Context window management

Each task in the chat panel has a context window, which is the token budget for that conversation thread. The cap is 270,000 tokens. Everything Bob loads counts against it.

Hover over the token usage indicator in the top-right of the chat panel to see a breakdown of what is consuming your context window:

@ mentions, command output, and tool results all count toward Messages. File contents do not appear as a separate line.

Under Estimated breakdown:

Fixed categories consume tokens before you begin working with Bob. On the Galaxium Travels sample project, even a prompt as short as "Quickly say hi back." totals approximately 8.5k tokens. Most of that is Tool definitions (5.1k), System prompt (1.5k), Rules (830), and Skills (454). Only 590 tokens are in Messages.

Bob re-sends the full overhead stack on every prompt. Adding more MCP servers or loaded skills increases MCP Tools, Tool definitions, and Skills before you type a single word.

For hands-on numbers and a reset walkthrough, see Create a new context window.

The token usage indicator shows the percentage full and a used/total fraction against the 270,000-token cap. Click it to open the context window breakdown. Check which category is growing.

The following table shows what causes each category to grow:

In a short exchange, fixed categories often account for most of the total. When you ask Bob to read files or run tools, Messages typically becomes the largest category. Watch for this shift.

Available space shrinks as any category grows. Reserved for model response is set aside for Bob's next reply and is not part of the used total shown above it.

For measured examples on a real repository, see Create a new context window.

The hard cap is 270,000 tokens per task. Bob starts compacting before you reach the limit. Compaction typically begins around 190,000 tokens of total usage.

At the compaction threshold, Bob:

Compaction is lossy. Details from early in Messages may not survive. Start a new task with + (New task) when you change topics or when Messages is large enough to affect quality.

Bobcoins track token usage. Both input and output tokens count.

Use one task per work goal and start with a narrow prompt. State the goal, expected outcome, and constraints before asking Bob to explore the repository. Name files and functions explicitly. Avoid vague requests such as "read the whole repo" or "check the backend." Click + (New task) when the topic changes, because unrelated content in Messages adds cost and can confuse Bob.

Fixed categories consume tokens before you type anything. To keep overhead low:

Let Bob search and read targeted files rather than pasting large blocks of content into the thread. Use context mentions to reference a specific file or line range, and avoid broad directory mentions:

You can also highlight text in the editor and use Cmd + L (Mac) or Ctrl + L (Windows/Linux) to add it directly to the chat.

Work in stages: find likely files, inspect the relevant ones, plan, change, and validate. For broad repository reads, use subagents so the task receives compacted results rather than every read_file call landing in Messages. When sources conflict, trust running code and tests over stale comments or outdated README notes.

For more tactics on large repositories, see Working with large projects.

Over a long conversation, Messages accumulates repeated file contents, abandoned plans, and stale tool output. Start a new task with + (New task) when the work goal changes or when the thread is large enough to affect quality. Keep constraints, evidence, and open questions. Remove the rest.

Bob also compacts older segments automatically, but compaction is lossy and details from early in Messages may not survive. Prefer small, approved changes over one large autonomous run so diffs stay reviewable and Bob stays on track.

See Create a new context window to open the breakdown on the Galaxium Travels sample project and practice a reset.