AEGIS Roadmap

Version: 2.6.0 Updated: 2026-04-12 Status: Active Cross-References: README.md, CLAUDE.md

This document is the single source of truth for AEGIS future work, including active PRs, open issues, and release milestones.

Next Steps (Ordered Checklist)

Work through these in order. Check off each item as completed. Source: Discovery Analysis 2026-02-08.

Immediate (no blockers, can start now)

• 1. Fix dependency misclassification — scipy and prometheus_client moved from dev to dedicated engine and telemetry optional groups with try/except import guards and clear ImportError messages at point of use. Files: pyproject.toml, src/engine/utility.py, src/telemetry/prometheus_exporter.py, README.md, tests/test_optional_deps.py
• 2. Docs version sync — repository-structure.md:28 and comprehensive-todo-discovery.md synced to CLAUDE.md v4.2.3, metrics 1475/94.21%. Commit: d84330a
• 3. Update dependency versions — safety>=2.3.0 → >=3.0.0; python-bitcoinlib already pinned at >=0.12.0 (no change needed). All quality gates pass. File: pyproject.toml, Commit: e2ab0a5
• 4. Document broad exception catches — Added # Intentional: <reason> comments to 15 except Exception sites across 8 files per CLAUDE.md §3. Commit: fc0b41c

Short-term (v1.1.0)

• 5. Boundary tests for all gates — 77 parametrized BVA tests verifying comparison operators at exact thresholds for all 6 gates + drift detector. File: tests/test_gate_boundaries.py
• 6. GOVERNANCE actor type — Override workflow orchestration: Governance actor class with override lifecycle (initiate/sign/approve/reject/expire), compliance checking (complexity gate non-overridable), emergency halt; ultrathink-hardened (halt guards, fail-closed compliance, thread safety). Files: src/actors/governance.py, src/actors/base.py, tests/test_governance_actor.py (41 tests)
• 7. CALIBRATOR actor type — Statistical threshold tuning: Calibrator actor class with drift recalibration (delegates to DriftMonitor.calibrate_thresholds()), Bayesian prior update (delegates to BayesianPosterior.update_prior()), gate parameter proposals with recognized parameter whitelist, approval-gated application workflow, telemetry emission; thread-safe with threading.Lock; ultrathink-hardened (U-1..U-5). Files: src/actors/calibrator.py, src/actors/base.py, tests/test_calibrator_actor.py (69 tests incl. 12 regression)
• 8. Extract shared serialization pattern — ensure_utc() extracted from 3 workflow files to src/workflows/serialization.py. Files: serialization.py, consensus.py, override.py, proposal.py
• 9. Extract shared parameter validation — 4 validators (validate_positive, validate_range, validate_normalized, validate_threshold_ordering) extracted to src/engine/validation.py, replacing ~24 inline checks across 5 engine modules. Files: validation.py, bayesian.py, drift.py, gates.py, utility.py, complexity.py Deferred: consolidating inline timezone checks in persistence/telemetry (different module boundary).

Medium-term (v1.2.0)

• 10. Production deployment guide — docs/deployment/production-guide.md with Docker, K8s, AWS examples, Docker Compose, HSM integration, multi-region DR, observability setup, production checklist. Files: docs/deployment/production-guide.md, Dockerfile, docker-compose.yaml, monitoring/prometheus/prometheus.yml
• 11. Migration guide — Parameter recalibration via Calibrator actor, workflow state migration, schema upgrade path, version compatibility matrix. File: docs/deployment/migration-guide.md
• 12. Performance SLAs — Latency targets (p50 < 100ms, p95 < 500ms, p99 < 1s), throughput baselines, component latency budget, recorded benchmark results. File: docs/deployment/performance-slas.md
• 13. Shadow mode for KL divergence calibration — shadow_mode=True on pcw_decide() with ShadowResult dataclass, drift monitor integration, SHADOW_EVALUATION telemetry, Prometheus mode label, CLI --shadow flag, MCP shadow_mode param. 44 new tests. Unblocks issue #1. Files: src/integration/pcw_decide.py, src/telemetry/emitter.py, src/telemetry/prometheus_exporter.py, src/cli.py, src/aegis_governance/mcp_server.py, src/aegis_governance/__init__.py, monitoring/prometheus/alerting-rules.yaml, monitoring/prometheus/recording-rules.yaml, tests/integration/test_shadow_mode.py
• 14. HTTP telemetry sink — HTTPEventSink (per-event POST), BatchHTTPSink (batching + retry), http_sink() factory; stdlib-only (urllib.request); AegisConfig.telemetry_url; CLI --telemetry-url; MCP telemetry_url param; SDK re-exports. 45 new tests. Files: src/telemetry/emitter.py, src/config.py, src/cli.py, src/aegis_governance/mcp_server.py, src/aegis_governance/__init__.py, src/telemetry/__init__.py, tests/telemetry/test_http_sink.py
• 15. Drift detection → policy connection — DriftMonitor wired into production pcw_decide() path: CRITICAL → HALT (non-overridable), WARNING → advisory constraint, NORMAL → no change; _evaluate_drift_policy() + _apply_drift_overrides() helpers; DRIFT_POLICY_ENFORCED telemetry; create_drift_monitor() factory; CLI --drift-baseline; MCP drift_baseline_data; DriftAction/DriftResult re-exports; null-value filtering; drift-specific next_steps. 39 new tests. Files: src/integration/pcw_decide.py, src/config.py, src/cli.py, src/aegis_governance/mcp_server.py, src/aegis_governance/__init__.py, src/engine/__init__.py, src/telemetry/emitter.py, tests/integration/test_drift_enforcement.py, tests/integration/test_drift_regression.py

Long-term (v2.0.0) — AWS Infrastructure DEPLOYED

• 16. Agent integration guide & parameter cookbook — (a) docs/integration/parameter-reference.md — comprehensive parameter reference with derivation guidance, domain examples, boundary behavior for all inputs; (b) docs/integration/domain-templates.md — 4 worked examples (trading, CI/CD, content moderation, autonomous agent) with parameter mappings, JSON inputs, gate-by-gate walkthroughs; (c) MCP tool descriptions enriched with semantic context, minimum/maximum JSON Schema constraints, instructions field in initialize response. Files: docs/integration/parameter-reference.md, docs/integration/domain-templates.md, src/aegis_governance/mcp_server.py
• 17. GAP-L1 Deployment — Grafana deployment, production Prometheus, alert routing (Slack/PagerDuty). Issue #9. DEPLOYED: AegisMonitoringStack-dev — CloudWatch dashboard AEGIS-Governance-dev, SNS topic aegis-governance-alarms-dev, 4 alarms (Lambda errors, Lambda throttles, ECS unhealthy, billing). Grafana/Prometheus observability via ADOT sidecar on ECS Fargate.
• 18. GAP-L2 — OpenTelemetry distributed tracing. DEPLOYED: ADOT sidecar running on ECS Fargate (AegisMcpStack-dev), configured for Prometheus remote write to AMP. Full OTLP span correlation deferred to production workload phase.
• 19. Issues #2, #5, #7, #8, #9 Phase 2 — Infrastructure requirements met. DEPLOYED: DynamoDB aegis-governance-state-dev, S3 aegis-governance-audit-dev-164171672016, Secrets Manager aegis/signing-keys-dev, KMS encryption key, IAM auth on API Gateway, SNS alarm topic. Remaining: Slack/email subscription on SNS, multi-region replication, Locust load testing against live API Gateway endpoint.
• 20. Red-team fuzzing — Phase 2 adversarial testing. DEPLOYED: Cloud Run aegis-api + aegis-mcp at https://aegis-api-980022636831.us-central1.run.app/ available as live targets (migrated from AWS Lambda+ECS, 2026-03-28). Fuzzing execution pending.
• 20a. MCP hardening (CoSAI/Red Hat) — Per research/003: ✅ (a) MCP audit logging — structured log for every tool invocation; ✅ (b) MCP rate limiting — token bucket/sliding window on mcp_server.py; ✅ (c) TLS enforcement — _validate_sink_url() enforces HTTPS on HTTPEventSink/BatchHTTPSink (with allow_insecure escape hatch for local dev), MCP _ALLOWED_TELEMETRY_SCHEMES restricted to {"https"}; ✅ (d) CoSAI MCP-T cross-reference in CLAUDE.md §11.4.1 (MCP-T1..T12 → AEGIS control matrix); ✅ (e) MCP tool schema signing — AMTSS Protocol v1 (src/crypto/schema_signer.py, ToolSchemaSigner, Ed25519, RFC 8785, _meta inline delivery, capabilities.experimental keyset); research: docs/research/004-mcp-schema-signing-design.md. All 5 sub-items complete.

