Detailed Analysis
A Reddit user's complaint about Claude sessions repeatedly "resuming" or resetting mid-conversation highlights a recurring pain point in how Anthropic's Claude Code and related interfaces manage context windows and token consumption. The poster describes a scenario where each new prompt triggers what looks like a cold restart of the session—context appears to double from roughly 200,000 to 400,000 tokens—even after running the `/compact` command, which is specifically designed to condense conversation history and reduce token overhead. The practical effect is that the user is being billed or metered (in this case against a third-party wrapper called "Fable" usage) as if the entire prior conversation is being reprocessed on every turn, rather than the incremental exchange the user actually typed.
This kind of behavior points to a structural issue rather than a one-off bug: large language model APIs, including Claude's, typically charge based on total tokens processed per request, which includes the full conversation history sent back to the model for context. If a client application isn't properly truncating, caching, or compacting that history—or if a bug causes the app to resend the entire session transcript instead of just new messages—costs can balloon rapidly without any corresponding increase in useful work. The user's frustration is compounded by the fact that Claude's own compaction tool, meant to summarize and shrink context, appeared ineffective in this instance, suggesting either a client-side implementation problem or an edge case in how compaction interacts with session state persistence.
The mention of "Fable" is notable because it signals this isn't necessarily a direct Anthropic first-party product issue but potentially a third-party application or wrapper built on top of Claude's API. This distinction matters enormously for troubleshooting and accountability: token bloat and unexpected billing can stem from how a downstream developer implements context management, session resumption, and caching—areas where Anthropic provides tools (like prompt caching and the Claude Code CLI's compact feature) but cannot fully control third-party integration quality. Users often conflate "Claude is broken" with "the app built on Claude mishandled the API," and support teams for either the platform or the wrapper may struggle to diagnose which layer is at fault, especially when the symptom (rapid session/context reset) looks identical from the user's vantage point regardless of where it originates.
More broadly, this complaint reflects a persistent tension in the AI assistant ecosystem between long-context capabilities and cost transparency. As models support increasingly large context windows, the risk grows that inefficient session management—whether from bugs, poor caching strategies, or unclear resumption logic—can silently multiply costs for end users. It also underscores the importance of robust, responsive support channels when usage-based billing is involved, since users have limited visibility into token-level accounting and must rely on either the model provider or the app developer to explain and fix unexpected spikes. As more consumer-facing products layer subscription or credit-based pricing on top of raw API costs, incidents like this are likely to recur until context management, compaction, and billing transparency become more standardized and better tested across the ecosystem.
Read original article →