diagnose

Start from the file, command, or URL named in the reproduce notes, then walk from the symptom to the source: follow stack frames for exceptions, trace return values back to handler boundaries, or identify the rendering component for wrong HTML. Read the candidate code in full before pinning a location.

Read each frame in order starting from the deepest application frame, skipping framework internals, and confirm the call sequence matches what the reproduce stage actually executed. The diagnosis should name a single line or tightly scoped block, not just a file.

Recurring culprits include missing locale filters on content queries, unsafe SQL identifier interpolation, pagination cursor off-by-ones, missing awaits on ignored promises, non-null assertions hiding undefined handling, missing permission checks, and Lingui t called at module scope.

No. The skill is strictly read-only: no edits, no test runs, no commits, and no writes to GitHub. It only reads code and returns a root cause, confidence rating, and hypothesis notes for a separate fix stage to act on.

Confidence is low when multiple plausible causes exist that cannot be ruled out without instrumentation, or when the right area is identified but no specific bug is visible. Low and medium diagnoses include hypothesis notes describing alternative causes and what would distinguish them.