Model policy

Runtimes and models per role are set in botgumi.config.ts (agents). This page holds the defaults and the rubric the orchestrator uses to rate tasks.

Role	Default runtime and model
Orchestrator	claude · Fable 5.1 (claude-fable-5-1). Falls back to Opus 5.5 (claude-opus-5-5) when tokens run low
Worker, high	claude · Opus 5.5 (claude-opus-5-5)
Worker, mid and low	claude · Sonnet 5.5 (claude-sonnet-5-5)
Reviewer, primary	codex · any Codex model (the config shows an example)
Reviewer, second	claude · Opus 5.5 (claude-opus-5-5)

Per-track overrides (agents.overrides.<track>) replace any of these for one track.

#Tokens low

The orchestrator switches to its fallback when its runtime reports under 20% of the token budget left (AgentRuntime.remainingBudget()), or at the first usage-limit warning. It records the switch in the run log.

#Tiers

Tier	Examples
high	architecture; API contract design and change requests; payments and entitlements; auth and sessions; AI pipeline and spend guard; data migrations; hard bugs (no clear repro, cross-track, concurrency)
mid	screens from an approved hi-fi spec; endpoints that follow an existing pattern; tracking-plan wiring; seeds and fixtures
low	copy and localization (en + zh-Hans); docs and checklists; config; small fixes

The tier also sets review depth: high-tier PRs get both reviewers; mid and low get one.