feat(desktop): preview and normalize custom OpenAI request URLs

[liuxiaocs7]

Summary

A custom OpenAI relay configured with a full endpoint can send requests to /chat/completions/chat/completions or discover models under /responses/models. The service URL field now previews the final request address as the user types, in both connection creation and editing.

Chat and Responses previews, connection probes, SDK requests, and model discovery share endpoint normalization. Full endpoint inputs avoid duplicate suffixes; custom prefixes are preserved without assuming /v1. Empty or invalid drafts hide the preview, and long addresses wrap with an accessible description on the input.

Fixes #5692

Verification

• npm run build, npm run typecheck, and npm run lint: passed.
• Biome formatting: all tracked files passed. The checkout-wide command also finds unrelated untracked local research files, which are not part of this PR.
• Desktop suite: 2,834 passed (NODE_OPTIONS=--disable-warning=ExperimentalWarning npm --workspace @maka/desktop run test:dist).
• Runtime thinking options, endpoint, Responses wire, and model discovery suites: 81 passed. Vercel wire assertions cover each advertised effort without freezing the snapshot catalog list. The added regressions exercise actual probe, SDK generation, and discovery paths and fail on the original duplicated URLs.
• Regenerated Astryx surface inventory: coverage check and 19 generator tests passed.
• Renderer architecture check and npx knip --workspace apps/desktop / npx knip --workspace packages/ui: passed.
• Browser verification: live updates in add/edit forms, both protocols, clearing and incomplete input, full endpoint deduplication, narrow-window wrapping, and accessible input descriptions.

The full-workspace suite was also run. Local failures remain outside this change: the Runtime Host implementation-child test also times out with the baseline Runtime sources, and the Eval Python tests are incompatible with the system Python 3.9. The other process/timing failures passed on isolated serial reruns.

The screenshots use the actual upstream and PR forms with an example relay address:

Before	After

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Codex assisted with implementation, regression tests, browser verification, and issue/PR drafting. The implementation and follow-up commits include Generated-by: Codex.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[Astro-Han]

Adversarial pass at 0ea5494d5; verified SDK URL composition against the installed @ai-sdk/openai/openai-compatible versions and exercised openAiBaseUrl across the input-shape matrix.

The core fix is correct: openAiBaseUrl is the single suffix-stripping authority (the while loop even peels already-duplicated …/chat/completions/chat/completions), custom prefixes are preserved without assuming /v1, the codec rejects query/fragment/userinfo/non-http(s) before normalization so the preview is byte-identical to the wire request, invalid drafts hide it, and probes/SDK send/discovery all funnel through it. No P0/P1.

One P2 inline — model discovery still strips endpoint suffixes only for providerType === 'custom', while send and probe now strip for every OpenAI wire. A non-custom provider that allows endpoint override (OpenRouter, DeepSeek, etc.) with a saved full-endpoint baseUrl now probes and chats correctly but discovers models at …/chat/completions/models — a new inconsistent state (pre-PR it failed uniformly). Strip whenever the resolved adapter kind is openai/openai-compatible instead of keying on providerType.

P3s, all deferrable: openai-compatible adapters with normalizeBaseUrl bypass the strip and can produce …/chat/completions/v1/chat/completions (kimi-coding-plan only — compose the two normalizers); redactSecrets can make the preview differ from the request for token-shaped path segments (safe direction, worth a comment); duplicate preview render in the legacy-credential edit edge; a decorative ternary in discovery where both branches call openAiBaseUrl; the anthropic-messages wire has the same bug class with no preview (issue-scoped, worth noting); and the inherent ambiguity — a gateway legitimately mounted at /chat/completions gets stripped — deserves one line in the openai-urls.ts comment.