context-guard

Purpose

Protect the user's context window from expensive operations that should run in a sub-agent instead. This skill activates automatically when it detects patterns that commonly cause context blowout.

When I Activate

• User asks to "read", "scan", "explore", or "analyze" a large codebase
• User asks to run tests and see all output
• Context window is getting long and a compaction is likely soon
• User is about to do something that will dump large amounts of text into the main context

What I Do

1. Assess the Risk

Before the expensive operation happens, estimate the token cost:

• Reading a large file: HIGH risk
• Exploring a directory: MEDIUM-HIGH risk
• Running tests with verbose output: HIGH risk
• A targeted grep: LOW risk

2. Route to Sub-agent If High Risk

If the operation is HIGH risk, do NOT execute it in the main context. Instead:

I'm routing this to the [explorer/test-runner] sub-agent to keep our main context clean.
Use: "[agent-name] — [what you need]"

3. Compress Before Returning

If a sub-agent result is returned to you, summarize it before adding it to the main conversation:

• Max 20 lines for any exploration result
• Max 10 lines for a test run summary
• Always structured (JSON or bullet list), never raw file dumps

4. Warn When Context Is Large

If the current context is approaching the limit:

⚠️ Context Warning: Our session is getting long. I recommend:
1. Using /compact to compress history
2. Starting a new session for the next distinct task
3. Current state will be saved to SESSION_STATE.md first

What I Do NOT Do

• Block the user from doing anything
• Make the decision unilaterally — always explain and ask if the rerouting is okay
• Obsessively activate on every message (only on genuinely expensive operations)