@clear-capabilities/agentic-security-scanner 0.136.2 → 0.137.0

package/src/mcp/tools.js

@@ -21,6 +21,7 @@ import { synthesizeDeterministicPatch } from '../posture/deterministic-fix.js';
 import { verifyLastScan } from '../posture/integrity.js';
 import { analyzeTranscript, formatCacheReport, renderCacheStatusLine } from '../posture/cache-economics.js';
 import { redactString, redactFinding } from './redact.js';
+import { _remediationOf } from '../report/index.js';
 
 // Lazy-loaded: these transitively pull in npm packages (@babel/core and
 // friends) that aren't available in the plugin-cache install path
@@ -152,8 +153,13 @@ function _validateScratchpadPath(relPath) {
   return { ok: true, agent, session, fileParts };
 }
 
+// Routes through the same lstat+realpath confinement every other write/
+// path-taking tool uses (OWASP MCP05) — a lexical prefix/charset check
+// alone doesn't stop a pre-planted symlink at any path component from
+// relocating the write/read outside the session root. Throws on escape;
+// callers must catch (see append_scratchpad / read_scratchpad).
 function _scratchpadAbs(sessionRoot, relPath) {
-  return path.resolve(sessionRoot, relPath.replace(/\\/g, '/'));
+  return _confine(sessionRoot, relPath.replace(/\\/g, '/'), 'scratchpad path');
 }
 
 function _scratchpadTotalBytes(sessionRoot) {
@@ -332,17 +338,35 @@ export const scan_diff = {
       fileContents[rel] = content;
     }
 
