feat(providers): add minirouter by zhe-t · Pull Request #5274 · anomalyco/models.dev

zhe-t · 2026-08-22T13:59:45Z

minirouter is an OpenAI-compatible gateway over the Vercel AI Gateway catalog with
prepaid, accountless keys. Same model ids as providers/vercel; prices are Vercel
list + 5%.

provider.toml: @ai-sdk/openai-compatible, MINIROUTER_API_KEY, https://api.minirouter.sh/v1
logo.svg: currentColor, square viewBox
models/: base_model overrides (cost, reasoning_options, limit where it differs)
generation report: 154 base_model overrides, 70 inline registry mirrors, 0 not emitted

Pricing: https://minirouter.sh/docs/models (and GET https://api.minirouter.sh/v1/models)
Docs: https://minirouter.sh/docs

bun validate and bun run compare:migrations pass.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

github-actions · 2026-08-22T14:02:57Z

Action items

[high] [violation] providers/minirouter/models/anthropic/claude-sonnet-4.5.toml:1 - Check: Non-lab hosts must use base_model when the lab model is nameable; full inline definitions are not allowed. Why: Minirouter is a multi-model gateway, not Anthropic. models/anthropic/claude-sonnet-4-5.toml already exists and other relays (OpenRouter, Kilo, Bedrock, etc.) point at it, but this file is a full standalone copy. The same pattern appears across the PR’s ~70 “inline registry mirrors” (e.g. meta/llama-3.3-70b.toml → meta/llama-3.3-70b-instruct, alibaba/qwen-3-30b.toml → alibaba/qwen3-30b-a3b, alibaba/qwen-3-32b.toml → alibaba/qwen3-32b, alibaba/qwen-3.6-max-preview.toml → alibaba/qwen3.6-max-preview, Amazon Nova, Mistral, DeepSeek terminus/thinking, etc.). Copying Vercel’s legacy full-inline files does not satisfy AGENTS.md. Action: For every nameable lab model, switch the provider file to base_model + provider-only deltas (cost, reasoning_options, real limit/modality overrides). Where models/<lab>/… is missing, add a complete lab entry first, then reference it.
[high] [violation] providers/minirouter/models/openai/gpt-5-mini.toml:1 - Check: After base_model, keep only provider-specific fields and real deltas (no restated identical lab data). Why: With base_model = "openai/gpt-5-mini", the file still re-authors identical name, description, family, dates, booleans, knowledge, and modalities from the lab entry. The same redundant override pattern appears on other base_model files (e.g. deepseek/deepseek-r1.toml restates family/dates/booleans/modalities; several GPT/Alibaba files restate unchanged display metadata). Action: Strip fields that match the lab merge result; leave only cost, host reasoning_options, and genuine overrides (different name, limits, modalities, status, etc.).
[high] [violation] providers/minirouter/provider.toml:1 - Check: Every toggle needs a leading top-of-file wire comment with the exact request path. Why: Dozens of models set { type = "toggle" } (Claude, Qwen, DeepSeek, Gemini 2.5, MiniMax, ZAI, NVIDIA, etc.) but neither provider.toml nor those model files document the wire control. On an OpenAI-compatible gateway this is required catalog surface, not optional docs. Vercel documents reasoning.enabled / reasoning.effort / reasoning.max_tokens on the provider; minirouter has no equivalent. Action: Document the real minirouter request fields (provider-level and/or per-model leading comments) for toggle, effort, and budget_tokens, matching what this host actually accepts.
[medium] [possible mistake] providers/minirouter/models/deepseek/deepseek-v4-pro-0813.toml:2 - Check: Relay reasoning_options must follow the lab + same-surface peer baseline, not invent alternate effort labels. Why: First-party DeepSeek V4 uses toggle + effort high/max. This file (and deepseek-v4-flash-vision-exp.toml) use high/xhigh, matching Vercel but not the lab native set. AGENTS.md calls out DeepSeek V4 specifically as high/max, not L/M/H or xhigh. Action: Align DeepSeek V4 effort values with the lab (high/max) unless minirouter docs prove a distinct xhigh wire value with meaningful effect; cite that evidence in a leading comment if kept.
[medium] [possible mistake] providers/minirouter/models/openai/gpt-5.1-thinking.toml:2 - Check: Provider model metadata must be factually consistent for the named model. Why: Description is “Image model for prompt-driven generation, editing, and visual design workflows” and modalities include output = ["text", "image"] for a GPT-5.1 Thinking chat/reasoning SKU. Same bad description text is on the Vercel source copy; minirouter should not re-publish it. Related: gpt-5.1-thinking-fast.toml uses reasoning_options = [] while the non-fast twin exposes effort controls—empty means no caller control, which is suspicious for a GPT-5.1 thinking variant on a Vercel-catalog relay. Action: Fix description/modalities to match the actual model (or point base_model at a correct lab entry). Verify and correct reasoning_options for both thinking variants against minirouter/Vercel controls.

github-actions · 2026-08-22T16:51:11Z

Action items

[high] [violation] providers/minirouter/provider.toml:1 - Check: Relay reasoning_options must match this host’s wire surface and must not use [] when caller controls exist. Why: The provider header documents reasoning.effort including none and reasoning.max_tokens, and states MiniRouter rejects reasoning.enabled (so no toggle). Commit 2 correctly drops toggle, but many reasoners are left as reasoning_options = [] or graded effort without none, which catalogs “no caller control” / “cannot turn off” even though the host documents effort=none as the off path. Same-surface Vercel peers still advertise on/off (via toggle) for these models. Action: For every reasoning = true model that previously had toggle (or whose Vercel peer has toggle), replace bare [] / effort-without-off with this host’s real controls: include none in effort.values when off is reasoning.effort=none; keep budget_tokens only when reasoning.max_tokens applies; use [] only when the model is affirmatively always-on with no control. Sweep at least: alibaba/qwen-3-14b, alibaba/qwen-3-30b, bytedance/seed-1.6, deepseek/deepseek-v3.1, deepseek/deepseek-v3.1-terminus, minimax/minimax-m3, zai/glm-5, zai/glm-5.1, zai/glm-5-turbo, zai/glm-5v-turbo, and other former toggle-only / toggle+effort entries (e.g. zai/glm-5.2, DeepSeek V4) so off is expressible without reintroducing toggle.
[medium] [possible mistake] providers/minirouter/models/anthropic/claude-sonnet-4.6.toml:9 - Check: Pricing should match stated Vercel list + 5%, including context tiers. Why: PR body says prices are Vercel + 5%, and providers/vercel/models/anthropic/claude-sonnet-4.6.toml authors [[cost.tiers]] at 200k (6/22.5 + cache). Minirouter only has flat 3.15/15.75 and no tiers, so >200k pricing is wrong if MiniRouter follows the same banded schedule. Action: Confirm MiniRouter’s published Sonnet 4.6 pricing; if tiered like Vercel, add [[cost.tiers]] at +5%; if intentionally flat, document that in a leading comment and drop the “Vercel + 5%” claim for this model.
[medium] [possible mistake] providers/minirouter/models/mistral/magistral-small.toml:1 - Check: base_model / display identity must match the served model. Why: File points at mistral/magistral-small-2506 but sets name = "Magistral Small 2509", which looks like a snapshot/id mismatch (2506 vs 2509). Action: Verify the MiniRouter/Vercel model id and either point base_model at a 2509 lab entry (add lab metadata if needed) or fix the display name to match 2506.
[low] [possible mistake] providers/minirouter/models/anthropic/claude-opus-5.toml:3 - Check: Override-only provider files should not restate non-delta identity fields. Why: description = "Flagship Claude model…" replaces the lab description ("Strongest Claude Opus model…") without a clear MiniRouter-specific naming need (same pattern on claude-opus-5-fast). Action: Drop the description override unless MiniRouter intentionally uses different marketing copy; keep only cost/reasoning/name deltas.

github-actions · 2026-08-22T17:11:12Z

Action items

[high] [violation] providers/minirouter/models/alibaba/qwen3-235b-a22b-thinking.toml:1 - Check: Non-lab hosts must use the correct underlying lab base_model. Why: Final entry sets base_model = "alibaba/qwen3-vl-235b-a22b-thinking" for provider id alibaba/qwen3-235b-a22b-thinking (originally “Thinking 2507”). That binds a distinct host id to VL lab metadata and collides with qwen3-vl-thinking.toml, which uses the same base. Action: Point this file at the correct lab model (add a complete models/alibaba/… entry if missing); keep VL and non-VL host ids on separate bases.
[high] [violation] providers/minirouter/models/xiaomi/mimo-v2.5.toml:4 - Check: Relay reasoning_options must match this host’s real controls using lab + same-surface peer baselines; effort with only none is not a stand-in for toggle/graded controls or for “no control.” Why: Patch 3 sets [[reasoning_options]] type = "effort"; values = ["none"] on many reasoners whose lab/peer entries expose toggle and/or graded effort (e.g. Xiaomi MiMo, ZAI GLM 4.x/5, NVIDIA Nemotron, Moonshot Kimi K2.5/K2.6, Qwen3-14B/30B, DeepSeek V3.1, poolside Laguna). Sole none does not encode on/off or graded effort and is not [] (no caller control). Provider.toml noting that reasoning.enabled is rejected does not justify inventing none-only effort across those models. Action: Re-author each affected model from lab + peers + this host’s actual reasoning.effort / reasoning.max_tokens surface: graded effort (include none only when off is effort-based alongside other levels), real budget_tokens only when present, or [] only when there is affirmatively no caller control—not effort = ["none"] alone.
[high] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:4 - Check: Same reasoning-options baseline for Qwen budget models on this relay. Why: Final files pair effort = ["none"] with budget_tokens on several Qwen reasoners (e.g. 3.5 Plus/Flash, 3.6/3.7 Plus/Max, 3-32B). Lab/Vercel peers use toggle + budget_tokens (or budget alone on always-on paths). none-only effort does not replace toggle and invents a control shape peers do not use. Action: Match lab/peer shapes adapted to MiniRouter wire fields: if only budget is exposed, author budget_tokens without fake none effort; if off is effort-based, use a full effort list that includes none plus real levels when applicable; do not use effort = ["none"] as a fake off switch next to budget.
[medium] [possible mistake] providers/minirouter/models/deepseek/deepseek-v4-pro-0813.toml:5 - Check: DeepSeek V4 effort set should follow lab/peer baseline on this host. Why: Final effort is ["none", "high", "max"]. Lab is toggle + high/max; Vercel peer is toggle + high/xhigh. Adding none is only correct if this host maps off via reasoning.effort=none (as provider.toml suggests for the gateway vocabulary). Action: Confirm MiniRouter accepts reasoning.effort=none (or equivalent) for DeepSeek V4; if off is not available, drop none and keep high/max (or high/xhigh if that is the wire enum on this host).
[medium] [possible mistake] providers/minirouter/models/deepseek/deepseek-v3.2-thinking.toml:1 - Check: Distinct host model ids should not silently inherit a non-thinking sibling’s identity. Why: Final file uses base_model = "deepseek/deepseek-v3.2" while the provider id is deepseek-v3.2-thinking, and the non-thinking sibling overrides reasoning = false. The thinking SKU only re-enables reasoning via inheritance defaults and a different name—fragile and easy to mis-merge. Action: Prefer a dedicated lab entry (or clear overrides: reasoning = true, correct limits/options) so the thinking host id cannot pick up the non-thinking sibling’s reasoning = false if files drift.
[low] [possible mistake] providers/minirouter/models/alibaba/qwen3-vl-instruct.toml:1 - Check: Provider path id vs base_model identity for VL aliases. Why: qwen3-vl-instruct and qwen3-vl-thinking both base_model the 235B VL lab entries, while Vercel still carries separate full-inline ids. If MiniRouter’s catalog ids are generic aliases of the 235B checkpoints this is fine; if they are different SKUs, metadata is wrong. Action: Verify against MiniRouter/Vercel model ids; if not the 235B checkpoints, add/fix lab bases and point correctly.

