Auto-summarize Agent Panel chats on context limit for endless conversations to give a better dx Roadmap fitEnquiry

[k5602]

Zed's Agent Panel hits hard context limits (e.g., 120k-200k tokens) and errors out (HTTP 403), forcing users to manually "New from Summary" or restart threads. This breaks flow in long agentic sessions, unlike VS Code Copilot Chat or cli tools, which auto-summarize/condense history to continue seamlessly.

Solution Proposal
Add optional auto-summarization:

• Monitor token usage (already tracked).
• At ~80-90% limit, trigger LLM to summarize history (e.g., via a "summarize_thread" tool or prompt).
• Replace/append condensed summary to context, reset counter, and continue in same thread (with UI toggle: "Enable Auto-Continue").

Pros:

Matches user expectation for "endless" chats, boosting productivity and Improve DX.
Builds on existing "New from Summary" logic and token counters.
Configurable to avoid perf hits (e.g., opt-in, model-specific).

Cons:

• Extra LLM calls add latency/cost (mitigate with small models or local LLMs).
• Risk of summary drift/loss (add user confirmation prompt).

This aligns with requests for better context mgmt (e.g., #33198, Cursor-like agents in #24924). Happy to prototype a PoC, or scope a PR Thoughts? Roadmap fit?​

[dcrdev]

This would be great to have

[pleibers]

i would also suggest to not only summarize when a token limit is about to be hit but also after sub tasks are done/output that becomes irrelevant so the context does not get bloated in the first place