Explore Skill

Deep codebase exploration and Q&A using local LLM for file discovery and Sonnet for detailed analysis.

Step 1: Classify the Query (Local LLM, Zero Claude Tokens)

Offload exploration-type classification to the local LLM:

echo "[user's question]" | bash ~/.claude/hooks/llm-commands.sh classify-query

Expected output: one of explanation, usage-search, architecture, location, data-flow. If Ollama is unavailable, fall back to matching the question against these patterns directly:

Step 2: Local LLM File Discovery (Zero Claude Tokens)

Use local LLM to extract search keywords from the user's question:

echo "[user's question]" | bash ~/.claude/hooks/llm-commands.sh generate-slug

Then use shell tools for initial file discovery (zero Claude tokens):

# Search for relevant files using keywords (use appropriate file types for the project)
grep -rl "[keyword]" .
# Or use glob patterns for structural exploration
find src/ -type f | head -50
# For architecture queries, get an overview of shared utilities
bash ~/.claude/hooks/scan-shared-utils.sh

For larger codebases, use local LLM to build a relevance map:

# Pipe discovered file contents for relationship analysis
bash ~/.claude/hooks/pre-review.sh code

If Ollama is unavailable, proceed to Step 3 with shell-only discovery results.

Step 3: Sonnet Deep Analysis (Sub-agent)

Launch a Sonnet sub-agent for detailed code tracing:

You are a senior engineer performing codebase exploration.

Query type: [explanation/usage-search/architecture/location/data-flow]
User's question: [original question]

Initial file discovery:
[File list and/or local LLM analysis]

Task:
1. Read the most relevant files from the discovery results
2. Based on query type:
   - explanation: Trace the implementation step by step, explain each layer
   - usage-search: Find all callers, build a dependency graph. **Derive the set from the defining primitive, not from the discovery keywords.** For "find every X" / "all callers of Y" / "everywhere that does Z", do not answer from the keyword-matched or "obvious" subset of files surfaced in Step 2 — identify the primitive that *defines* membership (the function/decorator/route-export/table/call whose presence makes a site a member) and `grep -rlE '<primitive>' <code roots>` across the whole tree to get the complete set. Then include *indirect* members the symbol grep alone misses: operations reached via a parent cascade, raw queries that bypass the ORM/guard layer, and calls routed through an aliased or wrapped caller. A keyword/path-scoped enumeration silently drops members under non-obvious paths — state the primitive you grepped so the completeness of the set is auditable. **Any count in the answer ships the command that reproduces it (R29).** "N call sites", "only M places do X", a coverage or occurrence figure — the number arrives in a frame ("derived from the code") that the reader, and every later round that restates it, treats as verified. Give the exact command beside the figure or omit the figure; a count summed over a subset of the tree reconciles with nothing and is the recurring shape.
   - architecture: Map the directory structure, identify key patterns
   - location: Pinpoint the exact file, line, and context
   - data-flow: Trace data through each transformation layer
3. Follow references to related files (imports, function calls, type definitions)
4. Build a clear, structured answer with file:line references

Output format:
- Start with a one-paragraph summary
- Then provide detailed explanation with code references
- End with a relationship diagram (ASCII) if applicable

If sub-agents are unavailable, perform the analysis directly.

Step 4: Synthesize and Present

Review Sonnet's analysis for completeness and accuracy:

• Verify referenced files exist (automated): Pipe the sub-agent output through verify-references.sh to catch stale or missing file:line references in bulk before manual review:

  echo "[sub-agent output]" | bash ~/.claude/hooks/verify-references.sh

  Any MISSING or OUT-OF-RANGE line from the summary is a stale reference — remove or correct it.

• Verify function/type existence: For each named function, type, or constant still in the output after the automated check, grep to confirm it exists at the stated location. Sub-agents may hallucinate names or locations even when the file:line resolves.
• Fill in any gaps the sub-agent may have missed
• Resolve any ambiguities

Present the answer to the user with:

• Clear structure (summary → details → diagram)
• Clickable file:line references (verified to exist)
• Follow-up suggestions for deeper exploration