nsauditor-ai-agent-skill 0.1.32 → 0.1.34

package/CHANGELOG.md

@@ -4,6 +4,57 @@ Release notes for **`nsauditor-ai-agent-skill`** — installable knowledge packa
 
 ---
 
+## 0.1.34 — Catalog refresh: EE 0.7.3 R-CRITICAL hotfix closing 2 production bugs surfaced by EE 0.7.2 dogfood scan against operator's GCP test infra (cross-version google-auth-library fragmentation broke SA impersonation chains [R-CRITICAL — 100% false-clean impact on free-trial/gmail GCP customers + business GCP customers with no-long-lived-SA-keys policy]; GOOGLE_CLOUD_PROJECT_ID env-var alias silently skipped [R-MEDIUM]; +14 new tests across 2 new suites incl. regression pin replicating gax 5.x grpc adapter idiom; plugin count UNCHANGED at 24; coverage matrix UNCHANGED at 10/4/33; EE regression 5782/5782 across 900 suites; 67-session 100% green streak preserved; twenty-fourth consecutive trio-publish)
+
+**Trio-publish institutionalization continued.** Paired with EE 0.7.3 + CE 0.1.67 — **twenty-fourth consecutive trio-publish across EE + CE + agent-skill in a single session** (0.4.5–0.7.3).
+
+### Headline — R-CRITICAL hotfix surfaced by EE 0.7.2 dogfood scan within 30 minutes of the 0.7.2 trio publish
+
+EE 0.7.3 closes 2 production bugs that shipped silently in EE 0.7.0–0.7.2. Both bugs surfaced when running `nsauditor-ai scan --plugins 1025 --compliance soc2` against operator's GCP test infra immediately after the 0.7.2 trio publish.
+
+**Gap #2 (R-CRITICAL)**: cross-version `google-auth-library` fragmentation. EE's `utils/gcp_auth.mjs` resolved `9.15.1` at the top level (hoisted via `googleapis@^144`); `@google-cloud/resource-manager@^6` bundles nested `10.6.2` + `google-gax@5.x` whose grpc adapter calls `headers.forEach((value, key) => ...)` expecting WHATWG Headers instance. 9.x returns plain object → `.forEach` undefined → TypeError → `2 UNKNOWN: Getting metadata from plugin failed with error: headers.forEach is not a function`. Plugin 1025's conservative classifier emitted `gcp-iam-project-unreadable` LOW + walkthroughRequired, masking the fact that ALL 7 dims silently skipped. Production false-clean impact: ~100% on any impersonation-using deployment in EE v0.7.0–0.7.2.
+
+**Fix**: NEW `_wrapAuthClientHeadersShim` in `utils/gcp_auth.mjs` monkey-patches the Impersonated instance's `getRequestHeaders` to coerce 9.x's plain-object return into a Headers instance via `new Headers(plainObject)`. 10.x returns pass through unchanged. Version-agnostic, future-proof. +8 new tests including a regression pin that exactly replicates the gax 5.x grpc adapter idiom — catches any future shim regression at unit-test time.
+
+**Customer-segment impact:**
+- **GCP free-trial / gmail customers** — impersonation is the ONLY working credential model when `iam.disableServiceAccountKeyCreation` is enforced (Google's "Secure by default"). Pre-0.7.3 100% false-clean. **Post-0.7.3 audit works end-to-end.**
+- **Business GCP customers with no-long-lived-SA-keys security policy** — many enterprise security teams mandate impersonation as their auth model. Same impact. **Post-0.7.3 audit works.**
+- **Business GCP customers using JSON keyfiles or pure ADC** — unaffected (R-CRITICAL specific to impersonation injection; pure-ADC + keyfile paths use the nested 10.x auth chain entirely).
+
+**Gap #1 (R-MEDIUM)**: operators following the `gcloud auth application-default login` setup convention (which writes `GOOGLE_CLOUD_PROJECT_ID` with `_ID` suffix) saw silent skip with `[plugin 1025] No GCP_PROJECT_ID configured`. Extended `loadConfig` + `preflight` from 2-way OR to 3-way OR: `opts.projectId > GCP_PROJECT_ID > GOOGLE_CLOUD_PROJECT > GOOGLE_CLOUD_PROJECT_ID`. +6 new tests covering all precedence paths + preflight failure-reason enumeration + end-to-end run() with env-only resolution.
+
+### Dogfood validation post-fix
+
+Re-ran the scan with both fixes applied. **8 findings emitted** (was 1 false-clean LOW pre-fix): 5 PASS + 2 MEDIUM + 1 LOW. All 7 dims exercise via the impersonated `nsauditor-readonly` audit SA. `accessDeniedByApi.listPolicies: 1` confirms the 0.7.2 R2-MED-13 counter wiring works end-to-end against real GCP.
+
+### Regression preserved
+
+EE full regression: **5782/5782 across 900 suites** (was 5768/5768 across 898 suites at 0.7.2; +14 tests / +2 suites). **67-session 100% green streak preserved.** Plugin count UNCHANGED at 24. Coverage matrix UNCHANGED at 10/4/33.
+
+---
+