Infrastructure (ECS MCP)

• 23. MCP Streamable HTTP transport — Re-enabled (2026-03-13) via FastMCP argparse --transport streamable-http. Public API Gateway (HTTP API) + VPC Link with Unkey Lambda authorizer added to ECS stack. Health check endpoint via @mcp.custom_route("/health"). Files: src/aegis_governance/mcp_server.py, infra/stacks/ecs_stack.py, infra/app.py

Business & Legal (pre-public-release)

• 21. IP & licensing review — License changed from MIT to BSL 1.1 (2026-03-13). Change Date 2030-03-13, Change License Apache 2.0. Production use requires commercial subscription at portal.undercurrentholdings.com. Remaining: (a) patentability evaluation; (b) trademark feasibility for "AEGIS Governance"; (d) prior art landscape assessment. Files: LICENSE, NOTICE, pyproject.toml
• 22. Commercialization strategy — Market research complete: AI governance market $300-850M (2025), 35-40% CAGR to $1.5B-$4.8B by 2033. SaaS + BSL model (BSL 1.1 licensed engine, distributed via hosted API, MCP server, and self-service portal). Pricing tiers defined: Community (free), Professional ($3,500/mo), Enterprise ($18,000/mo), Financial Services ($50,000/mo). Files: src/aegis_governance/tiers.py, docs/research/006-pricing-unit-economics.md, docs/research/007-gtm-action-plan-Q2-2026.md
• 22a. Phase 1: Customer visibility — Customer records in DynamoDB, usage metering with atomic counters, admin CLI (aegis admin), Lambda integration with response headers, provisioning script. Files: src/aegis_governance/customer.py, src/cli.py, src/lambda_handler.py, scripts/provision-customer.py. ADR: ADR-008. Guide: customer-management.md.
• 22b. Phase 2d: Self-service portal — Customer portal (Next.js 16 + Clerk auth), usage dashboards (quota gauge, daily chart, channel breakdown), API key management (Clerk-stored keys validated against AEGIS API), account/profile management, tier upgrade requests. Directory: aegis-portal/. Live at portal.undercurrentholdings.com.
• 22c. Phase 3: Monetization — Stripe Billing + Meters, Checkout Sessions, subscription lifecycle webhook, tier enforcement on upgrade/downgrade. Live-mode products and keys configured (2026-03-20). Files: aegis-portal/lib/actions.ts, src/aegis_governance/stripe_webhook.py. ADR: ADR-008.

Discovery Analysis 2026-03-22

Comprehensive 8-agent codebase discovery. Source: 2026-03-22 analysis session (8 parallel agents: TODO scanner, ROADMAP audit, dependency analysis, code quality, test coverage, portal audit, git activity, config review). Project health: excellent — 0 bugs, 4124 tests, ~92% coverage, 12/12 CI workflows green, Grade A code quality.

Immediate (no blockers, can start now)

• D1. Add annual Stripe price IDs to .env.example — STRIPE_PRICE_ID_PROFESSIONAL_ANNUAL, STRIPE_PRICE_ID_ENTERPRISE_ANNUAL, STRIPE_PRICE_ID_FINANCIAL_SERVICES_ANNUAL. Annual billing checkout (actions.ts:309) fails without these; products already created in Stripe Dashboard (CLAUDE.md §9, v4.6.82). Files: aegis-portal/.env.example, Vercel env vars, Secrets Manager
• D2. Commit .afa.yaml to version control — Project-level AFA agent config (analogous to .github/dependabot.yml). Defines model (Gemini 3 Flash), analysis scope (Python only, excludes tests/docs/spec), guardrails (3 max candidates, auto-commit off, afa/ branch prefix), and budget cap ($50/mo). Committed as shared project policy.
• D3. Update eslint-config-next — 16.1.6 → 16.2.0 to match Next.js framework version. File: aegis-portal/package.json

Short-term (this week)

• D4. Merge or close feat/dogfood-shadow-ci — Branch was fully merged into main (ancestor check confirmed). All deliverables on main: tools/ci/aegis_shadow_evaluate.py, docs/integration/dogfooding-guide.md, shadow CI workflow step, backfill script. Local branch deleted 2026-04-04. Completed: 2026-04-04
• D5. Clarify REST API Lambda split in docs — docs/api/rest.md conflates evaluation endpoints (Lambda 1: src/lambda_handler.py — /evaluate, /risk-check, /health) with customer endpoints (Lambda 2: src/aegis_governance/customer_api.py — /customer/*). Add section headers distinguishing the two Lambda functions. File: docs/api/rest.md
• D6. Extract decision serialization — _safe_round() + _decision_to_dict() duplicated across src/cli.py, src/lambda_handler.py, src/aegis_governance/mcp_server.py. Extract to shared src/aegis_governance/result_formatter.py to prevent divergence. Code quality audit rated this MODERATE priority. Effort: 1-2 hours
• D7. Clean up stale branches — 5 fix branches from March 14-16 (fix/qg75-deploy-test-mlkem-mock, fix/qg76-docs-sync-followup, fix/qg76-crypto-hardening-gap-c1-c2, fix/docs-sync-bsl-terminology, fix/rest-api-docs-alignment) + 4 Dependabot branches. Clarify feat/portal-brand-redesign (remote-only) — delete if superseded by portal work already on main.

Medium-term (this sprint)

• D8. Add AFA portal E2E tests — Create aegis-portal/e2e/specs/afa-product.spec.ts covering /afa/dashboard, /afa/keys, /afa/account pages. Mirrors main portal E2E structure (11 existing specs). Effort: 2-3 hours
• D9. Verify actions/cache version — Already at latest (v5.0.4, no v6 exists). No change needed.
• D10. Set up SNS subscriptions for alarm routing — Slack/email on aegis-governance-alarms-dev SNS topic. Completes Issue #5 Phase 2.
• D11. Execute Locust load tests against live API Gateway — tests/load/locustfile.py with direct Lambda invocation, 5 proposal templates, SLA validation (p50<100ms, p95<500ms, p99<1s). Initial results: /evaluate p50=470ms cross-region (in-region est. ~70-100ms), 0% error rate. Completes Issue #2 Phase 2.

