Feature request: graceful performance degradation preferred over silent context compaction
Resolved 💬 3 comments Opened May 5, 2026 by jodybrownell Closed Jun 4, 2026
The Problem
When a conversation approaches the context limit, Claude Code silently auto-compacts the conversation — summarizing earlier messages to make room. The user gets no warning, no prompt, no choice. Their working context is wiped without consent.
This happened to me mid-session while building a production application. I lost a multi-day conversation's worth of working context — exact code discussions, decision rationale, nuanced back-and-forth — with zero warning. I'm paying for this context window. Having it silently taken away is unacceptable.
What Should Happen Instead
Performance degradation should be preferred over compaction. Specifically:
- Never auto-compact without explicit user consent. Ask the user: "Context is getting deep — do you want me to compact now, or keep going?" Wait for a yes.
- If the user says no, degrade gracefully:
- Slower responses are acceptable
- Reduced tool parallelism is acceptable
- Shorter responses are acceptable
- Losing working context without consent is NOT acceptable
- Give the user visibility into context usage. A simple indicator (percentage used, messages remaining, something) so they can make informed decisions about when to compact on their own terms.
- If compaction is truly unavoidable (hard limit), warn with enough lead time for the user to run
/compactthemselves and control what gets preserved.
Why This Matters
- Users pay for context window capacity. Silent compaction removes what they paid for.
- For professional users building production software, the conversation context IS the working memory. It's not disposable.
- Trust in the tool requires predictable behavior. Silent data loss — even "just" conversation data — destroys trust.
- The user's context is their property. Wiping it without consent is a UX failure regardless of the technical justification.
Environment
- Claude Code CLI
- Claude Opus 4.6 (1M context)
- Long multi-day session with heavy tool usage (hundreds of tool calls, large diffs)
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