Auto-Condense

Automatic context management through LLM-based condensation

What is it

Auto-condense is a context management system that automatically compresses conversation history when it approaches the model's context window limit.

Instead of losing early messages, Skycode uses an LLM to create a concise summary of the previous context. This enables long sessions without losing important information.

How it works

1. Skycode tracks token count in the current session
2. When context approaches the threshold, condensation triggers
3. The LLM analyzes the history and creates a compact summary: key decisions, modified files, current task state
4. The summary replaces old messages — recent messages remain as-is

When to enable

• Long sessions — refactoring, migration, series of related tasks
• Models with small context — 8K–32K tokens
• Complex projects — many files, frequent context switches

When not needed

• Short tasks (1-2 messages)
• Models with large context (200K+) on small tasks

Configuration

Settings → Context → Auto-Condense

Threshold

The trigger threshold is set via the context window bar in the task header. Click on the bar to set a marker — when context usage exceeds this mark, automatic condensation will occur.

Manual condensation

In addition to automatic condensation, you can trigger it manually — the AI uses the condense tool to summarize the current context at any point.