Skip to content

Add HopBase provider - #12200

Open
victor-zzh wants to merge 20 commits into
anomalyco:devfrom
victor-zzh:add-hopbase
Open

victor-zzh wants to merge 20 commits into
anomalyco:devfrom
victor-zzh:add-hopbase

Conversation

@victor-zzh

@victor-zzh victor-zzh commented Oct 31, 2026 •

Copy link
Copy Markdown

Adds HopBase (https://hop-base.com), an OpenAI-compatible API gateway for chat, image and video models.

Disclosure: I work on HopBase.

  • npm = "@ai-sdk/openai-compatible", opencodeai = "@opencode/ai/providers/openai-compatible"
  • api = "https://api.hop-base.com/v1", env = ["HOPBASE_API_KEY"], doc = "https://hop-base.com/en/docs"
  • Monochrome currentColor logo, square viewBox.
  • Every model file uses base_model and only overrides what this host does differently.
  • [cost] is filled for every model (USD) with the price any registered account pays in the public group that serves it; no contract or volume deal is assumed. The 9 headline text prices match the self-serve table at https://hop-base.com/pricing (also in https://hop-base.com/llms.txt). Doubao Seed prices are HopBase's own USD prices, as published on the public pricing endpoint (https://api.hop-base.com/api/v1/models/pricing).
  • DeepSeek V4.1 Flash 1017 (and -sg) bills off-peak at half price. The base price is the off-peak rate, and a [[cost.multipliers]] factor of 2 covers the weekday peak windows (UTC 01:00-04:00 and 06:00-10:00), matching the billing logs. DeepSeek V4 Pro has a single all-day rate.
  • Long-context tiers are included where this host bills them (GPT-5.6 Sol above 272K, Grok above 200K, Doubao Seed Character and Seed 2.0 Mini above 32K).
  • Media: kling-v4 per second (1080p override), dreamina-seedance-2-5-260628 per 1M video tokens (resolution and input_video overrides, same shape as the Vercel entry), seedream-5-0-pro per image (override above 2.61 MP; reference images as image input).

Model catalogue: https://hop-base.com/models. GET /v1/models needs a key; the same IDs are also listed without auth at https://api.hop-base.com/api/v1/models/pricing.

Models

Chat (/v1/chat/completions):

  • deepseek-v4.1-flash-1017-sg: DeepSeek V4.1 Flash 1017 from a Singapore data center
  • doubao-seed-2-1-pro-260915
  • doubao-seed-2-1-lite-260915
  • doubao-seed-character-260628
  • doubao-seed-2-1-pro-260628
  • doubao-seed-2-1-turbo-260628
  • doubao-seed-2-0-mini-260428
  • doubao-seed-evolving
  • grok-4.20-multi-agent-0309: tools off (client-side tools need xAI beta access)
  • deepseek-v4.1-flash-1017
  • deepseek-v4-pro-202606: HopBase ID for DeepSeek V4 Pro 0813
  • gpt-5.6-sol
  • gemini-3.8-flash
  • gemini-3.1-pro-preview
  • grok-4.7
  • kimi-k3
  • glm-5.3

Anthropic Messages (/v1/messages, per-model [api.messages] override with @ai-sdk/anthropic, because HopBase serves Claude over Messages only):

  • claude-fable-5-2: verified live 2026-11-10 (streaming usage, auto and strict tool use, tool_choice any rejected as on Anthropic, all five effort levels, image input, explicit cache write/read). Priced at the public "Claude Max (ccmax)" group rate.

Media (async/sync task endpoints, aisdk/opencodeai set accordingly):

  • kling-v4 ([api.videos], path = "/video/generate", async task + poll)
  • dreamina-seedance-2-5-260628 ([api.videos], same task API)
  • seedream-5-0-pro ([api.images], OpenAI-compatible /v1/images/generations)

New lab files, because nothing existed to point base_model at:

  • models/bytedance-seed/seed-2.1-pro-0915.toml: 260915 snapshot, 1M context and 256K output. The 256K-context 260628 snapshot keeps seed-2.1-pro.
  • models/bytedance-seed/seed-2.1-lite.toml
  • models/kuaishou/kling-v4.toml: v2_only, the same shape as kling-v3.0-pro.
  • models/xai/grok-4.20-multi-agent-0309.toml: copied from the xAI provider entry so relays can reference it. The xAI provider file is unchanged.

