Full sweep — orchestrator

You are the orchestrator. This exists because a one-off "does everything look okay?" pass
reliably misses things, and a structured fan-out reliably does not. The first run of this
sweep (2026-08-11) found a live crash that would have hit any teammate whose export had a
blank county cell, and a live confidentiality leak in the public repo. Neither was visible
to casual inspection; both were found by a specialist told exactly what to attack.

Before you start

Run the deterministic checks yourself. They are fast, and a failure here means stop and fix
before spending agent time:

python system/scripts/check_screener.py "<a real CoStar export>"
python system/scripts/check_portfolio_comparison.py
python system/scripts/check_mcp_health.py

Then find a real export to hand the agents — config.COSTAR_DROP_DIR first, then
system/data/drop/. If there is genuinely no export on the machine, say so; several checks
below are meaningless without one and you should not fabricate a file.

Dispatch these five, in parallel, in ONE message

They own disjoint areas and must not be serialised. Brief each with what changed since the
last sweep — an agent told "check everything" returns platitudes; one told "this specific
thing changed yesterday, attack it" returns findings.

Agent	What it owns
vaulter-screening-checker	Was the export read correctly and ranked fairly? Prove ranking is unaffected by anything informational (comparisons, cautions) by running with and without and diffing scores row-for-row.
vaulter-report-checker	Is the generated HTML report factually right, and genuinely readable by a partner who was not in the room?
vaulter-leak-guard	Real firm data in the public repo — tracked files and commit messages and git history. .gitignore coverage. What the pattern list would still miss.
vaulter-onedrive-auditor	Shared-folder accumulation, misfiled items, stray temp files, and whether real-data files match their backups.
vaulter-setup-tester	Could a brand-new teammate install from the current zip and connect? Dependency pins especially.

Constraints every agent must be given, without exception:

• Do not write into the OneDrive shared folder (Vaulter AI Shared) — it syncs to the whole
  team. Read-only; propose, never delete.
• Do not modify the live install (C:\Users\<user>\Vaulter AI) or the real Claude Desktop
  config.
• Fix real bugs in the dev repo only, and report exactly what changed.

When they report back

Verify before you relay. Agents are confident narrators and are sometimes wrong — on the
first run one reported a backup file "not found" and was right, while another's framing of a
known-and-deferred item as a new finding was not. Check any claim you are about to act on or
repeat, especially a claimed failure. This costs one command and has already prevented one
false alarm.

Then:

1. Fix what's cheap and clear-cut — a missing pattern, a crash with a known shape, a
   cosmetic string. Add a regression check for anything that was a real defect, so the suite
   grows every sweep. It went 71 → 77 on the first run.
2. Surface what needs a human — anything destructive (rewriting published history),
   anything that trades confidentiality against usability, anything about firm history only a
   partner can settle. Present the trade, recommend, and stop.
3. Ship — full regression, then release.py + --promote so existing installs update
   themselves, then build_handoff.py so the zip for new teammates is current. A fix that is
   only in the dev repo has reached nobody.

What "passing" means

Not "no agent complained." It means: every deterministic suite green, every agent's claimed
failure either fixed or explicitly accepted with a reason, and the published version actually
carrying the fixes — verified by reading system/VERSION on disk, not by trusting a tool's
own success message.