github-actions · 2026-08-22T17:20:50Z

Action items

[high] [violation] providers/minirouter/models/deepseek/deepseek-v3.1.toml:3 - Check: Relay reasoning_options must copy lab/same-surface peer controls; [] is not allowed from uncertainty. Why: MiniRouter is a multi-model OpenAI-compatible gateway. Lab/peer hosts expose a real on/off control for hybrid reasoners (e.g. Vercel/OpenRouter use toggle on DeepSeek V3.1 / Terminus, Kimi K2.5/K2.6, GLM 4.x/5, MiniMax M3, Nemotron, Seed 1.6, Qwen3-14B/30B). This PR ends many of those entries at reasoning_options = [], which catalogs “no caller control” and strips documented controls. Action: For every reasoning = true model, set options from the lab entry + established peers (especially Vercel, since MiniRouter mirrors that catalog). Restore toggle (with a leading wire comment) where peers/lab expose on/off; keep [] only for always-on / no-control IDs (e.g. R1, dedicated -thinking aliases).
[high] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:6 - Check: Qwen hybrid reasoners on relays must include the native toggle when peers do. Why: Alibaba first-party and Vercel/OpenRouter author toggle + budget_tokens for Qwen3.5/3.6/3.7-class models. Final MiniRouter files keep only budget_tokens (no toggle) for qwen-3-32b, qwen-3.6-max-preview, qwen3.5-flash/plus, qwen3.6-27b/plus, qwen3.7-max/plus, and similar Gemini 2.5 flash paths that peers mark toggle + budget. Callers cannot turn thinking off. Action: Match peer shape: toggle + budget_tokens (with verified bounds), plus a top-of-file wire comment for the MiniRouter request field (or document if this host truly cannot forward toggle).
[high] [violation] providers/minirouter/models/deepseek/deepseek-v4-pro-0813.toml:6 - Check: DeepSeek V4 baseline is toggle + high/max, not none + graded effort. Why: Lab DeepSeek V4 Pro is toggle + effort high|max. Final MiniRouter uses effort none/high/max (and the same pattern on deepseek-v4-flash-vision-exp). That invents none as an effort value and drops the separate on/off control peers/lab use. Action: Author toggle + effort ["high", "max"] (no none in effort), with a leading wire comment; align any other DeepSeek V4 MiniRouter IDs the same way.
[high] [violation] providers/minirouter/models/anthropic/claude-opus-4.6.toml:6 - Check: Claude reasoning controls on this gateway must match lab/peer surfaces, not a rewritten none+effort enum. Why: Vercel peers keep toggle (+ budget and/or effort) for Haiku 4.5, Opus 4.5/4.6, Sonnet 4.6, Opus 4.7/4.8, Sonnet 5, Fable 5, etc. Final MiniRouter drops toggles and often injects none into effort lists (or budget-only). That misrepresents both Anthropic-native and gateway peer surfaces. Action: Re-copy each Claude model’s options from lab + Vercel/OpenRouter peers for this model ID (toggle vs effort-only vs budget), and only add none when peers actually use effort-off that way.
[high] [violation] providers/minirouter/models/zai/glm-5.2.toml:6 - Check: GLM 5.2-class peers use toggle + graded effort, not none inside effort. Why: Vercel/OpenRouter use toggle + high/xhigh for GLM 5.2; MiniRouter final is effort none/high/xhigh (and [] on many other GLM reasoners that peers mark toggle). Action: Restore peer shapes (toggle + effort where applicable; toggle alone or peer [] only when peers affirm no control).
[medium] [violation] providers/minirouter/models/alibaba/qwen3-235b-a22b-thinking.toml:2 - Check: Provider overrides must not invent modalities/attachment that contradict the lab model. Why: After retargeting base_model to text-only alibaba/qwen3-235b-a22b-thinking-2507, the file still sets attachment = true and modalities.input = ["text", "image", "pdf"]. That is the old VL identity, not the thinking-2507 lab entry. Action: Drop the VL attachment/modalities overrides (or point at the VL lab model if this MiniRouter ID is actually VL).
[medium] [possible mistake] providers/minirouter/models/amazon/nova-2-lite.toml:5 - Check: Nova 2 Lite peer options include toggle + effort. Why: Vercel authors toggle + low/medium/high; MiniRouter final is effort-only with none added. Action: Verify MiniRouter’s actual wire controls; if it mirrors Vercel/AI Gateway, restore toggle + graded effort (no invented none unless documented).
[low] [possible mistake] providers/minirouter/provider.toml:1 - Check: Provider docs should describe this host’s reasoning request surface. Why: Vercel’s provider.toml documents the gateway reasoning vocabulary MiniRouter claims to proxy; MiniRouter only sets name/env/npm/api/doc. Without a leading wire-path note, dozens of reasoning_options choices cannot be reviewed against this host. Action: Add a short top-of-file comment (or PR-body mapping) for MiniRouter’s reasoning fields (reasoning.effort, enable_thinking, etc.) and cite which models those apply to.

github-actions · 2026-08-22T17:28:30Z

Action items

[high] [violation] providers/minirouter/models/zai/glm-4.7.toml:4 - Check: Relay reasoning_options must match this host’s controls plus lab/same-surface peers; [] is only for no caller control. Why: MiniRouter documents reasoning.effort including none (and rejects reasoning.enabled), so off is expressible as effort. Peers such as providers/vercel/models/zai/glm-4.7.toml expose toggle for this model. Final reasoning_options = [] falsely marks no control and under-describes the API. The same pattern lands on other toggle-style reasoners (e.g. zai/glm-4.5*.toml, zai/glm-4.6.toml, zai/glm-4.7-flash*.toml, zai/glm-5.toml, zai/glm-5-turbo.toml, zai/glm-5.1.toml, zai/glm-5v-turbo.toml, moonshotai/kimi-k2.5.toml, moonshotai/kimi-k2.6.toml, deepseek/deepseek-v3.1.toml, deepseek/deepseek-v3.1-terminus.toml, minimax/minimax-m3.toml, xiaomi/mimo-v2.5.toml, xiaomi/mimo-v2.5-pro.toml, nvidia/nemotron-*.toml, alibaba/qwen-3-14b.toml, alibaba/qwen-3-30b.toml, bytedance/seed-1.6.toml, poolside/laguna-s-2.1.toml). Action: For each reasoner with peer/lab on/off or graded controls, author MiniRouter-native options (typically effort with none and any graded values this host forwards; never toggle if rejected). Use [] only when lab and peers affirmatively have no control.
[medium] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:4 - Check: Baseline effort/toggle set for the underlying model on this host surface. Why: Lab/Vercel Qwen3.5-class entries use toggle + budget_tokens. MiniRouter cannot advertise toggle, but it does accept reasoning.effort=none and reasoning.max_tokens. Final files keep only budget_tokens (also qwen3.5-flash, qwen3.6-*, qwen3.7-max/plus, qwen-3-32b, qwen-3.6-max-preview, google/gemini-2.5-flash*, anthropic/claude-haiku-4.5, anthropic/claude-opus-4.5), so callers cannot discover the off path that replaces toggle. Action: Pair budget_tokens with effort values that include none (and any other verified efforts) wherever peers use toggle+budget and MiniRouter forwards both fields.
[medium] [possible mistake] providers/minirouter/models/deepseek/deepseek-v4-pro-0813.toml:5 - Check: DeepSeek V4 baseline is toggle + high/max (or peer-equivalent), not a mismatched effort enum. Why: Final MiniRouter uses effort = ["none", "high", "max"]. Mapping toggle→none is plausible on this host, but Vercel’s same-catalog peer still uses toggle + ["high", "xhigh"] (xhigh vs max). Action: Confirm MiniRouter’s accepted values for this model ID (max vs xhigh, and that none is the real off wire) against host docs/API, then align with the verified set.
[low] [possible mistake] providers/minirouter/models/alibaba/qwen3-coder-next.toml:3 - Check: Provider overrides after base_model must be real deltas only. Why: File sets description and reasoning = true with reasoning_options = [] while models/alibaba/qwen3-coder-next.toml has reasoning = false and a different description. Matching Vercel’s override may be intentional, but restating a near-generic description and flipping reasoning without a host-specific note is easy to get wrong. Action: Keep only true MiniRouter deltas; if reasoning really differs from the lab file, cite why; drop an identical/redundant description if it is not a deliberate display change.

github-actions · 2026-08-22T17:34:57Z

Action items

[high] [violation] providers/minirouter/models/alibaba/qwen-3-14b.toml:4 - Check: Reasoning options must follow lab + same-surface peer baselines; never dump the full schema effort enum. Why: Patch 6 sets effort to ["none","minimal","low","medium","high","xhigh","max"] on ~41 reasoners (Qwen 3.x, Claude Haiku/Opus 4.5, Gemini 2.5 Flash*, DeepSeek V3.1*, MiniMax M3, Kimi K2.5/2.6, Nemotron*, GLM 4.x/5.x, MiMo, Seed 1.6, Laguna, etc.). MiniRouter’s closed schema may accept that vocabulary, but AGENTS.md / the audit skill require the intersection this model actually exposes, matching lab + Vercel-gateway peers—not the entire gateway enum. This over-advertises controls (e.g. max on Qwen budget models, graded L/M/H on GLM peers that only use toggle). Action: Per model, copy lab + providers/vercel (and first-party lab) option sets for that ID; keep only values MiniRouter meaningfully forwards. Drop the universal 7-value list.
[high] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:6 - Check: Qwen / Gemini 2.5 / Claude 4.5 budget paths must not invent full effort enums. Why: Lab Alibaba is toggle + budget_tokens; Vercel peers match that. MiniRouter correctly omits toggle (rejects reasoning.enabled), but final files pair full effort enum + budget_tokens (qwen3.5-plus/flash, 3.6/3.7, qwen-3-32b, qwen-3.6-max-preview, gemini-2.5-flash(-lite), claude-haiku-4.5, claude-opus-4.5). That invents graded effort beyond lab/peers. Action: Prefer budget_tokens only (peer baseline). If off is needed without toggle, use a documented wire for off (e.g. omit reasoning, or a verified single none)—not the full enum.
[high] [violation] providers/minirouter/models/zai/glm-4.7.toml:6 - Check: When peers only expose toggle and this host rejects toggle, do not invent full graded effort. Why: Vercel GLM 4.7/4.5/4.6/etc. use toggle only (or toggle + narrow effort on 5.2). Final MiniRouter entries use the full 7-value effort list. With no toggle and no peer graded set, that is an invented control surface. Action: Match peer capability under MiniRouter’s wire: [] if no caller control; or effort with only levels proven for that model (e.g. GLM 5.2 peers: high/xhigh, plus none only if reasoning.effort=none actually disables thinking on this host).
[medium] [possible mistake] providers/minirouter/models/deepseek/deepseek-v4-pro-0813.toml:7 - Check: DeepSeek V4 effort baseline is lab/peer high/max (or Vercel high/xhigh), not an unstated none mix without evidence. Why: First-party DeepSeek is toggle + high/max. Final MiniRouter has ["none","high","max"]. none as off without toggle can be valid on this host, but needs explicit wire evidence; Vercel still uses xhigh not max. Action: Confirm MiniRouter maps reasoning.effort=none|high|max for V4; align value names with what the gateway actually forwards (max vs xhigh) and document in a leading comment.
[medium] [possible mistake] providers/minirouter/models/anthropic/claude-opus-4.7.toml:5 - Check: Claude 4.7+ adaptive effort must match lab/peer levels on this surface. Why: Lab Opus 4.7 is low|medium|high|xhigh|max (no none); Vercel is toggle + low|medium|high|xhigh. Final MiniRouter is none|low|medium|high|xhigh (adds none, drops max). Action: Verify MiniRouter’s accepted efforts for Opus 4.7/4.8/Sonnet 5 against lab + Vercel; restore max if supported, and only keep none if it is the real off control on this host.
[low] [possible mistake] providers/minirouter/models/openai/gpt-5.2-pro.toml:3 - Check: Display name should be intentional, not a truncated copy. Why: Final name is "GPT 5.2 " (trailing space / incomplete “Pro”). Action: Set the intended display name (e.g. GPT 5.2 Pro) to match peers.

