Model Catalog Refresh

Skill: /model-catalog-refresh

Model preference: #investigate (per horizon_aios_model_prefs.md; overridable by a prompt directive).

Fetch live model data from provider documentation and return a structured catalog the user (or the model-prefs skill) uses to populate or validate horizon_aios_model_prefs.local.md. This is a reference document — you, the agent, perform the fetches and parsing at runtime with your web/bash access. Do not rely on training-cutoff knowledge of model ids or prices; the entire point is live data.

Quick Reference

• Purpose: produce a dated, structured catalog of current models + pricing across Anthropic, OpenAI, Google Gemini, and Ollama, and diff it against an existing model-preference config.
• Triggers: "refresh model catalog", "update model groups", "check current models", "what models are current", "validate my model config", /model-catalog-refresh.
• Companion: /model-prefs consumes this output to edit the gitignored extend file. This skill fetches truth; that skill writes config.

When to invoke

Whenever the user wants an up-to-date picture of available models to configure or sanity-check their groups — e.g. before defining #lowcost/#highcap, after a provider releases a new model, or to confirm a member id is still valid.

Providers and fetch strategy

Fetch each provider's MODEL LISTING and PRICING separately — they are always two different pages. Prefer an API/CLI when available over scraping.

1. Anthropic

• Models: https://platform.claude.com/docs/en/about-claude/models/overview
• Pricing: https://platform.claude.com/docs/en/about-claude/pricing
• Changelog / notices (for suspensions): check the docs changelog and any banner.
• Extract: full model id string, tier (haiku/sonnet/opus/fable), input & output $/MTok, context window, known aliases, deprecation or access-suspension notices.

2. OpenAI

• Preferred: if OPENAI_API_KEY is set, GET https://api.openai.com/v1/models and parse directly — more reliable than the docs page.
• Models (fallback): https://developers.openai.com/api/docs/models
• Pricing (scrape; no pricing API): https://openai.com/api/pricing
• Extract: model id, family (gpt-5.x / gpt-oss), open-weight flag, input/output $/MTok, context window, recommended-for notes.

3. Google Gemini

• Models: https://ai.google.dev/gemini-api/docs/models
• Changelog (deprecations): https://ai.google.dev/gemini-api/docs/changelog
• Extract: exact versioned model id (Google uses date suffixes — capture them), tier (flash-lite/flash/pro), pricing (note tiered-by-context-length), GA vs preview status, shutdown notices.

4. Ollama

• Preferred: if ollama is on PATH and running, ollama list for what is already pulled, then ollama show <model> per model for metadata.
• Fallback (no local ollama): fetch https://ollama.com/library and take the top models by pull count in each relevant category (coding, reasoning, fast/small).
• Extract: exact pull tag (e.g. qwen2.5-coder:7b), category, VRAM requirement, parameter count, notable capability notes.

Output schema

Return a plain-text block (not JSON — this lands in config-adjacent context):

Model Catalog —

Anthropic

| model_id | alias | tier | input $/MTok | output $/MTok | ctx | status |

OpenAI

| model_id | family | open_weight | input $/MTok | output $/MTok | ctx | status |

Google

| model_id | tier | input $/MTok | output $/MTok | ctx | ga_or_preview | status |

Ollama (local-available / library-top)

| tag | category | vram | params | notes |

Deprecation / Access Alerts

List any model found suspended, sunset-announced, or access-restricted.

Config Diff (only if a config was provided — see Diff behavior)

Per group in the current config, flag: - member ids that no longer appear in provider docs - newer models that better fit the group's intent - pricing changes vs the member's prior cost

Diff behavior

If the current Horizon AIOS ## Model Groups block is already in context when invoked, run the Config Diff automatically. If it is not, prompt once: "Paste your current ## Model Groups block to get a config diff." Do not invent a config to diff against.

Freshness

• Stamp the output with the fetch date.
• State, per provider, whether you got live data or fell back (scrape / cache / unavailable).
• If a page is unreachable, say so explicitly for that provider — never return stale data silently as if it were current.

Populating the config

After presenting the catalog, offer to hand the relevant ids to /model-prefs (or do it directly if the user asks) to update horizon_aios_model_prefs.local.md. Use runtime-qualified members: claude:<id|alias> for Anthropic, ollama:<tag> for local models. Prefer Anthropic aliases (haiku/sonnet/opus/fable) over pinned full ids unless the user wants a specific version. Never write the base horizon_aios_model_prefs.md — user choices go only in the extend file.

Known Gotchas

• Ollama tags often omit the size suffix — always use the full tag (:7b, not :latest); :latest silently pulls whatever the registry defaults to.
• Google model ids frequently carry date suffixes — never store bare gemini-3.5-flash without verifying the exact stable string.
• OpenAI open-weight models (gpt-oss-*) run locally via Ollama/LM Studio AND are available via the OpenAI API — record both surfaces; they differ in price.
• Anthropic access suspensions (e.g. Fable 5, mid-2026) may not show on the models overview page — check the changelog and any banner before trusting availability.
• Pricing and model-listing pages are separate fetches for every provider — do both; a listing without prices is an incomplete catalog.

What this skill must NOT do

• No executable code embedded as the mechanism — you fetch at runtime; the file is instruction.
• No hardcoded model ids or prices — live fetch only.
• No assumption about which runtime/harness you are in.

Notes for the executing agent

• Run independent provider fetches in parallel where possible; one provider being down must not block the others.
• This skill is read-only toward provider sources and the base config; the only thing it may write is the gitignored extend file (and only when the user asks).
• Brains may invoke this — it touches public docs and the user's own extend file only; no privileged paths.