feat(providers): add Parasail serverless inference catalog

[Kaihuang724]

Summary

• Add Parasail as an @ai-sdk/openai-compatible provider backed by https://api.parasail.io/v1.
• Include the serverless model catalog with base_model references to canonical lab metadata.
• Author cost and reasoning_options for publicly priced / controllable SKUs; mark intentionally unpriced entries.

Data sources and claim mapping

Claim	Source
Catalog completeness / model IDs	https://api.parasail.io/v1/models (live list)
Missing/correct lab metadata	models/ directory + first-party lab entries
GPT-OSS reasoning surface	reasoning_effort = low|medium|high (per OpenAI GPT-OSS docs / peers)
DeepSeek V4 reasoning surface	thinking.type = enabled|disabled, reasoning_effort (DeepSeek first-party + OpenRouter peers)
Qwen3.5 reasoning surface	enable_thinking / thinking_budget (Qwen first-party + OpenRouter peers)
Qwen3.8-27B reasoning surface	enable_thinking / reasoning_effort = low|medium|xhigh (OpenRouter peers)
Kimi K3 reasoning surface	thinking.type + output_config.effort = low|high|max (OpenRouter peers)
Kimi K2.7 Code	Always-on; no caller control (Moonshot first-party + OpenRouter peers)
GLM-5.1/5.2, MiMo-V2.5, MiniMax-M3	thinking.type = enabled|disabled (first-party docs + OpenRouter peers)
Gemma 4 reasoning surface	Toggle (thinking.type) per Google first-party + OpenRouter peers
Gemma-3 on Parasail modalities	Prior full Parasail definition was text-only/attachment=false; overridden in provider files
Peer-sourced USD/MTok costs	OpenRouter endpoints (cited in leading TOML comments) where Parasail pricing is not published

Intentionally unpriced on Parasail (public pricing not available)

• GPT-OSS 120B Fast
• Resemble TTS English
• MythoMax 13B
• Sao10K Lunaris 8B

Provider changes

• Converted all nameable SKUs to base_model references and created missing lab stubs where needed.
• Removed BGE-M3 (embedding SKU outside the chat catalog semantics).
• Dropped duplicate/wrong models/muse/glimmer-30b and models/mlabonne/unslopnemo-12b stubs in favor of meta/muse-glimmer-30b and thedrummer/unslopnemo-12b.
• Updated reasoning_options and wire comments to match the lab/peer control surface.
• Added cost from OpenRouter peers for unpriced public SKUs with a leading source comment.

Test plan

• bun validate passes
• CI validation passes

Generated with Devin

[github-actions]

