def _normalize_empty_agent_response(
|
||||
"⚠️ Session storage was temporarily unavailable, so this "
|
||||
"turn was stopped to protect your conversation history. "
|
||||
"Your message should already be saved — please send it again in a moment.")
|
||||
if any(p in error_str for p in (
|
||||
"context", "token", "too large", "too long", "exceed", "payload")) or (
|
||||
"400" in error_str and history_len > 50):
|
||||
if is_overflow:
|
||||
return (
|
||||
"⚠️ Session too large for the model's context window.\n"
|
||||
"Use /compact to compress the conversation, or /reset to start fresh.")
|
||||
if response:
|
||||
return response
|
||||
return (
|
||||
f"The request failed: {str(error_detail)[:300]}\n"
|
||||
"Try again or use /reset to start a fresh session.")
|
||||
if response:
|
||||
return response
|
||||
|
||||
api_calls = int(agent_result.get("api_calls", 0) or 0)
|
||||
if agent_result.get("interrupted"):
|
||||
|
||||
+24
-15
@@ -43,6 +43,29 @@ if TYPE_CHECKING: # string annotations only; never imported at runtime (cycle)
|
||||
logger = logging.getLogger("gateway.run")
|
||||
|
||||
|
||||
_CONTEXT_OVERFLOW_ERROR_PHRASES = (
|
||||
"context length", "context size", "context window",
|
||||
"maximum context", "token limit", "too many tokens",
|
||||
"reduce the length", "exceeds the limit",
|
||||
"request entity too large", "prompt is too long",
|
||||
"payload too large", "input is too long",
|
||||
)
|
||||
|
||||
|
||||
def is_context_overflow_failure_result(agent_result: dict, history_len: int) -> bool:
|
||||
"""One verdict for "this failed turn is a context overflow", shared by transcript persistence
|
||||
(#1630 skip) and the user-facing reply so the two can never disagree.
|
||||
|
||||
Multi-word phrases (not bare "exceed"/"token") avoid matching "rate limit exceeded" or
|
||||
"invalid authentication token"; a bare 400 only counts on a long session."""
|
||||
if not agent_result.get("failed"):
|
||||
return False
|
||||
if agent_result.get("compression_exhausted"):
|
||||
return True
|
||||
err = str(agent_result.get("error") or "").lower()
|
||||
return any(p in err for p in _CONTEXT_OVERFLOW_ERROR_PHRASES) or ("400" in err and history_len > 50)
|
||||
|
||||
|
||||
class GatewayTurnMixin:
|
||||
"""Agent-turn execution for GatewayRunner (see module docstring)."""
|
||||
|
||||
@@ -1498,14 +1521,6 @@ class GatewayTurnMixin:
|
||||
except Exception as e:
|
||||
logger.debug("Watch queue drain error: %s", e)
|
||||
|
||||
_CONTEXT_OVERFLOW_ERROR_PHRASES = (
|
||||
"context length", "context size", "context window",
|
||||
"maximum context", "token limit", "too many tokens",
|
||||
"reduce the length", "exceeds the limit",
|
||||
"request entity too large", "prompt is too long",
|
||||
"payload too large", "input is too long",
|
||||
)
|
||||
|
||||
def _hmwa_classify_turn_failure(self, agent_result, history, session_entry):
|
||||
"""Classify a finished turn for transcript persistence. Returns
|
||||
``(agent_failed_early, hidden_reasoning_incomplete, is_context_overflow_failure)``.
