claude-code-internals

Claude Code Internals Explorer

Procedures for analyzing Claude Code's standalone binary.

Quick Start

Note: All script paths are relative to this skill's directory (where SKILL.md resides).

Delegate all binary exploration to subagent:

Task tool (subagent_type: Explore)
Prompt: "Run scripts/find_installation.sh to get binary path.
        Then search for [keyword] using: strings $BINARY | grep -E '[pattern]' | head -50
        Return: version, matching lines with context (-B2 -A2)."

The subagent handles token-heavy strings output; main context receives only summarized findings.

Zero-Context Mode (Recommended)

Problem: Subagent output is recorded in session logs. Binary analysis can cause session file bloat → SIGTRAP crashes.

Solution: Write prompt to temp file, then execute via Headless CLI wrapper. Script internals are not logged; only final stdout is recorded.

Workflow

1. Write prompt to temp file (supports multiline)
2. Execute scripts/analyze-binary.sh <prompt_file> [allowed_tools]

Script Interface

The script injects BINARY_PATH automatically.

Example

# Use unique ID in filename for parallel independence
PROMPT_FILE="/tmp/internals_prompt_$.txt"

# 1. Write prompt to temp file
cat > "$PROMPT_FILE" << 'EOF'
Search the binary for beta headers.

Instructions:
1. Use: strings "$BINARY_PATH" | grep -E "anthropic-beta|20[0-9]{2}-[0-9]{2}" | head -30
2. List all matching headers with their apparent status
3. Note any patterns suggesting feature flags
EOF

# 2. Execute with tool restriction
scripts/analyze-binary.sh "$PROMPT_FILE" "Bash,Read,Glob,Grep"

# 3. Cleanup (optional - /tmp clears on reboot)
rm -f "$PROMPT_FILE"

Dynamic Prompting

The agent constructs prompts dynamically based on investigation context.

File naming convention: /tmp/internals_prompt_<unique_id>.txt

• Use $ (PID) or UUID for parallel execution independence
• Multiple concurrent investigations won't conflict

PROMPT_FILE="/tmp/internals_prompt_$(date +%s%N).txt"

cat > "$PROMPT_FILE" << 'EOF'
Investigate context management settings.
1. Search for contextWindow, warningThreshold, errorThreshold
2. Find related numeric constants (5-6 digit numbers)
3. Return: setting name, value, surrounding context
EOF

scripts/analyze-binary.sh "$PROMPT_FILE"

Tool Inheritance

Agent's tool restrictions are passed to Headless CLI via --allowedTools:

Why this works: Bash subagent executes the script as an external process. Only the script's stdout appears in session logs, not the intermediate 350K+ lines from strings.

---

⚠️ Warning (Legacy Mode): If using direct subagent delegation below, subagent output is recorded in session logs. Large outputs (>1MB) can cause session file bloat, leading to SIGTRAP crashes on new Claude sessions. Always use | head -N to limit results.

Workflow

1. Delegate Binary Exploration

Call Task tool with:

• subagent_type: Explore
• prompt: Specify search target and expected output format

Example delegations:

# Feature investigation
Find Claude Code binary using scripts/find_installation.sh.
Search for "anthropic-beta" using: strings $BINARY | grep -E "anthropic-beta|20[0-9]{2}-[0-9]{2}" | head -30
Return: version, matching lines (max 30).

# Settings discovery
Run find_installation.sh, then search for setting patterns:
strings $BINARY | grep -E "autoCompact|permission|default.*:" | head -50
Return: setting names and apparent default values (max 50 lines).

The subagent executes commands and returns summarized findings.

2. Subagent Search Commands (Reference)

Note: These commands are for subagent execution, not main context.

# Get binary path using find_installation.sh
source scripts/find_installation.sh
# Sets: BINARY_PATH variable

# Search with context (ALWAYS limit output!)
strings "$BINARY_PATH" | grep -B2 -A2 "pattern" | head -50

# Cache for multiple searches (within subagent session)
strings "$BINARY_PATH" > /tmp/claude-strings.txt
grep "pattern1" /tmp/claude-strings.txt | head -30
grep "pattern2" /tmp/claude-strings.txt | head -30
# Remember to cleanup: rm /tmp/claude-strings.txt

3. Investigation Types

4. Analyze & Document

1. Use grep -B2 -A2 for context
2. Compare with references/known-features.md
3. Update known-features.md with new discoveries

Common Investigations

Resources

• scripts/find_installation.sh: Locate Claude Code binary
• references/search-patterns.md: Comprehensive search patterns by category
• references/known-features.md: Baseline of known features for comparison

Resource Map

Tips

1. Cache strings: strings $BINARY_PATH > /tmp/claude-strings.txt once, grep multiple times
2. Start narrow: Specific terms before broad exploration
3. Chain searches: Find one string, search nearby for related code
4. Clean up: Remove /tmp/claude-strings.txt after investigation