A /usage readout traced 79 percent of one developer's Claude Code usage to sessions past 150k context

The readout is bucketed by context size, and it pays to read that literally. /usage reported that 79 percent of his usage occurred at over 150k context ^[2]. The unit there is context, not prompts. A meter that can slice usage that way is sensitive to how much context each request carries. The post describes Claude Code working from previous conversation history, files it has read and tool results, as well as the latest sentence typed ^[5].

During an end-to-end check, the post lists the same loop repeated: open a page, inspect elements, click a button, fill in a form, check the state after navigation, then fix something and try again ^[8]. Relatively large tool results land in context at those steps, which the author treats as a necessary cost of testing ^[8]. His complaint is about afterwards, when continuing in the same session makes subsequent processing heavier ^[13].

The three lines overlap. Add them and you get 176 percent ^[11]. A single browser session can sit above 150k context, spawn subagents and be attributed to the playwright server at the same time, so at least 76 points of that total is the same usage counted more than once ^[11]. He read the three lines as three separate problems at first, then decided they were connected ^[15]. An operator cannot assume that clearing after a browser test recovers 36 percent.

The overage itself has a different cause in his own telling. He wrote that "Fable usage was what had caused me to go beyond the included usage and consume additional credits" ^[3], and, on the browser server, "That did not mean Playwright MCP itself was the reason I had exceeded my usage limit" ^[4]. What confused him first was that his plan usage still had plenty of room while the additional credits he had configured were exhausted ^[1]. The post calls itself "a short learning log about how I got there" ^[10].

/clear resets the session, so whatever the next task genuinely needs gets read again ^[14]. So the rule he settled on is conditional. "I do not clear the session mechanically after every test," he wrote ^[7]. The test he applies is "Will I need this context for the next task?" ^[6].

For his percentages to mean anything elsewhere, several things have to match: browser loops long enough to push a session past 150k context ^[2], subagents carrying part of the work ^[2], and a plan that meters included usage separately from top-up credits ^[1]. The 150k figure is his tool's bucket boundary. The gap he names, between context he has stopped needing and context the session is still carrying, travels without any of those conditions ^[9].