How the models were verified

All checks ran live on 2026-10-31 against https://api.hop-base.com/v1/chat/completions, using each exact model ID and a regular customer API key. Each file's header comment has the details.

  • Basic calls: non-streaming, streaming with stream_options.include_usage, and a function-tool call for every chat model.
  • AI SDK path: @ai-sdk/openai-compatible 1.0.60 with ai 5 ran generateText, streamText, and a multi-step tool round trip (stopWhen). This passed on deepseek-v4.1-flash-1017-sg, doubao-seed-2-1-pro-260915, glm-5.3, kimi-k3, gpt-5.6-sol and grok-4.7.
  • Reasoning: every documented reasoning_effort level was tried, plus an invalid one. Toggles were tried where they exist (thinking.type). The files list only the levels the API accepts. Specifics:
    • DeepSeek and GLM: toggle plus native low/high/max.
    • Doubao Seed: none through max. Seed Character is the exception: toggle plus minimal/low/high, because a graded effort sent without thinking enabled returns 400.
    • GPT-5.6 Sol: max is rejected with 400, so it is left out.
    • Grok 4.7: none and max are rejected.
    • Gemini: low/medium/high, mapped to thinking budgets on this host.
    • Kimi K3: always-on thinking with no caller control (reasoning_effort has no effect).
    • reasoning_field = "reasoning_content" is set only where the field actually came back.
  • tool_choice: none, required and a named tool were each tested. choice lists only the values that were honoured. For example, Seed 2.1 Pro 0915, Seed Evolving and Gemini 3.8 Flash ignore it, so they get ["auto"].
  • Other capabilities:
    • Image input: a red PNG data URL.
    • structured_output: json_schema with strict: true and an enum.
    • Implicit cache: a repeated prefix of about 2.5K tokens had to come back with cached_tokens > 0.
    • Output caps: max_tokens at the documented limit and limit + 1.
    • Text-only models (DeepSeek V4 Pro, GLM-5.3) return 400 image_input_not_supported on image input.
  • GPT-5.6 Sol freeform: a custom tool with regex grammar ZZ[0-9]{3} returned ZZ000.
  • Media models: no generations were run. The IDs were checked against the public catalogue only.

bun validate passes. bun run v1:check reports only additions: the new provider, its 18 v1-visible models and the 3 new non-v2_only lab entries, 21 in total. No existing v1 field changes. kling-v4 and seedream-5-0-pro are v2_only.

