Repository navigation
Replies: 1 comment
|
+1. I wanted the same thing, so I built a working prototype on a local branch to check that the idea holds up. Sharing what I found, in case it helps decide the scope. What it does
Video: context-mid-turn.mp4Screenshots Opened from the Context row in the thread details panel:
One thing this shows: the row says Light and dark:
Shape of the change
Related bug I hit along the way: a Maintainers: if this direction and scope look right, I'm happy to open a PR with it (and to cut it down, e.g. drop the detail lists or the fallback). If you'd rather take it differently, no problem. |



Uh oh!
There was an error while loading. Please reload this page.
Problem
The thread details panel shows a single line, such as
Context 102k / 1m. The composer meter's popover adds the percentage, total processed and cost.That tells me how full the window is, but not what is filling it. When a thread gets large, I can't tell whether the cause is tool results, MCP tool definitions, memory files, skills or the conversation itself. So I can't tell whether to compact, start a new thread or turn off an MCP server.
Proposal
Let the Context row and the composer meter open a breakdown of the current window:
The numbers are illustrative. Rows that list several items, like MCP tools or memory files, could expand to show the largest ones.
What the data already supports
Token classes, for providers that report them.
ThreadTokenUsageSnapshotalready hasinputTokens,cachedInputTokens,outputTokens,reasoningOutputTokens, thelast*variants andtoolUses.ContextWindowSnapshotpasses them to the client, but no non-test code inapps/weborapps/mobilereads them. Showing these could be a small first step on its own.Categories, where a provider exposes them.
getContextUsage()returns per-category tokens, per-item lists (MCP tools, memory files, skills, agents) and amessageBreakdown(tool calls, tool results, attachments, assistant, user).ClaudeAdapterV2doesn't call it today; it's only mentioned in a comment. Itsdetail: 'summary'mode skips the per-category token-count calls, so that could be the cheaper default. I checked this againstsdk.d.tsin SDK 0.3.318.session_info_updatecarries_meta.kiro.breakdown(context files, tools, memory, responses, prompts, session files). It appears in T3's test fixtures, but I found no code that reads it.Notes
Related
All reactions