datatug / datatug-cli / 42363576401

100%

push

github

fix: apply sensitivity labels to saved, chat and datatug:// reads (#589)

Saved and project-scoped reads now apply Meaning Graph sensitivity labels. Before this change, the demo's `agent` role got real `Customer.Email` values back from a saved query.

Found by the security review of the semantic-discovery spec (datatug/backstage#190, plan task 1). On `origin/main` the demo agent reads real emails through:
- a saved federated TugQL query (`--project --query`, `database:` leaves);
- a saved single-source TugQL or native SQL query;
- `datatug chat --role agent` (saved queries and the model's own TugQL);
- `--db datatug://<project>/<source> … --no-policies`.

Each path now refuses with `FIELD_PROTECTED` (exit 5), or masks where the `--project --source` read already masks.

- **Executor-level label guard:** every `secureread.Executor` read entry point checks labels first. A reflection test fails if a new read method skips the check.
- **Federated reads:** the check runs per leaf, before any row is read, and walks nested queries. A wildcard over a labelled leaf counts as projecting it.
- **Every principal:** saved queries and `datatug://` reads apply labels whatever the policy profile, including projects with no `policies/` and `--no-policies`.
- **Native SQL:** refused on a source with a protected label, and when the source can't be opened to check.
- **Chat:** labels are looked up by catalog id, so catalogs sharing one URL get the union of their labels.
- **Behaviour change for owners:** pending read-time suggestions now mask columns (for example `FirstName`) in `datatug://` reads, as `meaning-sensitivity` and the `--project --source` read already do. Unmask a column by labelling it `none` or dismissing the suggestion.
- **Not changed (follow-ups):**
  - `db copy` / `db export` from `datatug://` still copy labelled values; the reviewer rated this Medium and pre-existing;
  - `serve` exec/run_query keeps its agent-header rule;
  - plain HTTP sources skip t... (continued)

487 of 487 new or added lines in 13 files covered. (100.0%)

98558 of 98559 relevant lines covered (100.0%)

1.01 hits per line