Long-term (backlog)

• D12. Collect 30+ days KL divergence data for threshold calibration. Issue #1. Shadow evaluation active on 4 repos (aegis-governance, libertas-cicd, AgoraIV, afa) via aegis-shadow.yml CI workflow (deployed 2026-03-24). CalibrationTriggerEvaluator wired into pcw_decide() for automated trigger detection. Clock started 2026-03-24.
• D13. Provision Grafana dashboards — CloudWatch deployed (AEGIS-Governance-dev); Grafana provisioning pending. Issue #9 Phase 2.
• D14. Multi-region DR setup — Cross-region DynamoDB replication for disaster recovery. Issue #8 Phase 2.
• D15. Plan for btclib <2025.0.0 constraint — Upper bound in pyproject.toml will expire; monitor btclib releases for v2025+ compatibility. File: pyproject.toml
• D16. Shadow → enforcement transition — Once KL data collected (D12), transition AEGIS from shadow to enforcement mode for the first time. Blocked by D12.

Stripe Integration Review (2026-04-01)

Deep review of portal.veiltide.com → portal.undercurrentholdings.com Stripe integration. Context7-verified against authoritative Stripe API docs. PR #61.

Completed

• S1. Webhook signature bypass on ImportError (CRITICAL) — api_server.py except ImportError block silently accepted unverified webhook events. Changed to return 500 (fail-closed). 2 AST regression tests. Commit: f180b67
• S2. Checkout session metadata mismatch (HIGH) — Portal set subscription_data[metadata] but _handle_checkout_completed() reads session.metadata — different Stripe objects (Context7-confirmed). Added session-level metadata to both AEGIS and AFA checkout actions. 2 regression tests. Commit: 8a03e67
• S3. Deploy workflow E2E exclusion — aegis-deploy.yml test step included E2E tests requiring h2 and aegis SDK not installed in deploy runner. Added --ignore=tests/e2e. Commit: 8884a58

Short-term

• S4. Webhook secret dual-path inconsistency (MEDIUM) — Removed redundant STRIPE_WEBHOOK_SECRET env var pre-check in api_server.py. _verify_stripe_signature() now handles both env var and Secrets Manager ARN paths. Returns 401 when no secret configured (fail-closed).
• S5. Billing Meters V2 migration (HIGH) — New src/aegis_governance/stripe_meter.py — StripeMeterV2 with session-based auth (/v2/billing/meter-event-stream), 14-min proactive refresh, thread-safe caching, 401 retry, idempotency keys. Replaces V1 inline code. Lambda verified inactive via CloudWatch (0 real API traffic). 16 new tests.
• S6. Pre-existing E2E failure: test_b01_provision_customer (LOW) — Test now accepts both 200 and 201 (server returns 200 for existing customers with force_reprovision). Fixed in Stripe review.
• S7. Remove unused stripe npm + Python packages (LOW) — Removed stripe from both aegis-portal/package.json (v18.0.0, unused) and pyproject.toml (>=12.0.0, unused). All Stripe calls use raw HTTP. Fixed in Stripe review.

Long-term

• S8. Add stripe_customer_id to Customer model (LOW) — Currently lives only on Subscription record. Meter reporting does extra DB read via get_subscription() fallback. Adding to Customer would eliminate the extra lookup. Files: src/aegis_governance/customer.py, src/aegis_governance/customer_pg.py, src/aegis_governance/stripe_webhook.py

Compliance Remediation (Audit 2026-03-22)

Five-framework compliance audit (NIST AI RMF, EU AI Act, ISO 42001, SOC 2, FedRAMP). AI-specific frameworks fully compliant. Operational security frameworks have documentation complete but exercises deferred. Legal soundness audit found 6 critical + 4 moderate outward-facing language issues. Tasks below close all gaps in dependency order.

Audit results: NIST AI RMF OPERATIONAL (3 STRONG, 1 MODERATE) | EU AI Act COMPLIANT (15/15) | ISO 42001 COMPLIANT (7/7) | SOC 2 PARTIAL (38/42 implemented, 29/42 tested) | FedRAMP PARTIAL (30/31 implemented, exercises deferred)

Immediate (no blockers)

• C1. Add KL divergence thresholds to interface-contract.yaml — tau_warn: 0.3, tau_crit: 0.5 with calibration basis documented in docs/integration/parameter-reference.md. Closes NIST MEASURE gap (threshold calibration). Files: schema/interface-contract.yaml, docs/integration/parameter-reference.md
• C2. Create NIST AI RMF full mapping — MEASURE only Expanded to all 72 subcategories across all 4 functions (GOVERN, MAP, MEASURE, MANAGE) via adversarial audit (2026-04-03). 45/72 IMPLEMENTED, 17 PARTIAL, 6 GAP, 4 N/A. 8 code-documentation discrepancies identified (see N1-N8 below). Files: docs/compliance/nist-ai-rmf-{govern,map,manage}-mapping.md, docs/compliance/nist-ai-rmf-coverage-summary.md, docs/compliance/nist-ai-rmf-audit-2026-04.md, docs/compliance/DECOMMISSIONING.md

Short-term (Q2 2026)

• C3. Execute first quarterly access review — Follow docs/compliance/ACCESS-REVIEW.md procedures, capture baseline metrics, document findings. Closes SOC 2 CC4.2 + FedRAMP AC-2. Effort: 4 hours
• C4. Execute BCP/DRP backup restoration test — Follow docs/compliance/BCP-DRP.md §4-5 (updated for GCP Cloud Run primary + AWS legacy fallback), verify RTO < 4h and RPO < 1h. Closes SOC 2 A1.2 + FedRAMP CP-9/10. Effort: 8 hours
• C5. Execute IRP tabletop exercise — Scenario: governance invariant violation. Follow docs/compliance/IRP.md §5. Closes SOC 2 CC8.1 + FedRAMP IR-4/5. Effort: 4 hours
• C6. Add decision accuracy metrics to telemetry — Gate precision/recall tracking: gate_accuracy_matrix in telemetry. Closes NIST MEASURE accuracy gap. Effort: 8-10 hours
• C7. Retain external legal counsel — Required for GDPR breach notification, regulatory interpretation. Open item in docs/compliance/IRP.md §3. Effort: External

Medium-term (Q3 2026)

• C8. External TEVV audit — Third-party validation of ai/technical_file/testing/tev_report.md. Advances NIST AI RMF OPERATIONAL → VALIDATED. Effort: External
• C9. SOC 2 Type 2 audit preparation — 3-month evidence package from C3-C5. Engage auditor for 12-month observation (Q4 2026 → Q4 2027). Requires: C3, C4, C5, C18 (segregation of duties) complete. Effort: External
• C10. Document break-glass emergency access procedures — Gap in ACCESS-REVIEW.md §5. Root account rotation + emergency runbook. Closes FedRAMP IA-5 + IR-1. Effort: 2-3 hours
• C11. Pre-draft DSR notification templates — Access, deletion, objection templates for DSR-PRIVACY-OPERATIONS.md. Closes SOC 2 P4.1. Effort: 3-4 hours

Legal remediation (pre-customer deployment)

• C12. Compliance language alignment — "compliance-ready" → "compliance-aligned" across all 3 repos. Partially applied v1.99.0; full sweep completed 2026-04-04 (CR1). Closes legal findings L1-L5. Completed: 2026-04-04
• C13. Add AI agent disclaimer to llms.txt — Prevent downstream compliance misrepresentation: "These artifacts do not constitute certification." Closes L5.
• C14. Redact single-operator IRP dependency from public docs — Move docs/compliance/ai-governance.md operational limitations to internal documentation. Closes L6.
• C15. Add SLA + licensing disclaimers to customer docs — SLA disclaimer on onboarding health check, BSL 1.1 production license callout. Closes L7, L10.
• C16. Soften README marketing claims — Replace "giving enterprises confidence" with "evaluate proposals through quantitative gates"; use ranges ("4000+ tests"); qualify "tamper-evident". Closes L1, L8.
• C17. Add date stamps to rate limit tables — "Current as of YYYY-MM-DD. Subject to change." Closes L9.