Action items

• [high] [violation] providers/parasail/models/parasail-gpt-oss-120b.toml:1 - Check: Non-lab hosts must use base_model when the lab model is nameable. Why: Parasail did not create GPT-OSS; models/openai/gpt-oss-120b.toml already exists, but this file is a full inline definition. Action: Switch to base_model = "openai/gpt-oss-120b" and keep only Parasail deltas (cost, reasoning_options, real limit overrides).
• [high] [violation] providers/parasail/models/parasail-gpt-oss-20b.toml:1 - Check: Non-lab hosts must use base_model. Why: Same issue for GPT-OSS 20B; lab metadata is already at models/openai/gpt-oss-20b.toml. Action: Use base_model = "openai/gpt-oss-20b" with override-only provider fields.
• [high] [violation] providers/parasail/models/parasail-deepseek-v32.toml:1 - Check: Non-lab hosts must use base_model. Why: DeepSeek V3.2 is a lab model (models/deepseek/deepseek-v3.2.toml, reasoning = true, tool_call = true), but this entry fully inlines and sets reasoning = false / tool_call = false, dropping shared lab facts. Action: Use base_model = "deepseek/deepseek-v3.2" and only override true Parasail deltas; set reasoning_options from lab/peer DeepSeek V3.2 controls if reasoning is exposed.
• [high] [violation] providers/parasail/models/parasail-qwen3-coder-next.toml:1 - Check: Non-lab hosts must use base_model. Why: Lab metadata already exists at models/alibaba/qwen3-coder-next.toml. Action: Use base_model = "alibaba/qwen3-coder-next" and keep only cost/limit overrides.
• [high] [violation] providers/parasail/models/parasail-qwen3-vl-235b-a22b-instruct.toml:1 - Check: Non-lab hosts must use base_model. Why: Lab metadata already exists at models/alibaba/qwen3-vl-235b-a22b-instruct.toml. Action: Use that base_model and keep only provider-specific fields.
• [high] [violation] providers/parasail/models/parasail-gemma3-27b-it.toml:1 - Check: If the lab model is nameable, add models/ and point base_model at it. Why: Gemma 3 27B IT is a Google model already cataloged on multiple hosts, but there is no models/google/… entry and this file is full inline. Action: Add a complete models/google/gemma-3-27b-it.toml (or the correct lab id) and make this provider file override-only.
• [high] [violation] providers/parasail/models/parasail-qwen3vl-8b-instruct.toml:1 - Check: Nameable third-party models need lab metadata + base_model. Why: Qwen3 VL 8B Instruct is an Alibaba model served elsewhere, but no models/alibaba/ entry is added and the provider file is fully inlined. Action: Add complete lab metadata under models/alibaba/ and reference it via base_model.
• [high] [violation] providers/parasail/models/parasail-gpt-oss-120b.toml:14 - Check: Do not invent budget_tokens unless the host exposes a real reasoning-budget field. Why: Established gpt-oss peers (Groq, OpenRouter, Together, etc.) use effort low/medium/high only; no reasoning budget. Action: Drop budget_tokens on both gpt-oss files; keep effort ["low", "medium", "high"] unless Parasail docs prove a budget API.
• [high] [violation] providers/parasail/models/parasail-gpt-oss-20b.toml:14 - Check: Same budget_tokens policy for gpt-oss. Why: Same invented budget control as the 120B entry. Action: Remove budget_tokens; author effort-only options matching peers.
• [high] [violation] providers/parasail/models/parasail-gemma-4-26b-a4b-it.toml:4 - Check: On relays, [] means no caller control, not uncertainty; copy lab/peer controls when they exist. Why: First-party Google and OpenRouter expose toggle for Gemma 4 IT reasoners; empty options misrepresent the control surface. Action: Replace [] with lab/peer options (at least { type = "toggle" }) if Parasail forwards the control, or document/prove no control before keeping [].
• [high] [violation] providers/parasail/models/parasail-gemma-4-31b-it.toml:4 - Check: Relay reasoning baseline must match lab/peers. Why: Same empty-[] issue as Gemma 4 26B; lab/peers use toggle. Action: Align reasoning_options with Google/OpenRouter Gemma 4 IT controls, or prove Parasail has none.
• [high] [violation] providers/parasail/models/parasail-glm-5.toml:4 - Check: Relay must not use [] when the lab exposes controls. Why: First-party Zhipu glm-5 uses toggle. Action: Copy lab toggle (with wire comment) if forwarded; do not leave [] from incomplete verification.