+## 0.1.33 — Catalog refresh: EE 0.7.2 Move B pure-test functional patch closing 5 deferred 0.7.1 reviewer-pass coverage gaps (+50 new tests across 6 new suites; no production code changes; no plugin emissions changed; no soc2.json changes; no new SDK deps; plugin count UNCHANGED at 24; coverage matrix UNCHANGED at 10/4/33; EE regression 5768/5768 across 898 suites; 66-session 100% green streak preserved; twenty-third consecutive trio-publish)
+
+**Trio-publish institutionalization continued.** Paired with EE 0.7.2 + CE 0.1.66 — **twenty-third consecutive trio-publish across EE + CE + agent-skill in a single session** (0.4.5–0.7.2).
+
+### Headline — pure-test functional patch (no plugin/soc2.json/SDK changes)
+
+EE 0.7.2 is a pure-test functional patch closing the 5 test-coverage gaps deferred at 0.7.1's reviewer pass — bundled with the staged `peerDependencies.nsauditor-ai` bump (`^0.1.40` → `^0.1.65`) queued at 0.7.1 post-publish per `[[npm_tarball_replacement_trap]]` discipline. Plugin 1025 GCP IAM Project-Level Auditor's 7-dim coverage shipped in EE 0.7.1; this cycle backfills the test surface around it without altering production behavior.
+
+### Test additions — 50 new tests across 6 new suites
+
+- **R2-MED-7 BFS edge cases (+17)** — `_detectGcpImpersonationPaths` exercised against multiple disjoint cycles (Island A doesn't bleed into Island B), disconnected subgraphs, terminate-at-first-admin (multi-admin chain), parallel branches to distinct admins, depthCap exact-match + one-short + =1 boundaries, per-PATH visited Set semantics, malformed edges (null / missing-to / non-string-to), nonexistent edge targets, cycle through admin, self-loop on start, edge label fallback chain (label → displayName → key), fractional depthCap, parallel edges to same admin with different `via`.
+- **R2-MED-13 counter wiring (+15 parameterized)** — 5 v2 apiName strings × 3 counter classes: `projects.roles.list` + `projects.serviceAccounts.list` + `projects.serviceAccounts.keys.list` + `projects.serviceAccounts.getIamPolicy` + `listPolicies` × throttle-retry + access-denied + wall-budget-exhausted. Closes the institutional contract "every API surface increments the right counter key" — prior cycle tested v1's `getIamPolicy` directly but only indirect coverage of v2 apiNames via run() integration.
+- **R2-LOW-16/17 helper edges (+10)** — `_saEmailFromName`: trailing slash → `""`, leading slash → segment-after, only-slash → `""`, multiple slashes (lastIndexOf semantics), control-char strip BEFORE slash detection. `_parseIso8601ToMs`: positive `+HH:MM` offset (yields earlier UTC ms), negative `-HH:MM` offset (yields later UTC ms), date-only string (UTC midnight), fractional-seconds + Z, finite-return for well-formed-with-offset.
+- **R2-HIGH-4 SDK loader graceful-degradation contract (+8)** — direct unit tests for `_loadGoogleApisIamAdminSdk` + `_loadOrgPolicySdk` missing-dep error branches. Both SDKs are in optionalDependencies and NOT installed in the EE working tree by default. The institutional contract: loader throws with dep-name + `Cannot find package` cause; run()'s catch handler converts to single-warning skip of the affected dim cohort.
+- **R2-MED-12 buildGcpAuthOptions real-SDK fallback (+3)** — exercises the `deps._googleAuthLibrarySdk || await _loadGoogleAuthLibrarySdk()` fallback path that all other buildGcpAuthOptions tests bypass via dep injection. Uses `crypto.generateKeyPairSync` to write a valid PKCS#8 SA JSON keyfile to tmpdir, then verifies the real google-auth-library returns a real `Impersonated` instance with documented `targetPrincipal` + `lifetime` shape.
+
+### Regression preserved
+
+EE full regression: **5768/5768 across 898 suites** (was 5715/5715 across 892 suites at 0.7.1; +53 tests / +6 suites). **66-session 100% green streak preserved.** Plugin count UNCHANGED at 24. Coverage matrix UNCHANGED at 10/4/33.
+
+---
+
 ## 0.1.32 — Catalog refresh: plugin 1025 GCP IAM Project-Level Auditor EXTENDED to v2 (3 dims → 7 dims) — paired with EE 0.7.1 trio-publish (EE-RT.22 v2 R2 expansion closing all 4 v1-deferred dims; +4 new dims: custom-role permission audit + SA key custody + SA impersonation graph BFS + Organization Policy constraint enumeration; NEW `utils/gcp_auth.mjs` helper honors `GOOGLE_IMPERSONATE_SERVICE_ACCOUNT`; **17 same-session reviewer folds = NEW HIGH-WATER MARK** vs 0.7.0's 12 (1 R-CRITICAL EE-RT.20 class recurrence catch + 7 R-HIGH + 8 R-MEDIUM + 1 R-LOW(+1 grouped)); plugin count UNCHANGED at 24; +22 new soc2.json mappings; new SDK deps `googleapis` + `@google-cloud/org-policy` in optionalDependencies; twenty-second consecutive trio-publish)
 