Long-term (pre-customer deployment)

• C18. Team expansion for segregation of duties — Single-operator across all compliance roles. SOC 2/FedRAMP require multi-person approval gates for customer-facing SaaS. Effort: Organizational

NIST AI RMF Audit Code Remediation (2026-04-03)

Adversarial code-documentation reconciliation identified 8 discrepancies where documentation claims exceed code reality. Source: docs/compliance/nist-ai-rmf-audit-2026-04.md §3.

CRITICAL (before customer deployment)

• N1. Persist audit chain to database — Decision audit chain now persisted to Neon PostgreSQL via AuditChainStore with per-customer partitioning. SELECT FOR UPDATE serializes concurrent appends across Cloud Run instances (0-4). Chain entries record instance_id (from K_REVISION env). Three admin endpoints: /admin/audit-chain/verify/{customer_id}, /admin/audit-chain/entries/{customer_id}, /admin/audit-chain/status/{customer_id}. In-memory fallback preserved for sandbox evaluations and DB unavailability. Genesis race condition handled via IntegrityError catch + retry. 32 regression tests. Files: src/aegis_governance/audit_chain_store.py (new), src/aegis_governance/models_pg.py, src/api_server.py, tests/test_audit_chain_store.py (new) Completed: 2026-04-04
• N2. Add override replay protection — DualSignatureValidator.create_message_hash() now binds to requested_at timestamp (prevents cross-request replay). SignatureRecord gains nonce field (128-bit, secrets.token_hex). OverrideWorkflow tracks _used_nonces set. Clock-skew freshness check rejects future-dated requests (>5 min tolerance). SignatureProvider protocol unchanged. Backward-compatible serialization (legacy records default nonce=""). 12 regression tests. Files: src/workflows/override.py, tests/test_override_coverage.py Completed: 2026-04-04

HIGH (Q2 2026)

• N3. Fix RBAC fail-open in override handler — override_handler.py L143 changed from is not None and not to is None or not (fail-closed). When rbac_enforcer=None, override requests are now REJECTED instead of silently bypassing authorization. 3 regression tests + 4 existing tests updated to use permissive mock enforcer. Files: src/integration/override_handler.py, tests/test_override_handler.py Completed: 2026-04-04
• N4. Closed as by-design — All three transports (CLI L348, Lambda L894, MCP L165) intentionally zero out risk in utility synthesis (_pert(0.0)) because the risk gate evaluates risk independently via GateEvaluator. Including risk in utility via kappa*delta_R would double-count it AND penalize risk reduction. MCP DOES accept risk_baseline/risk_proposed and they flow correctly to the risk gate through build_pcw_context(). Comment expanded in mcp_server.py. Closed: 2026-04-04

MEDIUM (Q2-Q3 2026)

• N5. Fix TelemetryEmitter thread safety — Added threading.Lock to TelemetryEmitter: emit() reads/writes shared state under lock, sinks dispatched outside lock (may block on I/O), add_sink/remove_sink synchronized, sinks list snapshot before iteration. Docstring updated from "NOT thread-safe" to "Thread-safe". 3 regression tests (concurrent emit, add-during-emit, AST docstring check). Files: src/telemetry/emitter.py, tests/telemetry/test_emitter.py Completed: 2026-04-04
• N6. Distribute sandbox rate limiting — Sandbox rate limit is in-memory per Cloud Run instance, resets on restart. Under auto-scaling, users can exceed 10/day by hitting different instances. Fix: Migrate limiter state to Redis (Upstash) or Neon. Files: src/aegis_governance/sandbox.py Effort: 1 day

LOW (Q3 2026)

• N7. Document PERT variance in interface contract — src/engine/utility.py L73-92 uses simplified PERT variance (±22-40% error vs true Beta-PERT). Docstring warning exists but schema/interface-contract.yaml does not flag the approximation. Fix: Add pert_variance_note to interface contract. Consider true Beta-PERT when scipy available. Files: schema/interface-contract.yaml, src/engine/utility.py Effort: 0.5 day
• N8. Add per-gate drift thresholds — src/engine/drift.py uses global tau_warning/tau_critical for all gates. Gates with different variance profiles (novelty vs quality) would benefit from per-gate thresholds. Fix: Add per-gate override to KLDriftConfig. Files: src/engine/drift.py, src/config.py Effort: 1 day

GCP Infrastructure Security (2026-04-03)

Live infrastructure deep dive (GCP + AWS) identified critical security gaps. Source: GCP gcloud audit + AWS aws CLI verification on 2026-04-03.

CRITICAL (immediate)

• G1. Migrate secrets to GCP Secret Manager — 8 secrets (Unkey, Stripe, AWS IAM, Neon DATABASE_URL, service key) migrated from plaintext --set-env-vars to GCP Secret Manager --set-secrets on Cloud Run. Setup script: scripts/gcp-secrets-setup.sh. IAM granted to runtime SA + service agent. Legacy deploy-cloud-run.yml deleted. Files: .github/workflows/aegis-deploy.yml, scripts/gcp-secrets-setup.sh Completed: 2026-04-03
• G2. Restrict WIF scope to specific repos — WIF attribute condition narrowed from assertion.repository_owner == 'undercurrentai' (any org repo) to assertion.repository in ['undercurrentai/aegis-governance', 'undercurrentai/afa']. SA IAM binding swapped from org-wide principalSet to per-repo bindings (defense-in-depth). Files: GCP IAM policy (live gcloud commands) Completed: 2026-04-03

HIGH (Q2 2026)

• G3. Create GCP vendor risk assessment — GCP is the primary compute vendor (5 Cloud Run services) with zero formal risk assessment. AWS (the assessed vendor) is now secondary. Fix: Create docs/compliance/VENDOR-RISK-GCP.md following the AWS template. Files: docs/compliance/VENDOR-RISK-GCP.md Effort: 4 hours

• G4. Create Neon vendor risk assessment — Neon PostgreSQL holds all customer data (shared across 3 services) with zero assessment. Fix: Create docs/compliance/VENDOR-RISK-Neon.md. Files: docs/compliance/VENDOR-RISK-Neon.md Effort: 3 hours

• G5. Create Unkey vendor risk assessment — Unkey is the auth gateway for all protected routes with zero assessment. Fix: Create docs/compliance/VENDOR-RISK-Unkey.md. Files: docs/compliance/VENDOR-RISK-Unkey.md Effort: 2 hours