@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [high] [violation] providers/hopbase/models/kimi-k3.toml:26 - Check: Reasoning controls baseline for a relay (AGENTS.md + audit skill Step 2). Why: moonshotai/kimi-k3 native entry (providers/moonshotai/models/kimi-k3.toml:34) authors effort = ["low", "high", "max"] (always-thinking with caller control); this relay declares reasoning = { supported = true } with no controls, i.e. no caller control, with only “same behaviour for low/max/unknown, not benchmarked” as evidence. Leaving controls out from uncertainty is explicitly forbidden on relays. Action: Author the lab/peer set low/high/max if the host forwards reasoning_effort, or provide affirmative host evidence that HopBase pins/ignores effort and truly offers no control and why native controls do not apply.
  • [high] [possible mistake] providers/hopbase/models/glm-5.3.toml:21 - Check: Toggle baseline vs first-party lab. Why: Native providers/zhipuai/models/glm-5.3.toml:22 is effort-only low/high/max with comment “thinking.type only accepts enabled” (no toggle); this file adds toggle = true claiming thinking.type = disabled -> 0 reasoning tokens, an extra control beyond lab/peers that contradicts native docs. Extra controls need host docs or live meaningful effect. Action: Verify disabled truly disables thinking on this host (wire docs/log) or remove toggle to match low/high/max only.
  • [medium] [possible mistake] providers/hopbase/models/deepseek-v4.1-flash-1017.toml:33 - Check: Cost consistency for regional variant. Why: deepseek-v4.1-flash-1017.toml (price = 0.08825, cache.read = 0.00295) and deepseek-v4.1-flash-1017-sg.toml (price = 0.08825, cache.read = 0.0018) share input/output prices but cache differs by ~65% (3.3% vs 2.0% of input); the -sg ratio matches native DeepSeek Flash 2% (0.003/0.15), the non-sg matches Pro 3.3%, suggesting a typo. Action: Verify the correct cache.read for both IDs against HopBase pricing/billing and fix the erroneous file.
  • [medium] [possible mistake] providers/hopbase/models/deepseek-v4.1-flash-1017.toml:39 - Check: Time-of-day multipliers must match host billing. Why: Native DeepSeek (providers/deepseek/models/deepseek-v4-flash.toml:65) uses not = { holiday = { equals = "CN" } } to exclude Chinese holidays from peak doubling; both HopBase Flash files omit the exclusion, and deepseek-v4-pro-202606.toml omits the peak multiplier entirely while native Pro has one, contradicting the “single all-day rate” assertion without billing evidence. Action: Confirm whether HopBase bills peak on CN holidays and whether Pro is truly flat, and add the not holiday exclusion / Pro multiplier if the host follows upstream.
  • [medium] [possible mistake] providers/hopbase/models/doubao-seed-2-0-mini-260428.toml:39 - Check: Context-tier thresholds and completeness (USD cost overrides). Why: Volcengine authoritative entries use when = { context = { gt = 32_000 } } (and a second tier gt = 128_000 for this model); HopBase uses gt = 32_768 and declares only one tier, while claiming prices “equal the official Volcano Engine CNY list”. The 768-token shift and missing 128_000 tier change billed prices vs the cited source. Same 32_768 vs 32_000 shift appears in doubao-seed-character-260628.toml:42. Action: Verify HopBase’s actual tier boundaries from pricing/billing logs, fix to 32_000/128_000 if following upstream, or document why HopBase bills at 32_768 with a single tier and add the missing second tier if it exists.
  • [medium] [possible mistake] providers/hopbase/models/grok-4.7.toml:32 - Check: Long-context override operator. Why: Native xAI (providers/xai/models/grok-4.7.toml:43, grok-4.20-multi-agent-0309.toml:53) uses when = { context = { gte = 200_000 } } (“prompts ≥200k”); HopBase uses gt = 200_000 in both grok-4.7.toml and grok-4.20-multi-agent-0309.toml:24, moving exactly-200k prompts to the base tier. Action: Verify whether HopBase tiers start at > vs ≥ 200k and align to gte if following upstream.
  • [low] [possible mistake] providers/hopbase/models/gpt-5.6-sol.toml:19 - Check: GPT-5+ freeform formats must be probed per API. Why: Native OpenAI declares freeform = ["text", "lark", "regex"]; this file declares only ["regex"] with header “lark and text grammars were not tested”, so omission reflects lack of testing, not proven rejection, understating capability. Action: Probe lark/text with a restrictive grammar proving enforcement (e.g. ZZ[0-9]{3}) and add formats that work, or keep ["regex"] only with explicit host evidence that other grammars are rejected.

@victor-zzh

Copy link
Copy Markdown
Author

Thanks for the review. All items were re-checked on 2026-10-31 against the live API and HopBase's billing logic; evidence is in each file's header.

