adlee-was-taken 7d9988caf3 test: make the HTTP fakes exercise the boundary instead of modelling it ...

The mojibake in 8518114 survived a green suite, and not by omission. There
is a test called `test_a_stream_is_proxied_verbatim_including_the_telemetry
_comments` which passed for months against a proxy that was decoding the
stream as ISO-8859-1 and re-encoding it as UTF-8. It asserted three ASCII
substrings, and a test that only ever sees ASCII cannot observe a charset
bug.

Two separate lies, both now fixed.

The fakes. `iter_lines` yielded `self._lines` unchanged in both modes, so a
test holding str lines got those str back even under decode_unicode -- a
model of a stream that had ALREADY been decoded correctly. A fake that
hands back the right answer cannot reproduce a decode bug. They now treat
the wire as UTF-8 bytes whatever the test wrote, derive `encoding` from
Content-Type using requests' own `get_encoding_from_headers`, and default
to the charset-less `text/event-stream` OpenRouter really sends. The
hostile case is the default, so a broken passthrough can no longer look
correct.

The assertions. The verbatim test now compares BYTES against every line the
provider emitted, and the shared STREAM_LINES sample carries a raw UTF-8 em
dash, so the whole streaming surface fails if the proxy ever goes back to
decoding and re-encoding.

Measured, by reverting 8518114 and running the suite:

    before this commit   1 test caught it (the one written for it)
    after                2, incl. the verbatim test that had been lying

The dedicated regression test also gets simpler -- it no longer needs a
bespoke response subclass, because the shared fake is now faithful enough.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U

2026-09-08 14:05:16 -04:00

8.5 KiB

Raw Permalink Blame History

View Raw