Repository navigation
Add httpx.SSLContext configuration. - #3022
Conversation
897a5de to
88f2ab7
Compare
This comment was marked as resolved.
This comment was marked as resolved.
|
Yes, I agree that removing them completely is not what the user expects. Or should we deprecate them anyway? Not obvious to me. |
This comment was marked as resolved.
This comment was marked as resolved.
|
Looking good.
I'd suggest you don't spin time on that until we've had a clear discussion around the "API finessing" and "Documentation refresh" areas mentioned in #947 (comment). I like this change, but we'll want to make it as part of a coherent & consistent 1.0 refresh. |
httpx.SSLContext configuration.
|
Okay, shall we take the docs from this comment... #3007 (comment) and include them in the PR? |
I like it |
|
Have we decided to remove or deprecate the old arguments? |
Co-authored-by: T-256 <132141463+T-256@users.noreply.github.com>
Dropping `trust_env` from `httpx.SSLContext()` API.
|
Thanks so much @karpetrosyan - going to merge this into "version-1.0" now, and then apply remaining changes incrementally. Fantastic work!! 😎 |
|
Sorry, I couldn't find time to finish all the tidy-ups. |
|
Thank you so much for your time & dedication @karpetrosyan. 🙌 |
Try it live calls the GitHub API from the browser and runs the pipeline in front of you. Fetching, linking and filtering are the production logic ported from the Python, and they reproduce its numbers exactly: on encode/httpx#3022, 17 files and 112 comments in, 50 kept at 45 percent, same as the Python run. Extraction is the one stage that is not the real thing. The shipped product calls a small fine-tuned model there; this runs rules so the page needs no key and costs nothing. That is stated under every result rather than left for someone to discover. No server anywhere. An optional token raises the GitHub rate limit and never leaves the tab. Finding nothing is presented as a real outcome, because most pull requests record no reasoning at all and measuring how often is the first research question. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Without a key and without a model on their machine, the page could only show rule-based candidates, which is the weakest thing it does. "See a finished record" replays five decisions read out of the real encode/httpx#3022 thread by a person and checked against the final diff. They run through the same renderer as a live extraction, so the shape is the real one, and every record carries origin "confirmed" rather than "extracted". That is the only example of the distinction anywhere in the app, and the card now says which it is rather than telling a verified record that nobody confirmed it. context is null on all five. The thread argued the choices but never stated the condition behind them. Inventing one would make the page look better and the system worth less. Six were verified by hand; five are shown. "ssl.SSLContext() is constructed with an explicit protocol argument" had nothing weighed against it, and the gate says that is not a decision. Exempting our own demo from our own rule would be the cheapest possible way to stop believing the rule, so the banner says a sixth was dropped and why. Also: a local model server only answers pages it has been told to allow. Verified that Ollama returns 403 to an OPTIONS from the deployed origin while accepting localhost, and that the deployed page's fetch therefore fails before any model runs. The interface now says so before Run is pressed, with the OLLAMA_ORIGINS command, instead of letting it surface as a bare network error. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Discussed in #3007
TODO:
Migration examples
Enabling and disabling verification
Use a different set of certificates
Client side certificates
Using alternate SSL contexts