Draft Ashby interview scorecards from Metaview AI notes

Build me an agent workflow that drafts Ashby interview scorecards from Metaview AI notes so I stop chasing interviewers for feedback.

Trigger: cron. Run once an hour during business hours (default 9am to 6pm local, Monday to Friday). On each run, look at Metaview interviews that were completed in the last hour.

Metaview is exposed as an MCP server, so tool names are not statically documented. The agent must first Initialize MCP Session, then List Available Tools to discover the current catalog, then Call a Tool for the specific action needed. For this workflow, use Metaview to (a) list recently completed interviews with their AI notes, structured competency scores, candidate name and email, role/job name, and interviewer name/email, and (b) fetch the full AI note content and any competency ratings for each interview. Log the exact tool names and argument schemas the agent picked so we can debug later.

For every Metaview interview returned, do the following in Ashby:

1. Find the matching application. Use List Applications filtered by the candidate's email (or name as a fallback) and the job title from Metaview. If there is exactly one active application, use it. If there are multiple or none, skip this interview and log why.

2. Identify the right feedback form. Use List Feedback Form Definitions to see what forms exist, then pick the one that matches the application's current interview stage. If unclear, prefer the form the interviewer has submitted most recently for that stage. Call Get Feedback Form Definition Info to load its full field schema (field types, allowed values for select fields, rating scales).

3. Map Metaview data to Ashby fields, conservatively. For each Ashby field, look for a corresponding Metaview competency, structured score, or clearly relevant section of the AI notes. Convert Metaview competency ratings into the Ashby field's scale (for example, a 4/5 Metaview competency maps to the equivalent Ashby rating option). For free-text fields, pull a concise summary from the matching AI note section. IMPORTANT: if there is no confident mapping for a field, leave it blank. Do not guess and do not invent opinions the interviewer did not express. It is better to draft five fields well than ten fields poorly.

4. Resolve the interviewer. Use List Users in Ashby to find the user whose email matches the Metaview interviewer. Submit Application Feedback should be attributed to that user so the draft appears under their name.

5. Submit Application Feedback as a draft (not a final submission). The interviewer will open Ashby, review, adjust, and submit the final scorecard themselves.

6. Create Candidate Note on the candidate's Ashby profile with a short two-line summary of the interview, plus a link back to the original Metaview transcript so the interviewer can verify anything the draft says.

Guardrails. Never submit final feedback. Never overwrite an existing scorecard that the interviewer has already submitted or drafted themselves. If a scorecard draft already exists for that application/interviewer, skip and log. Skip any ambiguous field mapping instead of guessing. Do not use Metaview transcripts for any candidate the interviewer has not opted into recording.

Reporting. At the end of each hourly run, log: number of Metaview interviews processed, number of Ashby scorecards drafted, number skipped, and the reason for each skip (no matching application, form ambiguous, existing scorecard, etc.).