[PCcoding666]

Summary

Refs #234. This implements the shared-connection portion of the BigModel RFC. A user can select an existing mainland BigModel account when adding models in another category. In settings, a short account identifier distinguishes two credentials on the same endpoint. Edits and deletions apply to the selected category. Rotating a key updates the selected account; the other account keeps its own key.

The selected category retains its protocol and models, including its own proxy settings. Existing standalone records can be reused without changing their model IDs.

Evidence

Two accounts on the same endpoint	Reusing the selected account without entering its key again

Chromium exercised the settings form against the local API with two fixture image accounts. In each language I selected a different account and saved three chat models with its existing credential. Reopening the catalog showed both original image records and the chat models under the selected account. Chinese account picker · Saved chat models

Verification

• 51 focused Python tests and all 767 Node tests passed. Production build, bilingual checks, generated contracts and import boundaries passed.
• Full Python suite with isolated OpenCode state: 1,620 passed, 100 skipped. With the host's existing OpenCode state, one runtime-catalog test failed on both this branch and clean main.
• In a temporary compatibility check, the preset definitions from main and Add BigModel GLM-Image generation for canvas and Agent #251, Add BigModel CogVideoX-3 video generation #254 and [feat] providers: add Zhipu BigModel GLM-TTS and voice management (#238) #276 matched the intended account across all 16 source/destination pairs. A separate FastAPI/persistence check using the video protocol definitions from Add BigModel CogVideoX-3 video generation #254 exercised all 16 pairs with two same-endpoint accounts, including reload and category-scoped deletion.
• These checks used synthetic credentials. No real BigModel generation was submitted. The full multi-provider-PR build and the four-modality media-generation/adoption workflow remain unverified. They can be exercised once the outstanding modality PRs land.

Prepared and tested with Codex assistance.