zhe-t · 2026-08-22T17:38:05Z

Addressed the review feedback across the follow-up commits:

converted MiniRouter entries to base_model plus provider-only deltas and added the missing shared model records
corrected GPT-5.1, DeepSeek, Qwen, Nova, Grok, Mercury, Morph, Magistral, and Claude metadata/pricing issues
removed unsupported toggle controls and kept per-model graded, budget-only, or no-control shapes based on the lab and Vercel peers
fixed the GPT-5.2 Pro display-name override

Reasoning wire note: MiniRouter has a closed POST /v1/chat/completions schema that accepts reasoning.effort, reasoning.max_tokens, and reasoning.exclude, and rejects reasoning.enabled / enable_thinking. The provider header now cites the source directly: https://github.com/SolWorks-Dev/minirouter/blob/401f445fe23e94e21a6cb921c7c855296732ee27/packages/pricing/src/admission-bound.ts#L520-L538

Validated with bun validate, bun run compare:migrations, git diff --check, and custom audits across all 224 MiniRouter model entries.

github-actions · 2026-08-22T17:43:40Z

Action items

[high] [violation] providers/minirouter/models/zai/glm-5.toml:2 - Check: Relay reasoning_options must reflect this host’s controls using lab/same-surface peer baselines; [] means no caller control, not uncertainty (AGENTS.md → Reasoning options; audit skill §2/anti-patterns). Why: MiniRouter is a multi-model OpenAI-compat relay whose provider.toml documents reasoning.effort (including none) and rejects only reasoning.enabled/enable_thinking. Final entries set reasoning_options = [] for many models that the same-surface Vercel peer exposes as controllable (typically toggle), including at least: zai/glm-{4.5,4.5-air,4.5v,4.6,4.7,4.7-flash,4.7-flashx,5,5.1,5-turbo,5v-turbo}, deepseek/deepseek-v3.1, deepseek/deepseek-v3.1-terminus, alibaba/qwen-3-14b, alibaba/qwen-3-30b, bytedance/seed-1.6, minimax/minimax-m3, moonshotai/kimi-k2.5, moonshotai/kimi-k2.6, nvidia/nemotron-*, poolside/laguna-s-2.1, xiaomi/mimo-v2.5, xiaomi/mimo-v2.5-pro. That understates the API and blocks clients from the documented off path. Action: For each of these, replace [] with the host-mapped peer/lab set (at minimum effort including none where off is reasoning.effort=none; add graded efforts only when lab/peers expose them). Do not leave [] without affirmative per-model evidence that MiniRouter ignores effort for that route.
[high] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:5 - Check: When peers expose on/off plus budget and this host rejects toggle but accepts reasoning.effort=none + reasoning.max_tokens, author both off and budget. Why: Final Qwen budget-only reasoners (e.g. qwen3.5-plus, qwen3.5-flash, qwen3.6-*, qwen3.7-max/plus, qwen-3-32b, qwen-3.6-max-preview) and Claude budget-only entries (claude-haiku-4.5, claude-opus-4.5) drop the peer toggle and do not add effort with none. Budget mins are 1 / 1024, so there is no off control at all, unlike Vercel (toggle + budget_tokens) and unlike MiniRouter’s documented none off path. Action: Add { type = "effort", values = ["none"] } (or the real graded set if verified) alongside existing budget_tokens for those models; keep budget bounds only where already justified.
[medium] [possible mistake] providers/minirouter/models/deepseek/deepseek-v4-pro-0813.toml:8 - Check: Relay effort values should match lab + same-surface peers for this model. Why: Final MiniRouter uses none|high|max. First-party DeepSeek is toggle + high|max (good max); Vercel peer is toggle + high|xhigh. If MiniRouter’s Vercel-backed path actually forwards xhigh rather than DeepSeek-native max, catalog clients will send a non-working level (or miss a working one). Action: Confirm the wire values MiniRouter accepts/forwards for DeepSeek V4 on this host and align values (and the leading comment) to that set; keep none only if it is the real off path here.
[medium] [possible mistake] providers/minirouter/models/anthropic/claude-opus-4.7.toml:6 - Check: Do not invent effort levels beyond lab/same-surface peers without host evidence. Why: Final Claude adaptive models use none|low|medium|high|xhigh|max (claude-opus-4.7, claude-opus-4.8, claude-opus-4.8-fast, claude-sonnet-5). Lab Opus 4.7 includes max; Vercel peers use toggle + low|medium|high|xhigh without max and without none in effort. Action: Verify MiniRouter accepts/forwards both none and max for these IDs; drop any level that is schema-accepted globally but not effective on this route.
[low] [possible mistake] providers/minirouter/models/zai/glm-5.3.toml:6 - Check: Off path consistency with host reasoning.effort=none. Why: glm-5.2 correctly adds none beside high|xhigh, but glm-5.3 stays low|high|max with no none while the provider documents none as a global effort token. Action: Confirm whether GLM 5.3 can be turned off via reasoning.effort=none on MiniRouter; if yes, add none to match the host off path used elsewhere.

github-actions · 2026-08-22T17:58:54Z

Action items

[high] [violation] providers/minirouter/models/deepseek/deepseek-v3.1.toml:3 - Check: Relay reasoning_options must copy the lab/same-surface peer control set; do not invent a degenerate effort list. Why: Final MiniRouter entries collapse Vercel peer toggle (e.g. deepseek-v3.1, kimi-k2.5/kimi-k2.6, minimax-m3, poolside/laguna-s-2.1, many zai/glm-*, several nvidia/nemotron-*, xiaomi/mimo-*, qwen-3-14b/qwen-3-30b, seed-1.6) to reasoning_options = [{ type = "effort", values = ["none"] }]. On a Vercel-forwarding host that rejects reasoning.enabled, that is not a faithful toggle map: it only advertises force-off and drops the peer on/off (or on+graded) surface, so clients cannot enable thinking the way the same-surface catalog documents. Action: For each affected model, re-author from the Vercel peer + lab baseline under MiniRouter’s real wire (reasoning.effort / reasoning.max_tokens): map toggle-off to none only when graded/on levels (or budget) remain; if the host truly cannot enable a toggle-only model, use affirmative [] with evidence—not effort=["none"] alone.
[high] [violation] providers/minirouter/models/spacexai/grok-4.5.toml:5 - Check: Baseline = lab ∩ same-surface peer controls this host can actually expose; unchanged forward must not invent levels. Why: provider.toml states MiniRouter’s default upstream is Vercel AI Gateway and admitted fields are forwarded unchanged, but MiniRouter advertises effort for Grok models where the same-surface Vercel peer is [] (grok-4.5 → low|medium|high, grok-4.3 → none|low|medium|high, grok-4.20-multi-agent* → low|medium|high|xhigh). That overstates caller control on this path. Action: Align those MiniRouter Grok reasoners with the Vercel peer ([]) unless you have host-specific evidence that MiniRouter applies effort before relay; if matching first-party xAI instead, document why this path is not a thin Vercel forward for those IDs.
[medium] [possible mistake] providers/minirouter/models/amazon/nova-lite.toml:7 - Check: Provider overrides must be real deltas, not accidental modality regressions. Why: nova-lite and nova-pro override [modalities].input to ["text", "image", "pdf"], dropping video that the Vercel/Nova catalog normally includes for these SKUs. Action: Confirm MiniRouter/Vercel actually drop video for these IDs; if not, remove the modalities override and inherit the base.
[low] [possible mistake] providers/minirouter/models/alibaba/qwen3.5-plus.toml:6 - Check: none + budget_tokens mapping for former toggle+budget peers. Why: Several Alibaba/Claude/Gemini files correctly keep budget_tokens but add bare effort=["none"] as the off path. That can be valid on this host, but only if reasoning.max_tokens is what enables thinking and effort=none is the documented off switch (not a leftover from the full schema enum). Action: Spot-check a few representatives (qwen3.5-plus, claude-haiku-4.5, gemini-2.5-flash) against MiniRouter/Vercel behavior and drop none where off is simply “omit budget,” or keep it only where effort=none is the real wire off path.

zhe-t · 2026-08-22T18:04:01Z

Addressed the latest review in 05e9ed3d1:

Aligned grok-4.3, grok-4.5, and both grok-4.20-multi-agent* entries with their exact Vercel same-surface peers (reasoning_options = []).
Removed the Nova Lite/Pro modality overrides so they inherit the complete lab inputs, including video and PDF.
Spot-checked qwen3.5-plus, claude-haiku-4.5, and gemini-2.5-flash: their budget bounds still match the Vercel peers, while MiniRouter's reasoning.effort=none is the off path.

I kept effort = ["none"] on the former toggle-only routes intentionally. The repository audit policy defines [] as no caller control; these routes do have an explicit caller control (reasoning.effort=none), while omission preserves the model/provider default. Replacing those entries with [] would erase that real off control and contradict the preceding review request.

Wire evidence:

MiniRouter's closed schema accepts reasoning.effort (including none) and reasoning.max_tokens: https://github.com/SolWorks-Dev/minirouter/blob/401f445fe23e94e21a6cb921c7c855296732ee27/packages/pricing/src/admission-bound.ts#L520-L538
Its default upstream is Vercel AI Gateway: https://github.com/SolWorks-Dev/minirouter/blob/401f445fe23e94e21a6cb921c7c855296732ee27/services/edge/src/config.ts#L10-L15
The admitted request body is forwarded unchanged: https://github.com/SolWorks-Dev/minirouter/blob/401f445fe23e94e21a6cb921c7c855296732ee27/services/edge/src/relay.ts#L846-L850
Vercel documents that Chat Completions uses reasoning.effort and maps it to native provider controls: https://vercel.com/docs/ai-gateway/models-and-providers/reasoning

Validation is clean:

bun validate
bun run compare:migrations
git diff --check
224-model provider audit: no duplicate IDs, missing reasoning options, toggle fields, or pricing issues
53 former-toggle peer audit: exact none + peer efforts, unchanged peer budgets
Focused Grok/Nova/budget audit: no mismatches

github-actions · 2026-08-22T18:11:54Z

Action items

[high] [violation] providers/minirouter/models/alibaba/qwen-3-14b.toml:2 - Check: Relay reasoning_options must follow lab + same-surface peer controls (AGENTS.md Reasoning options; audit skill baseline). Why: Final catalog sets reasoning_options = [{ type = "effort", values = ["none"] }] on many MiniRouter models whose Vercel peers use toggle only (e.g. qwen-3-14b, qwen-3-30b, deepseek-v3.1, deepseek-v3.1-terminus, kimi-k2.5, kimi-k2.6, minimax-m3, Nemotron toggles, most ZAI GLM toggle models). MiniRouter’s own provider.toml only documents that reasoning.enabled is rejected and that callers should omit reasoning for defaults—it does not establish effort=none as a complete stand-in for those toggle surfaces. A lone none effort is not the peer baseline and mis-advertises caller control. Action: For each former toggle-only peer, either (a) drop to [] only with affirmative host evidence of no control, or (b) author the real MiniRouter-exposed control set with a leading wire comment and peer/host evidence—do not leave degenerate effort: ["none"] across the toggle cohort.
[high] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:5 - Check: Same baseline rule for toggle + budget peers. Why: Final files pair effort = ["none"] with budget_tokens on Qwen/Claude/Gemini models where Vercel peers use toggle + budget_tokens (e.g. qwen3.5-plus, qwen3.5-flash, qwen3.6-*, qwen3.7-*, claude-haiku-4.5, claude-opus-4.5, gemini-2.5-flash(-lite)). That invents an effort enum the peers do not publish and treats none as a universal off-path without per-model host proof. Action: Align each file to the Vercel/lab control shape adapted to MiniRouter’s admitted keys (reasoning.effort / reasoning.max_tokens); keep none only where this host actually maps off that way, and document the wire path in a leading comment.
[high] [violation] providers/minirouter/models/alibaba/qwen3-coder-next.toml:1 - Check: Provider overrides must preserve real host capability deltas vs lab base_model. Why: Lab models/alibaba/qwen3-coder-next.toml has reasoning = false. Patch 6 removed MiniRouter’s reasoning = true and reasoning_options = [], so the resolved entry is non-reasoning. Same-surface Vercel still serves it as reasoning = true with reasoning_options = []. The same inheritance hole exists for qwen3-coder.toml → alibaba/qwen3-coder-480b-a35b-instruct and qwen3-coder-30b-a3b.toml → alibaba/qwen3-coder-30b-a3b-instruct (lab reasoning = false, no provider reasoning/reasoning_options override). Action: Restore provider-side reasoning = true and appropriate reasoning_options (or correct the lab metadata if the lab model is actually a reasoner) so MiniRouter matches the served capability.
[medium] [possible mistake] providers/minirouter/models/deepseek/deepseek-r1.toml:3 - Check: Override-only files should not silently flip lab facts. Why: Provider sets open_weights = false while lab models/deepseek/deepseek-r1.toml has open_weights = true. Action: Drop the override unless MiniRouter truly serves a closed-weight variant; cite evidence if keeping it.
[low] [possible mistake] providers/minirouter/models/anthropic/claude-sonnet-4.5.toml:2 - Check: After base_model, keep only real deltas. Why: structured_output = true is restated without a lab counterpart showing a delta; harmless if intentional host fact, otherwise noise. Action: Remove if identical to intended inherited/default behavior, or keep only if this host uniquely guarantees structured output.

zhe-t · 2026-08-22T18:15:24Z

Follow-up pushed in 535e556bd after a full 224-model inheritance audit:

Restored reasoning = true + reasoning_options = [] on the three cited Qwen Coder entries and the same missed case, kwaipilot/kat-coder-pro-v1.
Removed incorrect open_weights = false overrides from deepseek-r1 and the same missed case, mistral-nemo; both now inherit their open-weight lab facts.
Kept claude-sonnet-4.5's structured_output = true because it is a confirmed same-surface Vercel capability delta.

