nsheaps/binary-inspection

Binary Inspection Skill

Use the binary-inspector agent from the deep-research plugin to analyze binary files without modifying them.

When to Use

• Investigating a compiled program's internals (embedded strings, URLs, config keys)
• Understanding which shared libraries an executable depends on
• Finding function names and exported symbols in a library
• Examining the structure of an ELF binary (sections, segments, dynamic entries)
• Extracting readable content from any non-text file

How to Invoke

Spawn the binary-inspector agent with a clear goal and output path:

Agent(binary-inspector, "Inspect <path/to/binary>. Goal: <what you're looking for>. Save findings to docs/research/<topic>-binary-inspection.md")

Always specify:

1. The full path to the binary
2. What you're trying to find or understand
3. Where to save the findings

Common Tasks

Find hardcoded strings or config values

Agent(binary-inspector, "Inspect /usr/local/bin/claude. Goal: find any hardcoded channel names, URL allowlists, or API endpoint strings. Save findings to docs/research/claude-binary-strings.md")

Check linked libraries

Agent(binary-inspector, "Inspect /usr/lib/libssl.so. Goal: show all shared library dependencies and exported symbols. Save findings to docs/research/libssl-dependencies.md")

Identify file type and architecture

Agent(binary-inspector, "Inspect /path/to/unknown-binary. Goal: identify file type, architecture, and whether it is stripped. Save findings to docs/research/unknown-binary-identity.md")

Find exported function names

Agent(binary-inspector, "Inspect /usr/local/lib/libfoo.so. Goal: list all exported function names, especially those related to authentication. Save findings to docs/research/libfoo-symbols.md")

Example: Inspecting Claude Code for Channel Logic

One concrete use case is inspecting the Claude Code binary (/usr/local/bin/claude) to understand behavior that is not documented — for example, finding what channels or URLs are baked into the binary at build time.

Agent(binary-inspector, "Inspect /usr/local/bin/claude. Goal: find any strings related to channel names, API endpoints, authentication URLs, or allowlists. Pay special attention to strings containing 'api.anthropic.com', 'claude.ai', or channel identifiers. Save full findings to docs/research/claude-binary-channel-analysis.md and raw strings output to .claude/tmp/claude-binary-strings-raw.txt")

The agent will:

1. Run file to confirm it is a valid executable
2. Extract strings with strings -n 8
3. Filter for keywords relevant to channels and URLs
4. Check readelf -d for linked libraries
5. Attempt nm for symbol names (may be empty if stripped)
6. Save structured findings to the specified output path

Output

The binary-inspector saves a structured Markdown report to the path you specify. The report includes:

• File identity (type, architecture, linkage, stripped status)
• Strings of interest (filtered and annotated)
• Symbol analysis (if available)
• Library dependencies
• Key findings summary
• Gaps and limitations

Large raw outputs (full strings list, nm output) are saved as companion files in .claude/tmp/ to keep the main report readable.

Limitations

The binary-inspector performs static analysis only:

• Cannot trace runtime behavior (use strace, ltrace, or a debugger for that)
• Cannot recover function names from stripped binaries (strings and hexdump are the fallback)
• Cannot analyze obfuscated or packed binaries without unpacking first
• Requires standard Unix tools (file, strings, xxd, readelf, nm, ldd) — note if any are missing

After Inspection

Per the research-output rules, commit inspection findings to docs/research/ on the main branch. Binary inspection results have lasting value for understanding tool internals and should not be left in .claude/tmp/.