Code Search

Find code using natural language. Automatically routes to the right search backend.

Steps

1. Parse the user's query:

• If the user provided a query argument, use it directly
• Otherwise, ask what they're looking for

1. Call find_code with the query:

• Let focus default to auto
• If the user mentioned a specific directory or package, pass it in paths
• If the user wants test files included, set include_tests: true
• If the user is debugging search quality, set debug: true

1. Interpret the results based on confidence:

High confidence (strong matches from multiple sources):

• Show the top 3 candidates with file path, symbol name, and why it matched
• For the #1 candidate, show the snippet and enclosing function/component
• If the result has graph evidence, mention what was promoted/demoted
• Automatically call gather_context on the #1 candidate's symbol to provide ready-to-use context

Medium confidence (reasonable matches, single source):

• Show top 3 candidates with evidence
• Suggest the user refine their query or try a more specific term
• Offer to run find_pattern if the query has a structural shape
• Note any warnings (scope capped, partial results)

Low confidence (weak or no matches):

• Explain what was searched and why it didn't match well
• Check the IR: suggest using the exact function name if only NL tokens were used
• Offer to try find_pattern with an AST pattern instead
• If scope was capped, mention it and suggest narrowing with paths

1. Offer follow-up actions:

• "Want me to read the top result?" → Read the file
• "Want more context?" → Call gather_context on the top candidate's symbol
• "What calls this?" → Call call_hierarchy on the symbol
• "What breaks if I change this?" → Call impact_trace on the symbol
• "Is the API safe?" → Call api_guard on the file

1. If stats.partialResult is true or warnings is non-empty, mention it clearly so the user knows the search was incomplete.

Examples