Keysat
aec2b7775b
Fixes from the 2026-06-12 full-eval (P0 + two P1s); code-only, no schema change. Without these the "private CRM" premise was breachable on the LAN: - P0: the /assets/ route joined the request path onto FRONTEND_DIR without normalizing '..' (get_path/urlparse pass it through), so an unauthenticated GET /assets/../../data/crm.db read any file the process could — the LP DB, the JWT signing secret (-> admin-token forgery), the Gmail key. Add a realpath containment check that 404s anything resolving outside FRONTEND_ROOT. - P1: the LP-outreach drafter built its redaction Boundary with no ner_fn, so unknown people/firms in raw email bodies reached Claude in the clear. Pass the local-Qwen NER backstop (ner_fn=_ner_local), matching architect_grounding; fails closed via the existing scrub_unavailable path if the local model is down. - P1: get-by-id handlers leaked soft-deleted records by direct ID. Add deleted_at IS NULL to every get-by-id path — contacts, organizations, opportunities, lp_profiles — and to the nested related-data sub-selects in the contact/opportunity detail payloads, matching the list-handler convention. Bumps the package to v0.1.0:74 (utils.ts + versions/v0.1.0.74.ts + graph). Full report in EVALUATION.md; remaining P2/P3 triaged in AGENTS.md Current state.
29 lines
1.7 KiB
Markdown
29 lines
1.7 KiB
Markdown
---
|
|
paths:
|
|
- backend/redaction/**
|
|
- backend/mcp/**
|
|
---
|
|
|
|
# Redaction & the Claude privacy boundary
|
|
|
|
Read this before editing anything that sends data to a Claude model — the redaction layer or any MCP agent/tool path.
|
|
|
|
## The boundary
|
|
|
|
- `backend/redaction/` (`scrub.py` + `client.py`) is the **scrub → Claude → re-hydrate** boundary: `Boundary`, `SCRUB_BACKEND=local|gateway`, **fail-closed**.
|
|
- `SCRUB_BACKEND=gateway` routes scrubbing through Spark Control (caller-supplied dict). Local backend scrubs in-process. If scrubbing can't run, the call fails closed — it does not pass raw text through.
|
|
|
|
## Hard rules
|
|
|
|
- **Keep real LP data out of Claude.** Develop only on code/schema/synthetic-or-locally-redacted data. Route any real record substance through `backend/redaction` before it reaches a Claude model.
|
|
- **Never bulk-export the LP list** to any third party. Send only minimal, non-sensitive context to Claude.
|
|
- **Never call a Spark directly** — go through Spark Control (`SPARK_CONTROL_URL`).
|
|
|
|
## When adding a new Claude/MCP call
|
|
|
|
Trace the data path: any field carrying LP substance must cross `Boundary` first. A new MCP tool that reads CRM rows and hands them to a model without scrubbing is a leak — add it to the redaction path and extend the leak tests in `backend/redaction/test_*.py`.
|
|
|
|
A Claude path that sends **free-prose** LP content (email bodies, notes) must pass `ner_fn=_ner_local` to `Boundary` and **fail closed** if the local model is down — the dictionary+regex floor only tokenizes KNOWN CRM entities, so unknown people/firms in prose leak otherwise. See `backend/mcp/architect_grounding.py` (does it right) and `backend/mcp/outreach_agent.py`.
|
|
|
|
See also `docs/redaction-rehydration.md` and `docs/spark-control-scrub-endpoints.md`.
|