Code Investigate

/skill:code-investigate — Cross-Repo Code Investigation

Purpose: Search code for a requested pattern or behavior across one or more repositories and produce a structured, versioned Markdown report. Unlike the Explore subagent (single-scope, no report), this skill produces a durable investigation artifact with structured matches and confidence scores.

When to Use

• You need a durable, versioned investigation report — not an ephemeral search
• Scope spans multiple repositories
• Findings will be cited in a ticket, PR, or postmortem

When NOT to Use

• Single-repo, throwaway search — use the Explore subagent or Grep directly
• You already know the file and just need to read it — use Read
• You are debugging an active bug — use /skill:triage

Soft Pre-flight Config Check

This skill has a config file for defaults, but can run without it if arguments are provided.

Check ~/.claude/code-investigate.config.md:

• If it exists, load report_root and default_repo_roots as defaults.
• If it does NOT exist, check whether the user provided scope and output args. If yes, continue. If no:

  "No default config found. Provide search scope as an argument, or run /skill:configure code-investigate to set defaults."

Arguments

• <description> — what to find (exact pattern, API name, regex, or natural-language behavior like "send email" or "reads files")
• --repo <path> — search a single repo (default: current repo)
• --repos <path1,path2,...> — search an explicit list of repos
• --folder <path> — search all repos in a directory
• --output <path> — report output directory (default from config: docs/code-investigations/)
• --include-code — include matching code lines in the report
• --include-summary — include a short behavior summary per match
• --confidence <0.50-1.00> — minimum confidence threshold (default: 0.70)

Workflow

Step 1: Clarify Scope

Determine scope mode:

• repo — one repository (default: current)
• repos — explicit list
• folder — directory containing repositories

If unclear from arguments, ask once.

If not specified and not clear from context, ask:

1. "Include matching code lines in the report?"
2. "Add a short summary of what the code does at each match?"

Step 2: Build Search Plan

Convert the request into a search plan:

• Exact anchors — symbols, class names, method names, string literals, config keys
• Framework cues — common APIs and idioms for the behavior
• Structure clues — filenames, folder names, namespaces, service names
• Behavioral cues — patterns that imply the behavior even without keyword matches

For multi-repo scope, dispatch parallel searches using the Agent tool with one agent per repo.

Step 3: Search

Context-aware reading strategy: With 1M token context available, calibrate read depth to repo size:

• Small repos (<200 files): read all source files directly before grepping — comprehensive reads surface semantic matches that keyword search misses.
• Medium repos (200–2000 files): full read of entry points, public surfaces, and files matching structure cues; grep for anchors elsewhere.
• Large repos (2000+ files): layered approach below; sample by signal category rather than exhaustive read.

For each repo in scope, search using a layered approach:

1. Grep/Glob for exact anchors
2. Read files matching framework/structure cues (full read for small/medium per above)
3. Scan for behavioral matches in promising files
4. Rate each match with a confidence score

Match Schema

Each match is a structured object:

Step 4: Write Report

Write a versioned Markdown report:

Default path: <report_root>/<YYYY-MM-DD>--<HHMMSS>--<scope>--<slug>.md

Report structure:

# Code Investigation Report

- **Date:** YYYY-MM-DD HH:MM
- **Scope mode:** repo | repos | folder
- **Scope:** <human-readable scope>
- **Request:** <search request summary>
- **Repos searched:** <count>
- **Matches found:** <count>
- **Confidence threshold:** <threshold>
- **Include code lines:** yes | no
- **Include summaries:** yes | no

## Matches by Repository

### <repo-name>

#### <project-name>

| File | Lines | Type | Confidence | Why |
|---|---|---|---|---|
| path/to/file.ts | 42-58 | semantic | 0.85 | Calls sendEmail() with SMTP config |

<optional code block>
<optional summary>

## Summary

- Total matches: <N>
- High confidence (>=0.90): <N>
- Medium confidence (0.70-0.89): <N>
- Repos with matches: <list>
- Repos with no matches: <list>

Update <report_root>/index.md and <report_root>/latest.md if they exist.

Step 5: Report

"Investigation report written to <path>. Found matches across repos."

CRITICAL: Do Not Guess

• Do NOT fabricate matches. Every match must reference real code.
• Do NOT inflate confidence scores. If unsure, rate lower.
• Do NOT modify any source code. This skill is read-only.
• Do NOT skip repos in scope without reporting them as "no matches."
• Do NOT assume repo structure. Detect it.