feat:/fix:

[elan40213-eng]

Hi! I’ve tweaked a few features to suit my needs

Model aliases in /v1/models. You can assign aliases to specific models; they are now displayed under their own name, without the provider prefix. When {"model": "my-alias"} is sent from any OpenAI-compatible client, the router will resolve it internally and then return the alias name in the response and in every stream fragment. Aliases pointing to disabled models are excluded from the list, and the cache is cleared immediately after any alias change, rather than after the TTL expires.

Default system prompt. This new parameter contains text that is added to the system slot of any chat request if it does not already contain a prompt. It works with OpenAI, Claude and Gemini request formats and is triggered early enough to allow other system prompt features to be added after it. Disabled by default

Access Control List (ACL) for models at key level. API keys can now restrict which models they are permitted to call, in addition to existing restrictions by provider and combination. A selector is provided in the control panel for this purpose. The corresponding check is now applied universally, including on public endpoints without a configured key.

Caching of the Google project ID. The cached ID is now valid for 30 days instead of 1 hour, as it is tied to the account rather than the token. During a scheduled token refresh, the ID is retrieved from the database rather than being re-requested from the Google endpoint for registration.

Gemini CLI request handling. Fields rejected by the Cloud Code Assist API are removed before the request is sent. The ‘Thinking’ visibility is now mapped to a native configuration field rather than being ignored. Tool schemas are sent in a format that the API can actually read.

‘Thinking’ output formats for NVIDIA NIM. Each model family now uses its own native format for outputting reasoning results, rather than a single common format for all.

DuckDuckGo search. It has been registered, and request headers have been rewritten to match those of a real browser, so that DuckDuckGo no longer throttles its bandwidth.

Headroom detection. Checking for the presence of the Headroom binary previously blocked the control panel for over 20 seconds if it was not installed. The results are now cached, and the slow search path is only triggered when necessary.

Control Panel optimisation. The profile page has been rewritten to a smaller file, an ACL key selector has been added, and several user interface elements now support keyboard navigation.

Each new database column allows null values, and each new setting is disabled by default, so existing deployments will continue to work without any changes. Compiled and tested from start to finish on my VPS

[Vanszs]

hi @elan40213-eng thanks for the pr, i will review it asap, please keep the pr open

[mahdiwafy]

Found an issue while merging this locally: package.json adds @tailwindcss/oxide-wasm32-wasi under dependencies. That package is wasm32-only, and npm rejects it with EBADPLATFORM on x64 hosts (e.g. the Docker builder) — npm install fails outright there. The native oxide-linux-x64-musl binary is what's actually used at build time on those platforms, so the wasm variant should be skippable rather than required.

Suggested fix — move it to optionalDependencies (same pattern already used for better-sqlite3 in this repo):

   "dependencies": {
     ...
-    "@tailwindcss/oxide-wasm32-wasi": "4.3.0",
     ...
   },
   "optionalDependencies": {
+    "@tailwindcss/oxide-wasm32-wasi": "4.3.0",
     "better-sqlite3": "^12.6.2"
   },

Also worth double-checking on merge: this PR touches src/app/api/combos/route.js and src/lib/db/repos/combosRepo.js (adding alias) — #58 touches the same two files/functions (adding context_length). Both changes are compatible, just a straightforward text conflict for whichever merges second (left the same note on #58).