Update Grida product model presets, native adapter compatibility and the pinned public Models dependency; distinguish canonical facts from live GG offerings. Use for model selection, client catalogue updates and model-discovery drift.
77%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
Fix and improve this skill with Tessl
tessl review fix ./.agents/skills/ai-models/SKILL.md| Subject | Canonical owner |
|---|---|
| Model identities, labels, releases, labs, provider bindings, capabilities, regular prices | Public gridaco/models, consumed as the pinned @grida/ai-models release. |
| Currently served GG models, effective capabilities, prices and operation schemas | GG models, endpoints, operation schemas. |
| Product membership, defaults, tier choices and feature subsets | data/ai/presets.json. |
| Executable native/BYOK inputs and adapters | data/ai/native-inputs.json, packages/grida-ai/src/, crates/grida-ai/src/. |
| Library model/dimensions/normalization storage invariant | presets.json → library_embedding; coordinate existing vectors before changing it. |
Do not recreate an authored factual catalogue or GG allowlist in Grida. A public Models binding is a fact, not installed adapter support or GG availability. Normal GG additions using an implemented interface need no model-ID code change. A new transport, input contract or result format needs an actual client adapter.
@grida/ai/catalogue is the client compatibility projection plus product presets.
Its models values come from the external release; its catalog adds product
choices and the retained schema-1 installed-client protocol. It is not a second
facts authoring home. Unknown prices/estimates remain unknown, never zero.
Historical client estimates cannot authorize or settle hosted spending.
Read data/ai/README.md and inspect the full old/new
released Models diff before repinning. Review lower/removed/zero prices, billing
units, provider IDs, capabilities and retirements against regular-price evidence.
Do not copy promotional rates or infer price equivalence across providers.
Source corrections belong in the public Models project, not local overrides.
Product recommendations never replace an explicit saved selection. Defaults must remain active listed members; partial order must be deliberate and duplicate-free. BYOK, custom endpoints, Ollama and ChatGPT subscription identities retain their own selection and credential rules. Special background-removal/upscale tools remain separate feature contracts; no need to advertise them as general GG models.
Run from the repository root:
node scripts/ai-catalogue/generate.mjs
pnpm --filter @grida/ai build
node scripts/ai-catalogue/generate.mjs --bundle
node scripts/ai-catalogue/generate.mjs --bundle --check
pnpm --filter @grida/ai test
pnpm --filter @grida/ai typecheck
cargo test -p grida-ai --lockedCommit the pin, presets/native-contract changes and generated projections together.
Never hand-edit marked generated:* literals or package/crate JSON projections.
Use the CLI contract gate for native protocol changes; review intentional fixture
changes rather than regenerating a baseline to hide drift. Run editor typecheck
when touching its consumers.
GG discovery is anonymous, bounded and shared per client host; it carries no
execution credential. Respect cache expiry and unavailable state rather than
restoring retired GG models from a bundled list. The hosted renderer preserves
released binaries lacking the discovery bridge through their existing finite
native choices; never enter that compatibility path after a capable host's
empty, failed or expired read. BYOK/offline metadata remains
available independently from the immutable release pin. New hosts do not fetch
the legacy native catalogue. Keep /api/v1/models/catalog readable for older
clients using catalog.snapshot.legacy(); omit only shapes its schema cannot
represent, retaining exact facts in the current seed. Never treat that endpoint
as current GG inventory.
Signed-in product usage keeps organization-scoped authorization and the existing
native response contracts. It does not require a standalone GG customer API key.
Preserve accepted task identity and recovery after discovery changes; failed
result retrieval never justifies a second paid submission. Hosted billing,
reservations and settlement remain service-owned. Read
hosted-ai.md and the GRIDA-SEC-006
trust boundary for the native consumer contract.
Paid probes, production changes and deployments require their own authorization; a catalogue update alone does not authorize them.
gridaco/grida
219f40b
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.