Fix: Fixed models returning '404 model not found' when using :THINKING suffix

[stargazer617]

Bug

Sending messages to models with the :THINKING suffix returns a 404 model not found error.

For example, sending a test message to zai-org/GLM-5.1-TEE:THINKING returns:

Status Code:

404

Response Body:

{
  "error": {
    "message": "model 'zai-org/GLM-5.1-TEE:THINKING' not found",
    "type": "proxy_error"
  }
}

Whereas sending the same message to zai-org/GLM-5.1-TEE works as expected.

Root Cause

I believe this is because in e2ee_discovery.lua, the proxy calls:

local entry = map[model]
return check_confidential(model, entry)

But in the fetch_model_map() function, the model map is populated from the v1/models endpoint:

local map = {}
    for _, model in ipairs(data.data) do
        if model.id and model.chute_id then
            map[model.id] = {
                chute_id = model.chute_id,
                confidential = model.confidential_compute == true,
            }
        end
    end

The v1/models endpoint only returns the models' base names, without any variants or suffixes. For example, for GLM 5.1, the ID is zai-org/GLM-5.1-TEE. So if we make a request with the model zai-org/GLM-5.1-TEE:THINKING, the entry would be null, as the model map only contains zai-org/GLM-5.1-TEE, and not the suffixed variant.

Then, since entry is null, this line in check_confidential():

if not entry then return nil, "model '" .. model .. "' not found" end

returns model 'zai-org/GLM-5.1-TEE:THINKING' not found.

Fix

This pull request aims to fix this issue by stripping the :THINKING suffix from the model name before looking the model up in the model map.

[fstandhartinger]

looks good to me

[sirouk]

Superseded by #2 — moved the :THINKING handling to e2ee_round_trip
and dropped the per-model defaults. Thanks for catching the 404 and
for the back-and-forth. Your investigation is what made it clear
where the real chokepoint was.

[sirouk]

Opened #2 to supersede this — moved the :THINKING handling to
e2ee_round_trip and dropped the per-model defaults per the review
discussion. Thanks for catching the 404 and for the back-and-forth.
Your investigation is what made it clear where the real chokepoint
was. Feel free to close this out.