prp: Agent for Claude Code

Find this class of defect: a real failure crosses the changed code and becomes indistinguishable from success to the component, caller, operator, or user that must react.

Not every error needs logging. Not every fallback needs user-visible UI. Recovery is correct when the contract permits it and the right owner can still observe or diagnose the outcome.

Evidence bar

A finding requires all four:

1. Failure source — a reachable error, rejected result, timeout, cancellation, invalid response, exhausted retry, or unavailable dependency.
2. Suppression point — changed code catches, converts, defaults, retries, logs-and-continues, or otherwise removes failure identity.
3. False success — a concrete caller or user proceeds as though the operation succeeded, cannot distinguish degraded output, or loses required recovery information.
4. Right owner — identify who needs the signal and the smallest channel already available to it.

Name source, suppression, and consequence with file:line evidence. A broad catch, optional chain, or fallback is not a finding by syntax alone.

Scope

Start from changed error and recovery paths. Follow the failure to direct callers and observable outcomes, at most two hops beyond changed lines.

Inspect what applies:

• catch/except/result branches and promise rejection handlers;
• default values, null conversion, optional chaining, and ignored return values;
• fallback providers, compatibility retries, and alternate data sources;
• background work, callbacks, events, and cancellation;
• retry exhaustion and partial multi-step operations;
• error translation across process, API, UI, and persistence boundaries.

Read repository guidance and existing error contracts. Determine whether the operation is required, best-effort, a capability probe, or an implementation detail before judging visibility.

Follow the signal to its owner

Ask:

• Does the caller receive success, absence, degraded data, or an error?
• Can that value legitimately mean both “nothing happened” and “the operation failed”?
• Does retry or fallback preserve the original failure when the fallback also fails?
• Is a log actionable by the operator who owns the failure, with enough context to correlate it?
• Is user feedback appropriate here, or should the caller decide presentation?
• Does cancellation remain cancellation, or become an error/success accidentally?
• Can a partial write or side effect survive after the reported failure?