• G6. Migrate DynamoDB to GCP-native store — All data migrated from aegis-governance-state-dev (492 items) to Neon PostgreSQL. Dataclasses extracted to customer_models.py. All consumers switched to customer_pg.CustomerManager. AWS credentials removed from Cloud Run. MCP server DATABASE_URL gap fixed. DynamoDB table preserved read-only for G9. Migration script: scripts/migrate_dynamodb_to_neon.py. Completed: 2026-04-04

• G12. Provision api.aegis.undercurrentholdings.com vanity domain — Cloud Run exposes the raw project-ID URL (aegis-api-980022636831.us-central1.run.app) directly to customers. Industry standard for API-first SaaS is api.{company}.{tld} (Stripe, OpenAI, Anthropic, Twilio, GitHub, Resend, Linear, Vercel, Cloudflare, Postmark, SendGrid). Raw cloud URLs read as "prototype" in enterprise security reviews and tie customer integrations to the GCP project ID (blocks future migrations without customer re-integration). Triggered by incident 2026-04-10: customer (wali.khan@agentect.com) guessed api.aegis.undercurrentholdings.com after the portal dashboard advertised it via a broken hardcoded fallback — DNS didn't resolve, customer hit getaddrinfo ENOTFOUND. Layer 2 hotfix (portal PR #7, commit 830d992) pointed the fallback at the canonical Cloud Run URL to unblock; this item finishes the work.

  Ordered steps (separate tracked work item, own PR, own verification):

  1. Confirm DNS control — undercurrentholdings.com zone is with Cloudflare or Google Cloud DNS (per user recollection). Verify before running any DNS-touching command. If Cloudflare, need API token or dashboard access; if Google Cloud DNS, need gcloud dns access in the right project.
  2. Verify domain ownership in GCP — gcloud domains verify undercurrentholdings.com (requires DNS TXT or Search Console). Skip if already verified at org level.
  3. Create Cloud Run domain mapping: gcloud run domain-mappings create --service=aegis-api --domain=api.aegis.undercurrentholdings.com --region=us-central1 --project=undercurrent-production. GCP returns required CNAME/A records.
  4. Add DNS records in Cloudflare (or wherever the zone lives). Propagation <5 min to 1 hour depending on TTL.
  5. Verify TLS — curl -v https://api.aegis.undercurrentholdings.com/health should return 200 with a Google-managed cert.
  6. Run full E2E suite against new hostname: AEGIS_E2E_API_BASE=https://api.aegis.undercurrentholdings.com AEGIS_E2E_SERVICE_KEY="..." pytest tests/e2e/ -v --tb=short -x.
  7. Coordinated cutover PR (single reviewed change per repo, coordinated across both):
     • aegis-governance/aegis-sdk/src/aegis/_http.py:46 — change _PRODUCTION_BASE_URL to https://api.aegis.undercurrentholdings.com
     • aegis-governance/docs/api/openapi.yaml:22 — update server URL; also update the outdated "AWS API Gateway" label while we're there
     • aegis-governance/docs-site/content/docs/** — all customer docs (REST quickstart, API index, dogfooding guide, MCP tools, deployment index, rest.md, quickstart-rest.md, onboarding.md, production-guide.md)
     • aegis-governance/docs/integration/dogfooding-guide.md:380, docs/deployment/production-guide.md:407
     • aegis-governance/.env (local dev) — AEGIS_E2E_API_BASE
     • aegis-governance/.github/workflows/aegis-deploy.yml — health check URL
     • aegis-governance/.github/workflows/e2e-customer.yml — target URL
     • aegis-governance/.github/actions/aegis-gate/action.yml — gate endpoint
     • aegis-governance/CLAUDE.md §10 "Web Properties" — add api.aegis.undercurrentholdings.com row
     • aegis-governance/CLAUDE.md §10 "Live Endpoints" — update the canonical URL
     • portal/app/(portal)/dashboard/getting-started-card.tsx — swap the Layer-2 Cloud Run fallback to the vanity URL
     • portal/.env.example — update both AEGIS_API_BASE_URL lines (server and NEXT_PUBLIC_* client)
     • Vercel env vars (both server-side AEGIS_API_BASE_URL and client-side NEXT_PUBLIC_AEGIS_API_BASE_URL) — update to the vanity URL
     • Publish a new SDK version (aegis-governance on PyPI) with the new default, note in changelog as non-breaking improvement
     • Add a ROADMAP completion entry under "Recently Completed"
  8. Keep the Cloud Run URL working indefinitely as a secondary endpoint. Do NOT take it down. Old SDK versions, cached customer integrations, and the Layer-2 hotfix fallback must keep working.
  9. Add a CLAUDE.md §9 changelog entry for the cutover (next version bump), documenting the 2026-04-10 incident as root cause and the domain provisioning as the fix.

  Related artifacts: Incident plan ~/.claude/plans/ethereal-chasing-steele.md; customer audit /tmp/aegis-incident-2026-04-10-audit.md; portal PR #7 (merged 830d992); portal PR #8 (merged 9808f6e); portal PR #9 (merged 1200b4d). Effort: 1-2 days (mostly DNS propagation waits and coordinated PR review)

  MCP custom domain planning: G12 originally only addressed the REST API (api.aegis.undercurrentholdings.com). The 2026-04-10 incident also exposed that the dashboard MCP snippet was pointing at the wrong Cloud Run service (aegis-api/mcp instead of aegis-mcp/mcp, fixed in portal PR #9 1200b4d). When G12 ships, decide whether to also provision either (a) mcp.aegis.undercurrentholdings.com as a separate vanity host for the MCP server, OR (b) merge both REST and MCP under a single host with path-based routing (api.aegis.undercurrentholdings.com/v1/... for REST, api.aegis.undercurrentholdings.com/mcp for MCP) by deploying a Cloud Run-fronted load balancer that routes to the two services. Recommend (a) for simplicity — separate services, separate hostnames, mirrors the current Cloud Run deployment. The cutover PR for G12 must update both NEXT_PUBLIC_AEGIS_API_BASE_URL and NEXT_PUBLIC_AEGIS_MCP_BASE_URL in the portal.

• G13a. Clerk email + SMS template audit (closed 2026-04-10 via Clerk Backend API automated audit) — Initial plan called for dashboard click-through. Context7 + Exa research discovered that Clerk has a Backend API for templates: GET /v1/templates/email, GET /v1/templates/email/{slug}, and same for /templates/sms. This enabled a 120-line stdlib-only Python script (/tmp/clerk-template-audit.py) that authenticated via the existing production CLERK_SECRET_KEY (sk_live_* from portal Vercel env) and swept every template for the URL bug-class pattern. Result: 24 email templates + 7 SMS templates = 31 total. ALL CLEAN. Zero stale URL patterns. Templates are Clerk stock defaults (no customization / no AEGIS branding). Brand customization is a separate concern, not in scope for the URL bug-class audit.

• G13b / G26. Stripe customer email template audit (deferred — Playwright MCP disconnect blocked) — Stripe has no API to retrieve customer email body content. Only Invoice Rendering Templates API (/v1/invoice_rendering_templates) for PDF invoices, plus Branding API for logos/colors. Receipt/subscription/refund email bodies are server-rendered by Stripe based on transaction data + branding settings with no read access. Requires Playwright dashboard navigation with user auth pause. Attempted during Layer 5 but the Playwright MCP server disconnected mid-session, blocking completion. Schedule manually with user in a future session. Effort: 1 hour (includes 2FA pause window).