|
||||
@@ -1524,13 +1539,7 @@ class GatewayTurnMixin:
|
||||
# user turn so the conversation is preserved. (#7100)
|
||||
agent_failed_early = bool(agent_result.get("failed"))
|
||||
hidden_reasoning_incomplete = _is_gateway_hidden_reasoning_incomplete_turn(agent_result)
|
||||
_err = str(agent_result.get("error", "")).lower()
|
||||
# Multi-word phrases (not bare "exceed"/"token") avoid matching "rate limit exceeded".
|
||||
is_context_overflow_failure = agent_failed_early and (
|
||||
bool(agent_result.get("compression_exhausted"))
|
||||
or any(p in _err for p in self._CONTEXT_OVERFLOW_ERROR_PHRASES)
|
||||
or ("400" in _err and len(history) > 50)
|
||||
)
|
||||
is_context_overflow_failure = is_context_overflow_failure_result(agent_result, len(history))
|
||||
if is_context_overflow_failure:
|
||||
logger.info(
|
||||
"Skipping transcript persistence for context-overflow "
|
||||
|
||||
@@ -129,39 +129,32 @@ class TestGenericFailureRegression:
|
||||
|
||||
|
||||
class TestNonempty400EnvelopeOverflowReply:
|
||||
"""Failed turns that already carry the HTTP 400 envelope as
|
||||
``final_response`` must still get the session-too-large rewrite.
|
||||
|
||||
Production order is normalize then Telegram sanitizer. The empty-response
|
||||
rewriter used to return the 56-char envelope unchanged, and the sanitizer
|
||||
then replaced it with a generic 'provider failed' — so users never saw
|
||||
/compact even though the gateway already classified the turn as overflow.
|
||||
"""
|
||||
"""Failed turns that already carry the HTTP 400 envelope as ``final_response`` must still get
|
||||
the session-too-large rewrite (chat sanitizers would otherwise turn the envelope into a generic
|
||||
'provider failed' reply); every other failure with text keeps that text."""
|
||||
|
||||
_ENVELOPE = 'HTTP 400: {"object":"error","model":"deepseek-v4-flash"}'
|
||||
|
||||
def _failed_400(self):
|
||||
return {
|
||||
"final_response": self._ENVELOPE,
|
||||
"failed": True,
|
||||
"error": self._ENVELOPE,
|
||||
"api_calls": 1,
|
||||
}
|
||||
def _failed(self, text, error=None):
|
||||
return {"final_response": text, "failed": True, "error": error or text, "api_calls": 1}
|
||||
|
||||
def test_long_history_rewrites_envelope_to_session_too_large(self):
|
||||
response = _normalize_empty_agent_response(
|
||||
self._failed_400(), self._ENVELOPE, history_len=138,
|
||||
self._failed(self._ENVELOPE), self._ENVELOPE, history_len=138,
|
||||
)
|
||||
assert "context window" in response.lower()
|
||||
assert "/compact" in response
|
||||
assert self._ENVELOPE not in response
|
||||
|
||||
def test_telegram_sanitizer_keeps_overflow_rewrite(self):
|
||||
from gateway.run import _sanitize_gateway_final_response
|
||||
@pytest.mark.parametrize("text", [
|
||||
"Billing or credits exhausted: HTTP 429: You exceeded your current quota",
|
||||
"API call failed after 3 retries: HTTP 429 rate limit exceeded",
|
||||
"HTTP 401: invalid authentication token",
|
||||
])
|
||||
def test_non_overflow_failure_keeps_its_own_text(self, text):
|
||||
assert _normalize_empty_agent_response(self._failed(text), text, history_len=138) == text
|
||||
|
||||
rewritten = _normalize_empty_agent_response(
|
||||
self._failed_400(), self._ENVELOPE, history_len=138,
|
||||
)
|
||||
delivered = _sanitize_gateway_final_response("telegram", rewritten)
|
||||
assert "/compact" in delivered
|
||||
assert "provider failed after retries" not in delivered.lower()
|
||||
def test_curated_overflow_text_from_the_agent_survives(self):
|
||||
text = "Context compression timed out without reducing this conversation. No messages were dropped."
|
||||
result = {**self._failed(text), "compression_exhausted": True}
|
||||
assert _normalize_empty_agent_response(result, text, history_len=138) == text
|
||||
|
||||
Reference in New Issue
Block a user