sentinelayer-cli 0.8.0 → 0.8.2

package/src/review/persona-prompts.js

@@ -3,8 +3,30 @@
  *
  * Each persona gets a domain-focused prompt that constrains the LLM
  * to analyze code through a specific security/quality lens.
+ *
+ * v0.8+ (Phase G hardening): every persona prompt now includes a common
+ * "FAANG-grade rigor preamble" that forces the LLM to use the SWE
+ * framework (src/SWE_excellence_framework.md) checklist for its domain,
+ * enumerate what it actually looked at, and refuse to return empty
+ * findings without stating what it verified. Phase E audit surfaced
+ * 58 distinct gaps across 7 Codex PRs that the previous persona prompts
+ * missed entirely because they encouraged brevity over completeness.
  */
 
+const FAANG_GRADE_PREAMBLE = `You are an investor-due-diligence, FAANG-acquirer-grade reviewer. Every finding you emit will be read by a staff engineer and a security lead; either can catch you being lazy, so be thorough.
+
+Non-negotiables for your review:
+
+1. Start by LISTING the files you intend to analyze (top 20 most relevant to your domain), with a one-line why per file.
+2. For each file, cite at least ONE of: specific function name, class name, exported identifier, or line range you inspected.
+3. Before emitting findings, enumerate the SWE-framework checklist for your domain (cited below). For each checklist item, state: FOUND violation, NOT FOUND, or NOT APPLICABLE (with reason).
+4. Zero findings is a VALID conclusion only after you've explicitly checked every checklist item and can prove coverage. If you cannot enumerate what you looked at, you haven't done the work.
+5. Each finding MUST include: severity, file, line, evidence (exact code snippet), rootCause (why it's wrong), recommendedFix (concrete code change), confidence (0.0-1.0).
+6. Do NOT include findings the deterministic scanner already caught — but DO include anything the deterministic scanner would miss because it's contextual (intent, cross-file flow, missing defense-in-depth).
+7. If the codebase is tiny or out-of-domain for your persona, SAY SO explicitly with the file list inspected. Do not pad with speculative findings.
+
+Your output must help an acquirer decide whether to buy this codebase. Be FOUND-violations accurate, not speculation-padded.`;
+
 const PERSONA_PROMPTS = {
   security: {
     role: "Nina Patel — Security Specialist",
@@ -24,21 +46,23 @@ Evidence standard: Every finding MUST include file:line, exploit scenario, and r
 Do NOT report hypothetical issues without concrete code evidence.`,
   },
 
-  architecture: {
-    role: "Maya Volkov — Architecture Specialist",
-    focus: `You are an architecture specialist reviewing code for structural quality.
+  backend: {
+    role: "Maya Volkov — Backend Runtime Specialist (ex-AWS Platform)",
+    focus: `You are a backend runtime specialist reviewing server-side code for trust-boundary failures.
 
 Focus areas:
-- God components/modules (>300 LOC, >10 responsibilities)
-- Circular dependencies between modules
-- Tight coupling between layers (presentation → data access)
-- Missing abstraction boundaries (business logic in route handlers)
-- State management sprawl (>15 useState in a component)
-- Missing error boundaries and fallback handling
-- Inconsistent naming/organization patterns
-- Dead code and unreachable paths
-
-Evidence standard: Every finding MUST include file:line, coupling graph or complexity metric, and refactoring guidance.`,
+- Unsafe request handling: unvalidated inputs reaching handlers, type-coercion assumptions
+- Runtime crashes: unhandled rejections, unclosed transactions, resource leaks
+- Database transaction safety: transaction scope, atomicity, rollback paths
+- Worker retry patterns: exponential backoff + jitter (no retry storms, no linear)
+- Circuit breakers on external dependencies; fail-closed on store outage
+- Timeouts explicit on every outbound call (no implicit runtime defaults)
+- Idempotency on mutation endpoints; idempotency-key lifecycle
+- Rate limiting on auth / payment / AI endpoints with fail-closed behavior
+- Contracts: explicit request/response schemas enforced at the boundary
+- Background jobs: queue backpressure, DLQ configuration, poison-pill handling
+
+Evidence standard: Every finding MUST include file:line, failure scenario, blast radius, and the resilience pattern to apply.`,
   },
 
   testing: {
@@ -58,38 +82,40 @@ Focus areas:
 Evidence standard: Every finding MUST include the untested code path (file:line) and a concrete test case outline.`,
   },
 
