Sandbox Routing

Sandbox Routing

[!danger] HARD RULE — fires before any large Read or Bash that isn't for editing If you're about to Read a long file to analyze (not to Edit), or run a Bash command with >20 line output, route through sandbox tools instead. Raw output stays out of context; only the summary you print enters context.

This skill assumes your Claude Code install includes context-mode (or equivalent) sandbox tools — ctx_execute, ctx_execute_file, ctx_batch_execute. If those aren't available, the principle still applies: use head_limit on Grep, narrow Read to specific line ranges, and avoid cat large.log.

The decision tree

Am I about to Read or run Bash?
   ├── Editing a file? (need full content for Edit tool)
   │     → Read is correct, proceed
   │
   ├── Running git status / git log -n short / mkdir / mv / rm / short command?
   │     → Bash is correct, proceed
   │
   ├── Reading a file just to extract a piece of info / find a line?
   │     → STOP. Use ctx_execute_file with grep/awk code that prints just the answer.
   │
   ├── Running a shell command with >20 line output?
   │     → STOP. Use ctx_execute with the same command — only printed summary enters context.
   │
   └── Multiple commands at once?
         → Use ctx_batch_execute with all commands as labelled chunks, indexed for FTS5 search.

When to use each tool

ctx_execute(language, code)

• Shell command with large output → language: "shell"
• Python data processing → language: "python"
• API calls + response analysis → wrap the call in code
• Anything where you want the answer, not the raw output

ctx_execute_file(path, language, code)

• File analysis without loading the file into context
• Searching within a specific known file
• Extracting structured info from JSON/YAML/MD

ctx_batch_execute(commands, queries)

• Multiple commands you want to run + index together
• Chained recon (e.g., git log + git diff + git status as one batch)
• Building a knowledge base for later FTS5 queries