Recovery & Context

Last updated · 2026-09-05

Checkpoints and run recovery

A VibeSpace run advances through checkpoints: as each stage completes, the run state (conversation history, tool progress, pending approvals) is persisted to the local database.

In practice:

  • Quitting the app loses nothing: reopen VibeSpace and unfinished runs offer a resume entry point, continuing from the latest checkpoint;
  • Automatic reconciliation: on resume the app reconciles tool calls — what already ran and what was interrupted — so completed work is not repeated;
  • No duplicate approvals: operations you approved stay approved after recovery;
  • Cancel anytime: recovered runs can be cancelled like any other; their changes remain reviewable in the ChangeSet.

The context window panel

In long sessions the model's context window is a finite resource. The panel next to the timeline shows, in real time:

MetricMeaning
System promptTask setup and injected skill instructions
ConversationMessages exchanged so far
Source contentAttachments and referenced file contents
Tool resultsWhat tool calls returned
Total / maxCurrent usage of the window

Context compaction

As a session approaches the model's window limit, VibeSpace runs compaction automatically: history is summarized without losing key information so work can continue. Compaction events are annotated in the panel and timeline — no manual action needed.

Tips for long tasks

  1. One task, one goal: prefer splitting oversized jobs into multiple tasks over endlessly appending input;
  2. Watch total usage: when the panel nears the limit, start a fresh task and persist results into files first;
  3. Lean on recovery: an interruption (shutdown, quit) is not a failure — checkpoints let you continue instead of starting over.