Add Vercel AI Gateway as a first-class provider

[mundabra]

What

Adds the Vercel AI Gateway as a first-class provider: one key, 300+ models across every major creator, through the gateway's OpenAI-compatible endpoint. Filed as a reseller — same category as Together / Fireworks / OpenRouter (many labs' models behind one key, their model namespaces verbatim).

• registry.py — _compat() descriptor at the end of the reseller block: vercel: routing prefix, prefilled https://ai-gateway.vercel.sh/v1, AI_GATEWAY_API_KEY (Vercel's documented env var), recommended anthropic/claude-haiku-4.5
• matrix.py — six curated entries, one per major creator, agent-capable cost-efficient tiers. Ids, context windows, and vision flags are the gateway's live /v1/models values (315 models, checked 2026-08-05). Introduces _AGENTIC_VISION_COMPAT (vision without pdf=) and folds the two existing inline copies of that literal (Muse Spark, Kimi K3 via Together) into it
• GUI — provider-gallery logo (lobe-icons mark, vendored verbatim), order slot beside the other resellers, and a "get a key" link using Vercel's own documented dashboard deep-link
• README — joins the provider enumeration
• tests — reseller lockstep, prefix routing for slash-carrying ids (vercel:zai/glm-5.2), and matrix capability answers for creator/model-shaped ids

Design notes

• Vision entries omit pdf= on purpose: the gateway's catalog advertises pdf input, but its compat file-part schema (file.data + media_type) differs from the OpenAI shape this client emits (file_data data-URL), and it is unverified here — so PDFs keep the graceful local pdf_support fallback, the same posture as every existing vision-capable compat entry.
• Context windows are per-route: the gateway's catalog values differ from other resellers' caps for the same weights (GLM-5.2: 1M here vs 128k via Together); the matrix records what each route actually serves.
• Uncurated ids typed by hand (vercel:creator/model) route correctly and fall back to conservative capabilities, like every provider's custom ids.
• No provider-code changes: the request path is the same OpenAIProvider every compat vendor already uses.

Verification

• Live GET /v1/models catalog pulled 2026-08-05; every curated id, context window, and vision flag checked against it
• Test-first: with only the test changes applied, the suite fails 3 (descriptor, suggestions, lockstep); with the feature: 1112 passed, 1 skipped; GUI tsc --noEmit clean
• Routing traced end-to-end for slash-carrying ids (_provider_name/_bare split once on :; the GUI prefixed(), unconfigured-provider preview, and set_provider auto-add all attach the prefix). /v1/providers boot-checked in-process: vercel serializes with its own key slot and prefilled endpoint, ordered with the resellers
• The gallery mark render-checked against the sibling reseller logos on the light chip plate
• Live-verified with a real key (2026-08-05), through this descriptor's own build() path: non-stream and streaming completions plus a tool call on anthropic/claude-haiku-4.5 (finish tool_calls), and a second creator (zai/glm-5.2) — all returned correct results with usage counts. Key-verify can't false-positive — /v1/models is public but any wrong bearer returns 401

🤖 Generated with Claude Code