See an LLM context window the way the model does — one token sequence that grows every turn. Find buried instructions, spot the wrong turn, and know when to compact.
Read it as: each turn adds a layer. When the white total line climbs toward the dashed budget, that's your “time to compact” signal. The ▼ marks where output quality broke.
Read it as: the full token sequence, oldest on the left, the generation edge (where the model writes its next token) on the right. Block width = token count. An important instruction sitting far to the left is at risk of being ignored. Click any block to inspect it.
Read it as: where your token budget is actually going right now. Big tool-result or retrieved bars mean noise is crowding out signal.
Read it as: details for the block you clicked — including how far it sits from the generation edge.
Paste a JSON array of blocks: { id, type, label, tokens, enter, exit?, cached?, ref?, content? }. type ∈ system, tooldef, skill, user, assistant, thinking, tool_call, tool_result, retrieved, summary. enter = turn it entered; exit = turn it was evicted.