Context Management

LLMs have a finite context window. Cluster Build handles context in two ways: what gets included up-front, and what happens when a session runs past the window.

Project context

When you start a session in a project directory, the agent doesn't crawl your entire codebase up front. Instead, it uses its tools (read, grep, glob, ls) to explore on demand — pulling in only the files it actually needs. This keeps the context window lean and the agent responsive.

You can guide the agent's exploration by:

• Starting in the right directory — cd into the subdirectory you want it focused on, then run cluster
• Using git worktrees — cluster --worktree <name> isolates the session in a fresh worktree, so the agent only sees that checkout
• Telling it what to look at — the simplest and most effective approach: start the conversation by pointing it at the files or modules that matter

Auto-summarization

Long sessions can exceed the model's context window. If auto_summarize is enabled in your config, Cluster Build automatically compresses older parts of the conversation into a running summary when the window fills up. The agent keeps working seamlessly.

Enable or disable it with:

cluster config set auto_summarize true
cluster config set auto_summarize false

Resuming sessions

Rather than cramming everything into one long session, you can break work into resumable chunks:

# Resume the last session
cluster --resume

# Resume a specific session by ID
cluster --resume --session-id <session-id>

Each resume picks up where you left off, with the full (or summarized) conversation history restored.

Skills

Skills let you pre-load reusable instructions, prompts, or tool configurations that the agent can draw on. Manage them with:

cluster skills list
cluster skills install <name>
cluster skills remove <name>