• [high] [violation] providers/parasail/models/parasail-glm-51.toml:4 - Check: Same relay reasoning policy for GLM-5.1. Why: Lab glm-5.1 is toggle. Action: Match lab toggle or prove no control on Parasail.
• [high] [violation] providers/parasail/models/parasail-glm47.toml:4 - Check: Same relay reasoning policy for GLM-4.7. Why: Lab glm-4.7 is toggle. Action: Match lab toggle or prove no control on Parasail.
• [high] [violation] providers/parasail/models/parasail-kimi-k25.toml:4 - Check: Same relay reasoning policy for Kimi K2.5. Why: First-party Moonshot exposes toggle. Action: Use toggle (with wire comment) if Parasail supports it; do not author [] from uncertainty.
• [high] [violation] providers/parasail/models/parasail-kimi-k26.toml:4 - Check: Same relay reasoning policy for Kimi K2.6. Why: Lab entry is toggle. Action: Align with Moonshot/peers or prove no control.
• [high] [violation] providers/parasail/models/parasail-mimo-v25.toml:4 - Check: Same relay reasoning policy for MiMo V2.5. Why: First-party Xiaomi exposes toggle. Action: Match lab toggle or prove Parasail has no caller control.
• [high] [violation] providers/parasail/models/parasail-qwen3p6-35b-a3b.toml:4 - Check: Same relay reasoning policy for Qwen3.6 35B-A3B. Why: Alibaba first-party uses toggle + budget_tokens; empty [] contradicts the lab baseline. Action: Copy Alibaba/same-surface options Parasail actually exposes; do not use [] without affirmative no-control evidence.
• [high] [violation] providers/parasail/models/parasail-qwen3-235b-a22b-instruct-2507.toml:1 - Check: base_model must point at the correct underlying model; reasoning options must match that identity. Why: File points at alibaba/qwen3-235b-a22b (reasoning = true) and sets reasoning_options = [], but instruct-2507 is the non-thinking instruct variant on established hosts (e.g. OpenRouter reasoning = false). Action: Add/use the correct instruct-2507 lab entry (or override reasoning = false with no reasoning_options); do not inherit the hybrid/thinking base as an always-on reasoner.
• [high] [violation] providers/parasail/models/parasail-deepseek-v4-flash.toml:4 - Check: Every toggle needs a leading top-of-file wire-path comment. Why: Toggle is present with no comment documenting the request field Parasail accepts. Action: Add a leading comment such as # Toggle: … (and effort wire path) matching Parasail’s API.
• [high] [violation] providers/parasail/models/parasail-deepseek-v4-pro.toml:4 - Check: Toggle wire-comment requirement. Why: Same missing leading toggle comment. Action: Add top-of-file wire comments for toggle and effort.
• [high] [violation] providers/parasail/models/parasail-qwen35-397b-a17b.toml:4 - Check: Toggle wire-comment requirement. Why: toggle + budget_tokens without a leading wire comment. Action: Document exact Parasail fields (e.g. enable_thinking / thinking_budget) above the first key.
• [high] [violation] providers/parasail/models/parasail-qwen3p5-35b-a3b.toml:4 - Check: Toggle wire-comment requirement. Why: Same missing leading comment for toggle + budget. Action: Add top-of-file wire comments for the Parasail request syntax.
• [medium] [violation] providers/parasail/models/parasail-cydonia-24-v41.toml:1 - Check: Nameable third-party models should get lab metadata + base_model. Why: Cydonia 24B v4.1 is a known TheDrummer model on other providers, not unique to Parasail, but is fully inlined with no models/ entry. Action: Add complete models/thedrummer/… (or correct lab id) metadata and convert this file to override-only.
• [medium] [violation] providers/parasail/models/parasail-skyfall-36b-v2-fp8.toml:1 - Check: Same base_model rule for Skyfall. Why: Skyfall 36B v2 is also a known TheDrummer model hosted elsewhere, not Parasail-unique. Action: Add lab metadata and use base_model; keep only Parasail cost/limit deltas.
• [low] [possible mistake] .pr-review/pull-request.json - Check: Data-changing PRs should cite direct provider pricing/docs/API evidence. Why: Body claims pricing from Parasail’s OpenRouter-compatible models API and “documented” reasoning controls, but gives no URLs mapped to costs or wire fields. Action: Add first-party Parasail pricing/docs/API links and state what each citation supports (prices, reasoning fields per model family).

