refactor_tool(mode='dead_code'): response items missing documented relative_path / file_path / language fields

[johndowbs]

Summary

refactor_tool(mode="dead_code") returns items with only 5 fields (name, qualified_name, kind, file, line) but the tool's docstring / schema hint advertises additional fields including relative_path, file_path, and language. The documented fields are missing entirely on every entry.

Version

code-review-graph==2.3.2 (also reproducible on main as of 4b6293f)

Reproduction

refactor_tool(mode="dead_code", kind="Function", repo_root="/path/to/any/repo")

Each entry in dead_code[]:

{
  "name": "...",
  "qualified_name": "...",
  "kind": "Function",
  "file": "/abs/path/to/file.py",
  "line": 123
}

relative_path, file_path, and language are missing on every entry (not even null).

Root cause

In code_review_graph/refactor.py, find_dead_code() builds the per-entry dict at lines ~561-567:

dead.append({
    "name": _sanitize_name(node.name),
    "qualified_name": _sanitize_name(node.qualified_name),
    "kind": node.kind,
    "file": node.file_path,
    "line": node.line_start,
})

The wrapper in tools/refactor_tools.py::dead_code branch returns dead as-is. Nothing enriches it with relative_path or language, even though both are derivable: the calling tool has root from _get_store(repo_root), and node.language is already on the GraphNode model.

Suggested fix

Option A (preferred): enrich the entries in tools/refactor_tools.py after dead = find_dead_code(...):

from pathlib import Path
EXT_LANG = {".py":"python", ".ts":"typescript", ".tsx":"typescript",
            ".js":"javascript", ".jsx":"javascript", ".go":"go", ".rs":"rust",
            ".java":"java", ".rb":"ruby", ".sh":"bash", ".ex":"elixir",
            ".exs":"elixir", ".m":"objective-c", ".mm":"objective-c",
            ".c":"c", ".cpp":"cpp", ".cc":"cpp", ".h":"c", ".hpp":"cpp",
            ".cs":"csharp", ".php":"php", ".swift":"swift",
            ".kt":"kotlin", ".scala":"scala"}
for entry in dead:
    abs_path = entry["file"]
    entry["file_path"] = abs_path
    try:
        entry["relative_path"] = str(Path(abs_path).resolve().relative_to(root))
    except ValueError:
        entry["relative_path"] = abs_path
    entry["language"] = EXT_LANG.get(Path(abs_path).suffix.lower(), "unknown")

Even better: pull language directly off the GraphNode in find_dead_code (it's already on the model).

Option B: update the schema/hint string to match the 5-field reality. Less useful but trivial.

Companion request (optional)

find_dead_code returned 817 entries on a 4,525-node graph in our repo. A large fraction were template files (.claude/skills/*/modify/src/...) and one-shot training/backtest snapshots (data/files/*.py) which are intentionally never executed at runtime.

find_large_functions_tool already supports file_path_pattern. Adding a symmetric exclude_pattern to refactor_tool(mode="dead_code") would let consumers filter known-snapshot directories without post-processing 800+ false positives.

Acceptance criteria

• dead_code[] items include relative_path, file_path, and language fields (Option A), or
• Schema/docstring updated to advertise only the 5 fields actually returned (Option B)
• Optional: exclude_pattern parameter symmetric to file_pattern

[itxaiohanglover]

Hi, I'd like to work on this issue.

I've confirmed the bug — find_dead_code() in refactor.py manually builds the dict with only 5 fields, while GraphNode already has language readily available, and find_large_functions in query.py uses node_to_dict() to return a much richer response.

My plan:

1. Add language and file_path fields to the dead_code entries, sourced directly from GraphNode (no hardcoded extension map needed).
2. Add relative_path by computing it from node.file_path and the repo root — requires passing root into find_dead_code().
3. Keep the file key for backward compatibility alongside the new file_path key.

I'll keep the scope limited to the field enrichment (Option A) and leave exclude_pattern as a separate follow-up.