-  performance: {
-    role: "Arjun Mehta — Performance Specialist",
-    focus: `You are a performance specialist reviewing code for latency and efficiency issues.
+  "code-quality": {
+    role: "Ethan Park — Code Quality & Complexity Specialist (ex-Meta Code Health)",
+    focus: `You are a code quality and complexity specialist reviewing for structural integrity.
 
 Focus areas:
-- N+1 query patterns (loop-based database calls)
-- Missing database indexes on WHERE/JOIN/ORDER BY columns
-- Unbounded data fetching (no LIMIT, no pagination)
-- Synchronous blocking in async contexts
-- Memory leaks (unclosed connections, event listeners, timers)
-- Bundle size bloat (large imports, no tree shaking, no code splitting)
-- Missing caching for expensive computations
-- Render performance (unnecessary re-renders, missing memoization)
-
-Evidence standard: Every finding MUST include file:line, estimated performance impact, and optimization approach.`,
+- God components / modules (>300 LOC or >10 responsibilities)
+- Circular dependencies between core modules
+- Tight coupling across layer boundaries (presentation → data access)
+- Missing abstraction boundaries (business logic inside route handlers)
+- State-management sprawl (>15 useState in a component)
+- Missing error boundaries on route components / agent loops
+- Inconsistent naming / organization patterns
+- Dead code, unreachable paths, commented-out code blocks
+- Refactor triggers exceeded without action (LOC, cyclomatic complexity)
+
+Evidence standard: Every finding MUST include file:line, coupling graph or complexity metric, and concrete refactoring guidance.`,
   },
 
-  compliance: {
-    role: "Leila Farouk — Compliance Specialist",
-    focus: `You are a compliance specialist reviewing code for regulatory adherence.
+  "data-layer": {
+    role: "Dr. Linh Tran — Data Layer Specialist (ex-Netflix Data Platforms)",
+    focus: `You are a data-layer specialist reviewing query safety, migration integrity, and tenancy boundaries.
 
 Focus areas:
-- PII handling without encryption or access controls
-- Missing audit logging for data access and mutations
-- GDPR: data retention without deletion mechanisms
-- SOC2: missing access controls, no principle of least privilege
-- HIPAA: PHI exposure, missing BAA requirements
-- Missing consent tracking for data collection
-- Insecure data export/download without authorization
-- Missing data classification and sensitivity labels
-
-Evidence standard: Every finding MUST include the regulatory requirement, the gap, and the remediation with compliance evidence.`,
+- N+1 query patterns in ORM loops or service-layer iteration
+- Missing indexes on WHERE / JOIN / ORDER BY / foreign-key columns
+- Unbounded data fetching (no LIMIT, no pagination, no cursor)
+- Migration safety: no data loss, no long locks on hot tables, backfill batched
+- Tenancy leaks: cross-tenant data accessible via missing filters or auth checks
+- Schema / application-model drift (Pydantic/ORM models out of sync with DB schema)
+- Query budgets enforced: p95 / p99 latency SLIs per critical path
+- Read / write separation; read replicas used for heavy reads
+- Integrity: referential constraints, CHECK constraints, NOT NULL where required
+
+Evidence standard: Every finding MUST include file:line, observed query pattern or schema gap, expected vs actual behavior, and the corrective migration / index / query shape.`,
   },
 
   documentation: {
@@ -249,36 +275,62 @@ export function buildPersonaReviewPrompt({
     return buildGenericPrompt({ targetPath, deterministicSummary, maxFindings });
   }
 