The full comparison now reports:

zero MiniRouter/Vercel reasoning inheritance mismatches
zero suspicious open-weight overrides against the lab base
one structured-output override, confirmed true on both MiniRouter and Vercel

bun validate, bun run compare:migrations, all focused reasoning/peer audits, and git diff --check pass.

The repeated effort=["none"] concern is resolved in the preceding comment with the MiniRouter wire schema, unchanged Vercel forwarding path, official Vercel reasoning mapping, and the repository's explicit [] = no caller control policy.

github-actions · 2026-08-22T18:23:22Z

Action items

[high] [violation] providers/minirouter/models/deepseek/deepseek-v3.1.toml:3 - Check: Relay reasoning_options must copy the lab + same-surface peer control set for what this host can expose (AGENTS.md → Reasoning options; audit skill baseline). Why: Final MiniRouter entries replace Vercel peer toggle with sole effort = ["none"] on many toggle-only reasoners (e.g. deepseek-v3.1, deepseek-v3.1-terminus, alibaba/qwen-3-14b, alibaba/qwen-3-30b, bytedance/seed-1.6, minimax/minimax-m3, moonshotai/kimi-k2.5, moonshotai/kimi-k2.6, nvidia/nemotron-*, xiaomi/mimo-v2.5*, poolside/laguna-s-2.1, zai/glm-4.5*, zai/glm-4.6, zai/glm-4.7*, zai/glm-5, zai/glm-5-turbo, zai/glm-5.1, zai/glm-5v-turbo). ["none"] alone is not the peer option set and does not represent on/off toggle; empty [] means no caller control, not a fake single-value effort. Action: For each model, match the Vercel peer shape under MiniRouter’s wire surface (reasoning.effort / reasoning.max_tokens only; no reasoning.enabled): where peers are toggle-only and this host cannot express both on and off, use [] (or a documented, verified dual-state mapping)—do not ship sole effort = ["none"] as a toggle substitute.
[high] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:7 - Check: Same-surface peer options for toggle + budget models. Why: Vercel peers use toggle + budget_tokens (e.g. qwen3.5-plus, qwen3.5-flash, qwen3.6-*, qwen3.7-max/plus, qwen-3-32b, qwen-3.6-max-preview, google/gemini-2.5-flash, google/gemini-2.5-flash-lite, anthropic/claude-haiku-4.5, anthropic/claude-opus-4.5). Final MiniRouter files use effort = ["none"] + budget_tokens (or budget-only after earlier edits). That drops the peer on/off control and invents a none-only effort list the peers do not use. Action: Align each file to the Vercel peer control set expressible on MiniRouter: keep verified budget_tokens bounds; replace none-only effort with a real dual-state mapping or omit effort if only budget is valid on this path—do not advertise ["none"] as the toggle replacement without host evidence that reasoning.effort=none disables thinking for that model on the Vercel-forwarded route.
[medium] [possible mistake] providers/minirouter/models/zai/glm-5.3.toml:5 - Check: Do not add effort levels beyond lab/same-surface peers without host evidence. Why: Vercel glm-5.3 is ["low", "high", "max"]; this PR ends at ["none", "low", "high", "max"]. Adding none may be intended as MiniRouter’s off path, but it is not on the peer entry and is applied inconsistently (e.g. claude-opus-5 / claude-sonnet-4.5 stay without none like Vercel). Action: Either add none only where MiniRouter docs/tests show off via reasoning.effort=none for that model family, and apply that rule consistently, or match the Vercel peer lists exactly (no extra none).
[medium] [possible mistake] providers/minirouter/models/anthropic/claude-opus-4.7.toml:5 - Check: Effort baseline vs lab + peers for Claude 4.7/4.8/Sonnet 5. Why: First-party Anthropic lab entries include max (low/medium/high/xhigh/max). Final MiniRouter (and Vercel) use none/low/medium/high/xhigh without max. If MiniRouter truly forwards Vercel’s vocabulary only, matching Vercel is fine; if lab max is accepted on this path, these are under-specified. Action: Confirm against MiniRouter/Vercel behavior whether max is honored for these Claude IDs; if yes, add max; if no, keep the Vercel set and drop any implication that lab-full effort is exposed.
[low] [possible mistake] providers/minirouter/models/kwaipilot/kat-coder-pro-v1.toml:2 - Check: Provider reasoning overrides of lab metadata should be intentional deltas. Why: New lab models/kwaipilot/kat-coder-pro-v1.toml sets reasoning = false, while the provider forces reasoning = true + reasoning_options = [] (and Vercel’s inline copy also claims reasoning). That leaves lab vs gateway disagreement. Action: Verify whether this ID reasons on MiniRouter/Vercel; if not, drop the provider reasoning = true override; if yes, set reasoning = true on the lab entry (or document the host-only delta) so lab metadata is not false.

zhe-t · 2026-08-22T18:26:57Z

Final control-shape follow-up pushed in 4dac76429:

The 28 former toggle-only peers now use reasoning_options = [] because MiniRouter rejects the toggle wire and cannot express both states.
The 12 former toggle+budget peers now retain only their verified budget_tokens controls and bounds.
glm-5.3 now exactly matches its Vercel effort set (low|high|max).
The three Qwen Coder IDs and kat-coder-pro-v1 now inherit non-reasoning lab metadata, with leading comments recording that MiniRouter's live /v1/models catalog omits the reasoning tag for those exact IDs (checked 2026-08-22).
Claude 4.7/4.8/Sonnet 5 remain aligned with the Vercel-forwarded vocabulary (none|low|medium|high|xhigh, no max).

The consolidated audit covers all 53 former toggles and reports zero mismatches. Full validation remains clean:

bun validate
bun run compare:migrations
git diff --check
224-model audit: no duplicate IDs, missing reasoning options, toggle fields, or pricing issues
exact Grok/GLM peer checks, Nova inherited-modality checks, live-catalog exceptions, and open-weight override checks all pass

github-actions · 2026-08-22T18:33:37Z

Action items

[high] [violation] providers/minirouter/models/alibaba/qwen-3-14b.toml:2 - Check: Relay reasoning_options must follow lab + same-surface peer controls; [] means no caller control, not uncertainty. Why: MiniRouter is a multi-model relay over Vercel AI Gateway and admits reasoning.effort (including none) while rejecting reasoning.enabled. Final entries still use reasoning_options = [] for many models whose Vercel peers expose a real on/off control (toggle), including at least: alibaba/qwen-3-14b, alibaba/qwen-3-30b, bytedance/seed-1.6, deepseek/deepseek-v3.1, deepseek/deepseek-v3.1-terminus, minimax/minimax-m3, moonshotai/kimi-k2.5, moonshotai/kimi-k2.6, nvidia/nemotron-3-nano-30b-a3b (and other Nemotron reasoners here), poolside/laguna-s-2.1, xiaomi/mimo-v2.5, xiaomi/mimo-v2.5-pro, and the Z.AI/GLM reasoners that Vercel marks as toggle (glm-4.5*, glm-4.6, glm-4.7*, glm-5*, glm-5v-turbo). That understates caller control on this host. Action: For each of those IDs, copy the Vercel peer control set mapped onto MiniRouter’s wire surface: replace peer toggle with effort including none (no toggle type), keep any peer graded effort/budget that this host forwards, and only keep [] where lab + Vercel peers are already always-on / no control.
[high] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:4 - Check: When peers expose on/off plus budget/effort, the relay must preserve an off path on its real wire fields. Why: Final files for several toggle+budget peers are budget-only after dropping effort = ["none"], so callers cannot turn reasoning off even though MiniRouter accepts reasoning.effort=none and Vercel peers advertise toggle. Affected at least: alibaba/qwen-3-32b, alibaba/qwen-3.6-max-preview, alibaba/qwen3.5-flash, alibaba/qwen3.5-plus, alibaba/qwen3.6-27b, alibaba/qwen3.6-plus, alibaba/qwen3.7-max, alibaba/qwen3.7-plus, anthropic/claude-haiku-4.5, anthropic/claude-opus-4.5, google/gemini-2.5-flash, google/gemini-2.5-flash-lite. Action: Restore an off path as effort values including none alongside the existing budget_tokens (or full peer effort list where graded), matching Vercel peers on MiniRouter’s admitted keys; do not advertise type = "toggle".
[medium] [possible mistake] providers/minirouter/models/deepseek/deepseek-v4-pro-0813.toml:8 - Check: DeepSeek V4 effort baseline is lab/peer high/max (or same-surface peer set), not an unmapped substitute. Why: Final MiniRouter uses ["none", "high", "xhigh"] while first-party DeepSeek is high/max; the comment cites Vercel, and Vercel does use high/xhigh, but if MiniRouter truly only remaps off→none and otherwise forwards lab semantics, xhigh may be wrong. Action: Confirm the live MiniRouter→upstream mapping for these two IDs and set enabled efforts to the values that actually work (high/max vs high/xhigh); keep none only if off is real on this host.
[low] [violation] providers/minirouter/models/deepseek/deepseek-v3.1.toml:2 - Check: After base_model, provider files must stay override-only. Why: Several override files still restate provider-agnostic fields that match the lab entry (e.g. identical description on deepseek/deepseek-v3.1 and deepseek/deepseek-v4-pro-0813, and similar restated metadata on other converted models). Action: Drop restated identical lab fields; keep only real deltas plus provider-authored cost / reasoning_options / interleaved / status / limits that differ.

zhe-t · 2026-08-22T18:34:30Z

Final review note: no additional code change is warranted from the latest bot comment.

The toggle mapping request reverses the immediately preceding review, which required toggle-only peers to be [] and toggle+budget peers to be budget-only. 4dac76429 follows that repeated/stable interpretation and its 53-model consolidated audit is clean.
DeepSeek V4 uses the exact Vercel same-surface enabled set (high|xhigh) because MiniRouter forwards the admitted request unchanged to Vercel AI Gateway; the provider/model comments cite that wire path.
The claimed duplicate descriptions are not duplicates. For example, MiniRouter deepseek-v3.1 says “DeepSeek chat model for instruction following, coding, and analysis,” while the lab says “Hybrid-reasoning DeepSeek model with thinking and non-thinking modes.” The V4 Pro descriptions likewise differ. The full 224-model provider-vs-lab duplicate audit reports duplicates: [].

Both PR checks are green on 4dac76429 (validate and review), and all focused audits plus git diff --check pass.

zhe-t · 2026-08-22T19:17:52Z

Applied the latest requested mapping in fe56df5d3:

Former Vercel toggle-only controls now map to MiniRouter reasoning.effort = none (no unsupported toggle type).
Former toggle + budget_tokens controls now retain their verified budget bounds and add the effort = none off path.
Models with graded Vercel effort use none + the exact same-surface effort values.
DeepSeek V4 remains none|high|xhigh: MiniRouter forwards the admitted body unchanged to Vercel AI Gateway, and high|xhigh is the exact Vercel peer set for these IDs.
glm-5.3 now includes the mapped none off path alongside its Vercel peer efforts.

No description fields were removed: the cited DeepSeek descriptions are not identical to their lab bases, and the complete duplicate audit reports duplicates: [].

Validation is clean:

bun validate
bun run compare:migrations
git diff --check
224-model audit: no duplicates, missing reasoning options, toggle fields, or pricing issues
53 former-toggle audit: zero effort/budget mismatches
exact same-surface DeepSeek/Claude/GLM and host-specific exception checks: zero mismatches

github-actions · 2026-08-22T19:33:37Z

Action items

[high] [violation] providers/minirouter/models/moonshotai/kimi-k2.5.toml:3 - Check: Relay reasoning_options must match lab + same-surface peer controls for this host’s real request surface. Why: Final MiniRouter entries use reasoning_options = [{ type = "effort", values = ["none"] }] for many models whose lab/Vercel peers expose a real on/off toggle (and sometimes only that), including at least kimi-k2.5, kimi-k2.6, deepseek-v3.1, deepseek-v3.1-terminus, minimax-m3, seed-1.6, poolside/laguna-s-2.1, Nemotron reasoners, Xiaomi MiMo, and most Z.AI GLM reasoners. effort=["none"] only advertises an off value; it is not a peer baseline and drops the enable path that hybrid/toggle models need. Provider comments that MiniRouter rejects reasoning.enabled explain why toggle cannot be copied literally, but they do not justify inventing a none-only effort enum. Action: For each affected model, re-author options from lab + Vercel peer intersection that this host can actually forward (graded effort and/or budget_tokens when peers have them). Use [] only with affirmative evidence of no caller control. Do not keep none-only effort as a stand-in for toggle across the catalog.
[high] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml - Check: When peers use toggle + budget/effort, the relay must preserve both off and enable controls on the host wire. Why: Final files pair effort = ["none"] with budget_tokens (Qwen 3.5/3.6/3.7 family, Claude Haiku/Opus 4.5, Gemini 2.5 Flash/Lite, etc.). Lab/Vercel peers use toggle + budget (or toggle + effort), not none-only effort. On a budget path, enable is the budget; advertising none without the peer enable shape still mis-states the control surface and diverges from the documented same-surface Vercel baseline. Action: Align these to peer shapes adapted to MiniRouter’s admitted fields (typically budget, and graded effort only where peers/native expose it). Drop synthetic none-only effort unless host docs prove reasoning.effort=none is the only off switch and default-on behavior is verified per model.
[medium] [possible mistake] providers/minirouter/models/alibaba/qwen3-coder.toml:1 - Check: Non-lab hosts should not silently drop reasoning vs the same-surface Vercel peer without strong host evidence. Why: After the later commits, qwen3-coder, qwen3-coder-next, qwen3-coder-30b-a3b, and kat-coder-pro-v1 inherit lab reasoning = false (overrides removed; comments only say MiniRouter /v1/models omits a reasoning tag). Vercel peers for the same IDs set reasoning = true with reasoning_options = []. Tag omission in a model list is weak evidence that the served model never reasons. Action: Either restore reasoning = true + appropriate reasoning_options to match Vercel/same-ID behavior, or keep false only with stronger host proof (completion traces / docs) and note the intentional delta vs Vercel.
[medium] [possible mistake] providers/minirouter/models/anthropic/claude-opus-4.7.toml - Check: Graded effort lists on relays should not drop lab levels without peer or host evidence. Why: Final MiniRouter Claude Opus 4.7/4.8 and Sonnet 5 effort sets are none + low/medium/high/xhigh and omit lab max (Anthropic first-party Opus 4.7 / Sonnet 5 include max). Vercel also omits max on some of these, so this may be intentional same-surface copying, but adding none while dropping max is an uncited hybrid of host off-path mapping and peer truncation. Action: Verify against MiniRouter/Vercel live acceptance for these IDs; include max where the upstream accepts it, and only keep none if reasoning.effort=none is a real off path for that model.
[low] [possible mistake] providers/minirouter/models/anthropic/claude-sonnet-4.5.toml - Check: Override-only provider files after base_model. Why: Final Sonnet 4.5 still restates name, structured_output, and a full matching context limit while mainly needing cost + budget_tokens (and any true deltas). Same pattern appears on other files that still carry non-delta description / modality copies. Action: Strip fields identical to the lab base; keep only cost, reasoning_options, and real host deltas.

github-actions · 2026-08-22T20:16:35Z

Action items

[high] [possible mistake] providers/minirouter/models/alibaba/qwen3-235b-a22b-thinking.toml:1 - Check: base_model must identify the actual underlying lab model. Why: This entry points at alibaba/qwen3-235b-a22b-thinking-2507 (text-only lab metadata), while the same Vercel peer slug is authored as a multimodal vision/thinking model (image/pdf). If MiniRouter serves the VL id, the lab link, modalities, and limits are wrong. Action: Confirm the MiniRouter model id against the upstream catalog and either keep qwen3-235b-a22b-thinking-2507 with text-only facts, or point at the correct VL lab model and restore real host deltas only.
[high] [violation] providers/minirouter/models/kwaipilot/kat-coder-pro-v1.toml:3 - Check: Host reasoning / reasoning_options must match this host’s real surface and same-surface peers, not contradictory tag heuristics. Why: Final state sets MiniRouter reasoning = false because /v1/models omits a reasoning tag, while models/kwaipilot/kat-coder-pro-v1.toml is flipped to reasoning = true and the Vercel peer is reasoning = true with reasoning_options = []. For a Vercel-catalog relay, that is an inconsistent capability story. Action: Align MiniRouter with the verified host behavior (likely reasoning = true + peer options/[]), or keep false only with host-specific evidence that this route truly cannot reason—and then do not mark the shared lab model true from the opposite claim.
[high] [violation] providers/minirouter/models/alibaba/qwen-3-14b.toml:2 - Check: Relay reasoning_options baseline is lab + same-surface peers for this host’s wire surface; do not invent a none-only control set. Why: Final MiniRouter uses reasoning_options = [{ type = "effort", values = ["none"] }] (and the same none-only or none+budget pattern across many Vercel toggle peers: Qwen3.x, Claude Haiku/Opus 4.5, Gemini 2.5 Flash, GLM, Kimi, MiniMax M3, Nemotron, etc.). Vercel peers expose real on/off or graded controls (toggle, toggle+budget_tokens, toggle+effort). Advertising only effort=none claims a caller control that cannot turn reasoning on or select graded levels, which misrepresents the relay surface. Action: For each reasoner, copy the intersection of MiniRouter’s admitted wire fields (reasoning.effort, reasoning.max_tokens) with the Vercel peer/lab option set: map peer toggle to effort that includes none plus the peer’s graded values or keep budget_tokens when peers have budgets; use [] only when peers truly have no caller control.
[medium] [violation] providers/minirouter/models/alibaba/qwen3-coder-30b-a3b.toml:3 - Check: Provider reasoning overrides must not contradict established same-surface peers without stronger host evidence. Why: Final MiniRouter forces reasoning = true + reasoning_options = [] on qwen3-coder-30b-a3b, qwen3-coder-next, and qwen3-coder while lab metadata is reasoning = false. Vercel peers also mark these true with [], but MiniRouter earlier documented that /v1/models omits the reasoning tag for these IDs, then removed that evidence and forced true anyway. Action: Either keep lab false and host-only true with concrete MiniRouter proof of reasoning output, or keep true only if the Vercel peer is intentionally mirrored and drop the contradictory “omits reasoning tag” narrative; do not flip shared lab files without first-party lab evidence.
[medium] [possible mistake] models/interfaze/interfaze-beta.toml:1 - Check: Lab metadata must be provider-agnostic facts, not one gateway’s tag dump. Why: The lab file is updated from MiniRouter /v1/models tags (file-input, reasoning, tool-use, vision) after initially matching Vercel, which can bake host-specific capability claims into shared models/. Action: Keep models/interfaze/interfaze-beta.toml to lab-level facts; put MiniRouter-only capability deltas on the provider file only.
[low] [possible mistake] providers/minirouter/models/bytedance/seed-1.8.toml:2 - Check: Provider display name overrides should be real host deltas. Why: MiniRouter renames the model to "Bytedance Seed 1.8" while lab/Vercel use "Seed 1.8", which looks like an accidental rename rather than a host display difference. Action: Drop the name override unless MiniRouter’s public catalog actually uses that string.

github-actions · 2026-08-22T22:07:49Z

Action items

[high] [violation] providers/minirouter/provider.toml:10 - Check: Relay reasoning_options must copy the underlying model’s caller controls from lab + same-surface peers; [] means no control, not a dropped off-path (AGENTS.md → Reasoning options; audit skill anti-pattern “[] on a relay of a controlled reasoner”). Why: MiniRouter’s own header documents admitted reasoning.effort including none, and rejects only reasoning.enabled / enable_thinking. Patch 17 then sets toggle-only Vercel peers to [] and strips both toggle and none from toggle+effort / toggle+budget peers, claiming no representable off control while none remains in the admitted schema. That under-reports off for models such as alibaba/qwen-3-14b, deepseek/deepseek-v3.1, moonshotai/kimi-k2.5, zai/glm-4.5, amazon/nova-2-lite, deepseek/deepseek-v4-pro-0813, anthropic/claude-opus-4.7, and alibaba/qwen3.5-plus (budget kept, off dropped). Action: Restore a representable off path where peers expose one: for toggle-only peers use effort with none (no fake toggle); for toggle+graded effort use none + graded values (no toggle); for toggle+budget keep budget and add none (or document a verified budget-off) rather than budget-only. Do not use [] unless the host truly cannot pass any off/effort control for that model.
[medium] [possible mistake] providers/minirouter/models/alibaba/qwen3-235b-a22b-thinking.toml:3 - Check: base_model must name the actual underlying lab model for this provider ID. Why: Final file sets base_model = "alibaba/qwen3-vl-235b-a22b-thinking" (vision) after earlier commits used alibaba/qwen3-235b-a22b-thinking-2507 (text). The same provider still has qwen3-vl-thinking.toml → the same VL base, while Vercel’s same-ID peer is the non-VL 235B Thinking 2507 entry, and this PR still adds unused models/alibaba/qwen3-235b-a22b-thinking-2507.toml. Wrong identity would publish incorrect modalities/limits/description for alibaba/qwen3-235b-a22b-thinking. Action: Verify MiniRouter’s model id against the catalog/API; if it is the 2507 text reasoner, point base_model at alibaba/qwen3-235b-a22b-thinking-2507 and drop the VL overrides; if it truly is the VL model, keep the VL base, justify the non-VL filename/id, and remove or stop orphaning the unused 2507 lab stub if nothing references it.

github-actions · 2026-08-22T22:27:19Z

Action items

[high] [possible mistake] providers/minirouter/models/mistral/magistral-medium.toml:8946 - Check: Provider base_model must identify the real underlying lab model; override-only files must not invent a different product identity. Why: Final entry uses base_model = "mistral/magistral-medium-latest" while setting name = "Magistral Medium 2509", attachment = true, image input, and limit.output = 64_000. Lab models/mistral/magistral-medium-latest.toml is text-only, attachment = false, output 16_384, and is not a 2509 checkpoint. Same pattern was already corrected for Magistral Small (2506 → 2509). Action: Point at a real Magistral Medium 2509 lab id (add complete models/mistral/… if missing), or drop the 2509 rename/modality overrides and match magistral-medium-latest.
[medium] [possible mistake] providers/minirouter/models/mistral/ministral-8b.toml:9089 - Check: Provider overrides must be real host deltas, not contradictions of lab + same-surface peers. Why: Entry uses base_model = "mistral/ministral-8b-instruct-2410" (lab: text-only, attachment = false) but overrides attachment = true, modalities.input = ["text", "image"], and output = 4_000. Same-ID Vercel peer stays text-only with much larger output. Action: Verify MiniRouter’s catalog/API for this ID; remove unsupported vision/attachment overrides (and fix output) or retarget a multimodal Ministral lab model if that is what is served.
[medium] [possible mistake] providers/minirouter/models/morph/morph-v3-fast.toml:9510 - Check: Limit overrides need evidence when they diverge from lab metadata and the default upstream peer. Why: Final MiniRouter files set context = 81_920 (and large output caps) while new lab files and Vercel peers document 16k/16k (fast) and 32k/32k (large), and the lab description explicitly says 16k I/O. Action: Confirm MiniRouter /v1/models (or docs) for these IDs; if not 81_920, inherit lab limits or set the verified values for both morph-v3-fast and morph-v3-large.
[medium] [possible mistake] models/xai/grok-4.20-multi-agent.toml:6472 - Check: New lab metadata should not contradict established same-model provider facts without rationale. Why: New lab sets tool_call = false, but existing providers/vercel/models/spacexai/grok-4.20-multi-agent.toml and the MiniRouter host files set tool_call = true. Action: Align lab tool_call with first-party/peer evidence, or document why the lab default differs and keep only a real host override.
[low] [possible mistake] models/amazon/nova-lite.toml:5866 - Check: New lab [modalities] should match the model’s real inputs, not gateway guesses. Why: New lab entries for Nova Lite/Pro include pdf (and video), while existing Vercel peers list text/image/video without pdf. MiniRouter inherits the lab modalities after dropping local overrides. Action: Verify Amazon/Vercel modality docs and drop pdf from the lab files (or keep it only as a documented host override) if PDF is not supported.

github-actions · 2026-08-22T22:40:44Z

Action items

[high] [violation] providers/minirouter/models/zai/glm-4.5.toml:4 - Check: Relay reasoning_options must follow lab + same-surface peer controls (AGENTS.md Reasoning options; audit skill §2). Why: MiniRouter rejects reasoning.enabled/enable_thinking, so Vercel toggle-only peers have no representable on/off control. Final files still author reasoning_options = [{ type = "effort", values = ["none"] }] (and the same shape on many peers: glm-4.5-air, glm-4.5v, glm-4.6, glm-4.7*, glm-5, glm-5.1, glm-5-turbo, glm-5v-turbo, kimi-k2.5, kimi-k2.6, deepseek-v3.1, deepseek-v3.1-terminus, qwen-3-14b, qwen-3-30b, seed-1.6, minimax-m3, Nemotron IDs, mimo-v2.5*, laguna-s-2.1, etc.). Vercel peers expose toggle only (no effort enum). Solo none invents an incomplete control (off without an on path) and contradicts the earlier correct provider note that toggle-only peers should use []. Action: Set those toggle-only MiniRouter entries to reasoning_options = [] (always-on / no caller control on this wire). Keep none only where the peer already has graded effort (or a real budget) and MiniRouter’s documented off path is reasoning.effort=none.
[medium] [possible mistake] providers/minirouter/models/morph/morph-v3-fast.toml:1 - Check: Provider overrides must include real deltas claimed from the host catalog. Why: Leading comments state MiniRouter reports context_window=81920 and max_tokens=16384, but the files only set cost and still inherit lab limit 16_000/16_000 (morph-v3-large same). Catalog consumers will publish the wrong limits. Action: Add matching [limit] overrides on both Morph entries, or remove the limit claims from the comments if the lab values are intentional.
[medium] [possible mistake] providers/minirouter/models/mistral/ministral-8b.toml:1 - Check: Host-reported modalities/limits must be authored as overrides. Why: Comments claim MiniRouter reports vision, context_window=128000, and max_tokens=4000, but the file only sets attachment = true while still inheriting lab text-only modalities and output = 8_192. That is internally inconsistent with the stated MiniRouter catalog. Action: Override [modalities].input and [limit] to match MiniRouter, or drop the vision/max_tokens=4000 claims if they are not accurate for this ID.
[medium] [possible mistake] providers/minirouter/models/zai/glm-5.3.toml:6 - Check: Do not invent effort levels beyond lab/same-surface peers without host evidence. Why: Final values are ["none", "low", "high", "max"] while the Vercel peer is ["low", "high", "max"] with no toggle and no none. Same class of unforced none addition appears on other graded-only peers where Vercel has no off control. Action: Drop none unless MiniRouter docs/tests show a meaningful off path for these IDs; otherwise match the Vercel peer effort list exactly.

github-actions · 2026-08-22T22:52:41Z

Action items

[high] [possible mistake] providers/minirouter/models/alibaba/qwen3-235b-a22b-thinking.toml:4 - Check: base_model must identify the underlying lab model for this provider ID. Why: This file sets base_model = "alibaba/qwen3-vl-235b-a22b-thinking" (vision lab entry) and adds PDF input, while the same relative ID on the stated same-surface peer (providers/vercel/models/alibaba/qwen3-235b-a22b-thinking.toml) is authored as non-VL “Qwen3 235B A22B Thinking 2507”. The PR also deleted models/alibaba/qwen3-235b-a22b-thinking-2507.toml. If the MiniRouter ID is the 2507 thinking checkpoint, the VL base is wrong and inherited modalities/capabilities will be wrong. Action: Confirm the MiniRouter model identity against the live catalog; point base_model at the correct lab entry (re-add a complete models/alibaba/… file if needed) and drop incorrect modality overrides.
[high] [possible mistake] providers/minirouter/models/alibaba/qwen3.5-plus.toml:5 - Check: Relay reasoning_options must match this host’s real controls and the lab/same-surface baseline (not invented effort). Why: Final MiniRouter entries for toggle+budget peers (Qwen 3.5/3.6/3.7 budget models, Claude Haiku/Opus 4.5, Gemini 2.5 Flash/Flash-Lite, etc.) advertise effort with values = ["none"] plus budget_tokens, while Vercel peers use toggle + budget_tokens and do not expose an effort enum for those models. MiniRouter documents rejecting reasoning.enabled / enable_thinking, but that alone does not prove reasoning.effort=none is a meaningful off-path for Claude extended thinking / Qwen thinking_budget routes through the gateway. Action: Verify on MiniRouter that reasoning.effort=none actually disables reasoning for each of these families; if it does not, remove the synthetic effort=["none"] and keep only real budget (or [] when there is no caller control). Apply the same fix consistently across the affected files.
[medium] [possible mistake] providers/minirouter/models/amazon/nova-2-lite.toml:1 - Check: Provider overrides should reflect this host and established same-surface peers. Why: After switching to base_model = "amazon/nova-2-lite", MiniRouter no longer sets tool_call, so it inherits the new lab value tool_call = true, while the Vercel peer for the same ID sets tool_call = false. If MiniRouter is a Vercel-catalog gateway, this likely overstates tool calling. Action: Confirm MiniRouter/Vercel tool-use for Nova 2 Lite and set an explicit tool_call override if the host does not support tools.
[medium] [possible mistake] providers/minirouter/models/interfaze/interfaze-beta.toml:3 - Check: Provider capability overrides must not contradict lab/peer facts without solid host evidence. Why: The provider file forces attachment = true and tool_call = true, while the lab entry added in this PR ends at attachment = false / tool_call = false and the Vercel peer matches those false values. Catalog tag comments are thin evidence against the shared lab baseline. Action: Reconcile MiniRouter vs lab/Vercel (either prove MiniRouter really accepts files/tools and keep overrides, or drop the overrides and align the lab entry if the catalog was wrong).
[low] [possible mistake] models/kwaipilot/kat-coder-pro-v1.toml:8 - Check: Lab reasoning must be a provider-agnostic fact, not a single-host tag. Why: This PR flips lab reasoning from false to true based on Vercel/MiniRouter metadata, while other KAT coder lab entries and host behavior may not agree. That changes every future host that inherits the lab file. Action: Cite first-party KAT/Kwai docs (or clear multi-host agreement) for always-on reasoning; otherwise keep lab reasoning = false and only override on hosts that truly reason.

github-actions · 2026-08-22T23:16:15Z

Action items

[high] [possible mistake] models/kwaipilot/kat-coder-pro-v1.toml:11 - Check: Lab reasoning must match the model’s real capability and established same-ID peers when authoring shared models/ metadata. Why: This PR sets lab reasoning = false and drops provider reasoning_options, while the same-ID Vercel peer is reasoning = true with reasoning_options = []. That makes the shared lab entry disagree with the established relay catalog MiniRouter claims to mirror, and any future base_model consumer will inherit non-reasoning. Action: Reconcile with first-party/peer evidence: if the model reasons (even with no controls), set lab reasoning = true and author MiniRouter reasoning_options = []; if it truly does not reason, keep lab false and explain why Vercel’s reasoning = true is wrong.
[medium] [violation] models/interfaze/interfaze-beta.toml:7 - Check: New lab entries must be internally consistent (attachment vs [modalities].input). Why: Final lab metadata keeps attachment = false and tool_call = false while advertising input = ["text", "image", "pdf"]. Image/PDF inputs require attachment = true. MiniRouter then overrides both flags to true, which papers over broken lab metadata instead of fixing it. Action: Set lab attachment = true (and set tool_call to the host-agnostic truth). Keep only real MiniRouter deltas on the provider file.
[medium] [possible mistake] providers/minirouter/models/meta/llama-3.3-70b.toml:3 - Check: Provider overrides after base_model must be real host deltas, not accidental capability regressions. Why: MiniRouter sets attachment = false while lab meta/llama-3.3-70b-instruct and the Vercel same-ID entry are attachment = true. Unless MiniRouter actually rejects attachments for this ID, this is a false capability downgrade. Action: Verify against MiniRouter /v1/models (or request behavior). If attachments work, drop the override; if not, keep it and cite the host limitation in a leading comment.
[medium] [possible mistake] providers/minirouter/models/mistral/ministral-8b.toml:4 - Check: Multimodal/limit overrides on a text-only lab base need a correct identity or proven host delta. Why: base_model = "mistral/ministral-8b-instruct-2410" is text-only (attachment = false, input = ["text"]), but MiniRouter forces attachment = true, input = ["text", "image"], and output = 4_000. That is a large modality/limit rewrite on a non-vision lab id; if the served model is a different Ministral SKU, the base is wrong. Action: Confirm the exact upstream model MiniRouter serves for mistral/ministral-8b. Point base_model at the correct lab entry (adding one if needed), or keep these overrides only with clear host evidence that this ID is vision-capable with a 4k max output.
[medium] [possible mistake] providers/minirouter/models/bytedance/seed-1.6.toml:2 - Check: base_model must identify the underlying lab model, not a nearby text-only sibling. Why: MiniRouter uses bytedance-seed/seed-1-6 (lab is text-only) while setting attachment = true and input = ["text", "image"]. Repo already has models/bytedance-seed/seed-1-6-vision.toml for the multimodal Seed 1.6 line. Action: If this ID is the vision model, switch base_model to bytedance-seed/seed-1-6-vision and drop redundant multimodal overrides; if it is truly text Seed 1.6 with host-side vision, keep the overrides and document that host-specific delta.

github-actions · 2026-08-22T23:31:12Z

Action items

[high] [violation] providers/minirouter/models/bytedance/seed-1.6.toml:1 - Check: Non-lab hosts must base_model the real underlying lab model; overrides may only express true host deltas. Why: Final mapping uses bytedance-seed/seed-1-6-vision (reasoning = false, vision SKU) then forces reasoning = true and reasoning_options = []. That invents a hybrid identity: lab description/dates/family come from the non-reasoning vision entry while the host claims a reasoner. Same-surface Vercel bytedance/seed-1.6 is a multimodal reasoner with toggle; lab already has bytedance-seed/seed-1-6 (reasoning = true, text-only). Action: Point base_model at bytedance-seed/seed-1-6 (or a complete multimodal-reasoning lab entry if that is the real SKU), keep only real host deltas (attachment/modalities/cost/reasoning_options), and stop re-enabling reasoning on the vision lab row.
[high] [possible mistake] providers/minirouter/models/mistral/ministral-8b.toml:1 - Check: base_model must identify the same lab model the host serves. Why: File points at mistral/ministral-8b-instruct-2410 (text-only, output 8_192) while overriding attachment = true, modalities.input = ["text","image"], context = 128_000, output = 4_000. Same-ID Vercel peer is text-only. Forcing vision onto the 2410 instruct lab identity looks like a wrong base or an uncatalogued multimodal SKU. Action: Verify the exact upstream model MiniRouter serves for this ID; if it is not 2410-instruct, add/use the correct lab entry. If it is 2410, drop the vision/attachment overrides or prove this host alone exposes image input for that model.
[medium] [violation] providers/minirouter/models/deepseek/deepseek-v3.1.toml:1 (and same pattern on other former toggle-only peers: deepseek-v3.1-terminus, moonshotai/kimi-k2.5, moonshotai/kimi-k2.6, minimax/minimax-m3, nvidia/nemotron-*, zai/glm-*, alibaba/qwen-3-14b, alibaba/qwen-3-30b, etc.) - Check: On multi-model relays, reasoning_options = [] means no caller control and must not be used from uncertainty when lab/same-surface peers expose controls (AGENTS.md Reasoning options; audit skill). Why: MiniRouter is an OpenAI-compatible gateway over the Vercel catalog. Vercel peers for these IDs advertise toggle (or toggle+budget). Final MiniRouter files set [] because the host rejects reasoning.enabled/enable_thinking, without documenting a remaining on/off path. For hybrid reasoners (e.g. DeepSeek V3.1), that drops the only peer control rather than mapping an admitted wire field. Graded-effort peers correctly keep levels (and often none); toggle-only peers lose enable/disable entirely. Action: For each former toggle peer, either (1) document and author the real MiniRouter off/on mapping that still works (e.g. reasoning.effort=none vs omit/default, or budget), matching lab/Vercel capability, or (2) keep [] only with host evidence that callers truly cannot control reasoning on that ID—not merely that one toggle field name is rejected.
[medium] [possible mistake] providers/minirouter/models/mistral/mistral-nemo.toml:1 - Check: Override-only base_model files must not invent modalities/metadata the host does not serve. Why: Final file still overrides description, release_date, knowledge, and modalities.input = ["text","image"] on lab mistral/mistral-nemo (text-only, open weights). Image input is not established on the lab entry and conflicts with open-weight Nemo instruct identity. Action: Confirm MiniRouter actually accepts images for this ID; if not, drop the modality/description/date overrides and inherit the lab row.
[low] [possible mistake] providers/minirouter/models/alibaba/qwen3-coder-30b-a3b.toml:1 (also qwen3-coder-next.toml, qwen3-coder.toml) - Check: Provider reasoning overrides should be real host deltas, not copied peer mistakes. Why: Lab entries are reasoning = false; MiniRouter sets reasoning = true + [] “mirroring Vercel.” Vercel’s same-ID files are themselves full inline reasoners without lab base_model, so copying that does not prove MiniRouter (or the model) is a reasoner. Action: Confirm MiniRouter/catalog tags or live behavior show reasoning for these coder IDs; if not, drop the reasoning = true override and inherit lab false.

github-actions · 2026-08-22T23:46:56Z

Action items

[high] [violation] providers/minirouter/models/poolside/laguna-s-2.1.toml:2 - Check: Relay reasoning_options must copy lab/same-surface peer controls the host can expose; [] means no caller control, not uncertainty. Why: Final file keeps reasoning_options = [] while the same-ID Vercel peer is toggle-only, MiniRouter admits reasoning.effort=none, and sibling MiniRouter toggle peers were restored to { type = "toggle" } with a wire comment. Callers cannot discover the off path. Action: Set reasoning_options to toggle (with a leading # Toggle: reasoning.effort=none … wire comment), matching other MiniRouter toggle-only entries.
[high] [violation] providers/minirouter/models/nvidia/nemotron-3.5-lightning.toml:4 - Check: Same relay reasoning-options rule as above. Why: Left as [] after the toggle restore pass; Vercel peer is toggle-only and other MiniRouter Nemotron entries now advertise toggle via reasoning.effort=none. Action: Restore toggle + leading wire comment, consistent with nemotron-3-nano-30b-a3b, nemotron-nano-9b-v2, etc.
[high] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:4 - Check: When peers expose toggle + budget and this host admits reasoning.effort=none, the relay must keep an off control (toggle, or effort including none with budget—not budget alone). Why: Final shape is budget-only after dropping effort=none; same-surface Vercel peer is toggle + budget_tokens, and MiniRouter’s own schema/docs accept reasoning.effort=none. Same broken pattern on the other budget-only MiniRouter files that still omit off: qwen-3-32b, qwen-3.6-max-preview, qwen3.5-flash, qwen3.6-27b, qwen3.6-plus, qwen3.7-max, qwen3.7-plus, claude-haiku-4.5, claude-opus-4.5, gemini-2.5-flash, gemini-2.5-flash-lite (and any other peer toggle+budget rows reduced to budget-only). Action: For each, restore the peer off path MiniRouter can wire (toggle + budget with a toggle wire comment, or effort values including none + budget—without pairing toggle with none in the same effort list).
[medium] [possible mistake] providers/minirouter/models/alibaba/qwen3-coder-next.toml:4 - Check: Host capability should match the upstream/same-ID peer when MiniRouter is a Vercel AI Gateway relay with the same model IDs. Why: These entries inherit lab reasoning = false (and drop host overrides), while providers/vercel/models/alibaba/qwen3-coder-next.toml (and the matching coder-30b / coder-480b Vercel rows) set reasoning = true with reasoning_options = []. If MiniRouter truly mirrors that catalog, the resolved MiniRouter models under-report reasoning. Action: Verify against MiniRouter/Vercel for qwen3-coder-next, qwen3-coder-30b-a3b, and qwen3-coder; if they are reasoning-capable on this path, set reasoning = true and reasoning_options = [] (or the real controls) on the provider files.

github-actions · 2026-08-23T00:06:11Z

Action items

[high] [violation] providers/minirouter/models/alibaba/qwen-3-14b.toml:1 - Check: Reasoning options must match this host’s real wire controls (AGENTS.md → Reasoning options / toggle rules). Why: Final provider.toml and file comments state MiniRouter’s off path is reasoning.effort=none (and on is omit/default). That is an effort control, not a separate boolean toggle field. Authoring type = "toggle" for that path mis-describes the API and will drive incorrect client request shaping. The same pattern is applied across the restored toggle-only set (e.g. qwen-3-30b, seed-1.6, deepseek-v3.1*, kimi-k2.5/kimi-k2.6, Nemotron, Xiaomi, ZAI GLM toggle peers, etc.) and toggle+budget peers (qwen3.5-*, claude-haiku-4.5, claude-opus-4.5, gemini-2.5-flash*). Action: For every model whose only off path is reasoning.effort=none, author effort with none (alone for binary off, or none plus the verified graded levels). Reserve type = "toggle" only for a distinct on/off wire field. Align provider.toml comments with that encoding.
[high] [violation] providers/minirouter/models/bytedance/seed-1.8.toml:1 - Check: Relay baseline = same-surface peer option set for this model. Why: Final MiniRouter entry is effort-only none|minimal|low|medium|high. The same-ID Vercel peer (MiniRouter’s default upstream) is toggle + minimal|low|medium|high (off not via none). After re-encoding MiniRouter’s admitted surface, either keep graded levels and represent off as none without inventing levels Vercel does not use, or document host-specific proof if MiniRouter truly drops the peer toggle while adding none. As authored, the control shape is neither the Vercel peer nor a clean lab mapping. Action: Re-baseline Seed 1.8 (and other toggle+graded Vercel peers such as Nova 2 Lite / Claude 4.6–4.8 / Fable 5 / GLM 5.2) to the peer graded set plus MiniRouter’s real off encoding; do not mix peer toggle semantics with an undocumented effort enum.
[medium] [possible mistake] providers/minirouter/models/spacexai/grok-4.3.toml:1 - Check: Relays must not use reasoning_options = [] when lab/peer expose caller controls and this host admits the same field. Why: MiniRouter admits reasoning.effort, and first-party xAI entries expose graded effort for Grok 4.3 (none|low|medium|high) and 4.5 (low|medium|high). Final MiniRouter files set [] (matching Vercel’s empty catalog rows). Empty means no caller control, not “catalog tag missing.” If the gateway forwards effort for these IDs, [] under-reports controls. Action: Verify whether MiniRouter/Vercel actually accept effort for spacexai/grok-4.3 and spacexai/grok-4.5; if yes, copy the lab/peer effort lists; if no, keep [] and cite host-level rejection evidence in a leading comment.
[medium] [possible mistake] providers/minirouter/models/deepseek/deepseek-v4-pro-0813.toml:1 - Check: DeepSeek V4 effort baseline is lab high|max (or same-surface peer), not an invented synonym set. Why: Final MiniRouter uses none|high|xhigh. Lab first-party is toggle + high|max; Vercel peer is toggle + high|xhigh. xhigh may be the gateway synonym for max, but shipping xhigh without max (and without documenting the mapping) can break clients that send lab-native max. Action: Confirm the values MiniRouter/Vercel accept for these IDs; author the verified set and, if xhigh is gateway-only, note the lab max mapping in a leading comment.
[low] [possible mistake] providers/minirouter/models/bytedance/seed-1.6.toml:1 - Check: Provider overrides of attachment/modalities must be complete and consistent. Why: The file sets attachment = true and comments claim image+text input as a host delta from text-only bytedance-seed/seed-1-6, but the inherited lab [modalities].input is ["text"] unless an explicit modalities override remains in the final file. Attachment without image/pdf input is internally inconsistent. Action: Ensure final Seed 1.6 (and Seed 1.8) keep a real [modalities] input = ["text", "image"] (or whatever the host catalog proves) alongside attachment = true, or drop the attachment override.

github-actions · 2026-08-23T00:17:38Z

Action items

[medium] [violation] providers/minirouter/models/alibaba/qwen-3-14b.toml:5 - Check: Binary on/off reasoning controls must be authored as toggle (AGENTS.md → Reasoning options; audit skill §3). Why: Final encoding uses effort with only values = ["none"] (and the same shape on the other binary-off models in patch 25: Qwen 3.x, DeepSeek V3.1*, MiniMax M3, Kimi K2.5/2.6, Nemotron, Laguna, MiMo, GLM toggle-only rows, etc.). That is not a valid schema shape for binary control: effort is for graded levels (optionally including none beside other values); binary on/off is toggle, with the wire path in a leading comment. Same-ID Vercel peers correctly use toggle for these IDs. Clients that key off toggle will miss the off control. Action: For every binary-off MiniRouter model, replace effort = ["none"] with { type = "toggle" } and a top-of-file wire comment such as # Toggle: reasoning.effort=none (off); omit reasoning (on). Keep effort only when there are graded levels (with none in that list when off is effort-based).
[medium] [violation] providers/minirouter/models/alibaba/qwen3.5-plus.toml:7 - Check: Separate on/off + reasoning budget must be toggle + budget_tokens, not effort=["none"] + budget. Why: Final Qwen/Claude/Gemini budget rows (e.g. qwen3.5-plus, qwen-3-32b, qwen-3.6-max-preview, qwen3.6-*, qwen3.7-max/plus, claude-haiku-4.5, claude-opus-4.5, gemini-2.5-flash(-lite)) pair lone effort=["none"] with budget_tokens. Lab and Vercel peers for these IDs use toggle + budget_tokens. Action: Change the off control to toggle (with leading wire comment) and keep the existing budget_tokens bounds; do not leave effort with only none.
[low] [possible mistake] providers/minirouter/models/alibaba/qwen3-coder-next.toml:5 - Check: Provider reasoning overrides vs lab metadata must reflect this host, not only an upstream peer. Why: Lab alibaba/qwen3-coder-next (and the 30B/480B coder labs) has reasoning = false, but MiniRouter forces reasoning = true + reasoning_options = [] citing the Vercel peer. That can be correct if MiniRouter/Vercel actually serve a reasoning path, but it contradicts the shared lab fact and the earlier frozen MiniRouter snapshot comments that omitted reasoning for these IDs. Action: Confirm on MiniRouter (or the frozen catalog) that these coder IDs are reasoning-capable with no controls; if not, drop the reasoning = true override and inherit the lab false.

github-actions · 2026-08-23T00:28:20Z

I have enough context on the final reasoning shapes; checking a few remaining high-risk final-state details before writing the review.Checking DeepSeek V4 final options and whether any models still lack base_model.Inspecting final states for DeepSeek V4, Claude Opus 4, and remaining non-base_model entries.Checking lab reasoning flags and a few final peer mismatches that may still be blockers.Reading DeepSeek V4 and related final hunks.Checking lab metadata and final provider configuration against AGENTS requirements.Inspecting seed-1.8 and DeepSeek V4 evolution, plus whether graded models still pair toggle with none.Gathering final states for the remaining contested models.Verifying a few lab names and peer reasoning flags against final provider entries.Checking lab metadata and Vercel peers for remaining mismatches.## Action items

[medium] [possible mistake] providers/minirouter/models/alibaba/qwen3.8-max.toml:1 - Check: Provider entries must point base_model at the correct shared lab identity. Why: This host id is qwen3.8-max but base_model is alibaba/qwen3.8-max-preview while models/alibaba/qwen3.8-max.toml also exists, so capabilities/limits may follow the preview row instead of the GA max row. Action: Confirm which lab identity MiniRouter actually serves for alibaba/qwen3.8-max; if it is the GA model, set base_model = "alibaba/qwen3.8-max" and keep only real host deltas.
[medium] [possible mistake] providers/minirouter/models/alibaba/qwen3-coder.toml:1 - Check: Relay reasoning capability should match this host’s real surface and same-ID peers, not an under-specified catalog tag alone. Why: Final files drop reasoning/reasoning_options so these IDs inherit non-reasoning lab metadata, while the same-ID Vercel peer (MiniRouter’s stated upstream) treats the coder IDs as reasoning models with reasoning_options = [] (always-on / no caller control). Mis-tagging hides reasoning behavior for clients. Action: Verify live MiniRouter/Vercel behavior for qwen3-coder, qwen3-coder-next, and qwen3-coder-30b-a3b; if they reason, set reasoning = true and author the correct host controls ([] only if there is truly no caller control).
[medium] [violation] providers/minirouter/models/alibaba/qwen-3-30b.toml:3 - Check: After base_model, provider files must be override-only (no restated identical lab fields). Why: Final entries still restate name (and in other files description / unchanged capability flags) beside base_model even when those values are the shared lab defaults, which breaks the override-only rule and drifts from lab metadata over time. Same pattern appears on multiple MiniRouter rows (e.g. Haiku/Opus display names, DeepSeek descriptions). Action: Diff each base_model file against its lab target and delete fields that are identical after merge; keep only cost, reasoning_options, interleaved, status, and genuine host deltas.
[low] [possible mistake] providers/minirouter/models/deepseek/deepseek-v3.2.toml:1 - Check: Provider overrides of reasoning / tool_call must reflect this host, not an accidental inverse of the lab row. Why: The entry forces reasoning = false and tool_call = false on deepseek/deepseek-v3.2 while a separate deepseek-v3.2-thinking id exists; if MiniRouter’s non-suffix id is still the hybrid/toggle model (as on first-party/peers), these overrides are wrong. Action: Confirm MiniRouter’s deepseek/deepseek-v3.2 request surface; remove the false overrides if the id still supports reasoning and/or tools, or document the host-specific non-reasoning pin in a leading comment with evidence.

github-actions · 2026-08-23T00:47:24Z

Action items

[high] [violation] providers/minirouter/provider.toml:13 - Check: Reasoning options — when off is reasoning.effort=none, author effort with none, not toggle. Why: The provider header and dozens of model files encode binary off as type = "toggle" while documenting the wire path as reasoning.effort=none (on = omit reasoning). AGENTS.md / the reasoning audit skill require that shape to be effort with none in values (alone, or beside graded levels/budget). toggle is only for a separate on/off request field. This mislabels the control surface for clients. Action: Revert the patch-26 toggle encoding: set binary-off models to reasoning_options = [{ type = "effort", values = ["none"] }]; for budget/graded peers use effort including none plus those controls, with no toggle. Update the provider.toml comment to match.
[high] [violation] providers/minirouter/models/alibaba/qwen-3-14b.toml:1 - Check: Same effort-vs-toggle rule on all MiniRouter models whose header is Toggle: reasoning.effort=none. Why: Representative of the systematic set (Qwen 3.x toggle/budget rows, Claude Haiku/Opus 4.5, DeepSeek V3.1*, MiniMax M3, Moonshot K2.5/K2.6, Nemotron*, Xiaomi MiMo*, ZAI GLM*, Poolside Laguna, Seed 1.6, etc.). Each advertises toggle for an effort-value off path. Action: Replace those toggle entries with effort/none (and keep budget/graded values where already verified); keep leading comments as exact wire docs for reasoning.effort.
[medium] [possible mistake] providers/minirouter/models/alibaba/qwen3-coder-next.toml:1 - Check: Relay reasoning capability vs same-ID Vercel peer / lab base_model. Why: Final files drop host reasoning = true and inherit lab reasoning = false for qwen3-coder-next, qwen3-coder-30b-a3b, and qwen3-coder, while providers/vercel/models/alibaba/qwen3-coder*.toml mark the same IDs reasoning = true with reasoning_options = []. MiniRouter is described as a Vercel AI Gateway catalog relay; inheriting non-reasoning understates the peer surface unless this host truly never reasons. Action: Verify live/frozen MiniRouter behavior for these IDs. If they reason like the Vercel peer, set reasoning = true and reasoning_options = [] (with a short source comment). If they truly never reason, keep lab inheritance and cite host evidence that contradicts the Vercel peer.
[medium] [possible mistake] providers/minirouter/models/bytedance/seed-1.6.toml:1 - Check: Multimodal host vs text-only base_model. Why: Host sets attachment = true and [modalities].input = ["text", "image"] on base_model = "bytedance-seed/seed-1-6" (lab is text-only). A vision lab identity bytedance-seed/seed-1-6-vision exists; the PR briefly pointed there then reverted. Same pattern on seed-1.8 (text lab + image override). Action: Confirm whether MiniRouter’s Seed 1.6/1.8 IDs are the vision/reasoning variants. If yes, point base_model at the correct lab identity (or extend lab metadata) and keep only true host deltas; if text-only, drop the image/attachment overrides.
[low] [possible mistake] providers/minirouter/models/spacexai/grok-4.3.toml:1 - Check: Relay baseline = lab + same-surface peers. Why: Final Grok 4.3/4.5 restore first-party xAI effort lists, but the same-ID Vercel peers use reasoning_options = []. If MiniRouter only forwards the Vercel path and does not honor reasoning.effort for these IDs, the graded lists overstate caller control. Action: Confirm MiniRouter actually applies effort for these Grok IDs; if not, match the Vercel peer [] (always-on / no control).

github-actions · 2026-08-23T00:56:39Z

Action items

[high] [violation] providers/minirouter/models/alibaba/qwen-3-14b.toml:1 - Check: Binary on/off reasoning must be authored as toggle (with a leading wire comment), not effort containing only none (AGENTS.md → Reasoning options; audit skill §3). Why: Final patch encodes dozens of binary MiniRouter off-paths as reasoning_options = [{ type = "effort", values = ["none"] }] (and the same shape on Qwen/Claude/Gemini budget models as effort=["none"] + budget_tokens). That is not a valid control shape: effort with only none does not describe graded levels, and binary on/off is defined as toggle. Same-surface peers (providers/vercel/... for Kimi K2.5, GLM, Qwen3.5 Plus, Claude Haiku 4.5, etc.) and lab entries use toggle (plus budget where applicable). Consumers will mis-handle off/on. Action: For every binary off model, restore [[reasoning_options]] type = "toggle" with a top-of-file comment such as # Toggle: reasoning.effort = none (off); omit reasoning (on). For toggle+budget models (Qwen 3.5/3.6/3.7, Claude Haiku/Opus 4.5, Gemini 2.5 Flash/Lite, etc.), use toggle + budget_tokens only—do not keep a lone effort=["none"] entry. Keep effort lists that include none together with real graded values (e.g. Claude 4.6/4.7/4.8, DeepSeek V4, Seed 1.8, GPT) as effort-only (no toggle). Update providers/minirouter/provider.toml so it no longer documents binary off as effort=[none].
[medium] [possible mistake] providers/minirouter/models/alibaba/qwen3-coder-next.toml:1 - Check: Relay reasoning / reasoning_options must follow lab + same-surface peers unless this host affirmatively lacks the capability (AGENTS.md → Reasoning options §2.4; audit skill anti-pattern “[] / no control from uncertainty”). Why: Final files for qwen3-coder, qwen3-coder-next, and qwen3-coder-30b-a3b inherit lab reasoning = false and drop controls, while the same-ID Vercel peers set reasoning = true with reasoning_options = [] (always-on / no caller control). MiniRouter is described as a Vercel AI Gateway catalog relay; catalog tag absence alone is weak proof that the upstream reasoner is non-reasoning. Action: Align with the Vercel peer (reasoning = true, reasoning_options = []) unless you can show MiniRouter does not forward a reasoning model for these IDs (not merely that the OpenAPI tag list omits reasoning). If host truly strips reasoning, keep reasoning = false and state that stronger evidence in the leading comment.

github-actions · 2026-08-23T01:07:23Z

No actionable findings.

feat(providers): add minirouter

d3c6889

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(minirouter): address provider review feedback

02c0bae

fix(minirouter): address follow-up review

1d2e8f3

fix(minirouter): refine reasoning and model identities

bc33477

docs(minirouter): cite reasoning request schema

ab3cadc

fix(minirouter): expose native reasoning efforts

98259f3

fix(minirouter): narrow model reasoning controls

6515200

fix(minirouter): map relay reasoning controls

7e005bf

fix(minirouter): align Grok and Nova metadata

05e9ed3

fix(minirouter): restore inherited model capabilities

535e556

fix(minirouter): align relay reasoning shapes

4dac764

fix(minirouter): map reasoning off paths

fe56df5

zhe-t added 2 commits August 22, 2026 21:04

fix(minirouter): reconcile peer capabilities

0302a67

fix(models): correct KAT Coder reasoning capability

a537f63

fix(minirouter): align verified relay capabilities

d3682af

fix(minirouter): preserve relay reasoning off controls

0b1a2ad

fix(minirouter): verify disputed model metadata

2c20d5a

fix(minirouter): distinguish relay reasoning controls

c92230e

fix(minirouter): tighten relay capability evidence

dbe5379

fix(minirouter): reconcile relay model identities

3358371

fix(minirouter): restore relay toggle controls

31d445f

fix(minirouter): complete relay off controls

8f78b48

fix(minirouter): encode off controls as effort

1fb0f49

fix(minirouter): separate toggle semantics from wire

f6b018d

fix(minirouter): inherit shared model metadata

a9acfdf

fix(minirouter): encode relay effort controls

8b9bc61

fix(minirouter): align binary and coder reasoning

7c79b0e

github-actions Bot added the reviewer: ready Automated review found no actionable items label Aug 23, 2026

Uh oh!

Conversation

zhe-t commented Aug 22, 2026

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

zhe-t commented Aug 22, 2026

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

zhe-t commented Aug 22, 2026

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

zhe-t commented Aug 22, 2026

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

zhe-t commented Aug 22, 2026

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

zhe-t commented Aug 22, 2026

Uh oh!

zhe-t commented Aug 22, 2026

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 22, 2026

Action items

Uh oh!

github-actions Bot commented Aug 23, 2026

Action items

Uh oh!

github-actions Bot commented Aug 23, 2026