Per-session usage/cost metrics (tokens, cache, cost, context %)

[alexanderkreidich]

Consumers migrating from bespoke chat UIs (e.g. the appx canvas template) show per-session usage in the chat settings: input/output tokens, cache read/write, monetary cost, and context-window utilization. agent-client currently has no notion of usage at all — neither in the reducer/SessionState nor as a client helper.

The data is already on the wire: AssistantMessage.usage is part of the contract (REST history via getSessionMessages and SSE WireEvents). What's missing:

1. Aggregation — accumulate usage from assistant messages into SessionState (or expose a pure helper that folds a UiMessage[]/history into totals: input, output, cacheRead, cacheWrite, cost, context % of the active model's contextWindow).
2. Cost recalculation when usage.cost is 0 — custom LiteLLM-routed models often report zero cost; the canvas template recalculates from per-million-token rates. AgentModelRow from GET /v1/sessions/models doesn't expose cost rates, so either the rates get added to the agent-server models contract (cross-repo dependency) or the consumer passes a rates map into the helper.

Previously this was done by reading Pi session JSONL files from disk — impossible once sessions live inside the agent-server container under .pi-global/sessions/{projectId}/, so the client/contract path is the only one that survives the containerized architecture.

Suggested split: do (1) in this package now with an optional consumer-supplied rates map for (2); file the rates-in-models-endpoint question against agent-server.

[neuromaxer]

Would be cool to see session costs in appx and true that data is there. I wasn't planning on picking it up, so either will do later or you can give it a go

[alexanderkreidich]

i’ve already implemented this in create-appx-app. we can use that as the reference implementation and extract the reusable helper into agent-client if that makes sense.

[neuromaxer]

nice, makes sense, feel free to open PR in agent-client