Sync applied

Pull Ben's applied-jobs Google Sheet, normalize its messy rows, and write them into the triage dedup cache (data/corpus/applied.json) so the ranker never re-surfaces a job he has already applied to. Full reference: docs/operating/triage-applied-sync.md.

The sheet is free-form and inconsistent — company is often blank, the title lands in whichever column Ben pasted into (job-id OR description), and identity is sometimes only inferable from the link domain. That mess is why this step uses your judgment to normalize, not a column map.

1. Read the sheet

Load the Drive tool once: ToolSearch select:mcp__claude_ai_Google_Drive__read_file_content. Read profile/profile.yaml → applied_sheet for the id, then read_file_content(fileId: <that id>). (The id is never written down here — it is one user's Sheet, and a skill that ships must not carry it. If the read fails because it's a fresh/headless session with no Google connector, tell Ben and ask him to paste a CSV export instead — then normalize that.)

2. Normalize every non-empty row → JSON

For each row that has any of company / title / link, produce ONE object. Reconcile the mess:

• title — take the real role title, wherever it sits (job-id col like "Sr. Software Engineer, Studio", or the description col like "Software Engineer III"). Strip req numbers / locations mashed into it (e.g. Python developer (BBBH1688426) Georgia, USA → title "Python Developer", city "Georgia").
• company — use the hiring company col; if blank, infer from the link domain (careers-gotyto.icims.com → Tyto; recruiting.ultipro.com/HEA1015 → All One Health) or leave blank.
• city — only if clearly present (e.g. "Broad Run, VA", "Dayton"); else "".
• url — the raw link as-is (the tool re-normalizes LinkedIn/Indeed tracking cruft; don't hand-edit).
• confidence — high when company+title are clear (clean or confidently inferred); medium when one was inferred with some doubt; low when the row is genuinely ambiguous (a bare HN/company link, no real title). Only high/medium auto-block; low is surfaced for Ben, not silently hidden — so when in doubt, use low. A false dedup (hiding a job he never applied to) is worse than one resurfacing.
• note — one short phrase: how you inferred it, or why it's low.
• row — the sheet's No.

Write the array to a scratch file, e.g. <scratchpad>/applied-rows.json:

[{"row":29,"apply_date":"6/30","company":"NationMind LLC","title":"AI Engineer Developer","city":"","url":"","confidence":"high","note":"clean company+title"}]

3. Persist into the cache

Run from the repo root:

.venv/bin/python -m triage --sync-applied <scratchpad>/applied-rows.json

It computes the canonical keys (same composite_id the ranker uses, plus a normalized-URL key), REPLACES applied.json (the sheet is the source of truth), and prints how many were auto-blocked vs held for review.

4. Report to Ben

Give a tight summary: N records synced, N auto-blocked, and list the "held for review" rows (the low-confidence ones) by row number + best-guess so he can eyeball or clean them in the sheet. Mention that keeping the sheet's hiring company + title columns clean on new rows makes future syncs bulletproof.

Rules

• Read-only on the sheet — never write back to Google.
• The sheet fully replaces applied.json each run; don't try to merge by hand.
• The sync runs ONE WAY: the Sheet is Ben's own dashboard and nothing here ever writes to it.