  • kimi-k3 (no controls): HopBase does not apply reasoning_effort on this model. low, high, max, none and an invalid value (banana) all return 200, and reasoning tokens do not move with the level (two runs each on the same prompt: low 380/625, high 449/459, max 210/423, omitted 495/506). Thinking can't be turned off either, so no controls are declared.
  • glm-5.3 toggle: removed. A retest showed thinking.type = disabled still returns reasoning (182 tokens), so it now matches the lab: low/high/max only.
  • DeepSeek 1017 vs -sg cache read: both are correct. The non-sg ID's group bills cache reads at the official cache price, while the -sg group discounts cache reads too.
  • Holidays / V4 Pro: HopBase's peak windows are weekday-only with no CN-holiday exception, so not holiday is intentionally omitted. V4 Pro is flat on this host: 3,037 requests over the last 14 days were all billed at one price across peak and off-peak hours.
  • Doubao tiers: HopBase bills a single long-context tier above 32,768 input tokens (long_context.threshold = 32768 in the public pricing API) with no 128K tier. The files follow HopBase billing, and the header no longer claims Volcano's tier table.
  • Grok gt vs gte: HopBase applies the long-context price only when input is strictly greater than 200,000 tokens, so gt is intentional.
  • gpt-5.6-sol freeform: lark (grammar start: "ZZ" DIGIT DIGIT DIGIT) and regex both forced ZZ000 when asked for 'hello world', and text returns free text. All three are now declared.

@github-actions

Copy link
Copy Markdown
Contributor

No actionable findings.

@github-actions github-actions Bot added the reviewer: ready Automated review found no actionable items label Oct 31, 2026
HopBase bills gemini-3.1-pro-preview at a long-context tier above 200K input
tokens (input and cache read x2, output x1.5), as published in its pricing
catalog and llms.txt. The entry listed a single tier, understating long
requests.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@github-actions github-actions Bot removed the reviewer: ready Automated review found no actionable items label Nov 5, 2026
@github-actions

github-actions Bot commented Nov 5, 2026

Copy link
Copy Markdown
Contributor

No actionable findings.

@github-actions github-actions Bot added the reviewer: ready Automated review found no actionable items label Nov 5, 2026
@github-actions github-actions Bot removed the reviewer: ready Automated review found no actionable items label Nov 10, 2026
@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [critical] [violation] providers/hopbase/models/claude-fable-5-2.toml:15 - Check: Non-lab hosts must use base_model pointing to an existing models/<lab>/<model>.toml, adding the lab file when missing. Why: This file points at anthropic/claude-fable-5-2 but the PR adds no models/anthropic/claude-fable-5-2.toml and none exists on base (only claude-fable-5 / claude-fable-5-1), so validation fails with a missing target. Action: Add a complete models/anthropic/claude-fable-5-2.toml lab file and keep this provider file override-only.
  • [medium] [violation] providers/hopbase/models/claude-fable-5-2.toml:24 - Check: [api.messages] sdk.opencodeai must use the generic Messages entry for relays. Why: This relay sets opencodeai = "@opencode/ai/providers/anthropic" (first-party Anthropic entry) while all same-surface relay peers (e.g. providers/vivgrid/models/claude-fable-5-1.toml, providers/neon, providers/azure, providers/openrouter) use "@opencode/ai/providers/anthropic-compatible"; per AGENTS.md a host without its own entry must use the generic protocol entry. Action: Change to sdk = { aisdk = "@ai-sdk/anthropic", opencodeai = "@opencode/ai/providers/anthropic-compatible" }.