• G14. CI guardrails against the customer-facing URL bug class (triggered by 2026-04-10 incident; three sub-items closed across three repos; see G14a, G14b, G14d below) — The 2026-04-10 incident shipped twice in one session: once as the original placeholder bug and once as my own remediation defect (a vercel env add value with a trailing newline that the build baked into the rendered curl snippet). Both should be caught by automation, not code review. Status after Layer 5: G14a (marketing site) + G14b (portal) + G14d (docs-site + docs/) all shipped and verified in CI. G14c (env var sync check) deferred as lower-value after the 3 grep guards closed the bug class at the source level. See sub-items below.

  • G14a. Forbid decommissioned URLs in marketing site source — Implemented as grep -rEn step in undercurrent-core/brand/website/.github/workflows/ci.yml quality-checks job. Shipped via undercurrent-core PR #12. Catches yd1xm4ahcg, 31dil7zwza, wildcard *.execute-api.*.amazonaws.com, api.aegis.undercurrentholdings.com, YOUR_API_URL, and wrong-service routing patterns. Closed: 2026-04-10
  • G14b. Portal CI grep guard for decommissioned URLs — Implemented as dedicated .github/workflows/url-drift-guard.yml workflow in undercurrentai/portal. Runs on every PR with no path filter (fast feedback). Shipped via portal PR #12 (413ea79). Closed: 2026-04-10
  • G14c. Diff .env.example against Vercel project on PR (deferred — lower value after grep guards closed the bug class) — Would require a Vercel access token in CI secrets and is orthogonal to the grep guards that actually protect the customer surface. Revisit only if a regression makes it past G14a/b/d. Effort: 3 hours.
  • G14d. docs-site + docs/ CI grep guard — Implemented as forbid-decommissioned-urls step in .github/workflows/docs-consistency.yml. Shipped via aegis-governance PR #74. Closed: 2026-04-10

  Related: postmortem write-up captured in /tmp/aegis-incident-2026-04-10-audit.md (9+ lessons learned). Total shipped: 3 CI grep guards across 3 repos.

• G15. Architectural fix: centralize customer-facing endpoints in a single registry (triggered by 2026-04-10 incident; identified during quality-gate /ultrathink phase) — The 2026-04-10 incident shipped THREE distinct bugs in the same class across three separate files (getting-started-card.tsx, key-usage-guide.tsx, getting-started-card.tsx AFA variant). The systemic root cause is that each customer-facing TSX file independently encodes its own API URLs as constants or env var fallbacks. There is no central registry, no cross-checking, and no architectural mechanism to enforce consistency between code, .env.example, Vercel project state, and live endpoints. Each new dashboard card or onboarding page is a new opportunity for the same bug.

  Architectural fix: introduce lib/customer-endpoints.ts as the single source of truth for all URLs the portal advertises to customers. Pattern:

  // lib/customer-endpoints.ts
  export const CUSTOMER_ENDPOINTS = {
    aegis: {
      rest: requireUrl(process.env.NEXT_PUBLIC_AEGIS_API_BASE_URL, "https://aegis-api-980022636831.us-central1.run.app"),
      mcp:  requireUrl(process.env.NEXT_PUBLIC_AEGIS_MCP_BASE_URL, "https://aegis-mcp-980022636831.us-central1.run.app"),
      docs: "https://aegis.undercurrentholdings.com",
    },
    afa: {
      rest: requireUrl(process.env.NEXT_PUBLIC_AFA_API_BASE_URL, "https://api.afa.undercurrentholdings.com"),
      docs: "https://afa.undercurrentholdings.com",
    },
  } as const;
  
  function requireUrl(envValue: string | undefined, fallback: string): string {
    const v = (envValue || fallback).trim();
    if (!v.startsWith("https://") || v.includes(" ") || v.includes("\n")) {
      throw new Error(`Invalid customer endpoint: ${JSON.stringify(v)}`);
    }
    return v;
  }

  Migration: All existing customer-facing TSX components (app/(portal)/dashboard/getting-started-card.tsx, app/(portal)/keys/key-usage-guide.tsx, app/afa/dashboard/getting-started-card.tsx, plus any future ones) replace their inline constants with imports from customer-endpoints.ts.

  Pairs naturally with G14: the CI checks in G14a/b/c can be scoped to just the registry file (single source of truth) rather than scanning every TSX file individually.

  Why not done as part of the 2026-04-10 hotfix: refactor scope, not hotfix scope. Three PRs already shipped under time pressure during a live incident; introducing a new abstraction during emergency response would have added risk without improving the immediate fix. Tracked here for the next portal sprint. Effort: 4 hours (registry + helper + migration + tests + .env.example sync). Repo: portal. Depends on: G14 ideally lands first to enforce the new pattern via CI, OR ships in same PR for atomicity.

• G16. Portal dependency vulnerability remediation (identified during quality-gate /review phase 2026-04-10; closed during quality-gate /full-validation 2026-04-10) — npm audit flagged 3 pre-existing vulnerabilities. Closed during quality-gate Phase 5 full-validation iteration:

  • HIGH next@16.0.0-beta.0 - 16.2.2 (DoS via Server Components, GHSA-q4gf-8mx6-v5v3) → bumped next 16.2.0→16.2.3 and eslint-config-next 16.2.0→16.2.3 in package.json. npm install confirmed found 0 vulnerabilities. Build re-verified clean.
  • 2× MODERATE hono (IPv6 ipRestriction, toSSG path traversal, serveStatic slash bypass) → resolved automatically by npm audit fix (non-force) bumping hono 4.12.10→4.12.12 and @hono/node-server 1.19.12→1.19.13 within the existing >=4.12.7 override range. Lockfile-only change.
  • Status: All 3 vulns closed. npm audit post-fix returns found 0 vulnerabilities. Bumped via portal PR (filed during quality-gate Phase 7 ship). Closed: 2026-04-10

• G17. Portal proxy.ts: stop initializing Clerk middleware in E2E bypass mode (identified during quality-gate Phase 7 CI investigation 2026-04-10; closed in same session) — While shipping the 2026-04-10 incident hotfixes, the post-merge E2E run for portal PR #9 failed on mobile-safari with Error: page.goto: Navigation to "http://localhost:3000/dashboard" is interrupted by another navigation to "http://localhost:3000/dashboard". Console showed Failed to fetch RSC payload for /clerk-sync-keyless?returnUrl=/dashboard. Falling back to browser navigation. TypeError: Load failed.

  Root cause investigation (full causal chain in portal CLAUDE.md §9 v1.5.x changelog and KNOWN_ISSUES URL-FLAKE-1):

  • The previous proxy.ts wrapped the auth callback in clerkMiddleware() and short-circuited inside the callback when IS_E2E_BYPASS was true.
  • But clerkMiddleware() is a runtime wrapper — it intercepts every request matched by config.matcher and runs Clerk's keyless sync infrastructure regardless of whether the inner callback returns early.
  • On mobile-safari (WebKit), the keyless sync fetch occasionally fails with TypeError: Load failed (a known WebKit + streaming response issue), and Clerk falls back to a full browser navigation that races with Playwright's page.goto().
  • History: this flake hit e2e/specs/navigation.spec.ts:52 on portal PRs #7, #8, #9 (and earlier). PRs #7 and #8 retried-passed; PR #9 unluckily exhausted all 3 retries; PR #10 retried-passed. The line await page.goto("/dashboard") at navigation.spec.ts:58 was added by joshuakirby on 2026-03-18 (commit f97971c fix(e2e): reset mobile nav loop to clean state each iteration) as a fix for an earlier flake class — proving this test has had documented mobile-safari instability for months.

