See how full the agent's context window is

Every agent works inside a context window, and the fuller it gets the worse it works: it forgets earlier instructions, repeats itself, and at some point the agent compacts the conversation into a summary and loses detail. Sessionboxer shows where you are.

The gauge above the prompt box (a bar across the whole chat) reads 32.6k / 1M 3%: tokens in the window after the agent's last reply, the window's size, and the share used. Its bar goes from green when empty towards red as the window fills, and once more than half is used it turns red and says rotting — that is the point from which you should think about a fresh session, a fork from an earlier snapshot, or asking the agent to /compact. Next to it, ♻ 2 counts the compactions so far in this session, with a ⚠ as soon as there has been one, since the agent has already lost part of the conversation. Hover the gauge for the exact numbers and the session's cost so far (Claude Code reports one).

Each turn divider in the chat carries what that turn did to the window: +12.3k context · 3 model calls · in 2 / cached 30.1k / written 2.5k / out 610 · $0.04 — how much the window grew or shrank, how many times the model was called (one per tool round), the turn's input / cache-read / cache-written / output tokens, and its cost. A compaction shows up as its own marker, ♻ Context compacted 180k → 42k (automatic).

Click a compaction marker to see what the agent actually did: the dialog lists the messages that were compacted away (for Claude Code, the ones it carried into the new window word for word are flagged kept verbatim) next to the summary it became — the text that is, from then on, all the agent remembers of that part of the conversation. The details are read from the agent's own records in the box when you open the dialog (Claude Code's transcript file, Devin's session database; Codex's records cannot be read yet and the dialog says so), so the box must be running; a stopped session says so. Devin does not record the trigger or the size after compaction, and the dialog says that too rather than guessing.

Click the gauge, or Context in the header's ⋯ menu, for the Context pane: the occupancy, the compactions, a history of the window over the session (the half-way line marked, drops are compactions), and the breakdown by category, taken by asking the agent /context while it is idle: for Claude Code that is the system prompt, built-in tools, MCP tools (each of your servers' tools with its tokens), memory files (CLAUDE.md), skills, messages, free space and the autocompact buffer; for Devin the system prompt, tools, messages and free space, estimated. Take breakdown / Refresh breakdown asks again (a Context inspected: … marker appears in the chat; the exchange itself is not part of the conversation), and the raw report is a click away.