Model Catalog

Choose which model each kind of work uses, and fall back to the next model when one is rate limited, out of quota or unavailable. Built for agent steps run with @cachehit/agents (Kiro CLI).

You own the catalog: each role (for example implementer, reviewer, classifier) has ordered tiers of {model, costClass, toolProfile, effort}. resolveDispatch stores the tier a step should use for one (workItem, role, attempt). Given a failure class, or the newest failed @cachehit/agents invocation for the work item, it advances on rate-limit, quota and model-unavailable, stays on other failures, and parks when tiers or attempts run out. Results are idempotent per (workItem, role, attempt).

Quick Start

swamp extension pull @cachehit/model-catalog
swamp model create @cachehit/model-catalog catalog --global-arg 'catalog={
  "version": "1",
  "roles": {
    "implementer": {"tiers": [
      {"model": "claude-haiku-4.5", "costClass": "low", "toolProfile": "actor"},
      {"model": "claude-sonnet-4.6", "costClass": "medium", "toolProfile": "actor"}
    ]},
    "reviewer": {"tiers": [
      {"model": "claude-opus-4.8", "costClass": "high", "toolProfile": "readonly"}
    ]}
  }
}'

swamp model method run catalog resolveDispatch \
  --input workItem=WI-42 --input role=implementer --input instanceName=kiro
swamp data get catalog dispatch--WI-42--implementer--1 --json

Methods

• resolveDispatch: store the model, tool profile and effort for one attempt, applying a fallback when a failure signal is given or read.
• recordFallback: store only the fallback decision for a signal.

First release. resolveDispatch and recordFallback select a Kiro model per role from your catalog and fall back to the next tier on rate limits, quota and unavailable models, reading failures from @cachehit/agents.