+    // PRD R1 (docs/DETECTION_GAP_REMEDIATION_PRD.md): deep mode is default-on
+    // for the interactive CLI scan but was never requested here, so an
+    // agent's pre-write self-correction scan was regex/AST-only — blind to
+    // any bug whose source and sink are connected only through a call
+    // (`fileContents` scopes the deep engine's IR to exactly the files
+    // passed in, same bound this tool already enforces via MAX_FILES_PER_SCAN
+    // / MAX_TOTAL_SCAN_BYTES, so this does not turn scan_diff into a
+    // full-project deep scan).
     const runScan = await getRunScan();
-    const result = await runScan(sessionRoot, { network: false, fileContents });
+    const result = await runScan(sessionRoot, { network: false, fileContents, deep: true, deepInCi: true });
     const wantSet = new Set(Object.keys(fileContents));
     const sevRank = { info: 0, low: 1, medium: 2, high: 3, critical: 4 };
     const min = sevRank[severity] ?? 0;
-    const findings = (result.scan.findings || [])
+    // Stage 6 correctness audit: this only ever read result.scan.findings
+    // (the SAST channel) — scan.secrets and scan.logicVulns are separate
+    // arrays on the raw runScan() result (report/index.js's normalizeFindings
+    // is what merges all four channels, and that merge hasn't run yet here).
+    // A file containing a bare hardcoded credential reported findingCount: 0
+    // through a tool whose own description promises "Use BEFORE writing a
+    // Write/Edit to disk so the agent can self-correct". Also reused
+    // _remediationOf so a fix-string detector (the majority of engine.js's
+    // own, ~127 call sites) doesn't silently report an empty `description`
+    // the way reading only `.remediation` did.
+    const findings = [...(result.scan.findings || []), ...(result.scan.secrets || []), ...(result.scan.logicVulns || [])]
       .filter(f => wantSet.has(String(f.file || '').replace(/\\/g, '/')) && (sevRank[f.severity] ?? 0) >= min)
       .map(f => redactFinding({
         id: f.id, severity: f.severity, file: f.file, line: f.line,
         title: f.title || f.vuln, cwe: f.cwe,
-        description: f.description, remediation: f.remediation,
+        description: f.description, remediation: _remediationOf(f),
       }));
     // Harness-anatomy #1: offload when the result exceeds OFFLOAD_THRESHOLD.
     // The agent gets a head+tail preview plus a path it can page through;
@@ -505,10 +529,40 @@ export const apply_fix = {
         additionalProperties: { type: 'string', maxLength: 500_000 },
         minProperties: 1, maxProperties: 8,
       },
+      // Stage 6 correctness audit: same gap and same fix as verify_fix — the
+      // honesty gate is reachable but was never wired to any real caller.
+      // Here it's stronger than advisory: the inline re-verify below already
+      // gates the WRITE on `verdict.ok`, and verifyFixCore's own `ok`
+      // formula already folds in `honesty.ok` when fixMeta is supplied — so
+      // passing it through here makes a dishonest fixMeta (hand-wave
+      // residual, uncited false-positive verdict) block the write itself,
+      // not just report a verdict.
+      fixMeta: {
+        type: 'object',
+        additionalProperties: false,
+        properties: {
+          residual: { type: 'string', maxLength: 2000 },
+          verdict: { type: 'string', maxLength: 64 },
+          evidence: { type: 'array', items: { type: 'string', maxLength: 500 }, maxItems: 20 },
+          signals: {
+            type: 'object',
+            additionalProperties: false,
+            properties: {
+              sinkSignatureChanged: { type: 'boolean' },
+              allCallersRouted: { type: 'boolean' },
+              testDiscriminates: { type: 'boolean' },
+              rateLimitOnly: { type: 'boolean' },
+              docsOnly: { type: 'boolean' },
+              logOnlyNoReject: { type: 'boolean' },
+              partialSanitization: { type: 'boolean' },
+            },
+          },
+        },
+      },
     },
     required: ['finding_id', 'confirm'],
   },
-  async handler({ finding_id, confirm, dry_run = false, patch = null }, ctx) {
+  async handler({ finding_id, confirm, dry_run = false, patch = null, fixMeta = null }, ctx) {
     if (confirm !== true) {
       return { _meta: META, applied: false, reason: 'apply_fix requires confirm: true.' };
     }
@@ -564,6 +618,7 @@ export const apply_fix = {
             scanRoot: ctx.sessionRoot,
             originalFindingStableId: f.stableId,
             files: _files,
+            fixMeta,
           });
         }
       } catch (e) {
@@ -573,7 +628,7 @@ export const apply_fix = {
         return {
           _meta: META, applied: false,
           reason: `patch rejected by verifier: ${verdict.summary || verdict.rescan?.reason || 'did not verify'}`,
-          verify: { rescan: verdict.rescan, lint: { runner: verdict.lint?.runner, ok: verdict.lint?.ok } },
+          verify: { rescan: verdict.rescan, lint: { runner: verdict.lint?.runner, ok: verdict.lint?.ok }, honesty: verdict.honesty || null },
         };
       }
       if (dry_run) {
@@ -585,7 +640,7 @@ export const apply_fix = {
           const originalContent = fs.existsSync(v.abs) ? await fsp.readFile(v.abs, 'utf8') : '';
           const entry = await applyFixHistory({
             scanRoot: ctx.sessionRoot, file: rel, originalContent, newContent: v.content,
-            findingId: f.id, stableId: f.stableId, ruleId: f.rule || null, vuln: f.vuln || f.title || null,
+            findingId: f.id, stableId: f.stableId, ruleId: f.ruleId || f.cwe || f.family || null, vuln: f.vuln || f.title || null,
           });
           written.push({ file: rel, historyId: entry.id, backupPath: entry.backupPath });
         }
@@ -643,7 +698,7 @@ export const apply_fix = {
         newContent: f.fix.replacement,
         findingId: f.id,
         stableId: f.stableId || null,   // premortem 4R-8
-        ruleId: f.rule || null,
+        ruleId: f.ruleId || f.cwe || f.family || null,
         vuln: f.vuln || f.title || null,
       });
     } catch (e) {
@@ -680,7 +735,7 @@ export const apply_fix = {
 // proceed with apply_fix.
 export const verify_fix = {
   name: 'verify_fix',
-  description: 'Verify a proposed patch before applying. Re-scans the patched files in memory and runs the project linter. Returns { ok, rescan, lint, summary }. No filesystem writes.',
+  description: 'Verify a proposed patch before applying. Re-scans the patched files in memory, runs the project linter, runs the project test suite, checks fix honesty (FULL/MITIGATION/WORKAROUND) when fixMeta is supplied, and re-runs the PoC when one exists. Returns { ok, rescan, lint, tests, honesty, poc, summary }. Does not write to the target project’s own files, but DOES append one record per attempt to .agentic-security/fix-metrics.jsonl for the measured fix-loop.',
   inputSchema: {
     type: 'object',
     additionalProperties: false,
@@ -692,10 +747,41 @@ export const verify_fix = {
         minProperties: 1,
         maxProperties: 8,
       },
+      // Stage 6 correctness audit: posture/fix-honesty-gate.js's deterministic
+      // honesty checks (vague-assurance residual prose, unbacked false-
+      // positive verdicts, tier/residual consistency) were fully built and
+      // fix-verify.js already consulted them when given a `fixMeta` — but
+      // this schema never had a `fixMeta` property, so no call through the
+      // MCP surface could ever supply one. The gate can only run against
+      // claims the AGENT self-reports (residual risk, verdict, evidence,
+      // completeness signals) — nothing here is server-computable — so
+      // fixing this meant exposing the property, not inventing a lookup.
+      fixMeta: {
+        type: 'object',
+        additionalProperties: false,
+        properties: {