[github-actions]

Action items

• [critical] [violation] providers/parasail/models/parasail-gpt-oss-120b.toml:1 - Check: Non-lab hosts must use base_model for nameable lab models. Why: GPT-OSS 120B/20B/120B-Fast are OpenAI open-weight models (models/openai/gpt-oss-120b.toml, gpt-oss-20b.toml), but these files are full inline definitions. That skips shared lab metadata and invites drift (e.g. tool_call = false vs lab true, invented budget_tokens). Action: Switch to base_model = "openai/gpt-oss-…", keep only Parasail deltas (cost, verified reasoning_options, real limit overrides). Treat Fast as a host SKU of the same lab model unless it is a distinct unique-to-host weights/id.
• [critical] [violation] providers/parasail/models/parasail-qwen3-coder-next.toml:1 - Check: Non-lab hosts must use base_model. Why: Lab metadata already exists at models/alibaba/qwen3-coder-next.toml and models/alibaba/qwen3-vl-235b-a22b-instruct.toml, but Parasail redefines both fully. Action: Use base_model = "alibaba/qwen3-coder-next" and base_model = "alibaba/qwen3-vl-235b-a22b-instruct" with override-only fields.
• [critical] [violation] providers/parasail/models/parasail-qwen3-235b-a22b-instruct-2507.toml:1 - Check: base_model must point at the actual underlying lab model. Why: Filename/id is the Instruct-2507 SKU, but base_model = "alibaba/qwen3-235b-a22b" is the different hybrid reasoner (reasoning = true). The correct lab entry is alibaba/qwen3-235b-a22b-instruct-2507 (reasoning = false). Current merge wrongly inherits reasoner capabilities and then authors toggle + budget_tokens. Action: Set base_model = "alibaba/qwen3-235b-a22b-instruct-2507" and drop reasoning_options (invalid when reasoning = false).
• [high] [violation] providers/parasail/models/parasail-gpt-oss-120b.toml:14 - Check: Do not invent budget_tokens without a real reasoning-budget API. Why: GPT-OSS peers (Cerebras, Groq, etc.) expose effort low|medium|high only. budget_tokens is not a standard GPT-OSS control; AGENTS.md forbids inventing it. Same pattern on parasail-gpt-oss-20b.toml and parasail-gpt-oss-120b-fast.toml. Action: Remove budget_tokens unless Parasail docs/tests prove a reasoning-token budget field; keep verified effort values only, with a top-of-file wire comment.
• [high] [violation] providers/parasail/models/parasail-deepseek-v4-flash.toml:5 - Check: Every toggle needs a leading top-of-file wire-path comment. Why: Multiple Parasail reasoners use toggle with no wire comment (deepseek-v4-*, qwen35*, qwen3p5*, qwen3p6*, qwen38*). Sync/readers cannot tell the request field. Action: Add a leading comment naming the exact Parasail control (e.g. thinking.type, enable_thinking, reasoning_effort) for every toggled model.
• [high] [violation] providers/parasail/models/parasail-glm-52.toml:3 - Check: Relays must not use reasoning_options = [] when lab/peers expose caller controls. Why: Empty means no control, not uncertainty. Lab/peer baselines: GLM-5.2 effort high|max (Zhipu); GLM-5.1 toggle; MiniMax-M3 toggle; MiMo-V2.5 toggle; Kimi K3 toggle + low|high|max. Parasail sets [] on glm-52, glm-51, minimax-m3, mimo-v25, kimi-k3. Action: Copy the lab/same-surface control set Parasail actually forwards; only keep [] with affirmative evidence of no control.
• [high] [violation] providers/parasail/models/parasail-gemma3-27b-it.toml:1 - Check: Nameable third-party models need lab metadata + base_model. Why: Gemma 3 27B IT, Qwen3-VL-8B Instruct, Gemma 3 4B IT, UI-TARS 1.5 7B, BGE-M3, etc. are identifiable lab/community models but are fully inlined (or missing lab entries). Placeholder names/dates ("Googlegemma 3 4 B It", release_date = "2026-08-20") show stubbed metadata. Action: For each nameable model, add complete models/<lab>/<id>.toml if missing, then make the Parasail file override-only with base_model. Reserve full inline only for truly unique-to-Parasail aliases.
• [medium] [possible mistake] providers/parasail/models/parasail-gemma3-27b-it.toml:5 - Check: Inherited/authored capabilities must match the real model. Why: Gemma 3 27B IT is multimodal; this entry sets attachment = false and text-only modalities. Several stubs also look wrong for their class (e.g. embedding BGE-M3 as chat text→text; UI-TARS as text-only/no attachment). Action: After moving to base_model/lab files, correct modalities, attachment, tool_call, and limits from first-party model cards.
• [medium] [possible mistake] providers/parasail/models/parasail-glm-52.toml:1 - Check: Provider models should include cost unless intentionally unpriced. Why: Patch 2 drops or omits cost on many live catalog entries (dated DeepSeek SKUs, GLM-5.2, Kimi K3/K2.7-Code, MiniMax-M3, Qwen3.5-9B, Qwen3.8-27B, Trinity, Llama 3.2 3B, Mistral Nemo, plus most stubs) while the PR claims pricing from Parasail’s models API. Action: Author USD/MTok cost from Parasail pricing for each public model, or document per model that pricing is unavailable.
• [low] [possible mistake] .pr-review/pull-request.json - Check: Data-changing PRs should cite direct provider docs/pricing mapped to claims. Why: Body mentions the models API and “documented reasoning controls” but gives no URLs tied to pricing, limits, or wire fields for toggle/effort/budget. Action: Add first-party Parasail doc/API links and state what each supports (catalog IDs, $/MTok, reasoning request fields).

