test(e2e): shorten the new helper docstrings

This commit is contained in:
mateo-berri 2026-09-22 18:24:35 -07:00
parent 0d3be46456
commit e18fb87f69
4 changed files with 5 additions and 23 deletions

View file

@ -356,15 +356,8 @@ def tolerate_provider_rate_limit[R: BaseModel](
attempts: int = PROVIDER_RATE_LIMIT_ATTEMPTS,
sleep: Callable[[float], None] = time.sleep,
) -> Result[R]:
"""Bounded, opt-in retry of one call the PROVIDER throttled, for a deployment
with no fallback group whose provider key is shared with the suites that run
alongside this one. The transport never retries a 429 (see request_with_retry),
and this helper only absorbs the 429s the proxy relayed from the provider, which
litellm's exception mapping prefixes with its exception class; the proxy's own
limiter and budget 429s never carry that prefix and come back untouched, so
every assertion on those still sees them. The wait honors Retry-After when the
provider sent one and backs off geometrically otherwise, and every retry prints
so the throttling stays visible in the run log instead of vanishing into green."""
"""Retry a call up to `attempts` times while the proxy relays the provider's own 429;
any other outcome, the proxy's own 429 included, comes back at once."""
for attempt in range(1, attempts):
match issue():
case RateLimitedError(body=body, retry_after_seconds=retry_after) if PROVIDER_RATE_LIMIT_MARKER in body:

View file

@ -11,11 +11,6 @@ well-formed OCR document comes back. Per the e2e hard-fail contract, a case
fails when no proxy answers and also fails once a request reaches it: the proxy
fetches each provider's referenced secrets, so a
missing credential surfaces as a live provider error rather than silent green.
The Mistral key is shared with the suites that run alongside this one and the
deployment has no fallback group, so the OCR call tolerates a bounded number of
provider 429s the proxy relays (tolerate_provider_rate_limit); a 429 the proxy
issues itself still fails the case at once.
"""
from __future__ import annotations

View file

@ -377,13 +377,8 @@ class _AzureAnnotatedChatBody(BaseModel):
def azure_prompt_filter_skipped(resp: StreamingResponse) -> bool:
"""True when Azure answered without running its prompt content filter: the body
carries ``prompt_filter_results`` whose ``content_filter_results`` are all empty,
where a filter that ran records a verdict per category (jailbreak included).
Azure does this intermittently and it says nothing about the prompt, so a test
that needs the filter's verdict sends the prompt again instead of reading a
skipped filter as a pass. A body without the field (an OpenAI completion served
by a fallback, or an error body) is not a skipped filter."""
"""True when Azure's 200 recorded no prompt-filter verdict (every `content_filter_results`
empty), so the prompt has to be sent again."""
try:
annotated: Final = _AzureAnnotatedChatBody.model_validate_json(resp.body)
except ValidationError:

View file

@ -1,5 +1,4 @@
"""Unit tests for the reliability helpers that read provider bodies, driven with
the verbatim shapes the proxy relays; nothing here needs a proxy or a provider."""
"""Contract tests for the helpers that read provider bodies; no proxy or provider needed."""
from __future__ import annotations