update-models

name	update-models
description	Refresh elph-ai chat model catalogs from models.dev (origin) with optional live provider pricing, inject full thinkingLevelMap on every model, regenerate src/models/catalog.rs, and verify registration. Use when the user runs /update-models, asks to sync or regenerate model catalogs, update models.dev data, refresh pricing, or keep provider model lists current.

Update Models (models.dev origin)

Language

• In-chat reports follow the user's language.
• Docs/skill text and generated comments stay English.

Purpose

Regenerate crates/elph-ai/models/*.json and src/models/catalog.rs from models.dev as the sole catalog origin. Pricing prefers live provider APIs when keys/endpoints allow, then models.dev. Every model must include a complete thinkingLevelMap.

When to run

• /update-models
• "refresh model catalog", "sync models.dev", "regenerate models", "update pricing"

Prerequisites

• Network access (unless --offline with existing models/.cache/models.dev/api.json)
• Optional env keys for live pricing probes (OPENAI_API_KEY, OPENROUTER_API_KEY, …)
• No local pi clone required for chat catalogs

Workflow

1. Generate chat catalogs

cargo run -p elph-ai --bin generate-models -- chat
# or
make generate-models ARGS="chat"

Useful flags:

Flag	Effect
--offline	Use cached models.dev only
--no-live-pricing	Skip provider /models pricing probes
--no-regenerate-catalog	Write JSON only

2. Optional full pass (chat + image fixture path)

cargo run -p elph-ai --bin generate-models -- all --no-live-pricing

3. Verify

cargo test -p elph-ai --test providers --lib models
cargo check -p elph -p elph-ai

Confirm:

• thinkingLevelMap: complete=N incomplete=0 in generator output
• Verified … catalog providers are registered in builtin_providers()
• Spot-check anthropic.json, xai.json, a gateway (openrouter / kilo)

4. Summarize for the user

• Provider count / model count
• Any providers skipped (not on models.dev and no previous overlay)
• Remaining zero-priced models if any
• Reminder: do not hand-edit generated catalogs except intentional Elph overlays; re-run this skill after changes

Rules

• Origin: models.dev only — never seed chat catalogs from pi.
• Pricing: live API (when available) → models.dev → keep previous non-zero.
• Every model must have thinkingLevelMap with keys off|minimal|low|medium|high|xhigh|max (null = unsupported).
• Do not invent api / baseUrl from models.dev; preserve Elph factory overlays.
• Do not remove Elph-only gateway providers; preserve their model ids and enrich.
• New providers still need a factory in src/providers/builtin.rs (generator fails if unregistered).
• Schema contract for future user overrides: schemas/provider-schema.json (runtime merge not implemented yet).

Code map

Path	Role
bin/generate_models/main.rs	CLI
bin/generate_models/models_dev.rs	Fetch/cache models.dev
bin/generate_models/provider_sources.rs	Elph ↔ models.dev map
bin/generate_models/normalize.rs	Entry merge + cost fields
bin/generate_models/thinking_map.rs	Full 7-key maps
bin/generate_models/pricing.rs	Live pricing probes
bin/generate_models/chat.rs	Orchestration + catalog.rs
models/*.json	Embedded catalogs
schemas/provider-schema.json	Forward-looking override schema

Do not

• Commit without the user asking
• Run live pricing against unpaid keys in a loop
• Hand-edit hundreds of models when a generator fix exists