[github-actions]

Action items

• [high] [violation] providers/parasail/models/parasail-kimi-k27-code.toml:1 - Check: Relay reasoning_options must match the lab/peer control surface. Why: First-party providers/moonshotai/models/kimi-k2.7-code.toml and Moonshot provider docs mark K2.7 Code as always-on (reasoning_options = []; only thinking.type=enabled). This file invents toggle + effort low|high|max with a fabricated long_thinking wire path. Action: Set reasoning_options = [] (with an always-on comment) unless Parasail docs prove real off/effort controls on this host.
• [high] [violation] providers/parasail/models/parasail-kimi-k26.toml:1 - Check: Baseline effort/toggle set for Kimi K2.6. Why: Lab Moonshot K2.6 is toggle-only (thinking.type); no effort levels. This entry adds invented effort low|high|max and wrong wire comments (long_thinking, reasoning_effort). Action: Use toggle-only with # Toggle: thinking.type = enabled|disabled (or Parasail’s verified equivalent); drop effort unless Parasail documents it.
• [high] [violation] providers/parasail/models/parasail-kimi-k3.toml:1 - Check: Exact toggle/effort wire paths for Kimi K3. Why: Lab K3 uses thinking.type (+ adaptive) and output_config.effort = low|high|max. Comments claim long_thinking and reasoning_effort, which are not the lab surface and mislead consumers. Option shape (toggle + low/high/max) is fine; wire docs are not. Action: Replace leading comments with lab-accurate (or Parasail-verified) paths; do not invent field names.
• [medium] [violation] providers/parasail/models/parasail-glm-51.toml:1 - Check: Toggle wire comment must be the real request field. Why: Zhipu first-party documents thinking.type = enabled|disabled, not enable_thinking true|false. Action: Fix the leading comment to the Parasail-forwarded path (lab default: thinking.type), or cite Parasail if it truly uses a different field.
• [medium] [violation] providers/parasail/models/parasail-mimo-v25.toml:1 - Check: Toggle wire comment for MiMo. Why: Xiaomi documents thinking.type=enabled|disabled, not reasoning = true|false. Action: Correct the leading wire comment to the real path.
• [medium] [violation] providers/parasail/models/parasail-minimax-m3.toml:1 - Check: Toggle wire comment for MiniMax-M3. Why: minimax_reasoning is not a documented MiniMax or common-relay field; peers use lab toggle / chat_template_kwargs.thinking-style controls. Action: Document the actual Parasail request field or match a verified peer/lab path.
• [medium] [possible mistake] providers/parasail/models/parasail-llama-32-3b-instruct.toml:1 - Check: base_model must name the same underlying model. Why: Filename/API id is instruct, but base_model = "meta/llama-3.2-3b" points at the base (non-instruct) lab entry (tool_call = false, base description). Action: Point at an instruct lab SKU (add models/meta/…-instruct if missing) or justify why base metadata is correct for this host id.
• [medium] [possible mistake] models/gryphe/mythomax-13b.toml:12 - Check: New lab limit values must be model-accurate, not host defaults. Why: context = 131072 / output = 32768 are copy-pasted across several new community lab files (MythoMax, UnslopNemo, Glimmer, Lunaris). Classic MythoMax L2-13B is far smaller context; inflated lab limits will pollute every future base_model consumer. Action: Set lab limits from model cards/config; override only on Parasail if the host truly serves a longer window.
• [low] [possible mistake] providers/parasail/models/parasail-glm-52.toml:1 - Check: Provider cost when the host publishes prices. Why: Several live catalog entries ship without [cost] (e.g. GLM-5.2, Kimi K3, MiniMax-M3, Qwen3.5-9B, Qwen3.8-27B, GPT-OSS 120B Fast) while peers/labs publish USD/MTok and the PR claims an OpenRouter-compatible pricing API. Omission is allowed only when intentionally unpriced. Action: Add Parasail USD/MTok costs where public, or note in a leading comment / PR body which IDs are intentionally unpriced.
• [low] [possible mistake] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite first-party sources mapped to claims. Why: Body mentions the models API and pricing but gives no URLs or per-claim mapping for costs, limits, or reasoning wire fields. Action: Add direct Parasail docs/API/pricing links and state what each supports (catalog IDs, $/MTok, reasoning parameters).

