researching-code

Researching Code

Execute code research against the ChunkHound index and return synthesized findings. The skill picks the depth, sequences the queries, and returns the result.

Workflow

digraph researching_code {
    "Skill invoked" [shape=doublecircle];
    "Step 1: Detect depth and primitive directives" [shape=box];
    "Depth?" [shape=diamond];
    "Surface plan" [shape=box];
    "Broad plan" [shape=box];
    "Deep plan" [shape=box];
    "Plan uses ChunkHound?" [shape=diamond];
    "Step 2: Pre-flight (daemon_status)" [shape=box];
    "All hard gates pass?" [shape=diamond];
    "STOP — return structured failure" [shape=octagon, style=filled, fillcolor=red, fontcolor=white];
    "Step 3: Execute plan" [shape=box];
    "Step 4: Synthesize findings" [shape=box];
    "Return result" [shape=doublecircle];

    "Skill invoked" -> "Step 1: Detect depth and primitive directives";
    "Step 1: Detect depth and primitive directives" -> "Depth?";
    "Depth?" -> "Surface plan" [label="surface"];
    "Depth?" -> "Broad plan" [label="broad"];
    "Depth?" -> "Deep plan" [label="deep"];
    "Surface plan" -> "Plan uses ChunkHound?";
    "Broad plan" -> "Plan uses ChunkHound?";
    "Deep plan" -> "Plan uses ChunkHound?";
    "Plan uses ChunkHound?" -> "Step 2: Pre-flight (daemon_status)" [label="yes"];
    "Plan uses ChunkHound?" -> "Step 3: Execute plan" [label="no — searches no code"];
    "Step 2: Pre-flight (daemon_status)" -> "All hard gates pass?";
    "All hard gates pass?" -> "STOP — return structured failure" [label="no"];
    "All hard gates pass?" -> "Step 3: Execute plan" [label="yes"];
    "Step 3: Execute plan" -> "Step 4: Synthesize findings";
    "Step 4: Synthesize findings" -> "Return result";
}

Step 1: Detect depth and primitive directives

Pick surface, broad, or deep in this priority:

1. Explicit directive from the caller — phrases like "quick check", "surface", "deep dive", "full trace", "just locate X". Use it verbatim.
2. Question shape when no directive — surface for symbol-location questions ("Where is X defined?", "Is X used?", "Show me Y"); broad for subsystem questions ("How does Z work?", "What handles A?"); deep for multi-component traces, impact analyses, and full subsystem audits ("Trace data flow from A through B to C", "Audit all callers", "full impact map for refactoring X").
3. Default: broad. Surface drops context; deep wastes time.