claude-sdk-rs: Command for Claude Code

Initial Research — Conduct reconnaissance on a topic and report back

Conducts deep research into the codebase or external resources based on a detailed description. Reports back with highly structured, comprehensive context. Optionally captures this research directly into a pre-plan notes file if the --capture flag is provided.

Variables

$ARGUMENTS — a detailed description of the research topic, optionally including the --capture flag. Example: "--capture How does our current authentication flow handle token refresh?" Example: "Investigate the memory leak in the data processing pipeline"

Execution Model

Spawn a subagent (Agent tool) to execute all steps below. Pass the resolved $ARGUMENTS and this whole Instructions section in the subagent prompt. Return the subagent's result to the user.

Instructions

Step 1 — Parse arguments

1. From $ARGUMENTS, determine the core research goal.
2. Check if the --capture flag is present.

Step 2 — Conduct the research

1. Act as a senior engineer conducting a deep reconnaissance mission. Investigate the topic thoroughly by reading code, documentation, and existing architectural decisions — but reading is not the only mode. Where the subject can be run, run it once; that outranks reading about it. A command, a script, a test, a server endpoint observed actually executing settles a question that reading its source can only suggest. Read first to know what to run, then run it.
2. Critical Mandate: You must gather ALL important details during your investigation. Do not just summarize. You MUST collect:
   • File paths
   • Class/struct names, function signatures
   • Important snippets of relevant code
   • Any architectural nuances, constraints, or context that clarifies why things are the way they are.
   • Pointers on where to look to review/investigate further. This content must be EXTREMELY easy for the next agent or the user to dig into and know exactly what was discovered and how you reached your conclusions.