[github-actions]

Action items

• [high] [violation] providers/parasail/models/parasail-muse-glimmer-30b.toml:1 - Check: Third-party hosts must base_model the real lab identity; provider files stay override-only. Why: Parasail points at a new muse/glimmer-30b lab stub, but this catalog already defines Meta’s model as models/meta/muse-glimmer-30b.toml (reasoning, multimodal, 131k context). Peers (OpenRouter, Vercel, etc.) all use meta/muse-glimmer-30b. The stub sets reasoning = false, text-only modalities, and weak limits, so merge inherits the wrong model. Action: Delete models/muse/glimmer-30b.toml; set base_model = "meta/muse-glimmer-30b"; author Parasail cost and lab/peer-aligned reasoning_options (peers use effort low|medium|high|xhigh).
• [high] [violation] providers/parasail/models/parasail-gemma-4-26b-a4b-it.toml:4 / providers/parasail/models/parasail-gemma-4-31b-it.toml:4 - Check: Relay reasoning_options must follow the lab/same-surface baseline; [] means verified no caller control, not uncertainty. Why: First-party Google entries use toggle, and established relays (e.g. OpenRouter) also expose toggle. Parasail asserts “always-on; no caller control” with [] without host evidence, under-reporting controls. Action: Match the lab/peer set (toggle, with a leading wire comment) unless Parasail docs/API prove no on/off control—then keep [] with that evidence in a leading comment or PR body.
• [high] [possible mistake] providers/parasail/models/parasail-qwen38-27b.toml:5 - Check: Relay options must copy the underlying model’s native/peer control surface, not invent another Qwen generation’s shape. Why: Parasail uses toggle + budget_tokens (Qwen3.5-style). Same-surface peers for alibaba/qwen3.8-27b (OpenRouter, DeepInfra, Vercel) use effort (low/medium/xhigh, sometimes with none or toggle)—aligned with Qwen3.8 Max’s effort surface, not budget-only. Action: Align with Qwen3.8 lab/peer options for this host (effort and/or toggle as actually exposed); do not keep 3.5-style budget-only unless Parasail documents thinking_budget for this SKU.
• [medium] [violation] providers/parasail/models/parasail-glm-52.toml:1 (and other cost-less reasoners/hosts below) - Check: Provider models should author cost when public pricing exists (AGENTS: provider-side cost unless intentionally request-only). Why: PR claims pricing from Parasail’s models API, and earlier commits priced many SKUs, but several live catalog entries ship with no [cost] (e.g. parasail-glm-52, parasail-kimi-k3, parasail-kimi-k27-code, parasail-minimax-m3, parasail-qwen35-9b, parasail-qwen38-27b, parasail-gpt-oss-120b-fast, plus most community/embedding/TTS SKUs). Catalog consumers get incomplete pricing. Action: Add USD/MTok [cost] for every publicly priced serverless model, or mark/document intentional omission; cite Parasail pricing/API in the PR body mapped to these fields.
• [medium] [possible mistake] models/baai/bge-m3.toml:17 / providers/parasail/models/parasail-bge-m3.toml:1 - Check: Lab metadata must describe the real model modalities/limits. Why: BGE-M3 is an embedding model; the lab file uses chat-style output = ["text"] and output = 1 tokens, which misrepresents the SKU for API consumers. Action: Correct modalities/limits for an embedding model (or drop the entry if Parasail only exposes it outside the chat catalog semantics this repo models).
• [medium] [possible mistake] providers/parasail/models/parasail-gemma3-27b-it.toml:1 / models/google/gemma-3-27b-it.toml:14 - Check: Provider overrides must capture real host deltas after base_model. Why: New lab metadata is multimodal (attachment = true, image input). Parasail’s prior full definition was text-only/attachment = false. Without overrides, the resolved Parasail entry advertises vision if the host does not. Same risk for parasail-googlegemma-3-4b-it. Action: Confirm Parasail modalities; if text-only, override [modalities]/attachment on the provider files.
• [medium] [possible mistake] models/mlabonne/unslopnemo-12b.toml:1 - Check: Lab IDs should match the established model identity used elsewhere in the catalog. Why: Existing hosts catalog this community model under TheDrummer (openrouter/kilo thedrummer/unslopnemo-12b), while this PR invents mlabonne/unslopnemo-12b with placeholder-style limits. Action: Reuse or add the correct shared lab path peers already imply, and set limits from a real model card—not arbitrary values.
• [low] [possible mistake] pull request body - Check: Data-changing PRs should cite first-party pricing/docs/API with claim mapping. Why: Body mentions Parasail’s models API and docs root but does not map URLs to concrete costs, limits, or reasoning wire paths, so material values (especially missing prices and wire comments like thinking.type on MiniMax/MiMo/GLM/Kimi) cannot be reviewed. Action: Add direct citations and state what each supports (pricing per model, reasoning field names, catalog completeness).

