Agent Team Models
Use the companion agent-team-creator compiled runtime and read its
references/models-memory.md. Discover the currently configured provider/model
identifiers; do not infer availability, strength or price from a model name.
Ask for missing constraints in task terms: difficult reasoning, routine
implementation, mechanical edits, tools, vision, context size or cost. Reuse
existing preferences. Explain the three policy tiers as operator labels:
hard for difficult reasoning/review, medium for routine implementation, and
low for bounded mechanical work. They do not guarantee benchmark performance.
node <agent-team-creator>/scripts/cli.mjs models-discover --input discovery.json
node <agent-team-creator>/scripts/cli.mjs models-select --input selection.json
Discovery supports a configured OpenAI-compatible gateway (including liter-llm), or the explicit verified UAR/BossFang discovery endpoint. Supply credentials by environment variable reference. A successful model listing does not prove a successful inference. Offline operator-declared catalogs also work and are labeled as declarations.
Map gateway aliases explicitly to liter-llm catalog provider/model identities. Keep catalog provenance, freshness and per-token→per-million conversion. Annotate tiers explicitly. Missing capability or price data cannot satisfy a requirement or price ceiling; report no match instead of quietly relaxing constraints.
Policy order is team → role → requested skills in explicit order → task. Scalar values override; required capabilities accumulate. Explain the effective policy and why the selected model qualifies. This is a configured policy order, not a security boundary: a task override may intentionally change a budget. Stale prices are estimates, not a guarantee of current provider charges.
Save the concrete chosen ID in the role’s modelPolicy.model (or task policy
for a task-specific invocation) through a reviewed manifest/state update.
Re-run selection when the skills or task change. Skill policies influence
selection; they do not mutate a skill’s frontmatter or force every harness to
support per-skill model switching.
Read the selected harness’s native reference before binding it. Kimi has no role frontmatter model setting; DeepSeek team members lack a verified per-member override. Use supported invocation/global controls or report the limitation. Never add an invented model flag. Report policy, selected ID, unknown metadata, price age, and whether native application has actually been verified.