How a decision is made
- Every LLM call is routed, not just the first prompt of a session. In an agentic loop that means each tool call — and subagents, which have their own context, are routed too.
- The requested model matters. The model your agent asked for is the anchor: the router compares candidates against it and only moves the turn when the trade-off is worth it.
- Upgrades are allowed. If a turn needs a stronger model than the one requested, the router will move up a tier even though it costs more.
- Context is respected. The router estimates the context in each request and filters candidates using their provider-specific windows and an output reserve. When the conservative estimate excludes every candidate, the widest eligible models remain available for the provider to decide actual fit.
- Caching is priced in. Switching models evicts the provider-side prompt cache, so the router compares the expected gain of switching against the cache-eviction cost and stays put when staying is cheaper. Anthropic’s cache is modelled with a one-hour time to live (TTL), other providers with five minutes.
- Sessions can be pinned. Once a session is warm on a model, the router prefers to keep it there until the trade-off clearly changes.
The router is cache-aware, not a cache manager. It does not send keep-alive pings, warm caches on a schedule, or predict usage patterns to pre-warm them.
Routing priority
Under Models & routing → Routing priority, the quality-versus-price dial sets how aggressively the router trades quality for cost. The distribution preview shows how the current split shifts traffic across model tiers. Weave ships a tuned default; Reset to default returns to it.Subscription routing
The same panel picks how subscriptions and credits interact:
Strict pass-through until low takes a threshold (default 80%): the router stays out of the way until that share of the plan is used, then resumes routing so engineers keep working past their limit instead of being blocked.
Choosing which models can be used
Models
Models & routing → Models lists every model in the router’s roster. Uncheck one and the router stops selecting it — useful when a model isn’t on your organization’s approved list. Weave adds new models to the roster as providers release them; new entries are enabled by default, so re-check the list after a major release if you maintain an allowlist.Providers
Models & routing → Providers excludes an entire provider, including as a failover target. Requests are never sent through an unchecked provider.Preferred models
Preferred models let you lean toward specific models in priority order. The router still picks per request — a preference presses the scale on close calls rather than locking other models out. There is no strict priority ordering, and no way to require model A before model B is allowed.Overriding the router for one session
An unrecognized model ID in the force header fails the request with a 400 that names the value, and so does a model your organization has excluded. The router doesn’t fall back to automatic routing, so a typo can’t pass for a working pin. The
/force-model command is different: it replies that the ID isn’t recognized and keeps routing automatically.
Seeing the decision
The first routed response in a session, and any response where the serving model changes, carries an in-band marker (✦ Weave Router → <model>) plus the reason for the pick. Turns that stay on the same model don’t repeat it. Clients that display the route themselves (like pi) suppress it; programmatic callers can suppress it by sending X-Weave-Routing-Marker: off. Claude Code’s managed status line shows the response model, current selection, and an estimated cost difference, and the full history lands in the Router report.
Context indicators
Claude Code’s managed status line prefers the client’s own reported percentage, such as Context 58%. Codex’s integration uses the terminal title and has no client-native percentage or inline status bar through its lifecycle hooks. When native usage is unavailable, a recent session snapshot can show last Router ctx est. ~72k/128k. The numerator is a conservative whole-request estimate, including image estimates. The denominator is the effective window of the model/provider that served that request; some of that space is reserved for output. It is not remaining input capacity or the client’s compaction budget. Estimates never supply a percentage or warning cue. These displays refresh between turns through status-line or lifecycle events. Asynchronous telemetry can lag the latest turn, so snapshots are labeled last and expire after five minutes. Older clients or Routers omit unavailable context fields while preserving existing indicators. A model change alone does not compact a conversation: the client owns its transcript and compaction. Claude Code’s/compact is a client command.