You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 90e2eae
Browse filesBrowse the repository at this point in the historyBrowse files
- Safety diff-scoping is now a MULTISET subtraction (one base occurrence
consumes one head occurrence of the same rule + matched_pattern) — a PR
that adds a second identical injection still reports it; `safe` and
`risk_score` are recomputed when every threat is pre-existing so the
verdict doesn't stay stale-unsafe.
- PII exposures are diff-scoped the same way: (table, column) pairs already
exposed by the base are not reported as introduced.
- Validation is deliberately NOT diff-scoped: the engine validates fail-fast
(only the first error is reported), so base subtraction can hide genuinely
new breakage behind a pre-existing error — proven by test; re-reporting a
pre-existing error is the safe mode, documented inline.
- Pre-existing bug found while testing: the composite check with `base_sql`
+ a flat agent schema_context crashed with "missing field tables" — lintDiff
takes SchemaDefinition JSON; now normalized via `normalizeSchemaContext`
(exported from schema-resolver). This made every diff-scoped composite call
with a flat schema fail closed since the lintDiff wiring landed.
- Semantic CLI abstention severity: error when a schema was provided (the
analysis should have run), warning schema-less (abstains routinely).
- Policy tool telemetry includes advisory warnings (`policy_warning`).
- `pii-detector` gets the AGENTS.md namespace self-reexport.
Co-Authored-By: Claude Fable 5 <[email protected]>
0 commit comments