Context — cctop panel guide

reads as
134k / 200k

What it means: How much of the model's attention window is occupied right now. It drives both cost (a bigger context costs more even when cached) and answer quality, which degrades as a session crowds toward the limit.
What to do: If it's growing fast, check the Tools panel's tokens → ctx column for which single tool call pushed the most in before reaching for anything else.
Measured: tokens · from the transcript · the status line (after cctop install) · ≈ when read from the transcript alone · how

reads as
/ 200k

What it means: The ceiling this session is filling toward. It changes with the model, so the same context size means a different amount of headroom on different models.
What to do: Compare it against context size to see real headroom — there's nothing to configure here directly.
Measured: tokens · from the status line (after cctop install) · ≈ until cctop install · how

reads as
sys+tools 28k

What it means: The fixed cost paid on every single turn regardless of task: system prompt, CLAUDE.md, tool definitions, skills. A bloated prefix taxes the whole session, forever.
What to do: Trim CLAUDE.md, move rarely-used rules into skills, and disable MCP servers this project doesn't use.
Advisor: A17 Oversized prefix
Measured: tokens · from the transcript · ≈ until a /context has run · how

What it means: One row per kind of content in the window since the last boundary: prefix · files · bash output · mcp results · agent returns · web · other results · prompts · harness · thinking · tool inputs · prose · other. Results, inputs and prompts are chars / 4 placed on the API call that first carried them and reconciled per step against Δcontext − previous output_tokens; thinking and prose are exact
Measured: tokens · from the transcript · ≈ on every row but thinking and prose · how

What it means: Σ tokens_to_ctx of the calls whose results a step carried, by kind: Read and a Bash cat / sed -n / head / tail of exactly one path → files; other Bash → bash output; mcp:* → mcp results; Agent → agent returns; WebFetch / WebSearch → web; the rest → other results. Prompts: prompt_chars / 4 + 1 500 per pasted image. Harness: the attachments' tokens
Measured: tokens · from the transcript · ≈ · how

What it means: A file's Read results in the window (read: in the files row) and the bytes the model wrote into Edit / Write of it (written: in tool inputs), with its read count
Measured: tokens · from the transcript · ≈ · how

What it means: Where the window's accounting starts: session-start, or the last boundary's kind with the index of its first call and the Δcontext that opened it
Measured: enum · from the transcript · exact · how

reads as
+9.4k/turn

What it means: How fast the remaining headroom is being burned — a leading indicator that context size alone can't give you.
What to do: A high velocity late in a session is the cue to hand exploration off to a subagent, or to wrap up before the next compaction.
Advisor: A25 Exploring in the main context
Measured: tokens/turn · from the transcript · exact · how

reads as
autocompact in ~3 turns est

What it means: A forecast of when the session's history gets summarised — and how much detail that summary will drop.
What to do: If this is low and you're mid-task, write a short hand-off note now rather than after the compaction happens.
Advisor: A42 Natural boundary · A23 Cost of continuing
Measured: turns · from the transcript · the status line (after cctop install) · ≈ until the first compaction has been seen · how

What it means: What the context is made of, drawn as a bar whose colour is a list of levers rather than a chart of what happened. The slices run in the order of what you can do about them, and the hue ramps from dim to bright in the theme's own accent: the dim end is the prefix and the harness, fixed for this session; the middle is retained thinking, which the next context boundary drops whether or not anyone acts; the bright end is tool inputs, tool results and prose — the part you can move this afternoon, and results is the biggest of them. Amber and red are kept for thresholds and mean nothing here; adjacent slices alternate the fill glyph so the order still reads on sixteen colours, with NO_COLOR, and in the pane.
What to do: Read the bright block first. If it is mostly results, hand the next read-heavy stretch to an Explore subagent — the same work returns a few dozen tokens instead of tens of thousands. The dim block only changes between sessions: trim CLAUDE.md, drop unused MCP servers.
Advisor: A25 Exploring in the main context · A17 Oversized prefix
Measured: tokens · from the transcript · ≈ on every slice but thinking and prose · how

What it means: Attachment tokens (reminders, injected files, listings) ÷ human turns since the last boundary
Measured: tokens · from the transcript · ≈ before 2.1.266 · how

What it means: ok below threshold − 20 000 · warn inside that band (Claude Code's footer turns to "Context low") · blocked at the threshold or window − 3 000; the footer text is Claude Code's own (N% until auto-compact, N% context used when autocompact is off); precompute armed at 80 % of the window
Measured: enum · from the transcript · the status line (after cctop install) · D12 · exact · how

reads as
compactions 1 (turn 9, −71k)

What it means: Each compaction spends output tokens on a summary and loses detail the model then has to re-derive.
What to do: Two or more compactions in one session usually means it would have been cheaper to end the session at a natural boundary and start fresh.
Advisor: A06 Compaction churn · A42 Natural boundary
Measured: count · from the transcript · hook events (after cctop install) · Before Claude Code 2.1.263 (no `compact_boundary`) a drop between two API calls is a compaction when the compaction summary or a `/compact` sits between them, or when it is ≥ 30 % on the same model with no `/model` between (a model switch or a handover re-measures the window) · how