CodeRLM — Structural Codebase Exploration

You have access to a tree-sitter-backed index server. Use it instead of guessing with grep.

FIRST: Choose Your Mode

Is the question "how does X work", "trace X", or "follow the execution path of X"?

→ Delegate immediately to a subcall agent. Do not do any CLI searches first.

Agent(coderlm-subcall, "Trace how <X> works. Find the entrypoint, follow callers, and show the full execution path with code extracts.")

The subcall agent has a structured workflow for tracing and will return findings with code evidence. Using it keeps the full exploration out of your context window and produces more complete results than piecemeal CLI calls.

Is the question a targeted lookup ("where is X defined", "what calls X", "show me X")?

→ Use CLI commands directly (see below).

When to just Read a file directly. After 1-2 search or structure calls, if the relevant code is clearly concentrated in 1-3 files, Read those files directly instead of reconstructing them with granular CLI calls. Read is cheaper than 10+ impl/symbols/peek calls for a focused file. Only keep using CLI lookups when tracing across many files (5+) or when you need symbol-level precision that Read can't give you.

NEVER fall back to raw bash grep/find/cat. If a coderlm CLI command fails, use the coderlm fallback rules — not shell tools.

Setup

CLI=".claude/coderlm_state/coderlm_cli.py"

Session is auto-created by the plugin. If not, run python3 $CLI init.

Quick Lookups: One-Off Commands

For a single targeted lookup:

python3 $CLI search MyFunction
python3 $CLI impl MyFunction --file path/to/file.py
python3 $CLI callers MyFunction --file path/to/file.py

Multiple Lookups: Batch Mode

For 2-5 targeted lookups, use batch to minimize tool calls:

python3 $CLI batch --commands "
search MyFunction
impl MyFunction --file path/to/file.py
callers MyFunction --file path/to/file.py
"

Multiple Lookups: Python Exec Mode

For tracing a full execution path in one shot — search, then follow callers up the chain:

python3 $CLI exec --code "
# Find entrypoint, read its implementation, then walk callers up the chain
hits = search('serialize_response')
if hits.get('symbols'):
    sym = hits['symbols'][0]
    impl_(sym['name'], sym['file'])                    # read the function
    c = callers(sym['name'], sym['file'])               # find who calls it
    if c.get('callers'):
        caller = c['callers'][0]
        impl_(caller['name'], caller['file'])           # read the caller
        callers(caller['name'], caller['file'])         # one more hop up
"

Use exec mode when you need to: follow a call chain across 3+ hops, or make later queries depend on earlier results. One exec call replaces many individual Bash calls and keeps the full trace in a single result.

Available helpers in exec mode: search(), impl_(), callers(), tests(), grep(), symbols(), peek_file(), structure(), variables_list().

Commands

| Command | Purpose | |---------|---------| | search QUERY [--file FILE] | Find symbols by name | | impl SYMBOL --file FILE | Get full source code | | callers SYMBOL --file FILE | Find call sites | | tests SYMBOL --file FILE | Find tests referencing symbol | | grep PATTERN [--scope code] [--file FILE] | Regex search (--scope code skips comments) | | peek FILE --start N --end N | Read a line range | | symbols [--file FILE] [--kind KIND] | List all symbols | | structure [--depth N] | Project file tree |

Fallback Rules (coderlm only — not raw bash)

search returns 0 → use grep for the name
impl fails (404) → use peek with the file path
callers returns nothing → grep for the symbol name
Don't retry failures — pivot to an alternative

Output Requirements

Code extracts — Include actual source from impl or peek as fenced code blocks
File:line references — Cite specific locations for every claim
Chain of evidence — Show each hop when tracing execution paths

Inputs

This skill reads $ARGUMENTS. Accepted patterns:

query=<question> (required): what to find or understand
cwd=<path> (optional): project directory
port=<N> (optional): server port (default: 3000)