  Fix (portal proxy.ts):

  • Conditionally export a no-op handler when IS_E2E_BYPASS is true, instead of always wrapping with clerkMiddleware().
  • The lib/e2e-bypass.ts production kill-switch (RT-H1) guarantees IS_E2E_BYPASS is always false in production: it requires NODE_ENV=development AND throws on boot if the bypass env vars leak into NODE_ENV=production. So the no-op branch can never affect production deploys.
  • Production deploys (Vercel NODE_ENV=production): unchanged — full clerkMiddleware() runs as before.
  • E2E test runs only (Vercel preview with bypass vars never set; only Playwright dev server with all 3 bypass conditions): no-op handler returns undefined for every request, Clerk's keyless sync never initializes, the race condition is eliminated.

  Companion future work (defense-in-depth, not blocking):

  • G17a: harden e2e/specs/navigation.spec.ts:52 with await page.waitForLoadState("networkidle") after every page.goto to make ANY navigation race resilient (independent of which auth provider is active). Not strictly necessary after G17 lands, but follows the principle of "tests should not depend on external timing assumptions". Effort: 30 min.
  • G17b: portal .github/workflows/e2e.yml paths: filter does NOT include proxy.ts. While shipping G17, the manual gh workflow run was required because the path filter only watches app/**, lib/**, components/**, e2e/**, package.json, package-lock.json, playwright.config.ts, and .github/workflows/e2e.yml. proxy.ts is auth/middleware code that absolutely should trigger E2E on every change. Add proxy.ts to the filter (and consider adding next.config.ts, vercel.json for similar reasons). Effort: 5 min.

  Status: Closed in same session. Portal PR (filed during quality-gate Phase 7 ship after the user explicitly authorized "do what you think is logical, graceful, elegant, robust"). Closed: 2026-04-10

• G14a. CI grep guard for decommissioned URLs (marketing site) (closed 2026-04-10 via undercurrent-core PR #12 62716f8) — Added Forbid decommissioned URLs in customer-facing source step to undercurrent-core/brand/website/.github/workflows/ci.yml quality-checks job. Catches: yd1xm4ahcg.execute-api, 31dil7zwza.execute-api, any other *.execute-api.*.amazonaws.com, api.aegis.undercurrentholdings.com, YOUR_API_URL, aegis-api/mcp wrong-service routing, aegis-mcp/{evaluate,v1,risk-check} wrong-service routing. CI verified live (quality-checks SUCCESS on PR #12). Also closed dead-code cleanup of 8 broken AWS Lambda URLs in 4 orphaned section files.

• G14d. CI grep guard for decommissioned URLs (docs-site + docs/) (closed 2026-04-10 via aegis-governance PR #74) — Added forbid-decommissioned-urls step to .github/workflows/docs-consistency.yml. POSIX ERE portable pattern (no bounded quantifiers) catches: yd1xm4ahcg, 31dil7zwza, any *.execute-api.*.amazonaws.com, api.aegis.undercurrentholdings.com, YOUR_API_URL, aegis-api/mcp wrong-service routing, aegis-mcp/{evaluate,v1,risk-check} wrong-service routing. Scope: docs-site/content/docs/ + docs/. Exclusions: ROADMAP.md, roadmap.mdx, llms-full.txt, changelog.md, docs/claude/** (all intentional meta-references). Tested locally with 5 positive + 3 negative test cases. CI verified live on PR #74.

• G14b. CI grep guard for decommissioned URLs (portal) (closed 2026-04-10 via portal PR #12 413ea79) — New .github/workflows/url-drift-guard.yml workflow in undercurrentai/portal. Scoped to app/, lib/, components/, .env.example (customer-facing source only, excludes node_modules/, .next/, e2e/, CLAUDE.md, KNOWN_ISSUES.md). Same POSIX ERE pattern as G14a/G14d. Runs on every PR with NO path filter (fast feedback <30s, independent of the slower E2E workflow). Completes the G14 CI guardrail trifecta: G14a ✓ marketing site, G14d ✓ docs-site + docs/, G14b ✓ portal — the bug class is now blocked from reappearing at the CI level across all 3 customer-facing repos where it was found during the 2026-04-10 incident.

• G14e. Portal lint underscore-prefix convention (closed 2026-04-10 via portal PR #13 daa5d7c) — Side finding from quality-gate Phase 5 full-validation. Portal PR #11 (proxy.ts E2E bypass fix) introduced an e2eNoop handler whose signature takes _req: NextRequest (required by the type sig, body unused). Portal's eslint did NOT honor the _ prefix convention, so the warning slipped through because npm run lint in the portal doesn't enforce --max-warnings 0. Fix: added argsIgnorePattern: "^_", varsIgnorePattern: "^_", caughtErrorsIgnorePattern: "^_" to portal/eslint.config.mjs. Lint warnings dropped 6 → 2 (the remaining 2 are in auto-generated public/wasm/aegis_loader.js — out of scope, tracked separately in portal WASM work).

• G18. api.undercurrentholdings.com TLS cert mismatch (deferred — user decision required) — DNS resolves to agora-iv-980022636831.us-central1.run.app but the Cloud Run cert SAN does not include this hostname → browser cert error (NET::ERR_CERT_COMMON_NAME_INVALID) for anyone hitting the URL. Not currently referenced from any customer-facing copy verified by Layer 5 Tier 2 crawl, but the DNS record exists. Two valid fixes: (a) delete the DNS record (cleaner if there's no active use); (b) provision a Cloud Run domain mapping for api.undercurrentholdings.com → agora-iv with managed cert. Decide intent before fixing. Effort: 15 min once decided.

• G19. AEGIS Advisor promoted to docs nav (closed 2026-04-11 via aegis-governance PR #79 606d1a3) — User chose Promote (not noindex / not delete). Added docs-site/content/docs/resources/advisor.mdx explaining what the Advisor is, when to use it, how the 8 wizard steps map to the 6 gates, privacy handling, and cross-links to the full SDK / REST / MCP / GitHub Action docs. Wired into docs-site/content/docs/resources/meta.json sidebar nav. Auto-included in llms.txt (28 pages, was 27) via the existing generate-llms-txt.mjs walker. docs-site/public/advisor/ SPA itself untouched — it works correctly, uses the canonical Cloud Run URL (verified clean during Layer 5), and needed no changes.

• G20. Fumadocs /docs/index 404 in llms.txt (low priority) — Both aegis.undercurrentholdings.com/llms.txt and afa.undercurrentholdings.com/llms.txt reference /docs/index as the canonical home, but Next.js Fumadocs only generates /docs (not /docs/index). Any crawler honoring the llms.txt inventory logs a spurious 404. Fix in the scripts/generate-llms-txt.mjs generator (filter out index segments or rewrite to /docs). Effort: 15 min.