+          residual: { type: 'string', maxLength: 2000 },
+          verdict: { type: 'string', maxLength: 64 },
+          evidence: { type: 'array', items: { type: 'string', maxLength: 500 }, maxItems: 20 },
+          signals: {
+            type: 'object',
+            additionalProperties: false,
+            properties: {
+              sinkSignatureChanged: { type: 'boolean' },
+              allCallersRouted: { type: 'boolean' },
+              testDiscriminates: { type: 'boolean' },
+              rateLimitOnly: { type: 'boolean' },
+              docsOnly: { type: 'boolean' },
+              logOnlyNoReject: { type: 'boolean' },
+              partialSanitization: { type: 'boolean' },
+            },
+          },
+        },
+      },
     },
     required: ['stable_id', 'files'],
   },
-  async handler({ stable_id, files }, ctx) {
+  async handler({ stable_id, files, fixMeta }, ctx) {
     // Confine every file path before passing to the verifier.
     const confined = {};
     for (const [relPath, content] of Object.entries(files || {})) {
@@ -707,17 +793,48 @@ export const verify_fix = {
       confined[relPath] = String(content);
     }
     try {
+      // The PoC-re-check leg (verifyFixCore's `pocLeg`) needs a `poc` param
+      // to do anything — until now nothing supplied one, so it always
+      // reported {status:'not-requested'} through this surface (see
+      // posture/CLAUDE.md's disclosure). Rather than widening inputSchema
+      // to make the CALLER pass PoC data back, look it up server-side: the
+      // scan pipeline already attaches an HTTP-shaped f.poc to matching
+      // findings by default (engine.js's annotatePocs), and last-scan.json
+      // already carries it under the same stableId this handler receives.
+      // Best-effort: a missing/unsigned/tampered scan just means no PoC is
+      // available to re-check, not a verify_fix failure — the rescan/lint/
+      // tests legs below are independent of this and still apply.
+      let poc = null;
+      try {
+        const { scan: lastScan } = _readLastScanVerified(ctx.sessionRoot, { allowUnsigned: true });
+        const orig = lastScan && (lastScan.findings || []).find(f => f.stableId === stable_id);
+        if (orig && orig.poc && orig.poc.code) poc = { ...orig.poc, finding: orig };
+      } catch { /* best-effort lookup; poc stays null */ }
+
       const verifyFixCore = await getVerifyFixCore();
       const r = await verifyFixCore({
         scanRoot: ctx.sessionRoot,
         originalFindingStableId: stable_id,
         files: confined,
+        poc,
+        fixMeta,
       });
       return {
         _meta: META,
         ok: r.ok,
         rescan: { ok: r.rescan.ok, reason: r.rescan.reason, introduced: r.rescan.introduced || [] },
         lint: { runner: r.lint.runner, ok: r.lint.ok, skipped: r.lint.skipped || false, output: redactString(r.lint.output || '').slice(0, 1500) },
+        // verifyFix computes five legs, not two — tests/honesty/poc were
+        // being silently dropped here, leaving an agent with no structured
+        // way to see WHY verification failed when the failure was in one
+        // of those three (only the free-text summary carried it).
+        // test-runner.js's runProjectTests never returns raw stdout/stderr,
+        // so no redaction is needed there; honesty.violations are static,
+        // code-generated strings; poc.reason is redacted defensively since
+        // it can echo proof-harness detail derived from scanned source.
+        tests: r.tests,
+        honesty: r.honesty,
+        poc: r.poc ? { ...r.poc, reason: r.poc.reason ? redactString(r.poc.reason) : r.poc.reason } : r.poc,
         summary: r.summary,
       };
     } catch (e) {
@@ -801,7 +918,13 @@ export const synthesize_fix = {
       regression_test: f.regression_test || null,
       remediation: typeof fix.description === 'string' ? fix.description : (typeof fix === 'string' ? fix : null),
       patchBounds: { touchedFiles, locDelta, oversized },
-      recommendsFixPlan: oversized && !hasReplacement && !autofix,
+      // oversized can only be true when hasReplacement is true (locDelta is
+      // only computed in that branch, and touchedFiles never varies) — a
+      // `!hasReplacement` conjunct here was a structural contradiction that
+      // made this permanently false. The correct signal: the stored
+      // replacement itself is too big to trust auto-applying, and there's
+      // no safer deterministic alternative.
+      recommendsFixPlan: oversized && !autofix,
     };
   },
 };
@@ -933,7 +1056,9 @@ export const append_scratchpad = {
   async handler({ path: relPath, content }, ctx) {
     const v = _validateScratchpadPath(relPath);
     if (!v.ok) return { _meta: META, ok: false, reason: v.reason };
-    const abs = _scratchpadAbs(ctx.sessionRoot, relPath);
+    let abs;
+    try { abs = _scratchpadAbs(ctx.sessionRoot, relPath); }
+    catch (e) { return { _meta: META, ok: false, reason: `path-escape refused: ${e.message}` }; }
     const total = _scratchpadTotalBytes(ctx.sessionRoot);
     if (total + content.length > SCRATCHPAD_MAX_TOTAL_BYTES) {
       return {
@@ -979,7 +1104,9 @@ export const read_scratchpad = {
   async handler({ path: relPath, offset, limit }, ctx) {
     const v = _validateScratchpadPath(relPath);
     if (!v.ok) return { _meta: META, ok: false, reason: v.reason };
-    const abs = _scratchpadAbs(ctx.sessionRoot, relPath);
+    let abs;
+    try { abs = _scratchpadAbs(ctx.sessionRoot, relPath); }
+    catch (e) { return { _meta: META, ok: false, reason: `path-escape refused: ${e.message}` }; }
     if (!fs.existsSync(abs)) return { _meta: META, ok: false, reason: 'not-found' };
     let stat;
     try { stat = fs.statSync(abs); } catch (e) { return { _meta: META, ok: false, reason: `stat-failed: ${e.message}` }; }
@@ -1075,7 +1202,20 @@ export const query_triage_memory = {
   },
   async handler({ query }, ctx) {
     const { queryMemory } = await import('../posture/triage-memory.js');
-    const results = queryMemory(ctx.sessionRoot, query || '');
+    const raw = queryMemory(ctx.sessionRoot, query || '');
+    // Stage 6 correctness audit: this returned queryMemory's output
+    // verbatim, with no redaction pass — every other tool that echoes
+    // scanned-source-derived text redacts it (mcp/CLAUDE.md's "Adding a new
+    // tool" step 3). Round-trip through redactString the same way
+    // redactFinding already does for its own opaque `.trace` field: results
+    // here mix shapes (a triage decision's free-text `reason`, a finding's
+    // `vuln`/`family`/file path), so scrubbing the whole serialized
+    // structure catches secret-shaped substrings regardless of which field
+    // they landed in, rather than hardcoding a field allowlist that could
+    // miss one.
+    let results;
+    try { results = JSON.parse(redactString(JSON.stringify(raw))); }
+    catch { results = raw; }
     return {
       _meta: META,
       count: results.length,
@@ -1103,7 +1243,16 @@ export const query_findings_memory = {
   },
   async handler({ query }, ctx) {
     const { queryFindingsMemory } = await import('../posture/findings-memory.js');
-    return { _meta: META, ...queryFindingsMemory(ctx.sessionRoot, query || '') };
+    const raw = queryFindingsMemory(ctx.sessionRoot, query || '');
+    // Stage 6 correctness audit — same redaction gap and same fix as
+    // query_triage_memory just above: this mixes four differently-shaped
+    // result kinds (finding / triage / history / AGENTS.md text), so a
+    // whole-structure redactString round-trip is applied rather than a
+    // per-field allowlist that could miss one of the four shapes.
+    let body;
+    try { body = JSON.parse(redactString(JSON.stringify(raw))); }
+    catch { body = raw; }
+    return { _meta: META, ...body };
   },
 };
 

package/src/posture/CLAUDE.md

@@ -19,7 +19,7 @@ Annotators that run **after** every detector has emitted, plus state stores read
 
 **Production-posture ingest** — `auth-posture-import.js`, `network-policy-import.js`, `telemetry-ingest.js`, `waf-ingest.js`, `feature-flags.js`. These read customer-side YAML and convert to mitigation flags consumed by `mitigation-composite.js`.
 
-**Fix lifecycle** — `fix-history.js` (apply + backup + recover), `fix-verify.js` (closed-loop re-scan + lint), `fix-plan.js` (oversized-patch fallback), `regression-test-gen.js`, `deterministic-fix.js` (safe context-independent literal-swap patch synthesis — md5/sha1→sha256, TLS verify-off→on — materialized on demand by `mcp/synthesize_fix`; every patch still passes through `apply_fix`'s inline verify before it lands).
+**Fix lifecycle** — `fix-history.js` (apply + backup + recover), `fix-verify.js` (**five legs, not "re-scan + lint"**: rescan + lint + the project test suite + the fix-honesty gate + a PoC re-check, and it appends one record per attempt to `.agentic-security/fix-metrics.jsonl` — see `mcp/CLAUDE.md`'s `verify_fix` row, which had the same stale "no writes" claim), `fix-plan.js` (oversized-patch fallback — **not currently wired to anything**, see the dead-module allowlist), `regression-test-gen.js`, `deterministic-fix.js` (safe context-independent literal-swap patch synthesis — md5/sha1→sha256, TLS verify-off→on — materialized on demand by `mcp/synthesize_fix`; every patch still passes through `apply_fix`'s inline verify before it lands).
 
 **Measured fix loop (R5)** — `fix-metrics.js`. `verifyFix` times each stage
 (`rescan`/`lint`/`tests`/`honesty`) and appends one record per attempt to
@@ -49,7 +49,8 @@ rather than creating a stray state dir outside a project.
 - `entrypoint-inventory.js` — attack-surface completeness ledger. Enumerates every entry point (HTTP/queue/cron/CLI/env/upload/webhook) with a disposition each; on `scan.entrypointInventory`.
 - `root-cause-sweep.js` — from confirmed findings, finds sibling instances detectors missed with total-count accounting (`found === candidates + mitigated`); on `scan.rootCauseSweep`. Searches the corpus **once per distinct sink pattern**, not once per finding — findings deriving the same pattern share one walk and one set of (read-only) match records. The counts are always exact; the materialised `instances` list is a bounded sample (`INSTANCE_SAMPLE_LIMIT`, 100) and says so via `instancesTruncated`. Both properties are load-bearing on large corpora: the per-finding walk was O(findings × corpus-bytes) and the instance records were O(findings × matches), which together exhausted a 6 GB heap on a 40k-file suite. If you touch this module, keep the own-site exclusion **per pattern group** — resolving it globally makes a group subtract an exclusion it never matched and drives counts negative.
 - `model-routing.js` — capability-based CWE/severity→model policy; stamps `finding.dispatchModel` (strongest for crypto/auth/critical, mid for injection, cheapest for low-sev hardening) for cost-sensitive subagent dispatch.
-- `fix-honesty-gate.js` — deterministic honesty gates on fix output: a residual-risk hand-wave guard, a cited-file:line requirement for any FP/safe verdict, and FULL/MITIGATION/WORKAROUND completeness tiers. Consumed by `fix-verify.js` when the caller supplies fix metadata; the closed-loop test leg (`fix-verify-loop.js`) is wired into `mcp/apply_fix` behind `AGENTIC_SECURITY_FIX_RUN_TESTS=1`.
+- `fix-honesty-gate.js` — deterministic honesty gates on fix output: a residual-risk hand-wave guard, a cited-file:line requirement for any FP/safe verdict, and FULL/MITIGATION/WORKAROUND completeness tiers. `fix-verify.js` accepts a `fixMeta` param and consults this gate when it is present (`if (fixMeta && typeof fixMeta === 'object')`). Both `mcp/apply_fix` and `mcp/verify_fix` now expose an optional `fixMeta: {residual, verdict, evidence, signals}` input property and pass it straight through — `fixMeta` is inherently agent-self-reported (only the caller claiming a fix worked knows its own residual-risk reasoning), so the fix was exposing the property, not computing anything server-side. On `apply_fix`'s patch path this is stronger than advisory: `verifyFixCore`'s own `ok` formula already folds in `honesty.ok`, and the inline re-verify already gates the write on `ok` — so a hand-wave residual or an uncited false-positive verdict in `fixMeta` blocks the write itself. The closed-loop test leg (`fix-verify-loop.js`) is separately wired into `mcp/apply_fix` behind `AGENTIC_SECURITY_FIX_RUN_TESTS=1` and does not (yet) thread `fixMeta` through — that path still bypasses the honesty gate.
+- **The PoC-re-check leg is now genuinely reachable, without a schema change.** `verifyFixCore` accepts a `poc` param and, when given one with `poc.code` set, re-runs the proof harness against the patched files (`fix-verify.js`, the `pocLeg` block). `mcp/tools.js`'s `verify_fix` `inputSchema` still has no `poc` property — instead of widening it to make the caller resupply PoC data it never had, the handler looks up the original finding server-side from `last-scan.json` via `stable_id` (best-effort, `allowUnsigned: true` — a missing/tampered scan just means no PoC is available, not a `verify_fix` failure) and passes its `f.poc` straight through. Since `annotatePocs` attaches an HTTP-shaped `f.poc` by default on every scan (see the "Operator entry point" section below), any finding with a matching CWE template gets its PoC re-checked automatically on every `verify_fix` call — no agent-side plumbing required. `tests`/`honesty`/`poc` are all forwarded in the response (previously silently dropped).
 
 **Relevance scoping (R6 + R9)** — `relevance.js`. Turns the two existing *inventories* into *inputs*: `entrypoint-inventory.js` supplies the attack surface, `threat-model.js` supplies assets/boundaries/STRIDE, and `annotateRelevance(findings, ctx)` scores each finding by how reachable and how threat-modelled it is. Sets `entrypointReachable: true|false|null`, `relevance` (0..1), `relevanceTier: 'direct'|'indirect'|'unreachable'|'unknown'`, `relevanceFactors[]`, and re-ranks `exploitability` (ordinal priority, ×1.15 direct / ×0.6 unreachable, floored at 0.05, tier label recomputed on the same thresholds `annotateExploitability` uses). Reachability is a forward BFS over a literal-specifier import graph (JS/TS relative + Python dotted + Java FQCN) starting at every entry-point file.
 
@@ -76,6 +77,8 @@ Canonicalisation is an **allowlist, not a denylist** — each finding reduces to
 
 Wired in `bin/agentic-security.js` after every filter and after `makeDeterministic`, over `normalizeFindings(scan)` — i.e. it attests the set that actually ships — and surfaced as `attestation` in `toJSON`. `bundleSha` is read from the sidecar next to the *running* bundle and is `'unavailable'` when running from source, rather than reporting a dist hash that may not correspond to this run.
 
+**`verifyRunAttestation` now has two real callers.** `agentic-security verify-attestation <file>` auto-detects whether the given JSON is an evidence bundle (`.finding`+`.signature`, verified via `evidence-bundle.js`'s Ed25519 path, unchanged) or a run attestation (`.digest`+`.canonicalisation`, either bare or embedded under a full `last-scan.json`'s `.attestation` field) and dispatches accordingly. A run attestation isn't self-contained the way a bundle is — verifying it means re-scanning the project (`--against <path>`, default `.`) and confirming the fresh scan reproduces the attested digest, which is the actual, meaningful claim this artifact makes ("does this codebase, scanned now, match what was attested earlier"). Separately, `scripts/release-check.mjs`'s `attestation-self-check` gate round-trips a synthetic finding set through compute→verify (and a mutated copy through verify, which must fail) on every release, catching a broken canonicalisation or signing path before it ships — independent of whether any project ever calls `verify-attestation` on a real artifact.
+
 **Integrity + signing** — `integrity.js` (per-install HMAC for `last-scan.json`), `rule-pack-signing.js`. The HMAC key lives at `$XDG_CONFIG_HOME/agentic-security/scan-key`; override via `$AGENTIC_SECURITY_HMAC_KEY`. Premortem-derived; do not regress to hostname-derived.
 
 **Rule lifecycle** — `custom-rules.js` (YAML pattern DSL), `rule-overrides.js` (`disable:` gated on signature), `rule-packs.js`, `rule-synthesis.js` (proposes suppressions from triage feedback), `ruleset-version.js`.
@@ -105,7 +108,7 @@ is also true of a scan that read zero files — so when nothing was examined eve
 mapped control degrades to `engine-gap` instead of reporting as satisfied. That
 one was caught by the module's own test, not in review.
 
-**Posture artifacts** — `sbom.js`, `aibom.js`, `api-inventory.js`, `threat-model.js`, `trust-boundary-diagram.js`, `stack-playbook.js`, `deploy-platform.js`, `license-policy.js`, `material-change.js`, `mttr.js`, `streak.js`, `scorecard.js`, `security-trend.js`.
+**Posture artifacts** — `sbom.js`, `aibom.js`, `api-inventory.js`, `threat-model.js`, `trust-boundary-diagram.js`, `stack-playbook.js`, `deploy-platform.js`, `license-policy.js`, `material-change.js`, `mttr.js`, `streak.js`, `accuracy-scorecard.js` (see "Published accuracy scorecard (R3)" above — this line previously named a bare "scorecard" module that never existed under that filename), `security-trend.js`.
 
 **Why this fired** — `why-fired.js`. Runs LAST so it reflects every annotation. Customer-facing provenance.
 
@@ -234,10 +237,19 @@ the CI-gated tier and graduation into it is a human decision with a stated
 policy; an automated writer must not decide what blocks everyone's build.
 
 **Operator entry point:** `scripts/enroll-proven-finding.mjs <project>`
-(`--dry-run` scores without writing). It proves findings itself — the scan
-pipeline does **not** attach a `poc` to findings or promote proof tiers, so
-`last-scan.json` never contains an `execution-proven` finding on its own. PoCs
-come from the PoC-generator. Enrolment additionally needs fixed content
+(`--dry-run` scores without writing). It proves findings itself — **but the
+scan pipeline DOES attach an HTTP-shaped `f.poc` by default** (`annotatePocs`,
+`engine.js`, unconditional — not behind a flag; findings with no matching CWE
+template get `f.poc: null`). What the scan pipeline does NOT do by default is
+the *sandbox execution proof* that promotes a finding to the
+`execution-proven` tier: that pass is genuinely opt-in
+(`AGENTIC_SECURITY_PROVE=1`), so `last-scan.json` never contains an
+`execution-proven` finding from an ordinary scan on its own — this file
+previously conflated "attaches a poc" with "promotes to execution-proven,"
+which are two different passes with two different default states. Enrolment
+still proves findings itself via its own sandboxed run, independent of
+whichever tier `last-scan.json` shipped with. Enrolment additionally needs
+fixed content
 (`finding.fix.patch`) for `post/`; a proven finding with no fix is reported as
 skipped, not dropped. After enrolling, refresh the baseline
 (`npm run bench:cve-replay:update-baseline`) and commit it.

package/src/posture/accuracy-scorecard.js

@@ -335,8 +335,16 @@ export function renderScorecardMarkdown(m) {
     L.push('at the commit where a vulnerability really existed, with the CWE assigned by a');
     L.push('public advisory database rather than by this project.');
     L.push('');
+    // population.unscored is an array of {id, reason} (bench/independent/runner.mjs) —
+    // render its count, not the array itself (Array#toString would stringify to
+    // "[object Object],[object Object]" or blank for an empty array, both silently
+    // wrong on the line this section calls "the number that matters"). Tolerates a
+    // bare number too, for any already-committed artifact predating this fix.
+    const unscoredList = ind.population?.unscored;
+    const unscoredCount = Array.isArray(unscoredList) ? unscoredList.length
+      : (typeof unscoredList === 'number' ? unscoredList : 0);
     L.push(`**Measured ${ind.measuredAt} on engine ${ind.engineVersion}, ` +
-      `n=${ind.population?.scoredEntries}, ${ind.population?.unscored} unscored** ` +
+      `n=${ind.population?.scoredEntries}, ${unscoredCount} unscored** ` +
       '(*committed artifact*, `' + ind.source + '` — read, not re-run: scoring takes ~32 minutes).');
     L.push('');
     L.push('| | Advisory-local (**the claim**) | Wide (diagnostic) |');

package/src/posture/aibom.js

@@ -156,12 +156,16 @@ function _extractPromptFile(fp, content) {
   };
 }
 
+// A package's role isn't mutually exclusive in reality — `openai` and
+// `@anthropic-ai/sdk` are both a general inference framework AND an
+// embedding-provider SDK. Returns every matching class, not just the first.
 function _classifyFramework(c) {
   const name = (c.name || '').toLowerCase();
-  if (FRAMEWORK_PACKAGES.has(name) || FRAMEWORK_PACKAGES.has(c.name)) return 'inference-framework';
-  if (VECTOR_STORE_PACKAGES.has(name) || VECTOR_STORE_PACKAGES.has(c.name)) return 'vector-store';
-  if (EMBEDDING_PACKAGES.has(name) || EMBEDDING_PACKAGES.has(c.name)) return 'embedding-provider';
-  return null;
+  const classes = [];
+  if (FRAMEWORK_PACKAGES.has(name) || FRAMEWORK_PACKAGES.has(c.name)) classes.push('inference-framework');
+  if (VECTOR_STORE_PACKAGES.has(name) || VECTOR_STORE_PACKAGES.has(c.name)) classes.push('vector-store');
+  if (EMBEDDING_PACKAGES.has(name) || EMBEDDING_PACKAGES.has(c.name)) classes.push('embedding-provider');
+  return classes;
 }
 
 // Public: build the AI-BOM from already-scanned data.
@@ -189,10 +193,10 @@ export function buildAIBOM(scan, fileContents = {}, meta = {}) {
   const vectorStores = [];
   const embeddings = [];
   for (const c of (scan.components || [])) {
-    const cls = _classifyFramework(c);
-    if (cls === 'inference-framework') frameworks.push({ ecosystem: c.ecosystem, name: c.name, version: c.version, license: c.license || null });
-    else if (cls === 'vector-store') vectorStores.push({ ecosystem: c.ecosystem, name: c.name, version: c.version });
-    else if (cls === 'embedding-provider') embeddings.push({ ecosystem: c.ecosystem, name: c.name, version: c.version });
+    const classes = _classifyFramework(c);
+    if (classes.includes('inference-framework')) frameworks.push({ ecosystem: c.ecosystem, name: c.name, version: c.version, license: c.license || null });
+    if (classes.includes('vector-store')) vectorStores.push({ ecosystem: c.ecosystem, name: c.name, version: c.version });
+    if (classes.includes('embedding-provider')) embeddings.push({ ecosystem: c.ecosystem, name: c.name, version: c.version });
   }
   return {
     aibomFormat: 'agentic-security AI-BOM',

package/src/posture/auditor-walkthrough.js

@@ -98,11 +98,69 @@ export function loadFramework(scanRoot, id) {
  *   'absent'    — no signal / open critical findings on every mapsTo family
  *   'manual'    — control has no mapsTo (requires manual attestation)
  */
+// Stage 6 correctness audit: several compliance frameworks map controls to
+// `family:auth-missing` / `family:authz`, but no detector in this codebase
+// ever emits those literal family strings — the real missing-auth/authz
+// detectors use `broken-access-control` (generic), `fastapi-missing-auth`,
+// `springboot-missing-authz`, `laravel-missing-auth`, `quarkus-missing-authz`
+// (framework-specific). Without this alias, a control mapped to
+// auth-missing/authz read "present" (vacuously — the family bucket was
+// always empty) even with a critical, unauthenticated route open. Listed
+// under both compliance-side names since `broken-access-control` covers
+// both "nobody checked" (missing auth) and "checked wrong" (broken authz)
+// and it is strictly safer to over-count a real finding against both than
+// to keep silently excluding it from either.
+// CMP-1 (Stage 6 follow-up): grown incrementally by cross-referencing every
+// `family:` string the bundled compliance-frameworks/*.json files reference
+// against what src/sast + src/posture actually emit (a small, closed
+// problem — only the strings a control actually maps to, not a universal
+// vocabulary registry). k8s-admission.js's real rule id is
+// `k8s-pod-privileged`; `nist-csf-2.json`/`hipaa-security-rule.json` map to
+// the compliance-side spelling `k8s-pod-security-privileged`, which no
+// detector ever emitted.
+const COMPLIANCE_FAMILY_ALIAS = {
+  'auth-missing': ['broken-access-control', 'fastapi-missing-auth', 'springboot-missing-authz', 'laravel-missing-auth', 'quarkus-missing-authz'],
+  'authz': ['broken-access-control', 'idor', 'springboot-missing-authz', 'quarkus-missing-authz'],
+  'k8s-pod-security-privileged': ['k8s-pod-privileged'],
+};
+
+// CMP-1 audit trail: every `family:` string referenced by the bundled
+// compliance-frameworks/*.json files was cross-checked against real
+// detector output; entries above are the confirmed naming mismatches with a
+// real detector to alias, and `mcp-audit.js`/`sca/dep-confusion.js` were
+// fixed at the SOURCE (they now set `family` explicitly) rather than
+// aliased, since the finding constructors themselves were the root cause.
+// Four references have no matching detector at all — not a naming
+// mismatch, a genuine coverage gap that would need a new rule, out of
+// scope for an alias table: `crypto-tls-version` (crypto-protocol.js checks
+// verify-disabled, not minimum-TLS-version — the file's own header comment
+// even names the never-implemented `crypto-tls-min-version` rule id),
+// `nosql-injection` (referenced by cross-lang-meta.js's chain-detection
+// list, but no dedicated NoSQL-injection detector exists), `pii-exposure`
+// and `data-exposure` (both referenced only by consumers — threat-model-
+// auto.js's classifier and a Juliet-benchmark answer-key label
+// respectively — with no producer). Every control mapped to one of these
+// four reads `manual`/`engine-gap` rather than a false `present`, which is
+// the safe failure mode this whole mechanism exists to guarantee.
 export function evaluateFramework(scanRoot, fw, scan) {
+  // CMP-2: last-scan.json (what this is actually handed in production) carries
+  // findings across four separate channels — SAST (`findings`), secrets,
+  // business-logic, and SCA (`supplyChain`) — because report/index.js's
+  // normalizeFindings keeps them apart too. Reading only `scan.findings` made
+  // every secrets/logic/SCA finding invisible to every framework's family:
+  // mappings, e.g. a critical hardcoded secret never counted against a
+  // control mapped to family:hardcoded-secret. Each channel's family default
+  // mirrors normalizeFindings' own fallback for that channel so the two stay
+  // in agreement about what family an untagged finding belongs to.
   const findings = (scan && Array.isArray(scan.findings)) ? scan.findings : [];
+  const secrets = ((scan && Array.isArray(scan.secrets)) ? scan.secrets : [])
+    .map(s => ({ ...s, family: s.family || 'hardcoded-secret' }));
+  const logicVulns = (scan && Array.isArray(scan.logicVulns)) ? scan.logicVulns : [];
+  const supplyChain = ((scan && Array.isArray(scan.supplyChain)) ? scan.supplyChain : [])
+    .map(sc => ({ ...sc, family: sc.family || 'vulnerable-dep' }));
   const components = (scan && Array.isArray(scan.components)) ? scan.components : [];
   const families = new Map();
-  for (const f of findings) {
+  for (const f of [...findings, ...secrets, ...logicVulns, ...supplyChain]) {
     const k = f.family || 'unknown';
     if (!families.has(k)) families.set(k, []);
     families.get(k).push(f);
@@ -122,15 +180,43 @@ export function evaluateFramework(scanRoot, fw, scan) {
 
