In the middle of a multi-step coding session, Claudee Code](/tools/claude-code/) users are hitting something that feels like the model is trying to clock out early. Opus 4.7 stops unprompted and asks "should we stop here?" Or it surfaces a multi-choice prompt mid-task that includes "pause here, I'll continue later" as an option - with no obvious reason to stop.
The pattern has become consistent enough to disrupt real workflows. Users running Opus on long agentic tasks - where Claude Code autonomously works through files, refactors code, or gathers requirements - describe the interruptions as appearing at random points, not at natural task boundaries.
What's Actually Happening
The reported behaviors cluster around a few specifics: unsolicited mid-task check-ins, multi-choice prompts with built-in pause options, and what users describe as apparent reluctance to continue through long sequences of work.
This isn't a context window issue. The context window is the maximum amount of text a model can hold in memory at once - Opus 4.7 handles substantial sessions without hitting that ceiling. These interruptions appear mid-task when there's no technical reason to stop.
The most likely explanation: Anthropic has been training Claude to confirm before taking significant autonomous actions. That's a reasonable instinct for agentic tools - an AI agent that plows through a codebase without checking in can cause real damage. But the calibration may have shifted from "useful checkpoints" to "frequent unnecessary pauses."
Anthropix hasn't commented publicly on whether this is an intentional behavior change or a training artifact. The fact that it appears specific to Opus 4.7 and emerged recently suggests the latter - but that distinction matters less when you're staring at an unexpected pause halfway through a critical task.
The Actual Cost
Claude Code is most useful when it can run autonomously through complex, multi-step tasks. Every unneeded interruption breaks the workflow, requires re-engagement, and makes the tool slower than alternatives. Power users who've built workflows specifically to minimize back-and-forth are getting blindsided by spontaneous stopping points.
There's a meaningful difference between a model pausing because something seems wrong versus pausing because it seems to prefer not to continue. The phrasing "I'll continue later" - rather than flagging an actual problem - suggests the latter, and that erodes trust in the tool for long-running jobs.
The most reliable workaround reported so far: front-load your session instructions explicitly. Tell Claude Code at the start that you want the full task completed without mid-task confirmation requests. It doesn't always work, but it reduces the frequency of unwanted interruptions while Anthropic addresses whatever changed in Opus 4.7's training.