feat(provider): add CommandCode provider

[lincolnthalles]

CommandCode (https://commandcode.ai) is an OpenAI-compatible gateway that relays models from many labs (Anthropic, OpenAI, Google, DeepSeek, Alibaba, Moonshot, and others) under a single API and key.

The catalog didn't include it, so its models were unavailable to tooling that consumes models.dev.

This adds the provider with a compliant logo and 63 models.

Since CommandCode hosts other labs' models rather than creating its own, every entry is override-only: it points at the existing lab metadata via base_model and authors only what differs on this host - cost, reasoning controls, and reasoning side-channel.

Notes for review:

• Host classification: multi-model relay (@ai-sdk/openai-compatible with a custom api). Reasoning options are copied from each model's first-party lab entry rather than a generic GPT-style set (e.g., DeepSeek V4 = toggle + high/max, Qwen = toggle + thinking_budget, Tencent Hy = effort none|high with no toggle).
• Per-model pricing and documentation links are cited in the leading comment block of each file (the sync-safe position); host-specific quirks such as the 400K context cap on GPT-5.4/5.5 and DeepSeek's off-peak pricing are documented there too.
• Sync module is not viable: the models API is too poor, providing only model IDs and context limits.
• bun validate passes.

closes #3086 #4530 #5937 #6360

[github-actions]

Action items

• [medium] [possible mistake] providers/commandcode/models/gpt-5.4.toml:14 - Check: Context-tier pricing must match what this host charges when the standard rate already mirrors OpenAI. Why: Base rates match OpenAI GPT-5.4/5.5 (2.5/15 and 5/30 with the same cache reads), and both files still allow up to 400K context, so prompts in the 272K–400K band would hit OpenAI’s long-context tier on first-party and on peers such as Neon. These entries omit [[cost.tiers]] entirely, so catalog costs understate long-context usage if CommandCode mirrors OpenAI banding. Action: Add the 272K [[cost.tiers]] bands (OpenAI: 5.4 5/22.5/0.5, 5.5 10/45/1) or document in the leading comment that CommandCode flat-prices through the 400K cap.
• [medium] [possible mistake] providers/commandcode/models/gpt-5.6-luna.toml:15 - Check: Long-context cost tiers should carry the same cost keys the base band publishes when the host charges them. Why: Base [cost] includes cache_write (luna 0.25, sol 6.25), and OpenAI/Neon long tiers also publish cache_write (0.50 / 12.5). The 272K tiers here only set input/output/cache_read, so long-context cache pricing is incomplete versus the sibling gpt-5.6-terra entry and established peers. Action: Add cache_write on both long tiers (or cite CommandCode docs that long-context cache writes are uncharged/different).
• [medium] [possible mistake] providers/commandcode/models/thinkingmachines/inkling-small.toml:5 - Check: Relay reasoning_options = [] means affirmative no caller control, not uncertainty (AGENTS.md / audit skill). Why: The file claims always-on with no control, but same-model relays commonly expose effort (OpenRouter/Kilo none|minimal|low|medium|high|max, Vercel also xhigh), and first-party Inkling on this PR uses toggle + graded effort. Empty options will hide real controls if CommandCode forwards them. Action: Align with the host’s real wire surface (likely effort, and toggle if present) with a leading wire comment, or keep [] only with host-specific evidence that effort/toggle are rejected.
• [low] [possible mistake] providers/commandcode/models/tencent/Hy3.toml:6 - Check: Multi-model relays should copy the underlying model’s effort set from lab/same-surface peers. Why: hy3 / hy3-paid use ["none", "high"], but established peers (OpenRouter, Vercel, opencode-go) and HF-facing docs use none|low|high for Hy3. Dropping low understates the control surface if CommandCode passes reasoning_effort through. Action: Verify CommandCode’s accepted values; add low if supported, or cite host docs that only none/high work.

[github-actions]

Action items

• [high] [possible mistake] providers/commandcode/models/claude-haiku-4-5-20251001.toml:6 - Check: Reasoning options must match this host’s API surface, not treat Anthropic Messages and OpenAI-compatible chat as interchangeable (AGENTS.md Reasoning options / audit skill). Why: CommandCode is an @ai-sdk/openai-compatible multi-model gateway, but Haiku is authored with Anthropic Messages-style budget_tokens only. Same-surface peers (e.g. OpenRouter) expose Claude via gateway toggle/effort, not pure Messages budgets. Copying lab Messages controls can advertise a budget field this host does not accept. Action: Verify CommandCode’s actual Claude request shape; if it is chat-completions/openai-compatible only, replace budget_tokens with the host/peer control set (often toggle and/or effort) and document the wire path; keep budget_tokens only with evidence this path proxies Anthropic thinking.budget_tokens.
• [high] [possible mistake] providers/commandcode/models/claude-sonnet-4-6.toml:9 - Check: Same host-surface rule for Claude Sonnet 4.6. Why: Entry mixes lab Messages effort + budget_tokens on an openai-compatible relay. Established openai-compat peers commonly use toggle + effort without Messages budget (e.g. OpenRouter). Action: Confirm CommandCode’s Claude controls for Sonnet 4.6; drop invented/unproxied budget_tokens and align to lab∩same-surface controls this host actually forwards.
• [medium] [possible mistake] providers/commandcode/models/tencent/hy3-paid.toml:3 - Check: Distinct provider model IDs that share base_model need a distinguishing display name when the ID alone is not enough. Why: tencent/Hy3 (free, $0) and tencent/hy3-paid both inherit lab name "Hy3" with no override, so the paid SKU is indistinguishable in UI/catalog listings. Free/paid siblings elsewhere set an explicit name. Action: Add name = "Hy3 Paid" (or the host’s official label) on hy3-paid.toml; consider name = "Hy3 Free" on the free entry if that is how CommandCode brands it.
• [medium] [possible mistake] providers/commandcode/models/zai-org/GLM-5.toml:4 - Check: interleaved must reflect this host’s reasoning side-channel when present. Why: Lab zhipuai/glm-5 and same-surface peers author [interleaved] field = "reasoning_content", and CommandCode already sets that for GLM-5.3 / 5.3-Flash on this provider, but GLM-5 / 5.1 / 5.2 omit it. If CommandCode returns reasoning_content for those IDs too, clients will miss the side channel. Action: Verify response shape for GLM-5, GLM-5.1, and GLM-5.2 on CommandCode; add matching [interleaved] (and keep it off only if this host truly does not expose a side channel for those IDs).

[github-actions]

Action items

• [high] [violation] providers/commandcode/models/thinkingmachines/inkling.toml:1 - Check: Multi-model OpenAI-compatible relays must use same-surface peer controls, not a different lab API shape (AGENTS.md Reasoning options; audit skill Step 1–2). Why: CommandCode is @ai-sdk/openai-compatible (api …/v1), but this file authors Anthropic-Messages-style controls (thinking.type toggle + output_config.effort without none). Same-surface openai-compat peers (e.g. OpenRouter/Baseten/Vercel) expose effort including none for off, not that Messages pair. Consumers get the wrong wire surface and a false toggle. Action: Replace with the openai-compat peer set (typically effort including none, no Messages output_config toggle), document the real CommandCode field, and keep inkling / inkling-small consistent on this host.
• [medium] [possible mistake] providers/commandcode/models/claude-sonnet-4-6.toml:1 - Check: Toggle/effort wire comments must match this host’s actual request path; do not mix Anthropic Messages and OpenAI-compat surfaces. Why: Sonnet 4.6 documents thinking.type + output_config.effort, while claude-sonnet-5.toml documents thinking.type + reasoning_effort on the same gateway. Those cannot both be accurate for one OpenAI-compatible API. Action: Verify CommandCode’s Claude request fields once, then make every Claude entry’s leading wire comments (and option shape) match that single surface.
• [medium] [possible mistake] providers/commandcode/models/google/gemini-3.7-flash.toml:11 - Check: Host pricing should be internally consistent across the same model family unless a documented host exception exists. Why: Only gemini-3.7-flash sets cache_write = 0.08334; the other CommandCode Gemini Flash entries that bill cache_read omit cache_write. That usually means a copy error or an undocumentated exception. Action: Confirm CommandCode’s Gemini cache-write pricing; add cache_write to the other Gemini entries that support it, or remove it from 3.7 and note why only that model differs.
• [low] [possible mistake] providers/commandcode/models/Qwen/Qwen3.8-Flash.toml:1 - Check: Provider-authored interleaved should reflect this host’s reasoning side-channel when peers/lab expose reasoning_content. Why: Qwen3.8-Max and Qwen3.8-Max-0902 set [interleaved].field = "reasoning_content", but Qwen3.8-Flash and Qwen3.8-27B omit it even though Alibaba’s first-party Qwen3.8 Flash/Max entries both publish that side-channel. Action: Confirm CommandCode responses for Flash/27B; add the same interleaved block if reasoning_content is present, or note why Max-only.

[github-actions]

No actionable findings.