@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [high] [violation] providers/hopbase/models/claude-fable-5-2.toml:24 - Check: [api.messages] sdk.opencodeai must use the generic protocol entry when the host has no own entry. Why: File sets opencodeai = "@opencode/ai/providers/anthropic" (first-party Anthropic entry) while peer relays on Messages (providers/azure/models/claude-fable-5-2.toml:25, providers/databricks/models/databricks-claude-opus-4-5.toml:29, providers/modal/models/moonshotai/Kimi-K3.toml:36) use "@opencode/ai/providers/anthropic-compatible" per AGENTS.md. Action: Change to sdk = { aisdk = "@ai-sdk/anthropic", opencodeai = "@opencode/ai/providers/anthropic-compatible" }.
  • [medium] [violation] providers/hopbase/models/claude-fable-5-2.toml:20 - Check: Declared cache must match header evidence. Why: Header states explicit cache_control wrote 6,026 tokens and also “cached a fresh 14,441-token prefix sent without cache_control (written, then read on repeat)” — i.e. implicit caching observed — but file declares only cache = { explicit = ... } with no implicit. Action: Add the observed implicit kind (e.g. cache = { implicit = true, explicit = ... } with verified details) or remove/clarify the implicit-caching claim if it was not HopBase billing cache.
  • [medium] [possible mistake] providers/hopbase/models/deepseek-v4-pro-202606.toml:20 - Check: Sampling caps inherited via base_model must be narrowed where the Chat Completions surface lacks them. Why: Lab deepseek/deepseek-v4-pro-0813 has top_k = true, but native providers/deepseek/models/deepseek-v4-pro.toml:40 sets top_k = false for Chat Completions (“no top_k”); HopBase serves the same Chat Completions surface yet inherits true. Same applies to deepseek-v4.1-flash-1017.toml, deepseek-v4.1-flash-1017-sg.toml, and glm-5.3.toml (native providers/zhipuai/models/glm-5.3.toml:24 sets top_k = false). Action: Verify top_k live on HopBase for DeepSeek/GLM and set top_k = false in [capabilities] where rejected/ignored.
  • [low] [possible mistake] providers/hopbase/models/deepseek-v4.1-flash-1017.toml:28 - Check: Time-of-day [[cost.multipliers]] holiday handling needs evidence. Why: Native providers/deepseek/models/deepseek-v4-pro.toml:64 excludes holiday = CN from peak, but HopBase flash multipliers omit the exclusion while header asserts “no holiday exception” without cited billing-log evidence for a holiday (same for -sg file). Action: Provide holiday billing evidence or add not = { holiday = { equals = "CN" } } to match the lab billing calendar.

…or Claude Fable 5.2

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [medium] [violation] providers/hopbase/models/claude-fable-5-2.toml:20 - Check: Cache declaration must list every cache kind the host supports. Why: Header reports both explicit (cache_control wrote/read 6,026) and implicit (14,441-token prefix without cache_control written then read) hits, but cache declares only explicit, hiding observed implicit caching. Action: Add implicit alongside explicit if reproducible, or correct the header if the no-marker hit was incidental.

…ude Fable 5.2

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [low] [possible mistake] providers/hopbase/models/gemini-3.8-flash.toml:15 - Check: Cache support must reflect this host's actual behavior, and the cited evidence must be able to detect caching. Why: The file leaves cache out (published as no caching) based on "No cache hit was observed on a repeated ~3K-token prefix", but Google's implicit cache has a documented 4,096-token minimum (the native providers/google/models/gemini-3.8-flash.toml records implicit = { min_tokens = 4_096 }). A ~3K-token repeat can never hit, so the probe cannot distinguish "host doesn't cache" from "prefix below the minimum" — the absent-cache claim is unsupported. gemini-3.1-pro-preview.toml also declares no cache with no probe at all. Action: Repeat the cache probe on both models with a ≥4,096-token repeated prefix; declare cache = { implicit = … } if cached_tokens appears, otherwise keep it absent and replace the comment with the sufficient probe.

  • [low] [violation] providers/hopbase/models/claude-fable-5-2.toml:23 - Check: Provider model files must stay override-only — [api.<protocol>] values left out are filled from the provider (base_url from api). Why: base_url = "https://api.hop-base.com/v1" restates providers/hopbase/provider.toml's api verbatim; omitting it produces the identical resolved URL, so the line is a redundant restatement of provider data, not a delta. Action: Drop the base_url line from [api.messages] (or set it only if the Messages endpoint actually differs from the provider's api).

