gospel

Explain what $ARGUMENTS actually does by reading the real code paths, not by guessing, and report it back at the right altitude: the gist first, then the detail that matters.

1. Identify what you were asked about

$ARGUMENTS is one of: a file path, a symbol or function name, a concept, or a pasted error or stack trace. Figure out which before doing anything else:

• Path: read it directly.
• Symbol or function: Grep for its definition (not just call sites) and locate it.
• Concept: find where it lives in this repo via Grep/Glob, then read.
• Error or stack trace: read top-to-bottom, find the originating frame, and open the file and line it points at.

If the target is genuinely ambiguous, state your best interpretation and proceed.

2. Trace how it actually works

Follow the real code, not what the names imply. Read the definition, then follow what it calls, what calls it, and the data it touches, far enough to explain the behavior, no further. For a large or cross-cutting flow, dispatch the Explore agent to map the path rather than reading every file yourself. Note the actual conditionals, error handling, and edge cases; do not paper over branches you did not read.

3. For library, framework, or API questions, check current docs

If the question is about a third-party library, framework, SDK, CLI, or cloud service, your training data may be stale. Use the bundled documentation tools instead of relying on memory: Context7 for general libraries and frameworks, Microsoft Learn for Microsoft/Azure topics. Ground the explanation in what the docs say for the version in use.

4. For an error, find the root cause

Do not stop at the symptom. Identify the originating frame, explain why it fires (the specific state or input that triggers it), and distinguish the root cause from where it surfaced. Then give the concrete fix: the exact change, at the exact path:line.

5. Report at the right altitude

Lead with a one- or two-sentence plain-English summary, then the detail:

• What it is and does: the summary, first.
• How it works: the real flow, step by step, with clickable path:line references at each point that matters.
• Edge cases and gotchas: branches, failure modes, or surprises you saw in the code.
• The fix (errors only): root cause plus the concrete change to make.

Explain what the code does, not what it is probably supposed to do. If the code looks buggy or contradicts its own naming, say so. If you could not trace part of it, say that plainly rather than filling the gap with a guess.