A clearer view of context is not the same as exact attribution
The Turn Inspector mod makes Claude Code’s context window legible as a sequence of requests rather than a single growing token total. The thesis: turn-level visibility can help developers diagnose context growth, but block-level token attribution remains an estimate. The repository README describes a side pane with one row per model request, context-window usage, and expandable input/output blocks.
The mechanism: the mod observes turn and tool-call hook events, while splitting request-level usage across individual blocks by estimated size. The README says request input and output totals come from API usage and are exact; allocations to blocks such as thinking, text, and tool use are estimates because the engine reports one total per request. It also says the mod passes hook events onward unchanged. Why now: as Claude Code sessions accumulate tool round trips and earlier turns, a per-turn view offers a practical way to inspect where context grows. This is a visibility aid, not a change to the model’s context handling.
The evidence boundary: these claims come from the mod’s own GitHub README; they do not establish independent accuracy testing or prove that every displayed block can be measured precisely. For builders, the first implication: use the view to locate likely sources of context growth, while treating block-level figures as directional rather than exact. The first watch condition: look for independent tests or revisions that change how resumed-session estimates and block allocations are calculated.