Skip to content
Enso

Context

What it shows

Section titled “What it shows”

Everything pi sends the model with each request is its context: the system prompt, the tool schemas, the conversation so far and every tool result. The Context tab shows what the next request will be made of, and how that grew request by request.

It gives you two numbers, and does not choose between them:

  • pi’s estimate, written with ≈. pi counts characters and divides by four. It can be split by category, so the breakdown uses it.
  • The provider’s count, written plain. It is exact, but only exists for a request that was sent, and comes as one total.

How to open it

Section titled “How to open it”
  • Open a thread from the rail, then choose the Context tab at the top of the thread.
  • Or /context in the chat, which opens the Now card and the Context Browser in a dialog.

What it looks like

Section titled “What it looks like”

The Now and Trend cards

  1. pi’s context meter: the provider’s count for the last response, plus pi’s estimate of what has landed since, against the model’s window. On a thread with nothing running, this side shows the provider’s count alone.
  2. pi’s estimate of the next request, as a bar split into categories: system prompt, tool schemas, user messages, injected context, assistant messages and tool results.
  3. Trend: one bar per model request (Requests) or per prompt (Prompts). Total shows each bar’s whole makeup, Change shows what it added and removed.
  4. The compaction mark on a bar. The bar after a compaction is short, because a summary replaced what came before.

The card under the chart is the request you picked. Its In row lists what entered since the request before, and Open in browser opens it in the Context browser.

What to look for

Section titled “What to look for”

Look for requests that are almost all cache write. In the thread above, Request 9 is the first one after the switch from claude-sonnet-4-5 to claude-opus-4-1. Its bar is no taller than the one before it, but a new model has no cache, so the whole prefix was written again at the higher price: its card reads cache read 0 (0%). On the Stats tab, the same turn is a solid cache-write bar and the peak of Cost per turn. In the Context Browser’s Context Events, the same request carries a switch event. Three views, one cause.

The compaction, at Request 12, is the other pattern. Its card says compacted from 5.3K, and In shows a compaction summary where the conversation used to be. Its bar barely drops, and the breakdown says why: in this small thread most of the context is tool schemas and the system prompt, and a compaction summarises the conversation only. When a compaction frees less than you expected, look at what it cannot touch.

Expect the two numbers to differ. The provider counts real tokens and adds its own framing around the tools; pi counts characters. Compare each number with itself from request to request, rather than one with the other.

What it doesn’t do

Section titled “What it doesn’t do”
  • The breakdown divides pi’s estimate, not the provider’s count. A provider gives one total per request, so no split of it exists to show.
  • An estimate is characters divided by four. It is good for comparing requests with each other, not for predicting a bill.
  • It shows the current branch. A request on a branch you forked away from is not here.

Go deeper

Section titled “Go deeper”