Auto-fork thread with distilled context on rate-limit / quota cutoff

[Ephraim-9]

Problem

When working on large agentic tasks (especially multi-step refactors or research), sessions often hit model rate limits or session usage limits (e.g., Anthropic 5-hour quota).

By the time the limit window resets:

1. Prompt Cache is Cold: The 5-minute cache TTL has expired.
2. Context Penalty: Resuming the bloated thread (often 100k–180k tokens) immediately burns uncached input tokens on the very first "continue" prompt.
3. Pollution: The thread is cluttered with stale terminal output and dead ends.

Currently, the user must manually open a new thread and summarize the exact files, worktrees, and next steps to avoid burning tokens.

Proposed Solution

Add an "Auto-Fork to Fresh Thread" action (or banner when a provider reports quota_exceeded / session_limit):

1. Detect Rate-Limit Halt: Recognize when a provider session halts due to quota limits with a reset timestamp.
2. State Distillation: Generate a concise hand-off summary containing:
   • Target git branch / worktree
   • Modified / uncommitted files
   • Immediate next command or test failure
3. One-Click New Thread: A button on the rate-limit message: "Continue in New Thread with Compact Context". Clicking it opens a fresh thread pre-populated with the distillation, preserving the clean state while paying ~300 tokens instead of 150k.
4. Customizable Distillation Model (High Context / Low Cost):
   • The distillation step should not consume expensive flagship quota (e.g., Claude Opus / Sonnet).
   • Allow users to configure a dedicated, lightweight model for background distillation—such as Google / Antigravity harness (Gemini 3.8 Flash / Gemini 2.0 Flash), a local model, or any low-cost/high-context provider via ACP.
5. Optional Resumption Reminder: A desktop/app notification when the reset time arrives.

[Ephraim-9]

Added detail: the distillation runner should be customizable to use high-context, cost-effective models (like Gemini / Antigravity harness or custom ACP endpoints) so users don't burn expensive frontier model tokens just summarizing history.