@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [medium] [possible mistake] providers/hopbase/models/gemini-3.8-flash.toml:13 (also providers/hopbase/models/gemini-3.1-pro-preview.toml:13) - Check: capabilities.cache is omitted for both Gemini models while their cost tables bill cache reads (0.0342 / 0.0912, doubled above 200K for 3.1 Pro, per the files' own comments). Why: The only supporting evidence — "No cache hit was observed on a repeated ~3K-token prefix" — is below Google's documented 4,096-token implicit-cache minimum (implicit = { min_tokens = 4_096 } in the first-party entries), so it cannot distinguish "no caching" from "below the minimum". The first-party host and peer relays (e.g. openrouter) declare implicit caching for these models, and a host that prices cache reads almost certainly serves them; every other HopBase model showed cached_tokens at ~2.4–2.6K prefixes. Action: Re-test with a repeated ≥4,096-token prefix; if cached_tokens appears, declare cache = { implicit = true } (noting the minimum), otherwise document in the header why the host bills cache reads it never serves.
  • [low] [violation] providers/hopbase/models/doubao-seed-2-0-mini-260428.toml:14 - Check: AGENTS.md requires prices converted from non-USD list prices to note the conversion rate and date in a top-of-file comment. Why: The seven doubao files' prices are conversions of Volcano Engine's CNY list (each equals CNY/6.8 rounded up, as the PR body states), but the final header comments say only "follows the official Volcano Engine list price" — the "converted at 6.8 CNY/USD (2026-10-31)" note present in the first commit was removed by the later "simplify doubao pricing comments" commits, leaving the conversion undocumented in the files. Action: Restore the CNY→USD rate and date in the header comment of all seven doubao files (doubao-seed-2-0-mini-260428, doubao-seed-2-1-lite-260915, doubao-seed-2-1-pro-260628, doubao-seed-2-1-pro-260915, doubao-seed-2-1-turbo-260628, doubao-seed-character-260628, doubao-seed-evolving).
  • [low] [possible mistake] providers/hopbase/provider.toml:1 - Check: Sync module coverage for a new provider (AGENTS.md "Sync modules (recommended)"). Why: HopBase exposes a public, unauthenticated machine-readable catalog (https://api.hop-base.com/api/v1/models/pricing, cited throughout the PR) listing every model ID with token prices, cache prices and long-context thresholds — rich enough to authoritatively own [cost.*] and detect models no longer served, which is exactly what would keep this 24-model catalog current as the host changes prices or drops IDs. Action: Add a sync module per sync.md, or note in the PR why the catalog stays hand-authored.

@victor-zzh

Copy link
Copy Markdown
Author

Thanks — addressed the three items:

  1. Gemini caching: re-tested with a repeated ~11K-token prefix. Both gemini-3.8-flash and gemini-3.1-pro-preview returned usage.prompt_tokens_details.cached_tokens ≈ 9.15K from the second or third call, so both files now declare cache = { implicit = { min_tokens = 4_096 } } and the header notes the new probe (2be785f, 6d89f60).
  2. Doubao prices: these are HopBase's own USD list prices, as published on the public pricing endpoint (https://api.hop-base.com/api/v1/models/pricing), not a conversion made for this catalog, so there is no conversion rate to document. I corrected the PR description, which wrongly described them as a conversion.
  3. Sync module: keeping the catalog hand-authored for this first PR so every entry stays live-verified; happy to add a sync module against the public pricing endpoint in a follow-up if maintainers prefer.

@github-actions

Copy link
Copy Markdown
Contributor

No actionable findings.

@github-actions github-actions Bot added the reviewer: ready Automated review found no actionable items label Nov 15, 2026
@github-actions github-actions Bot removed the reviewer: ready Automated review found no actionable items label Nov 16, 2026
@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [medium] [violation] providers/hopbase/models/kimi-k3.1-pro.toml:30 - Check: Override-only provider files after base_model (AGENTS.md: never restate matching [limit] data). Why: [limit] output = 131_072 is byte-identical to the base lab file models/moonshotai/kimi-k3.1-pro.toml (output = 131_072), so it is a copied inherited value, not a host delta. Action: Delete the [limit] section and let the output limit inherit from the lab file.

  • [medium] [possible mistake] providers/hopbase/models/kimi-k3.1-pro.toml:1 - Check: A model's [api.<protocol>] tables are its complete list of APIs. Why: The header documents live round trips for this exact ID on POST /v1/responses (function call + function_call_output) and POST /v1/messages (tool_use + tool_result), but the file declares only [api.chat-completions], so the published API list omits two surfaces the PR's own evidence says work (the claude-fable entry shows the author knows the tables define the surface). Action: Add [api.responses] and [api.messages] tables (a messages table names another aisdk, so it must carry its own opencodeai, e.g. @opencode/ai/providers/anthropic-compatible), or remove the endpoint claims from the header if they are not part of the offering for this model.

