Detailed Analysis
A user report circulating on Reddit describes a specific and reproducible failure mode in Claude: conversations that approach long context lengths trigger an automatic "compacting" process, and once that process crosses roughly 90% completion, the session throws an error and halts rather than finishing the compaction and continuing the chat. The user notes they attempted the same workflow across multiple computers, browsers, and both the desktop and web clients, ruling out local environment issues, device-specific bugs, or browser incompatibilities as the root cause. This points toward a server-side or model-level issue rather than a client-side rendering or caching problem, which is a meaningful distinction for diagnosing the bug's true source.
Context compaction is a feature many AI chat products use to manage long-running conversations: rather than truncating older messages outright when a context window fills up, the system summarizes or compresses earlier turns to preserve continuity while freeing up token budget. This is especially important for Claude given its emphasis on long-context capabilities (with some models supporting context windows of 200K tokens or more) and its popularity among users who run extended coding sessions, research threads, or multi-step agentic workflows. When compaction fails partway through, users lose the ability to continue what may be hours of accumulated conversation, discussion, or in-progress work, which is a disproportionately painful failure compared to more graceful degradations like simple truncation or a soft warning that a new chat is needed.
The complaint also highlights a common pain point for users of frontier AI products more broadly: support responsiveness during acute technical incidents. The remark that "Claude support is sleeping" reflects frustration that's common when users hit an edge case that customer support teams haven't yet triaged or acknowledged, especially for bugs that may only affect a subset of sessions or a specific model version and thus aren't immediately visible on public status pages. For power users and developers building products atop Claude's API, an unpatched compaction bug can be more than an inconvenience — it can break long-running agent loops, coding assistants, or automation pipelines that depend on session continuity, making rapid diagnosis and communication from Anthropic important.
More broadly, this kind of report is emblematic of the operational growing pains that accompany the rapid scaling of consumer and developer-facing AI tools. As context windows grow and agentic, multi-turn use cases become more central to how people use Claude (versus quick single-turn Q&A), the infrastructure supporting summarization, memory, and session management becomes as critical to user experience as the underlying model's raw capabilities. Bugs at the edges of these systems — like a compaction routine crashing near completion — tend to surface unevenly, affecting heavy users first, which is exactly the profile of person likely to post detailed cross-platform bug reports on forums like Reddit. Anthropic's ability to quickly identify, communicate about, and patch such issues is a meaningful part of maintaining trust with its most engaged and technically sophisticated user base, even as the underlying model quality remains the primary draw.
Read original article →