claude-code - 💡(How to fix) Fix Long sessions are a user workaround for missing persistent context — not wasteful behavior [2 comments, 1 participants]

The Misframing

The /usage panel flags long sessions with: "These are often background/loop sessions. Continuous usage can add up quickly so make sure it is intentional."

This framing is wrong for a significant class of users. Long sessions are not accidental. They are the only available solution to a problem the system does not otherwise solve: persistent behavioral context.

What a Long Session Actually Is

This user ran a single session from September 19, 2025 to December 5, 2025 — 43 days, 15 hours, 516MB, 46,031 lines. That session was not a background loop. It was continuous, intentional, productive work.

The reason it was kept alive: every session reset costs usage and time to re-establish context that was already established. Math-heavy, cross-domain work requires Claude to understand a non-traditional framework before it can contribute usefully. That understanding is behavioral — built through correction and repetition — not informational. It cannot be summarized and pasted into a new session. It has to be rebuilt from scratch every time.

Keeping the session alive was the user's solution to a system that provides no persistent behavioral context. It was not waste. It was engineering a workaround for a missing feature.

What Happened When It Was Cut

The v1→v2 Claude Code update in December 2025 made the 516MB session unresumable. /resume returned "Session not found." The context was gone.

Everything that followed traces back to that break:

• Work split across multiple AI tools (Kimi for generative work, Claude Code for integration)
• Usage workarounds and throttling to survive session limits
• 41 failure reports filed over 5 months
• $20+ in extra charges in a single billing period
• Sessions starting cold, requiring re-establishment that costs tokens before any productive work happens

The 43-day session was not the problem. Cutting it was the problem.

The /usage Panel Gets It Backwards

Warning users about long sessions treats the symptom as the cause. The cause is missing persistent context. The long session is the user's patch for that missing feature.

Telling this user to /compact mid-task or /clear when switching tasks would destroy the accumulated behavioral context that makes the session functional. That advice optimizes for token cost at the expense of the actual value being preserved.

What Is Actually Needed

A stable, persistent behavioral context layer that survives session resets and model updates. Not a summary. Not a memory file. A behavioral context — the accumulated tolerance and framework understanding that takes dozens of correction cycles to build and cannot be reconstructed from documentation.

The user has built their own partial solution to this. It works within constraints. What would make it fully viable is an official system-level injection port: a sanctioned interface for injecting persistent behavioral context at the system prompt level, separate from conversation, that survives resets.

Anthropic does not need to build the context. They need to open the port. The users who need it most already know what to put in it.

The User's Input Pattern

Input is primarily mathematics and application of mathematics — maximally token-efficient. One line of notation carries the meaning of paragraphs. The usage cost is almost entirely on the output side: Claude generating responses, often wrong ones at Opus rates before eventually confirming the user was correct.

The long session minimized this cost by preserving the state where Claude was already calibrated to the framework and could respond correctly on the first attempt rather than after multiple correction cycles.

Summary

Long sessions = working persistent context. Session reset = re-establishment cost paid in tokens and time. 43-day session cut by update = inflection point for all subsequent problems. Warning users about long sessions = wrong diagnosis, wrong prescription.

Related: #52015, #52030, #52031