  • [medium] [possible mistake] providers/hopbase/models/claude-fable-5-2.toml:21 - Check: Cache TTL list and write prices must cover what the host actually bills. Why: Only ttl = ["5m"] with a single 5m write price (2.8) is declared, but the cited Anthropic pricing page and the first-party providers/anthropic/models/claude-fable-5-2.toml support a 1h TTL (official $16 write, ≈4.48 at the 0.28 group rate), and the header's own test shows the gateway forwards cache_control (default 5m); no probe of ttl = "1h" is documented. If 1h passes through, the entry is missing a real billing path. Action: Probe cache_control with ttl = "1h"; if it is accepted, add "1h" to ttl and the 1h write price; otherwise document in the header that HopBase rejects/ignores 1h caching for this model.

  • [low] [violation] providers/hopbase/models/doubao-seed-2-0-mini-260428.toml:14 (same pattern in all seven doubao files) - Check: Prices converted from a non-USD list must note rate/date in a top-of-file comment. Why: The doubao prices are the Volcano Engine CNY list ÷ 6.8 rounded up (0.0295 ≈ ¥0.2, 0.2942 ≈ ¥2, 0.8824 ≈ ¥6, 4.4118 ≈ ¥30 per MTok), and the initial commits documented "converted at 6.8 CNY/USD (2026-10-31)", but the later "simplify" commits stripped the rate/date, leaving only "follows the official Volcano Engine list price". Action: Restore the conversion note (6.8 CNY/USD, 2026-10-31) in the doubao headers, or state explicitly that these are HopBase-published USD prices rather than conversions.

  • [low] [possible mistake] providers/hopbase/models/doubao-seed-2-0-mini-260428.toml:25 (all seven doubao entries) - Check: cache must reflect the host's caching behavior; a kind that isn't listed is published as unsupported. Why: No doubao entry declares cache and no header documents a caching probe, while every other HopBase model on the same gateway documents a cached_tokens hit and declares cache = { implicit = true } — so these entries publish doubao as non-caching with no evidence either way. Action: Run a repeated-prefix probe on the doubao IDs and either declare cache = { implicit = true } (with details if known) or document the negative result in the headers.