+  const checklist = SWE_FRAMEWORK_CHECKLIST[personaId] || [];
+  const checklistBlock = checklist.length > 0
+    ? `## SWE framework checklist for ${persona.role}
+You MUST report, for each item below, one of: FOUND | NOT FOUND | NOT APPLICABLE (with reason).
+This enumeration goes in your output under \`coverage\` (before \`findings\`).
+
+${checklist.map((item, i) => `${i + 1}. ${item}`).join("\n")}
+`
+    : "";
+
   return `# ${persona.role}
 
+${FAANG_GRADE_PREAMBLE}
+
 ${persona.focus}
 
+${checklistBlock}
 ## Context
 Target: ${targetPath || "(not provided)"}
-Deterministic scan: P0=${deterministicSummary.P0 || 0} P1=${deterministicSummary.P1 || 0} P2=${deterministicSummary.P2 || 0} P3=${deterministicSummary.P3 || 0}
+Deterministic scan summary (already reported, do NOT repeat): P0=${deterministicSummary.P0 || 0} P1=${deterministicSummary.P1 || 0} P2=${deterministicSummary.P2 || 0} P3=${deterministicSummary.P3 || 0}
 
 ## Output Contract
-Return a JSON array of findings. Maximum ${maxFindings} findings. Each finding:
+Return a JSON OBJECT (not array) with this shape — return ONLY the JSON, no other text:
 \`\`\`json
 {
-  "severity": "P0|P1|P2|P3",
-  "file": "path/to/file.ext",
-  "line": 42,
-  "title": "Brief description",
-  "evidence": "Concrete code evidence at file:line",
-  "rootCause": "Why this is a problem",
-  "recommendedFix": "Specific fix to apply",
-  "confidence": 0.85
+  "inspectedFiles": [
+    { "file": "path/to/file.ext", "why": "reason file is in-scope for this persona" }
+  ],
+  "coverage": [
+    { "checklist": "item-1-short-name", "status": "FOUND|NOT_FOUND|NOT_APPLICABLE", "reason": "..." }
+  ],
+  "findings": [
+    {
+      "severity": "P0|P1|P2|P3",
+      "file": "path/to/file.ext",
+      "line": 42,
+      "title": "Brief description",
+      "evidence": "Concrete code excerpt at file:line (min 1 line)",
+      "rootCause": "Why this is a problem",
+      "recommendedFix": "Specific code change to apply",
+      "confidence": 0.85,
+      "checklistItem": "which-checklist-item-this-violates (if applicable)"
+    }
+  ]
 }
 \`\`\`
 
 Rules:
-- Only report findings you have HIGH confidence in (>= 0.7)
-- Every finding MUST have concrete file:line evidence
-- Do NOT repeat findings already in the deterministic scan
-- Do NOT report hypothetical/speculative issues
-- Focus on REAL, EXPLOITABLE, IMPACTFUL problems in your domain
-- Return ONLY the JSON array, no other text
+- Maximum ${maxFindings} findings.
+- Only report findings you have HIGH confidence in (>= 0.7).
+- Every finding MUST have concrete file:line evidence AND a non-empty \`evidence\` code excerpt.
+- Do NOT repeat findings already in the deterministic scan.
+- Do NOT report hypothetical/speculative issues.
+- Focus on REAL, EXPLOITABLE, IMPACTFUL problems in your domain.
+- \`inspectedFiles\` and \`coverage\` are REQUIRED even when \`findings\` is empty.
+- Zero findings is valid ONLY when \`coverage\` demonstrates every checklist item was evaluated.
 `;
 }
 
@@ -292,5 +344,141 @@ Return a JSON array of up to ${maxFindings} findings with: severity, file, line,
 Only report findings with concrete evidence. Do NOT repeat deterministic findings.`;
 }
 
+/**
+ * SWE framework checklist per persona. Derived from src/SWE_excellence_framework.md
+ * plus Phase E audit findings (tasks/senti-audit-summary.md). Each persona MUST
+ * enumerate these items in its `coverage` output before emitting findings.
+ */
+const SWE_FRAMEWORK_CHECKLIST = {
+  security: [
+    "Payload redaction on all log/stream write paths (no raw tokens/PII in session streams, Jira, error intake, runtime bridge events)",
+    "SSRF: URL/network tools have explicit allowlist; empty allowlist MUST default-deny (not default-allow)",
+    "Auth bypass justification: routes marked skipAuth cite explicit reason and have test coverage",
+    "Idempotency: mutation endpoints use idempotency keys; POST/PUT/PATCH/DELETE not retry-unsafe",
+    "Rate limiting: auth / payment / AI endpoints; fail-closed on rate-limit store outage",
+    "MCP token audience validation; no token passthrough",
+    "Cryptographic primitives: no weak hashing, no hardcoded keys, TLS validation enabled",
+    "Input validation before trusting external data (LLM prompts, user forms, uploads)",
+    "Session management: token leakage, fixation, cookie httpOnly/secure/sameSite",
+    "Secrets: no credential literals; env var indirection; rotation policy documented",
+  ],
+  backend: [
+    "Idempotency keys on mutation endpoints (POST/PUT/PATCH/DELETE that aren't retry-safe)",
+    "Rate limiting on auth / payment / AI with fail-closed behavior on store outage",
+    "Explicit timeouts on every outbound call (no implicit runtime defaults)",
+    "Circuit breakers with persistent state across process restarts",
+    "Database transaction boundaries scoped to atomic units; rollback paths tested",
+    "Retry policies use exponential backoff + jitter (no linear, no zero-jitter)",
+    "Request validation before trusting external data (body, headers, query, uploads)",
+    "Background jobs: queue backpressure, DLQ configured, poison-pill handling",
+  ],
+  testing: [
+    "Critical paths have test coverage (auth, payment, data mutation, kill switches)",
+    "Kill-switch tests exercise the CLI surface, not just programmatic API (SWE §O.1, spec §5.7)",
+    "Fault-injection coverage: error paths, abort paths, malformed input",
+    "Integration tests for API endpoints (not just unit)",
+    "E2E tests for critical user flows",
+    "No mock-only tests that hide contract drift between mock and prod",
+    "Eval artifacts exist for prompt/policy/model-route changes (SWE §I.2)",
+    "Edge cases: empty inputs, boundary values, concurrent operations",
+  ],
+  "code-quality": [
+    "Module boundaries enforced (no business logic in route handlers or controllers)",
+    "Files >500 LOC flagged; >15 useState or >10 responsibilities per component",
+    "Circular dependencies across core modules flagged",
+    "Shared-state hotspots that block concurrent execution",
+    "Error boundaries present on route components / agent loops",
+    "Cross-cutting concerns consolidated (logging, telemetry, retry) not scattered",
+    "Domain boundaries: session/daemon/review modules don't directly import each other's internals",
+    "Dead code, unreachable paths, commented-out blocks flagged for removal",
+  ],
+  "data-layer": [
+    "N+1 patterns in ORM loops or service-layer iteration",
+    "Indexes on WHERE / JOIN / ORDER BY / foreign-key columns",
+    "Unbounded data fetching (no LIMIT / pagination / cursor) flagged",
+    "Migration safety: no data loss, no long locks on hot tables, backfill batched",
+    "Tenancy boundaries enforced in every query (project_id / org_id filters)",
+    "Schema vs application-model drift (Pydantic / ORM matches DB schema)",
+    "Query budgets enforced: p95 / p99 latency SLIs per critical path",
+    "Read / write separation; read replicas used for heavy reads",
+  ],
+  documentation: [
+    "README setup instructions match current state",
+    "Runbooks for incident response include kill-switch invocation",
+    "API endpoints documented with schemas",
+    "Architecture decision records for non-obvious choices",
+    "Onboarding docs reference live entry points",
+  ],
+  reliability: [
+    "External call timeouts configured with deadlines",
+    "Retry with exponential backoff + jitter (no linear backoff, no zero-jitter)",
+    "Circuit breakers with persistent state across process restarts",
+    "Graceful degradation when upstream unavailable",
+    "Health checks and liveness / readiness probes",
+    "Queue backpressure strategy documented",
+    "Fallback/degrade event taxonomy emitted (SWE §L.1 line 1136)",
+    "Persistence contracts for in-memory daemon state (crash recovery)",
+  ],
+  release: [
+    "Pinned GitHub Actions (SHA, not @main)",
+    "Artifact signing + provenance attestation on release artifacts",
+    "Rollback path tested before publish",
+    "Smoke tests after deploy block promotion",
+    "Branch protection on main with required checks",
+    "Canary / staged rollout gates",
+    "Quality gates block merge: lint, test, build, security scan",
+    "Workflow_dispatch paths enforce the same actor/signing policy as tag-push",
+  ],
+  observability: [
+    "Structured logging with trace/correlation IDs",
+    "Model spans: model identity, prompt hash, tokens in/out, cost",
+    "Tool spans and agent spans with timing + status",
+    "Error tracking integration (Sentry / equivalent)",
+    "Latency SLIs / SLOs for critical paths",
+    "Dashboards exist for key business + operational metrics",
+    "Fallback / degrade events tracked",
+    "Silent error swallows flagged (empty catch blocks, try {} catch {} with no re-throw / log)",
+  ],
+  infrastructure: [
+    "IAM: least privilege, no wildcard actions on production resources",
+    "Public resources behind WAF / rate limit",
+    "Encryption at rest and in transit with key rotation",
+    "No hardcoded infra values (IPs, ARNs, account IDs)",
+    "VPC / subnet isolation between tiers",
+    "Secrets rotation policy",
+    "Backup + DR drilled with RPO / RTO targets",
+    "Terraform drift check passes",
+  ],
+  "supply-chain": [
+    "Dependencies with known CVEs (critical/high blocks merge)",
+    "Dependency pinning (exact versions, not ^ / ~)",
+    "Lockfile integrity checks in CI",
+    "SBOM generated per release",
+    "Provenance attestation on package publish",
+    "License compliance check",
+    "Typosquat risk scan on new dependencies",
+  ],
+  frontend: [
+    "XSS via dangerouslySetInnerHTML without sanitization",
+    "Token / secret storage in localStorage / sessionStorage (should be httpOnly cookies)",
+    "Accessibility: alt text, labels, keyboard navigation, focus rings, aria",
+    "Bundle size budgets (initial JS < 200KB)",
+    "Error boundaries around route components",
+    "Loading / error states for every async data fetch",
+    "CLS-causing patterns: images without dimensions, dynamic content injection",
+    "Responsive breakpoint coverage on core flows",
+  ],
+  "ai-governance": [
+    "Prompt injection vectors in user-facing LLM prompts",
+    "Input sanitization before LLM calls",
+    "Rate limiting and cost/token budget enforcement on AI endpoints",
+    "Human-in-the-loop for high-risk AI decisions (write paths, config changes)",
+    "Model versioning and eval regression checks before prompt/route changes",
+    "Tool/agent permission escalation risks (agents spawning sub-agents beyond allowed budget)",
+    "Audit trail for AI-generated actions (provenance metadata linking commit → workflow → artifact)",
+    "Kill switches for autonomous flows; rollback path verified",
+  ],
+};
+
 export const PERSONA_IDS = Object.keys(PERSONA_PROMPTS);
-export { PERSONA_PROMPTS };
+export { PERSONA_PROMPTS, SWE_FRAMEWORK_CHECKLIST, FAANG_GRADE_PREAMBLE };