[github-actions]

Action items

• [high] [violation] providers/parasail/models/parasail-qwen35-397b-a17b.toml:9 - Check: Reasoning options for multi-model relays must match lab/same-surface peers; do not invent budget_tokens. Why: Parasail is an OpenAI-compatible multi-model relay. Same-model OpenRouter peers for qwen3.5-397b-a17b and qwen3.5-35b-a3b expose toggle only; OpenRouter’s qwen3.6-35b-a3b is also toggle-only. Authoring budget_tokens + thinking_budget without Parasail evidence overstates caller controls. Action: Drop budget_tokens on parasail-qwen35-397b-a17b.toml, parasail-qwen3p5-35b-a3b.toml, and parasail-qwen3p6-35b-a3b.toml (keep toggle if Parasail actually forwards it), or cite Parasail docs/API showing a real reasoning-budget field on this host.
• [high] [possible mistake] providers/parasail/models/parasail-glm-52.toml:12 - Check: Provider cost must be this host’s USD/MTok pricing (or intentionally unpriced). Why: Several SKUs comment that Parasail pricing is unpublished, then copy OpenRouter peer prices (glm-5.2, gemma-3-4b-it, kimi-k3, kimi-k2.7-code, llama-3.2-3b-instruct, MiniMax-M3, mistral-nemo, muse-glimmer-30b, qwen3.5-9b, qwen3.8-27b, ui-tars-1.5-7b, unslopnemo-12b). That publishes another provider’s rates as Parasail’s. Action: Replace with Parasail-sourced prices (and cite them), or remove [cost] and mark intentionally unpriced like the other unpriced SKUs.
• [medium] [violation] models/the-drummer/cydonia-24b-v4.1.toml:1 - Check: Lab IDs must be stable and consistent for the same creator. Why: This PR adds models/the-drummer/* for Cydonia/Skyfall while also adding models/thedrummer/unslopnemo-12b.toml and pointing Parasail at thedrummer/unslopnemo-12b. Established OpenRouter paths use thedrummer. Split lab slugs fragment the catalog. Action: Use one lab id (prefer thedrummer to match existing peers), move Cydonia/Skyfall under it, and update Parasail base_model paths.
• [medium] [possible mistake] providers/parasail/models/parasail-minimax-m3.toml:1 - Check: Toggle wire comments must document this host’s real request field. Why: Multiple files claim Parasail accepts lab-native paths (thinking.type, enable_thinking, thinking_budget, output_config.effort) with no Parasail API evidence. MiniMax’s first-party entry has a toggle but no thinking.type docs; OpenRouter’s MiniMax-M3 is []. Wrong wire paths mislead clients. Action: Verify Parasail’s actual request surface and correct comments/options; use only fields Parasail documents or that live requests accept.
• [medium] [possible mistake] models/gryphe/mythomax-13b.toml:1 - Check: New lab metadata should use the established/nameable model identity. Why: Weights point at Gryphe/MythoMax-L2-13b, and OpenRouter already catalogs this as mythomax-l2-13b with 8k/4k limits. Parasail’s Lunaris SKU id (sao10kl3-lunaris-8b) also suggests L3 Lunaris, while the new lab is sao10k/lunaris-8b with different limits than OpenRouter’s l3-lunaris-8b. Action: Align lab ids/limits with the real HF/OpenRouter identities (mythomax-l2-13b, L3 Lunaris if that is the SKU) or justify Parasail-specific aliases with evidence.
• [low] [possible mistake] providers/parasail/models/parasail-trinity-large-thinking.toml:1 - Check: Missing public price should be explicit. Why: This reasoning model has no [cost] and no “intentionally unpriced” header, unlike other unpriced Parasail SKUs in the same PR. Action: Add Parasail pricing if known, or a leading comment that public Parasail pricing is intentionally omitted.
• [low] [possible mistake] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite first-party pricing/docs/API mapped to claims. Why: Body mentions an OpenRouter-compatible models API and catalog refresh but does not map Parasail URLs to concrete cost, limit, modality, or reasoning-control claims, so peer-copied prices cannot be reviewed. Action: Add direct Parasail pricing/docs/API links and state which fields each source supports.