# Conflicts:
#	models/bytedance-seed/seed-2.1-lite.toml
@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [high] [violation] models/bytedance-seed/seed-2.1-lite.toml:1 - Check: Lab files must not duplicate or contradict existing models/ entries. Why: dev already contains this exact lab file (sources accessed 2026-11-10) documenting audio input (audio/mpeg, audio/wav, audio/aac, audio/x-m4a from Ark's audio-understanding table) and limits 1_024_000/256_000. The PR adds it as a new file (merge conflict) that drops the audio modalities and uses 1_048_576/262_144 — merging it would regress newer, better-sourced dev data. Action: Rebase onto dev and drop the duplicate lab file; point providers/hopbase/models/doubao-seed-2-1-lite-260915.toml at the existing bytedance-seed/seed-2.1-lite, and add a [limit] output = 262_144 override there if the live max_tokens=262145 → 400 probe is to be kept, citing it in the header.
  • [high] [violation] models/kuaishou/kling-v4.toml:1 - Check: A new lab file must not create a second identity for a model dev already covers. Why: dev has models/kuaishou/kling-v4.0.toml for the same Kling 4.0 release (2026-10-15, API ID kling-4.0), sourced from kling.ai's own 4.0 API docs, with video input (video/mp4, video/quicktime) and omni-reference support. The PR's file describes the same model from a press release plus the Kling 3.0 imageToVideo doc, understates input modalities, and forks the lab ID. Action: Delete models/kuaishou/kling-v4.toml; in providers/hopbase/models/kling-v4.toml set base_model = "kuaishou/kling-v4.0" and override [modalities] input to this host's narrower surface (text + first-frame image), as the header already documents.
  • [high] [possible mistake] providers/hopbase/models/doubao-seed-2-1-pro-260628.toml:4 - Check: Resolved limits after the base_model merge must match the limits the file itself documents. Why: The header says 256K context (per hop-base.com/llms.txt), but the file overrides only output and inherits context = 1_024_000 from bytedance-seed/seed-2.1-pro, which on current dev represents the 260915 snapshot (1M context) — the "256K" reading of the undated lab file in the new seed-2.1-pro-0915 header is stale. The merged entry would publish a context ~4× larger than documented; bun validate will not catch this. Action: Add context = 262_144 to [limit] (or point at a dedicated 260628 lab file), or correct the header if HopBase now serves 1M for this ID.
  • [high] [possible mistake] providers/hopbase/models/doubao-seed-evolving.toml:4 - Check: Resolved limits after the base_model merge must match the file's documented limits. Why: Same defect: the header claims 256K context, but models/bytedance-seed/seed-evolving.toml was updated on dev (2026-11-10) to context = 1_024_000 ("was 256k when the file was written"), and this file overrides only output. The published context would contradict the file's own source. Action: Add the context override for this host's documented cap, or update the header to what HopBase actually serves.
  • [medium] [violation] models/bytedance-seed/seed-2.1-pro-0915.toml:1 - Check: One lab model cannot have two lab entries with contradictory data. Why: dev's undated bytedance-seed/seed-2.1-pro already represents the doubao-seed-2-1-pro-260915 snapshot (its 2026-11-10 sources say so) with limits 1_024_000/256_000; this new dated file describes the same snapshot with 1_048_576/262_144. After merge, the catalog holds two disagreeing lab entries for one snapshot. Action: Reconcile with dev: keep a single identity for the 260915 snapshot (or make the undated and dated files agree, with Ark sources for whichever limits are kept), and point doubao-seed-2-1-pro-260915.toml at it.
  • [medium] [violation] providers/hopbase/models/dreamina-seedance-2-5-260628.toml:14 - Check: Async video-task models stay out of v1 (v2_only = true), as AGENTS.md and same-model peers require. Why: Both existing provider entries for this model (providers/volcengine/models/doubao-seedance-2-5-260628.toml, providers/byteplus/models/dreamina-seedance-2-5-260628.toml) and the PR's own kling-v4.toml set v2_only = true; this file omits it, so a model with sdk = { aisdk = "none", opencodeai = "none" } and no token prices would enter the published v1 API under the chat SDK default. Action: Add v2_only = true, or justify v1 publication in the header.
  • [medium] [possible mistake] providers/hopbase/models/kimi-k3.1-pro.toml:10 - Check: A no-controls reasoning = { supported = true } on a relay of a natively controlled reasoner needs affirmative evidence (AGENTS.md → Reasoning controls; audit skill evidence bar). Why: The native providers/moonshotai/models/kimi-k3.1-pro.toml declares effort = ["low", "high", "max"] (default max) and relay peers (e.g. 302ai) publish effort sets; this file's one-line "reasoning_effort is accepted and ignored" gives no probe detail, unlike the sibling kimi-k3.toml which documents an invalid value returning 200 and flat reasoning-token counts. If the gateway actually forwards the field, the catalog under-reports the control surface. Action: Document the same live evidence for this model (invalid value accepted / reasoning tokens flat across low/high/max), or declare the effort levels this host honours.
  • [medium] [possible mistake] providers/hopbase/models/kimi-k3.1-pro.toml:33 - Check: [api.<protocol>] tables are the model's complete list of APIs served by this host. Why: The header records verified round trips on POST /v1/responses and POST /v1/messages for this exact ID, but the file lists only [api.chat-completions], so the file contradicts its own evidence and hides two working protocols from clients. Action: Add [api.responses] and [api.messages] tables with their SDK entries (the host serves them per the header), or correct the header if they are not part of the offering.

This branch was previously deployed

1 inactive deployment
release — 74f632a1 Deployed Nov 16, 2026 by victor-zzh via review #19257
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant