feat: add AI Code With (aicodewith) provider

[DengZY123]

Adds AI Code With as a provider: an API gateway serving Claude, GPT, Gemini, Grok, DeepSeek, GLM and Kimi models.

• OpenAI-compatible endpoint: https://api.aicodewith.ai/v1 (Claude models use the Anthropic format, GPT models the OpenAI Responses format, per the provider's api_format)
• Env: AICODEWITH_API_KEY
• Docs: https://docs.aicodewith.ai/zh/docs
• 37 models, all using base_model. Context/output/modalities come from https://api.aicodewith.ai/models; pricing (per 1M tokens) comes from https://aicodewith.ai/zh/models. CNY-priced models are converted to USD, with the rate noted in each file.
• bun run validate passes.

[github-actions]

Action items

• [critical] [violation] providers/aicodewith/models/claude-opus-4-6.toml:4 - Check: Relay reasoning_options must copy lab + same-surface peer controls, never [] from uncertainty. Why: All Claude entries set reasoning_options = [] with # Reasoning controls exposed by this relay are not documented; left unverified, while lab is reasoning = true and first-party providers/anthropic/models/claude-opus-4-6.toml has effort ["low","medium","high","max"] + budget_tokens, claude-sonnet-5.toml/claude-haiku-5-5.toml have toggle + effort ["low","medium","high","xhigh","max"], and providers/openrouter/models/anthropic/claude-opus-4.6.toml has toggle + effort + budget_tokens. [] asserts no caller control, contradicting evidence bar. Same applies to all other [] Claude files (claude-fable-5-1.toml:4, claude-fable-5.toml:4, claude-opus-4-7.toml:4, claude-opus-4-8.toml:4, claude-opus-5.toml:4, claude-opus-5-5.toml:4, claude-sonnet-4-6.toml:4, claude-sonnet-5-5.toml:4, claude-haiku-5-5.toml:5). Action: Replace [] with verified forwarded controls per host docs/test or lab+peer intersection; do not ship uncertainty [].
• [critical] [violation] providers/aicodewith/models/gpt-5.5.toml:5 - Check: GPT relay effort baseline. Why: Sets [] unverified while models/openai/gpt-5.5.toml is reasoning = true and providers/openai/models/gpt-5.5.toml + providers/openrouter/models/openai/gpt-5.5.toml both use effort ["none","low","medium","high","xhigh"] with [[cost.tiers]] size 272_000. Same for gpt-5.6-sol.toml:4, gpt-5.6-terra.toml:4, gpt-6-astra.toml:4, gpt-6-sol.toml:4, gpt-6.1-sol.toml:4 (peers show ...,"max"] e.g. providers/openrouter/models/openai/gpt-5.6-sol.toml:4-6). Action: Author peer-matching effort with none (no toggle when none present) and document wire field, or provide affirmative no-control evidence.
• [critical] [violation] providers/aicodewith/models/gemini-2.5-pro.toml:5 - Check: Non-GPT relay baselines (Gemini/DeepSeek/Grok/GLM). Why: Same unverified [] while labs are reasoning = true: providers/google/models/gemini-2.5-pro.toml uses budget_tokens 128..32768, providers/google/models/gemini-3.6-flash.toml uses effort ["minimal","low","medium","high"], providers/deepseek/models/deepseek-v4-pro.toml/deepseek-v4-flash.toml use toggle + effort ["low","high","max"], providers/xai/models/grok-4.5.toml/grok-4.7.toml use effort ["low","medium","high"]/["low","medium","high","xhigh"]. Applies also to gemini-3-pro-preview.toml:5, gemini-3.1-pro-preview.toml:5, gemini-3.5-flash.toml:5, deepseek-flash.toml:4, deepseek-v4-pro.toml:5, grok-4.5.toml:4, grok-4.6.toml:4, glm-5.1.toml:5, glm-5.3.toml:5. Action: Replace with lab/peer-matching options for this host's actual request surface; budget_tokens only with real reasoning-budget field.
• [high] [possible mistake] providers/aicodewith/models/claude-haiku-4-5-20251001.toml:4 - Check: reasoning = false override must reflect real provider delta. Why: Lab models/anthropic/claude-haiku-4-5-20251001.toml:7 is reasoning = true; same contradiction for claude-sonnet-5.toml:4 (lab true), deepseek-v4-flash.toml:5 (labs deepseek-v4-flash/deepseek-v4.1-flash true), gemini-3.6-flash.toml:6/gemini-3.7-flash.toml:6/gemini-3.8-flash.toml:6 (labs true), glm-5.2.toml:6 (lab true), grok-4.7.toml:5 (lab true), kimi-k2.6.toml:5/kimi-k2.7-code.toml:5/kimi-k3.toml:5 (labs true, e.g. moonshotai/kimi-k3.toml:7). If false were correct, reasoning_options would be correctly omitted, but no evidence is given that this Anthropic/OpenAI-compatible relay disables reasoning. Action: Remove reasoning = false to inherit true and add proper reasoning_options, or cite provider evidence that this host serves a non-reasoning variant.
• [high] [violation] providers/aicodewith/models/claude-haiku-5-5.toml:8 - Check: Context-based pricing must be [[cost.tiers]], not a comment. Why: Comment notes Above 100000 input tokens, higher long-context rates apply (input 0.5 / output 2.5) without tiers, while first-party providers/anthropic/models/claude-haiku-5-5.toml:20-24 encodes tier size 100_000. Same pattern in gpt-5.6-sol.toml:3/gpt-5.6-terra.toml:3/gpt-6-astra.toml:3/gpt-6-sol.toml:3/gpt-6.1-sol.toml:3 (272k tiers; cf. providers/openrouter/models/openai/gpt-5.5.toml:12-16, gpt-5.6-sol.toml:14-18) and grok-4.5.toml:3/grok-4.6.toml:3/grok-4.7.toml:3 (200k tiers; cf. providers/xai/models/grok-4.5.toml:12-16, grok-4.7.toml:13-17). Action: Encode noted higher bands as [[cost.tiers]] with input/output (+ cache where peer shows it) or remove the claim.
• [medium] [violation] providers/aicodewith/models/claude-fable-5-1.toml:16 - Check: Override-only provider files after base_model. Why: [limit] context = 1_000_000 output = 128_000 is identical to models/anthropic/claude-fable-5-1.toml:16-17 (same for claude-fable-5.toml:15-18 vs models/anthropic/claude-fable-5.toml:14-15); identical tables must be omitted, deltas only. Action: Delete the redundant [limit] block (keep genuine [modalities] delta if verified, and cost/reasoning_options).

[github-actions]

Action items

• [high] [violation] providers/aicodewith/models/claude-haiku-5-5.toml:1 - Check: Every toggle requires a leading top-of-file exact wire path. Why: All toggle files (claude-haiku-5-5, claude-sonnet-5, deepseek-flash, deepseek-v4-flash, deepseek-v4-pro, glm-5.1, kimi-k2.6) carry only a generic mirror providers/... (relay forwards the upstream request body) comment with no wire field, so on/off control cannot be verified. Action: Add the exact per-surface wire comment (e.g. # Toggle: thinking.type = enabled|disabled) for each host format, or remove toggle if forwarding is unverified.
• [high] [violation] providers/aicodewith/models/gemini-3-pro-preview.toml:5 - Check: Relay baseline must copy lab + same-surface peers; [] means affirmative no caller control, not uncertainty. Why: Lab google/gemini-3-pro-preview has reasoning = true with no Google provider entry, established relay providers/merge-gateway/models/google/gemini-3-pro-preview.toml exposes effort [low, high], but this file copies providers/anyapi/... [] with only a generic mirror comment. Action: Replace [] with the peer-supported effort set with host evidence, or provide affirmative proof this relay exposes no control.

[DengZY123]

Thanks for the review — all six items addressed in 57fc671:

1–3. Replaced every reasoning_options = [] with the controls from the corresponding first-party provider file, or a same-model peer where none exists (anyapi for gemini-3-pro-preview, 302ai for deepseek-flash). AICodeWith is a pass-through relay, so upstream request fields are forwarded unchanged.
4. Dropped the reasoning = false overrides; the supports_reasoning flag in our /models endpoint is inaccurate for several models and is being fixed, so entries now inherit reasoning = true.
5. Encoded long-context pricing as [[cost.tiers]] from https://aicodewith.ai/zh/models (Haiku 5.5 >100K, GPT-5.6/6.x >272K, Grok >200K).
6. Removed [limit] tables that match the base model; the remaining ones are real deltas.

bun run validate passes.

[DengZY123]

Follow-up for the latest CI highs in 57fc671:

1. Added leading # Toggle: thinking.type = … wire comments on every toggle model (claude-haiku-5-5, claude-sonnet-5, deepseek*, glm-5.1, kimi-k2.6), matching the Anthropic / DeepSeek / Zhipu / Moonshot request surfaces this relay forwards.
2. Replaced gemini-3-pro-preview empty reasoning_options with effort ["low","high"] from providers/merge-gateway/models/google/gemini-3-pro-preview.toml (anyapi's [] was uncertainty, not affirmative no-control).

bun run validate passes.

[github-actions]

Action items

• [high] [possible mistake] providers/aicodewith/models/kimi-k2.6.toml:1 - Check: Toggle wire path must match this host/lab API per AGENTS.md Reasoning options and audit skill. Why: File claims # Toggle: thinking.type = enabled|disabled, but same lab's providers/moonshotai/models/kimi-k3.toml states thinking.type is not supported and uses reasoning_effort = low|high|max; uniform thinking.type across Kimi/GLM files suggests unverified copy-paste, and wrong wire breaks toggle callers. Action: Verify Kimi toggle field on api.aicodewith.ai via host docs or live test and correct/remove toggle; cite wire source.
• [high] [possible mistake] providers/aicodewith/models/glm-5.1.toml:1 - Check: Toggle wire path must match lab/host API. Why: File claims thinking.type = enabled|disabled, but providers/zhipuai/models/glm-5.1.toml and glm-5.2.toml expose toggle/effort via reasoning_effort/reasoning_content with no thinking.type (DeepSeek/Anthropic use thinking.type, Zhipuai does not); relay forwarding claim alone does not prove this wire. Action: Verify actual GLM on/off field on this relay and fix comment/toggle or drop toggle if binary off is unavailable.
• [medium] [possible mistake] providers/aicodewith/models/grok-4.5.toml:12 - Check: [[cost.tiers]] size must be reachable within [limit]. Why: Final keeps [limit] context = 200_000 with tier = {type="context", size=200_000} (same in grok-4.6.toml, grok-4.7.toml); tier at exactly max context never triggers, while lab models/xai/grok-4.5.toml is 500_000 with tier 200_000 and peer providers/xai/models/grok-4.5.toml has no limit.context cap, indicating the 200_000 cap is likely a copied catalog max_tokens rather than context. Action: Verify relay context vs tier threshold from pricing/catalog and either inherit lab 500_000 context or correct/remove tier.
• [medium] [possible mistake] providers/aicodewith/models/gpt-5.5.toml:12 - Check: Provider [limit] overrides must be real deltas, not stale copies. Why: File keeps context = 400_000 vs lab models/openai/gpt-5.5.toml 1_050_000, while sibling relay files gpt-5.6-sol/terra, gpt-6-astra/sol correctly inherit 1_050_000 and add 272_000 tiers; isolated 400_000 matches older GPT-5 window and leaves no tier where peers have one. Action: Verify 400_000 against https://api.aicodewith.ai/models for this ID and drop [limit] if relay serves full 1_050_000 (adding 272_000 tier if long-context pricing applies).

[DengZY123]

Addressed the latest review in 1083842.

1. kimi-k2.6 toggle: Moonshot docs list thinking.type = enabled|disabled for kimi-k2.6 and mark reasoning_effort as "Not supported" (models-overview, thinking models); providers/moonshotai/models/kimi-k2.6.toml also uses toggle. Live test on api.aicodewith.ai (2026-10-13): every variant (default, thinking enabled/disabled, reasoning_effort none/low/high, plus /v1/messages) returned 503 "model unavailable", so it could not be re-verified on the relay. Kept toggle per lab wire and added the doc citation and the test note to the file.
2. glm-5.1 toggle: verified live on /v1/chat/completions. thinking.type=disabled -> 0 reasoning tokens, no reasoning_content; enabled/omitted -> reasoning_content with 120-195 reasoning tokens. reasoning_effort none/low/high still reasoned (116-160 tokens), so it is not an off switch. Kept toggle and added the result to the comment.
3. grok-4.5/4.6/4.7 context: the relay pricing data publishes long_context_threshold: 200000 (input 4 / output 12) for all three, so requests above 200K are billed. The /models catalog says context_window: 200000, which conflicts with that. Removed the context = 200_000 override so context inherits the xai lab value (500K) and the tier can apply; kept output 65_536.
4. gpt-5.5 context: https://api.aicodewith.ai/models still reports context_window: 400000, and the pricing data for gpt-5.5 has no long_context_pricing (gpt-5.6-sol has threshold 272000). Kept context 400_000 with no tier, and added input = 272_000 (context minus output, same split as the openai lab) because the inherited 922_000 was larger than the context.

bun validate passes locally.

[github-actions]

No actionable findings.