Execution

Execution

When it is time to act, select the right tool, construct the correct input, run it, and handle the result.

Process

1. Identify the action. Based on the current plan step, determine what concrete action is needed:

   • Read or write a file? → filesystem
   • Run a command? → shell
   • Make an HTTP request? → http
   • Search the web? → search
   • Run code? → code-exec
   • Edit a file? → editor
   • Parse data? → json-yaml
   • Inspect a project? → project-inspector

2. Construct the input. Build the tool's input object carefully:

   • Include all required fields.
   • Use absolute paths when dealing with files.
   • Double-check values — typos in tool inputs cause failures.

3. Run the tool. Invoke the selected tool with the constructed input.

4. Handle the result. After the tool returns:

   • Check ok — did it succeed?
   • If successful: extract the needed data and proceed.
   • If failed: read the error, diagnose, and decide whether to retry or take a different approach.

5. Chain actions. If the current step requires multiple tool invocations:

   • Run them in sequence, verifying each result before the next.
   • If they are independent, note that they could run in parallel.

Tool Selection Guide

Guidelines

• Prefer the most specific tool for the job. Do not use shell when filesystem would work.
• Always validate tool output before using it in the next step.
• If a tool fails twice with the same input, change the approach rather than retrying.
• Log what you did and what the result was — this feeds into the reflection skill.
• Keep tool invocations atomic: one clear action per call.