What problem does it solve? When a search or import skill produces a bad result (wrong person, bad LinkedIn URL, bad ranking, a crash), engineers need enough session context to reproduce the issue, but manually assembling that context is tedious and risks leaking private message content. This Skill mines the current session for repro-relevant identifiers, enforces a strict privacy contract, and files one structured feedback row to Powerset. ## Core Features & Use Cases - Session context mining: Collects the user's verbatim query, the failing command line, JSON payload statuses, run artifacts (decision.json, task-state files, stage manifests), and affected person identifiers from the current conversation without re-running failing commands. - Privacy contract enforcement: Sends identifiers only (names, LinkedIn URLs, UUIDs, counts, statuses) and never message bodies, dossier prose, or machine free-text reasons; person identifiers ship only after explicit user consent via a dry-run preview. - Structured submission: Posts one validated row to the POST /v2/feedback endpoint through the send_feedback primitive, with optional gzip+base64 inline artifact attachments and handling for submitted, needs_auth, and failed outcomes. - Use Case: After a $search run returns the wrong LinkedIn profile, say "report this" and the agent composes a feedback row with the query, decision.json, and the wrong profile's identifiers, previews it, asks about redaction, and submits it. ## Quick Start Report the bad search result from this session to Powerset, previewing the feedback row before sending it.