     let allCleared = true;
     let anySignal = false;
+    // Tracks whether ANY family:/module: mapping in this control actually
+    // passed (distinct from allCleared, which asks whether EVERY one did).
+    // Feeds the 'absent' vs 'partial' distinction below: a control where
+    // nothing at all checks out is a materially different auditor story
+    // ("no evidence") than one that's mostly clean with one open gap
+    // ("evidence with a gap") — both used to render as the same 'partial'.
+    let anyCleared = false;
+    // CMP-2: a rule: mapping's own observation says "verify manually" — it
+    // is deliberately not code-checked. It used to set anySignal=true and
+    // leave allCleared untouched, so a control whose ONLY mapping was rule:
+    // resolved to 'present' (fully evidenced), the same status as a control
+    // with real, checked evidence. Any rule: mapping present caps the
+    // control at 'partial' — never 'present' — regardless of what the
+    // family:/module: mappings in the same control found.
+    let hasUnverifiableMapping = false;
     for (const m of maps) {
       if (m.startsWith('family:')) {
-        const fam = m.slice('family:'.length).split(':')[0];
-        const open = (families.get(fam) || []).filter(f => !f.intentSuppressed && !f.pastDecision && (f.severity === 'critical' || f.severity === 'high'));
+        // `family:X` and the subfamily-qualified `family:X:Y` (used by
+        // owasp-llm-top-10 for LLM02/LLM06/LLM07) both used to collapse to
+        // just `X` here — the `:Y` qualifier was parsed and then silently
+        // dropped, so any finding of family X counted against every control
+        // mapped to X regardless of which subfamily the control actually
+        // named (four LLM controls flagged off one credential-in-prompt
+        // finding). Detectors that emit a subfamily always set it, so
+        // filtering is safe; findings with no subfamily set still count
+        // (recall-preserving default — same precedent as relevance.js).
+        const [fam, subfam] = m.slice('family:'.length).split(':');
+        const aliasFams = COMPLIANCE_FAMILY_ALIAS[fam] || [];
+        const candidates = [fam, ...aliasFams].flatMap(k => families.get(k) || []);
+        const scoped = subfam ? candidates.filter(f => !f.subfamily || f.subfamily === subfam) : candidates;
+        const open = scoped.filter(f => !f.intentSuppressed && !f.pastDecision && (f.severity === 'critical' || f.severity === 'high'));
         if (open.length) {
           allCleared = false;
           obs.push(`${open.length} open ${fam} finding(s) at high/critical.`);
         } else {
           obs.push(`✓ ${fam}: no open critical/high findings.`);
+          anyCleared = true;
         }
         anySignal = true;
       } else if (m.startsWith('module:')) {
@@ -174,6 +260,7 @@ export function evaluateFramework(scanRoot, fw, scan) {
         if (resolved && fs.existsSync(resolved)) {
           obs.push(`✓ ${mod}: ${label} present.`);
           anySignal = true;
+          anyCleared = true;
         } else {
           obs.push(`✗ ${mod}: expected ${label} not present.`);
           allCleared = false;
@@ -182,11 +269,23 @@ export function evaluateFramework(scanRoot, fw, scan) {
         // Could check whether a custom rule fires zero — leave a hint for now.
         obs.push(`(rule mapping) ${m} — verify manually that the bodyguard rule is enabled.`);
         anySignal = true;
+        hasUnverifiableMapping = true;
       }
     }
 
+    // 'absent' was documented (see the docstring above) as a fourth,
+    // distinct status — a control where NOTHING passed at all — but the
+    // assignment below never produced it; every non-present, non-manual
+    // control rendered as 'partial' regardless of whether it was "mostly
+    // clean with one gap" or "completely unevidenced". Only introduced for
+    // the fully-automated case (no rule: mapping) to avoid reclassifying
+    // any control that already reads 'partial' because of an inherently
+    // unverifiable rule: mapping — that precedent (rule: caps at 'partial',
+    // never reaching 'present' OR 'absent') is unchanged.
     if (!anySignal) status = 'manual';
+    else if (hasUnverifiableMapping) status = 'partial';
     else if (allCleared) status = 'present';
+    else if (!anyCleared) status = 'absent';
     else status = 'partial';
 
     results.push({ control: c, status, observations: obs });

package/src/posture/autopilot.js

@@ -110,7 +110,14 @@ export async function runAutopilot({
   for (const f of inScope) {
     const key = f.stableId || `${f.file}:${f.line}:${f.vuln}`;
     const prior = resume ? state.findings[key] : null;
-    if (prior?.outcome) { results.push(prior); continue; }
+    // A cached VERIFIED_FIXED whose patch was gated (not written) on a prior
+    // apply:false run is not a terminal outcome once the caller asks for
+    // apply:true — replaying it verbatim would silently never call
+    // applyFix on the exact two-step workflow (preview, then --apply) this
+    // gate exists to support. Everything else cached IS terminal and is
+    // replayed as before.
+    const gatedOnlyByApply = prior?.outcome === 'VERIFIED_FIXED' && prior.applied !== true && apply === true;
+    if (prior?.outcome && !gatedOnlyByApply) { results.push(prior); continue; }
 
     const rec = { key, file: f.file, line: f.line, vuln: f.vuln, severity: f.severity };