Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
AI provider model tiers
Tiers are active routing configuration, not display-only labels. Each preset (a stored provider record — see AI_PROVIDERS.md) maps Light (mechanical work), Medium (routine work), Heavy (complex work), and Ultra (exceptional frontier reasoning) to its own model. Tiers live on presets, not on harnesses or services: configure them in Models → Providers → Presets, on the preset's editor, and a tier name resolves on whichever preset a task names. Astra and Fable are examples of Ultra choices; use a model supported by the selected preset's service and plan.
A tier is an explicit request, never an automatic one. CoS runs a task on the tier or model the task, role, or stage names, and otherwise on the preset's Default Model — it does not pick a tier from the task description, priority, context size, or learning history. The CoS Learning view still reports per-tier success rates per task type as advice: act on it by pinning a tier on the task or schedule. Prompt stages use the same four names (Light, Medium, Heavy, Ultra) or Default; the legacy stage spellings Quick and Coding still resolve to Light and Medium. An unset tier on a preset inherits its Default Model, so a preset only needs the tiers it wants to map to a different model. Dispatch labels communicate capability to planning/claim agents; they are guidance, not permission to enable providers or spend on a new service.
Use model:ultra on an exceptional tracker task. In CoS task metadata, use
model: "ultra"; orchestration profiles accept the same tier names in each role's
model field. For example, an explicitly configured architect can request Ultra,
with a Medium implementer and Heavy reviewer. Exact model IDs still work and
remain appropriate for model-specific evaluations or compatibility requirements.
The tier resolves on the selected provider, including provider fallback.
Model capability and reasoning effort are independent: Ultra does not imply maximum effort. Provider defaults and scheduled jobs never upgrade to Ultra on their own. An unset Ultra mapping falls back to Heavy, then the provider default. Existing installs receive an optional Ultra field; migration offers Fable additively on standard Claude catalogs, selects Astra/Fable when advertised by the provider, and preserves all explicit Ultra pins. No migration calls a provider or starts AI work.
Prefer role/stage tier assignments for portable workflows; keep exact pins for intentional exceptions. Avoid bulk replacing installed stage pins or changing scheduled tasks without the user's instruction.