youdotcom-oss/risk-analysis-server/drive-sweeps

Drive risk sweeps from a chat client

The server exposes five tools. Work them in this order.

The tool set

The sweep protocol (two calls)

1. Start: call trigger_manual_sweep with profileId. It returns

immediately — the tool does NOT block: { task_id, status: "working", next: "poll with task_id every ~20s" }.

1. Poll: call trigger_manual_sweep again with task_id every ~20

seconds. While running: { task_id, status: "working", next: ... }. When done: { status: "completed", escalated, severity, reportId, threatProbability, usage, knowledgeHits } (or an error result with the failure reason). knowledgeHits counts licensed knowledge facts that reached the briefing — 0 is a valid result for news-shaped profiles. threatProbability is the raw triage noul behind the escalation decision; usage tallies the sweep's provider footprint (You.com search/contents calls, judgment tokens — Jev by default, or the sweep model's own JSON answers when RISK_JUDGE=qwen, which also tally a judgeMalformed parse-failure count).

1. Read the report even when not escalated. Below-threshold sweeps

still persist a low-severity clean-sweep report (signals reviewed, triage score) — a report isn't always for action, sometimes it documents inaction. Summarize it the same way via get_risk_report.

Typical duration is 2-3 minutes (agentic search loop + judgment gates + cloud-model synthesis; the RISK_JUDGE=qwen ablation arm runs longer, 7-12 minutes). Never assume a timeout means failure — the sweep keeps running server-side; keep polling.

Reading the report

• Call get_risk_report (optionally with report_id from the outcome) and

summarize the briefing: severity, key findings with source links, and recommended mitigations. Knowledge facts cite their licensed provider and an as-of date — surface those attributions when they appear.

• Briefings are GFM Markdown — cite the key findings with their source

links and list the mitigations; the report HTML/iframe era is gone.

Gotchas

• Profiles are tenant-scoped; ids are random UUIDs. After a fresh

conversation, list_risk_profiles is the only reliable way to get one.

• Sweeps are per-profile. Splitting a profile by location is a workaround

for slow runs, not a requirement — prefer one profile per risk domain and let the pipeline cap payloads itself.

• Failed sweeps return isError with the upstream error (e.g.

max_tokens_exceeded, rate limits). Report the error verbatim; do not retry silently in a loop.