Repository navigation
Carry reasoning and input-modality capability metadata through the neutral model catalog #366
Description
This work carries reasoning and input-modality metadata from models.dev through the
USAi catalog, the neutral model catalog (acq-neutral-model-catalog/v1, additive
optional fields), and the OpenCode kit's rendered provider config. Unknown stays
absent; false and text-only are preserved; no capability is inferred from model names.
Builder emission is opt-in (--include-capabilities) and no values are committed until
verified. Prime Agent work is out of scope (separate from #357).
Why it matters (observed, not a root-cause claim for the reported image failure):
- OpenCode 1.x (tag v1.18.35): a config-defined model with no
modalitiesresolves to
image=false, so images are disabled. The acq OpenCode kit used to fall back to
npm install -g opencode-ai, which resolves to 1.x. - OpenCode 2.0.30: an unlisted model defaults to text+image. A live run on 24 USAi models
with no modalities: 20 returned image-derived text (9 exact, 11 with misread characters; 1 refused on policy grounds; 3 unavailable or unusable); Claude 4.6+/Opus/Sonnet 5.x and
Gemini 3.x read a 6-digit synthetic image exactly; GPT-family models misread it;
gpt-5.6-terra refused; gemini-2.5-flash/-lite blocked globally; llama_4_maverick
unusable in OpenCode streaming. An explicit["text"]override made the two retested
models refuse images. - Explicit modalities therefore make image handling version-independent.
Acceptance: fixture regressions source -> catalog -> neutral snapshot -> render; schema
rollout compatible with consumers' schema copies (older validators reject new fields);
live synthetic-image procedure recorded before any values are committed
(sources.capabilitiesVerified).
Follow-up work is tracked as sub-issues: #539 (consumer schema adoption), #540 (live verification run), #541 (commit verified values).
Activity
[-]USAi catalog (later): promote reasoning + inputModalities into the catalog[/-][+]Carry reasoning and input-modality capability metadata through the neutral model catalog[/+]mogul commentedon Nov 10, 2026
🤖: AI agent (First Mate), posting on behalf of @mogul.
Closing as delivered. The metadata path is built, the first verified values are published, and the one remaining capability has its own tracking issue.
reasoningandinputModalitiesfields in the neutral catalog and the OpenCode renderer.inputModalities ["text","image"]published for 12 USAi models, with the live evidence committed alongside. The sub-issues Consumers adopt the additive neutral-catalog capability fields #539, Run live synthetic-image verification across USAi models #540 and Commit verified reasoning and input-modality values to the USAi catalog #541 are closed.reasoningstays absent for every model. OpenCode 2.0.30 does not carry the field through, so it could not be verified. That remaining work is tracked in #574, which stays open.