feat(provider): add Holver

[zenttzy]

Holver (https://holver.id) is an OpenAI-compatible AI model gateway. This PR registers it as a native provider so OpenCode users get it via /connect.

Endpoint & auth

• Base URL: https://api.holver.id/v1 (verified: public docs app BASE_URL, nginx api.holver.id, live GET /v1/models returns 401 without key with OpenAI-style error)
• Auth: Authorization: Bearer $HOLVER_API_KEY
• Adapter: @ai-sdk/openai-compatible (/v1/chat/completions, /v1/responses, /v1/messages; streaming, tools/tool_choice, response_format, reasoning_effort passthrough)
• Docs: https://holver.id/docs · Catalog: https://holver.id/models

Model mapping (19 models, all base_model override-only)
Anthropic (fable-5, fable-5.1, haiku-4.5, opus-4-8, opus-5), DeepSeek (4.1-flash, v4-pro), Google (3.1-pro via preview lab entry — no GA lab file exists yet, noted in header — 3.7/3.8-flash), Zhipu (glm-5.3-flash), OpenAI (5.6-luna/sol/terra, 6.0-astra, oss-120b), MiniMax (m3), Meta (muse-spark-1.3-contributor), Alibaba (qwen-3.8-max). Reasoning controls mirror each lab's first-party API.

Pricing
No fabricated pricing: Holver bills via IDR credit packages, not per-model USD rates, so cost is intentionally omitted (same precedent as agentrouter).

Testing

• bun validate → exit 0
• Built catalog contains holver with 19 models + logo
• OpenCode with local catalog: /connect flow shows Holver, /models lists 19 holver/* IDs
• Live inference holver/deepseek-4.1-flash via OpenCode returns exact reply; SSE streaming ([DONE]) and tool_calls (function + JSON args) verified against production

[github-actions]

Action items

• [medium] [violation] providers/holver/models/deepseek-4.1-flash.toml:133 - Check: Every toggle must have a leading top-of-file wire-path comment. Why: deepseek-4.1-flash, deepseek-v4-pro, minimax-m3, and qwen-3.8-max set { type = "toggle" } but only generic catalog/cost notes—no exact Holver request field (e.g. thinking.type, enable_thinking). Clients cannot know how to turn reasoning off/on on this host. Action: Add a leading # Toggle: … (and effort/budget wire notes where used) naming the real Holver request path for each of those four files.
• [medium] [possible mistake] providers/holver/models/claude-haiku-4.5.toml:96 - Check: budget_tokens only when this host exposes a real reasoning-budget field; do not treat Anthropic Messages and OpenAI chat-completions as the same surface. Why: Holver is registered as @ai-sdk/openai-compatible and the PR documents reasoning_effort passthrough, while Haiku’s lab/Anthropic control is thinking.budget_tokens. Same-surface peers often use toggle/effort, not budget, unless an Anthropic Messages path is verified. Action: Confirm Holver accepts a reasoning budget on the path clients will use; if not, replace with the controls Holver actually forwards (e.g. effort/toggle) or document the Messages wire path that makes budget valid.

[zenttzy]

Both points addressed in 6f1bae8. Verified against the Holver deployment source, not assumed:

Wire path (reasoning_effort only). Holver's OpenAI-compatible adapter sanitizes the request body against an allowlist (apps/api/src/gateway/adapters/ninerouter.js). reasoning_effort is in it; thinking, enable_thinking, and thinking_budget are not — they are dropped before anything reaches the routed upstream. So the lab toggle/budget controls flagged on deepseek-4.1-flash, deepseek-v4-pro, minimax-m3, and qwen-3.8-max were not reachable through this API. Those files are now effort-only, and minimax-m3 is [] (no caller-facing tiers on this surface). Every reasoning model now carries a leading top-of-file comment naming reasoning_effort as the actual Holver request field.

Haiku budget_tokens. You're right that this was inferred from the Anthropic surface rather than this host. Holver's /v1/messages translator builds the OpenAI body from an explicit field list (anthropicTranslate.js → buildOpenAIBody) that never includes thinking; on the chat path the allowlist above excludes it too. So budget_tokens was never forwarded. Replaced with the effort ladder, matching the same-surface OpenAI-compatible relay peer providers/daoxe/models/claude-haiku-4-5-20251001.toml.

Also confirmed empirically: thinking/enable_thinking sent to POST /v1/chat/completions are accepted with HTTP 200 and silently ignored (no error, no reasoning change), consistent with the strip-before-forward behavior.

bun validate passes (19 models).

[github-actions]

No actionable findings.