 **Trio-publish institutionalization continued.** Paired with EE 0.7.1 + CE 0.1.65 — **twenty-second consecutive trio-publish across EE + CE + agent-skill in a single session** (0.4.5–0.7.1).

package/SKILL.md

@@ -297,7 +297,7 @@ CE collision. CE reserves 001-099.
 
 **Plugin 1170 v3 (EE 0.6.6) SG→SG transitive chain reachability** — `aws-ec2-sg-perimeter-auditor` v3 extension. Pre-v3 each Security Group was audited in isolation; a SG with no direct public-CIDR ingress would emit the PASS-tier "no direct public-internet ingress CIDR rules" finding even if transitively reachable from the internet through a `UserIdGroupPairs` chain. v3 builds the SG-reference graph (`_buildSgReferenceGraph`), identifies public-CIDR roots (`_findPubliclyReachableSgs` — 0.0.0.0/0 / ::/0 ingress), and BFS-walks the graph (`_walkTransitiveReachability`) with cycle defense + depth cap (default 5, max 20) + per-target chain cap (default 10, max 100). 2-hop chains emit **HIGH**; 3+ hop chains emit **CRITICAL** (operator-blindness principle — deeper chains less likely to be noticed). Cross-VPC edges skipped (out-of-scope for v3 v1; INFO trailer). v3 v1 simplification: per-hop port-flow tracked but NOT intersected (`walkthroughRequired=true`). New operator opts: `skipTransitiveReachability` / `transitiveChainDepthCap` / `transitiveChainsPerTargetCap` / `transitiveChainSamplesPerFindingCap`. **v3 R-HIGH-1 fold**: BFS short-circuits enqueue past per-target cap (closes path-enumeration explosion on hub-and-spoke topologies — pre-fold the BFS kept cloning `path` and `visited` Sets and walking past the cap). **v3 R-LOW-2 fold**: depth-cap-hit surfaced separately from per-target-cap (closes silent-deep-truncation false-CLEAN class). 3 new soc2.json mappings under CC6.6 (transitive HIGH + CRITICAL + INFO truncation). **v3.1 EE 0.6.7 closes the edge-dedup R2-deferred item**: `_buildSgReferenceGraph` now dedupes edges by `(sourceGroupId, targetGroupId)` with `ports` aggregated as array of `{protocol, fromPort, toPort}`. Pre-fold a real-world ALB-fronting-app SG with 3 ingress perms on different ports (80/443/8080) referencing the same source SG emitted 3 distinct edges A→B; the BFS treated each as a separate chain, inflating `chainCount` 2-5× and exhausting per-target chain caps on noise. Post-fold the BFS sees exactly 1 chain per distinct (source, target) pair. `isCrossVpc` aggregation is AND-semantic — if ANY contributing pair is same-VPC, the merged edge is same-VPC (per `[[conservative_classifier_principle]]`: walk possibly-same-VPC chains rather than silently skip). Classifier port-render accepts both v3.1 array shape and v3 single-object shape (back-compat). **v3.1 R-MEDIUM-1 fold**: arrival-order independence locked with 2 regression fixtures + JSDoc tightening. **v3.1 R-LOW-1 fold**: partial-render contract on malformed port specs locked with 2 fixtures. **v3.1 R-LOW-2 fold**: `_portKeys` scratch-lifetime documented (MUST NOT escape).
 
-**EE SOC 2 substrate-evidence coverage (post-EE 0.7.1):** 10 covered controls (CC6.1 /
+**EE SOC 2 substrate-evidence coverage (post-EE 0.7.3):** 10 covered controls (CC6.1 /
 CC6.2 / CC6.6 / CC6.7 / CC6.8 / CC7.1 / CC7.2 / CC7.3 / C1.1 / C1.2) + 4 partial
 (CC6.3 / CC8.1 / A1.2 / PI1.5) + 33 OOS for static substrate scanning. Coverage matrix
 is institutionally honest: substrate-evidence depth grows release-over-release without

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nsauditor-ai-agent-skill",
-  "version": "0.1.32",
+  "version": "0.1.34",
   "description": "AI Agent Skill for NSAuditor AI — gives any AI coding agent built-in knowledge of NSAuditor's MCP tools, schemas, plugins, and security audit workflows.",
   "keywords": [
     "nsauditor",