update-models-metaai

Update src/modules/llms/server/openai/models/metaai.models.ts with latest model definitions.

Reference src/modules/llms/server/llm.server.types.ts and src/modules/llms/server/models.mappings.ts for context only. Focus on the model file, do not descend into other code. The vendor's protocol facts live in kb/modules/LLM-metaai-responses.md; the adapter deviations are the metaai row of _RSP_DIALECT_QUIRKS in src/modules/aix/server/dispatch/chatGenerate/adapters/openai.responsesCreate.ts.

Primary Sources: dev.meta.ai/docs. Appending .md to any docs URL returns raw markdown that plain curl fetches (no auth, no browser needed).

• Doc index: https://dev.meta.ai/docs/llms.txt (92 pages; one curl loop mirrors the whole set when a deep read is needed)
• Models, tiers, modalities, context windows: https://dev.meta.ai/docs/models.md
• Pricing and rate limits: https://dev.meta.ai/docs/pricing-rate-limits.md - Standard vs Contributor per-token tables, the search-grounding surcharge, per-image and per-hour prices, RPM/TPM per tier (page carries an unresolved "confirm these numbers" editorial note - Standard limits are wire-verified, the rest provisional)
• Reasoning effort ladder: https://dev.meta.ai/docs/reasoning.md
• Responses API: https://dev.meta.ai/docs/protocols/responses.md; full schema https://dev.meta.ai/docs/api-reference/responses/schemas.md (5,300 lines - grep it, don't read it)
• Capability pages to diff against interfaces / parameterSpecs: tool-calling.md, tool-search.md, search-grounding.md, structured-output.md, image-understanding.md, video-understanding.md, file-handling.md, image-generation.md, prompt-caching.md
• Release dates (the API carries none): the announcement posts on ai.meta.com/blog and research.meta.ai/blog, and the release table on Wikipedia's "Muse Spark" page. Known: Spark 1.1 2026-07-09, Muse Image 2026-07-07, Spark 1.2 2026-08-05, Spark 1.3 2026-09-02.
• There is NO Model API changelog (muse-code/changelog.md is the CLI's). Third-party trackers (llm-stats.com/models/muse-spark-1.3, openrouter.ai/meta/muse-spark-1.3) are fallbacks for dates and benchmarks only.

Do NOT use web search for facts. Fetch the URLs directly; search only to locate release announcements.

Known gaps and traps (verified 2026-09-02):

• GET /v1/models and GET /v1/models/{id} return only {id, object, created: 0, owned_by: 'meta'}: created is a constant 0 (useless for pubDate or ordering, despite models.md claiming otherwise) and there is no type/modality field. The list does not establish what a model IS - cross-check every id against models.md before curating it as chat.
• The list mixes families: muse-image-1.0 (image output over Responses; curated with LLM_IF_Outputs_Image + LLM_IF_HOTFIX_NoStream) and muse-voice-transcribe-1.0 (ASR on /v1/asr, 404 on /v1/responses; dropped via _METAAI_NON_CHAT_PREFIXES). A new non-chat family goes in that deny-list or gets curated with the right output interfaces.
• -contributor ids are the same model at ~12x/21x lower rates and Meta trains on prompts and completions: keep them hidden: true with the tradeoff in description; never a default pick.
• maxCompletionTokens is undocumented: 131072 is Meta's advertised figure (quickstart / coding-agents configs); the server accepts max_output_tokens up to 1M, so do not derive it from probes.
• Docs print keys as LLM|... while served keys are LLM_...: validateSetup accepts both, do not tighten.

Live endpoint (extra signal): If .env.api-keys has METAAI_API_KEY, scan the served list for new ids: curl https://api.meta.ai/v1/models -H "Authorization: Bearer $METAAI_API_KEY". Never commit or echo the key.

Probing tips:

• Effort domain: POST /v1/responses with {"model":"<id>","input":"pong","max_output_tokens":64,"reasoning":{"effort":"<v>"}}. A bad value 400s with unknown variant X, expected one of ... (the full server enum); a value the model rejects 400s with does not support "<v>" with this model. As of 2026-09-02 the enum is none|minimal|low|medium|high|xhigh, none is rejected on every Spark, max is not served yet (Meta announced it for 1.3 - re-probe). Omitted effort runs as high.
• Unknown TOP-LEVEL request params 400 with unknown parameter X - cheap discovery of new fields; unknown nested keys are silently ignored.
• tool_choice accepts only auto; truncation only disabled; logprobs are unreachable while reasoning is on.
• Temperature 2.0 is accepted but degenerate (rambling incomplete replies, occasional HTTP 500) - the sweep records 0..1.5; do not read a 500 there as an outage.
• End-to-end ablations through the real pipeline (tools, search, reasoning, image output): npx tsx tools/develop/aix-protocol-lab/lab.ts capture metaai-responses <hello|reason|fc|search|interleave> [--model <id>] [--no-stream]
• Parameter acceptance sweep: tools/develop/llm-parameter-sweep/sweep.sh --dialect metaai --key $METAAI_API_KEY --model-filter muse-spark writes llm-metaai-parameters-sweep.json; then /llms:verify-parameters metaai diffs it against the definitions.