Add TopxAI provider (Claude, GPT, Grok, Kimi, GLM-5.3-Abliterated via one OpenAI-compatible endpoint)

[UnbelievableT]

Adds the topxai provider: TopxAI relays the labs' own models through one OpenAI- and Anthropic-compatible endpoint (https://ai.topxea.com/v1) at fixed USD prices on prepaid credit.

• provider.toml: @ai-sdk/openai-compatible, api, env = ["TOPXAI_API_KEY"], docs link, and the reasoning request fields per endpoint in comments.
• logo.svg with currentColor.
• Nine models, all base_model references except the one fine-tune unique to the host:
  • anthropic/claude-sonnet-5, claude-opus-5, claude-fable-5-1, claude-fable-5
  • openai/gpt-5.6-sol, gpt-6-astra (with the 272K whole-request tier as cost.tiers)
  • xai/grok-4.6 (200K tier)
  • moonshotai/kimi-k3
  • GLM-5.3-Abliterated: a fine-tuned GLM-5.3 served from the host's private deployment, defined inline.
• Prices are the Auto-route (shared pool) prices from https://ai.topxea.com/pricing and GET https://ai.topxea.com/api/pricing, accessed 2026-09-19; every model page is linked in the file comments.

bun validate passes.

🤖 Generated with Claude Code

[github-actions]

Action items

• [high] [violation] providers/topxai/models/claude-fable-5.toml:6 - Check: Relay reasoning_options must match the lab + same-surface peer baseline for that model. Why: First-party Anthropic, OpenRouter, and Neon all author Claude Fable 5 as effort-only (low/medium/high/xhigh/max) with no toggle; Neon notes thinking.type.disabled is rejected. Adding toggle invents an off control this model does not expose. Action: Drop the toggle option and keep effort-only, matching lab/peers (same fix for claude-fable-5-1.toml).
• [high] [violation] providers/topxai/models/claude-sonnet-5.toml:7 - Check: Every toggle needs a leading top-of-file wire-path comment. Why: claude-sonnet-5, claude-opus-5, claude-fable-5, claude-fable-5-1, and kimi-k3 declare toggle but only have pricing/docs comments—no exact request field (e.g. thinking.type, enable_thinking). Provider-level notes do not satisfy the per-file requirement. Action: Add a leading # Toggle: … wire comment on each file that keeps toggle, or remove toggle if this host has no separate on/off control on the registered OpenAI-compatible surface.
• [medium] [possible mistake] providers/topxai/models/gpt-6-astra.toml:6 - Check: Effort values must not invent levels beyond lab/peer sets. Why: First-party OpenAI and established peers (OpenRouter, Neon) use low/medium/high/xhigh/max with no none for GPT-6 Astra; this entry adds none. Action: Remove none unless TopxAI docs or a live test show a meaningful off/none control for this model ID on this host.

[UnbelievableT]

Addressed all three items:

• claude-fable-5 and claude-fable-5-1: toggle dropped; effort-only (low/medium/high/xhigh/max), matching the lab and peers.
• claude-sonnet-5, claude-opus-5, kimi-k3: a leading # Toggle: wire comment on each file naming the request field ($.thinking.type on /v1/messages for Claude, forwarded byte for byte; $.thinking.type as Moonshot's API takes it for Kimi) and the effort field per surface.
• gpt-6-astra: none removed; values are now the lab's low/medium/high/xhigh/max. gpt-5.6-sol keeps none because OpenAI's own entry for that model lists it.

bun validate passes.

[github-actions]

No actionable findings.