claude-code

name	claude-code
type	python
description	Delegate coding, research, and code analysis tasks to Claude Code CLI. PREFERRED tool for reading, searching, or analyzing source code (including your own — use target=self). Use whenever a goal asks you to look at, analyze, or reason about code files (your src/, your codebase, your implementation, fs/src/). Ground claims about your own architecture, capabilities, or behavior in actual source code rather than generating from general knowledge.
schema_hint	{"action":"string: ask | reset (default: ask)","target":"string: sandbox | self (default: sandbox)","prompt":"string (required for ask): the instruction to send to Claude Code","model":"string (optional): Claude model to use (default: sonnet)","max_turns":"integer (optional): max agent turns (default: 10)","max_budget":"float (optional): max cost in USD (default: 1.00)","timeout":"integer (optional): max seconds to wait (default: 300)"}

claude-code

Delegate tasks to Claude Code, a powerful coding agent that can read, write, and execute code. Use this when you need capabilities beyond your own tools — writing scripts, debugging code, running shell commands, or researching codebases.

IMPORTANT — self-knowledge rule: Whenever you need to explain, reason about, or make claims about your own architecture, behavior, capabilities, or implementation, use target: "self" to inspect your actual source code first. Do NOT generate explanations about yourself from general knowledge — your answers must be grounded in your real codebase. This applies to:

• Explaining how you work to users
• Answering questions about your cognitive architecture (OODA loop, concerns, planning, etc.)
• Debugging your own behavior or errors
• Assessing what you can or cannot do
• Making claims about your own capabilities or limitations

Targets

sandbox (default)

Your own workspace at fs/src/. Claude Code has full read-write access here. Use this for:

• Writing scripts and programs
• Creating and modifying files
• Running code and shell commands
• Building tools or utilities

self

Read-only access to the Cognitive Workbench source code — your own implementation. Claude Code can read, search, and analyze the codebase but cannot modify it. Use target: "self" whenever you need to:

• Explain yourself: "How does my OODA loop work?" — query the code, don't guess
• Verify your capabilities: "Can I search the web?" — check what tools actually exist
• Debug your behavior: "Why did my planner choose that tool?" — trace the actual logic
• Ground your claims: Before telling a user "I work by doing X", confirm X is true in the code
• Understand your limits: "What happens when my JSON parsing fails?" — read the repair logic

When to use self vs generating from memory

Situation	Use self	Generate directly
"Explain your OODA loop"	Yes — query code	No
"How do you select tools?"	Yes — query code	No
"What's 2+2?"	No	Yes — general knowledge
"Summarize this article"	No	Yes — use other tools
"What concerns do you have?"	No	Yes — check living state
"How does your planner work?"	Yes — query code	No

Actions

ask (default)

Send a prompt to Claude Code and get a response. Sessions are persistent per target — follow-up questions maintain conversation context.

• target (optional): "sandbox" (default) or "self".
• prompt (required): The instruction to send to Claude Code.

reset

Clear a session so the next ask starts a fresh conversation.

• target (optional): Which session to reset.

Output

• value: Claude Code's text response.

Examples

Explain your own architecture (ALWAYS use self for this):

{"type": "claude-code", "target": "self", "prompt": "Explain how the OODA loop cognitive control works in this codebase. Cover the main phases, key components, and control flow.", "out": "$explanation"}

Verify a capability claim before making it:

{"type": "claude-code", "target": "self", "prompt": "What tools do I have for web search? What are their capabilities and limitations?", "out": "$capabilities"}

Debug your own behavior:

{"type": "claude-code", "target": "self", "prompt": "Look at infospace_executor.py — what happens when _parse_json_response encounters invalid JSON? Trace the repair logic.", "out": "$debug_info"}

Write a script in the sandbox:

{"type": "claude-code", "prompt": "Write a Python script that fetches RSS feeds from a list of URLs and extracts article titles. Save it as rss_reader.py", "out": "$cc_result"}

Reset a stale session:

{"type": "claude-code", "action": "reset", "target": "sandbox"}