[m-musikhin]

Summary

Two related changes in one PR (the second is a follow-up data fix to the first):

1. User-defined provider presets (~/.cligate/provider-presets.json)

Adding an OpenAI-compatible vendor previously required hardcoding a provider class. Now any vendor can be described declaratively in ~/.cligate/provider-presets.json (same schema as built-in presets: id, label, baseUrl, apiFormat, optional models/tiers/nativeModelPrefixes/passthroughModelMatch/pricing/validateModel) and it becomes a first-class provider everywhere: add-key form, model mapping with its own tier map, model discovery, protocol compatibility, pricing.

• Loader runs at module init; each invalid entry is skipped with a console warning (bad id shape, missing baseUrl, unknown apiFormat/tier key, built-in id collision, duplicates → first wins). A bad file never blocks startup.
• New preset field validateModel: for vendors without GET /models, key validation probes a 1-token chat request with this model (used by Zhipu).
• Generic openai_chat provider serves chat + Anthropic bridge (/v1/messages works for Claude Code via translation); Codex goes through the Responses→chat fallback.

2. Zhipu (GLM) / Moonshot (Kimi) / MiniMax migrated to presets + catalogs actualized

The three bespoke provider classes are replaced by declarative presets served by the generic OpenAI-compatible class (like qwen/openrouter before them). Behavior changes:

• Zhipu key validation now probes a 1-token chat request (validateModel: glm-4-flash) — previously there was no /models validation for Zhipu at all, so no regression.
• Zhipu/Moonshot/MiniMax keys now also serve /v1/chat/completions and the Codex Responses fallback (they previously only served /v1/messages).
• Each now has its own tier map and model list, so non-native incoming models are remapped instead of passing through verbatim, and the model-mapping page gives each its own editor instead of folding them into the OpenAI one (which the runtime never consulted for these keys — an existing UI/runtime contradiction, fixed in settings-page.js normalizeModelMappingProvider).
• Model catalogs actualized to Oct 2026 (ids verified against OpenRouter vendor listings: z-ai/glm-5.3, moonshotai/kimi-k3, minimax/minimax-m3): zhipu glm-5.3 family, moonshot kimi-k3/k2.7-code/k2.6, minimax MiniMax-M3/M2.7; pricing tables synced.

Test plan

• npm run test:unit: no new failures vs master (existing model-mapper.test.js failures are pre-existing on master and unrelated).
• New tests: tests/unit/provider-user-presets.test.js (subprocess-based, fresh CLIGATE_CONFIG_DIR per case: merge, collision, invalid entries, duplicates), tests/unit/model-mapping-provider-normalization.test.js (UI provider identity 1:1 with runtime), extended tests/unit/provider-presets.test.js (migrated presets addable, tier remaps, validateModel).
• Manual: user preset JSON → add-key form shows the vendor, mapping page shows its tier map, key validation, discovery.

🤖 Generated with Claude Code

[codeking-ai]

Thanks for this — the preset-driven approach is a clear improvement, and moving zhipu/moonshot/minimax onto the generic provider also fixes a real bug (the old classes inherited OpenAIProvider.sendAnthropicRequest, which POSTs to a /responses endpoint these vendors don't have). Related unit tests pass; the model-mapper.test.js failures are pre-existing on master. Merging.

A few follow-ups worth addressing (happy to take a follow-up PR, or I'll handle them):

1. User presets accept apiFormat values that can't work. userPresetProblem allows anthropic / gemini_native, but api-key-manager.js always builds presets with makeOpenAICompatibleProvider, and credential-selector.js only wires openai_chat presets into protocol surfaces. Such a preset can be added but is never routed. Suggest restricting user presets to openai_chat until other formats are actually supported.
2. Id-collision check only covers built-in presets, not the class-backed providers (openai, anthropic, gemini, azure-openai, vertex-ai, deepseek). A user preset with id: "openai" passes validation and shows in preset-driven UI/mapping, while api-key-manager silently keeps the real OpenAIProvider. Please reject those ids too (and probably ollama, which the local runtime owns).
3. gateway-route.js still hard-codes minimax, moonshot, zhipu in typesToTry alongside ...presetIdsByFormat('openai_chat'), so they're tried twice per attempt now. handleListProviders also doesn't list preset providers. Both should derive from the preset list.
4. Pricing entries removed for still-referenced models (glm-4-plus, glm-4-air, glm-4-airx, kimi-k2.5, kimi-k2-thinking, kimi-k2-thinking-turbo). Users who pin these via model mapping will now see $0 cost. Consider keeping legacy rows (pricing for unused models is harmless).
5. Minor: worth calling out in the changelog that non-native model ids sent to zhipu/moonshot/minimax keys are now tier-remapped instead of passed through verbatim — it's a behavior change for anyone relying on passthrough.