Updated Regolo.AI models

[Mte90]

This update adds/removes multiple Regolo models and sets new USD costs (EUR conversion), $0 free tiers, context/output caps, and reasoning controls, but the PR body only says models were added/removed—no first-party Regolo catalog, pricing, or API references.

[github-actions]

Action items

• [high] [violation] providers/regolo-ai/models/gemma4-31b.toml:6 - Check: Relay reasoning_options must follow lab + same-surface peers for that model, not a universal L/M/H set (AGENTS.md Reasoning options; audit skill). Why: Regolo is a multi-model relay. Lab/Google and peers (e.g. OpenRouter) use toggle only for Gemma 4 31B IT; this file invents effort low|medium|high. provider.toml already notes Regolo’s thinking docs demonstrate GPT-OSS and do not map controls to other models. Action: Set options from lab/peers and what this host actually exposes for Gemma (likely [{ type = "toggle" }] with a leading wire comment for thinking). Only keep L/M/H if Regolo docs/tests show meaningful graded effort for this model ID.
• [high] [violation] providers/regolo-ai/models/glm5.2.toml:6 - Check: Do not invent L/M/H or a toggle when lab/peers use a different control set. Why: First-party Zhipu GLM-5.2 is effort high|max (no toggle); OpenRouter peer is high|xhigh. This file uses toggle + low|medium|high, which matches neither lab nor established peers and treats the host like a generic GPT gateway. Action: Align with lab/peer baseline the host can forward (typically high/max or high/xhigh), or cite Regolo-specific evidence if their API truly exposes different controls for glm5.2.
• [high] [violation] providers/regolo-ai/models/qwen3.6-27b.toml:6 - Check: Qwen reasoning controls on relays must match lab/peer shape, not invented GPT-style effort. Why: Alibaba first-party is toggle + budget_tokens; OpenRouter peer is toggle only. This file adds effort low|medium|high with no budget, diverging from both. Existing Regolo Qwen entries also do not use L/M/H. Action: Use lab/peer-compatible options for this host (at least toggle if thinking works; add budget_tokens only if Regolo exposes a real reasoning budget). Do not keep L/M/H without host evidence for this model.
• [medium] [violation] providers/regolo-ai/models/gemma4-31b.toml:4 - Check: Every toggle needs a leading top-of-file wire-path comment (sync strips mid-file comments). Why: Toggle/effort comments sit after base_model in gemma4-31b.toml, glm5.2.toml, and qwen3.6-27b.toml, so they are not in the leading header block required by AGENTS.md. Action: Move the exact wire syntax (e.g. thinking / reasoning_effort) into the top-of-file comment block with the EUR→USD note, before the first key.
• [medium] [possible mistake] providers/regolo-ai/models/deepseek-ocr-2.toml:4 - Check: Provider cost must reflect real USD/MTok pricing (or intentional free). Why: input/output are 0.00 with no free-tier note, while another catalog host prices DeepSeek OCR 2 non-zero (e.g. Novita 0.03/0.03). No other existing Regolo models use zero token prices. Action: Verify Regolo pricing for deepseek-ocr-2 and either set the converted USD rates (with EUR note) or document why it is free.
• [medium] [possible mistake] providers/regolo-ai/models/faster-whisper-large-v3.toml:4 - Check: Cost and display identity for a third-party Whisper host. Why: Cost is 0.00/0.00 with no rationale, and with only base_model = "openai/whisper-large-v3" the resolved name stays "Whisper 3 Large" while the served ID is faster-whisper-large-v3. Action: Confirm Regolo’s price for this model; override name (and any real deltas) if the product is Faster-Whisper rather than OpenAI Whisper branding.
• [low] [possible mistake] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite provider pricing/docs/API and map each link to claims. Why: Body is only “We removed some models and added others,” so costs, limits, free models, Brick entries, and reasoning controls cannot be reviewed against first-party sources. Action: Add direct Regolo citations (pricing page, model list, thinking docs) stating what each supports.

[github-actions]

Action items

• [low] [possible mistake] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite pricing/docs/API in the PR body, with each citation mapped to what it supports. Why: This update adds/removes multiple Regolo models and sets new USD costs (EUR conversion), $0 free tiers, context/output caps, and reasoning controls, but the PR body only says models were added/removed—no first-party Regolo catalog, pricing, or API references. Several claims are therefore hard to review from the diff alone. Action: Add direct Regolo (and lab, where relevant) citations in the PR body and state what each supports (catalog membership, EUR list prices before conversion, free tiers, limits, thinking/reasoning_effort behavior).