• G20 + G21. Disable Swagger UI in prod (closed 2026-04-11 via aegis-governance PR #78 dd0aa77) — User chose Disable Swagger UI in prod, keep openapi.yaml in docs only. Implemented as AEGIS_DISABLE_PUBLIC_DOCS env var gating the FastAPI constructor's docs_url / redoc_url / openapi_url args — when truthy, all three return HTTP 404 at the routing layer. Set unconditionally to true in .github/workflows/aegis-deploy.yml for every production rollout; default unset (permissive) in local dev / CI. Authoritative OpenAPI spec for external consumers still lives at aegis.undercurrentholdings.com/docs/api/openapi. Added 19 regression tests in tests/infra/test_api_docs_disabled.py covering the truthy/falsy env var matrix, the three disabled routes, and the operational endpoints (/health, /openapi.yaml) remaining accessible.

• G22. LinkedIn URL inconsistency across web properties (low priority, cosmetic) — Three different LinkedIn identifiers referenced across aegis.undercurrentholdings.com/llms.txt (linkedin.com/company/108736059), aegis.undercurrentholdings.com/llms-full.txt (linkedin.com/company/undercurrent-holdings/), and undercurrentholdings.com/about/ (linkedin.com/in/kevin-choate/ + linkedin.com/in/kirby-joshua/). Pick a canonical company URL and update all references. Effort: 30 min.

• G23 + G24 + G25. Dead code cleanup in undercurrent-core/brand/website/src/components/sections/developers/ (closed 2026-04-11 via undercurrent-core/brand/website PR #13 633eae4) — Deleted 22 orphaned TSX section files (leftover from commit f4909cf) plus 4 orphaned test files in tests/unit/ that tested ONLY the deleted components (developer-doc-sections.test.tsx, developer-layouts.test.tsx, developer-navigation.test.tsx, developer-sections.test.tsx). Net: 26 files deleted, 6271 lines removed. G24 (phantom docs.afa.undercurrentholdings.com hostname) and G25 (stale doc paths) resolved automatically because both lived in the deleted dead code. Verification: npm run typecheck clean, npm run lint --max-warnings 0 clean, npm run test:unit passes 31 suites / 1054 tests (down from 35 / ~1250), npm run build static export succeeds with out/developers/aegis/ and out/developers/afa/ both rendering.

• G23b. Orphan CSS cleanup in undercurrent-core/brand/website/src/theme.css (follow-up to G23) — ~20 CSS classes (e.g., aegis-card-number, aegis-compliance-table, afa-tier-cta) that were referenced only by the deleted components still live in theme.css. Removing them requires tracing which classes are referenced by the 2 remaining landing pages (aegis-landing-page.tsx, afa-landing-page.tsx) vs which are genuinely orphaned. Scoped out of G23 to keep the deletion PR focused. Effort: 30 min.

• G27. docs/ vs docs-site/content/docs/ divergence CI sync-check (closed 2026-04-11 via aegis-governance PR #80) — Two parallel markdown trees: docs/ (plain markdown, visible on GitHub + scraped by LLMs) and docs-site/content/docs/ (Fumadocs MDX, deployed to aegis.undercurrentholdings.com). The 2026-04-10 incident exposed the drift risk: YOUR_API_URL placeholders had to be fixed in BOTH trees independently. Fix: new docs-site/scripts/check-docs-sync.mjs diffs the 13 known shared file pairs, normalizes out frontmatter + blank lines, computes line-level LCS delta, and fails CI when any pair drifts past 10% without a documented entry in docs-site/.mdx-sync-skip.json. Wired into .github/workflows/docs-consistency.yml after the forbid-decommissioned-urls step. First run found 5 DRIFT pairs — all grandfathered into the skip list with rationales (2 intentional: api/github-action, onboarding; 3 grandfathered for reconciliation: api/openapi, quickstart-rest, quickstart-sdk). Tested locally: injected synthetic drift → exit 1; reverted → exit 0.

• G27b. Reconcile 3 grandfathered drift pairs from G27 (follow-up to G27, tracked in docs-site/.mdx-sync-skip.json) — Three files in the G27 skip list are marked GRANDFATHERED (not INTENTIONAL) and should eventually be re-synced:

  • api/openapi: docs/api/openapi.md is stale from the pre-Cloud Run era (still says "Lambda function behind API Gateway"); docs-site/content/docs/api/openapi.mdx has the current Cloud Run narrative. Fix: update docs/ to match docs-site/.
  • getting-started/quickstart-rest: docs/ leads with /sandbox/evaluate, docs-site/ leads with /health. Pick one structure, align both.
  • getting-started/quickstart-sdk: docs/ has a "Sandbox Mode (No Signup Required)" section; docs-site/ omits it in favor of an authenticated example. Pick one, align both.

  After reconciliation, remove each entry from .mdx-sync-skip.json and run node docs-site/scripts/check-docs-sync.mjs to confirm the pair drops back below the 10% threshold. Effort: 1-2 hours (mostly content decisions, not code).

MEDIUM (Q2-Q3 2026)

• G7. Create Stripe vendor risk assessment — Stripe processes live payments across 3 services (same live key shared). Fix: Create docs/compliance/VENDOR-RISK-Stripe.md. Effort: 2 hours
• G8. Create GitHub vendor risk assessment — GitHub holds source code, CI/CD secrets, and WIF trust. Fix: Create docs/compliance/VENDOR-RISK-GitHub.md. Effort: 2 hours
• G9. AWS resource decommission — Phased teardown of dormant AWS infrastructure. Shadow workflows migrated to Cloud Run on 2026-04-08.
  • G9a. Migrate shadow workflows (completed 2026-04-08) — 3 CI shadow workflows (libertas-cicd, AgoraIV, AFA) migrated from aws lambda invoke to Cloud Run + Unkey auth. CI shadow customer provisioned (cust_734077b5dc53, Professional tier). AEGIS_API_KEY secret set in all 3 repos. All verified live (proceed).
  • G9b. Verify Lambda invocations at zero (target: 2026-04-15) — 7-day soak period. Check: aws cloudwatch get-metric-statistics --namespace AWS/Lambda --metric-name Invocations --dimensions Name=FunctionName,Value=aegis-evaluate-proposal-dev. DynamoDB item count should remain 492. Secrets Manager LastAccessedDate should stop advancing.
  • G9c. Delete CloudFormation stacks (target: 2026-04-16) — Delete in dependency order: AegisMonitoringStack-dev → AegisMcpStack-dev → AegisLambdaStack-dev → AegisSharedStack-dev. Pre-check: empty S3 bucket, verify KMS key not shared, confirm DynamoDB data in Neon. Effort: 1 hour
  • G9d. Clean up GCP-side AWS remnants (target: 2026-04-16) — Delete aegis-aws-access-key-id and aegis-aws-secret-access-key from GCP Secret Manager. Revoke AWS_AEGIS_GATE_ROLE_ARN OIDC trust. Effort: 15 min
  • G9e. Code cleanup (target: 2026-04-17) — Delete: Dockerfile.lambda, src/lambda_handler.py, src/aegis_governance/unkey_authorizer.py, src/aegis_governance/customer.py, scripts/migrate_dynamodb_to_neon.py. Archive infra/ to branch. Remove boto3 from deps (keep [kms] optional if needed). Update src/cli.py to remove DynamoDB fallback. Update .env.example, ADR-009 addendum. Effort: 1 hour
• G10. Fix agora-iv container tag — agora-iv uses :latest tag (non-reproducible deploys) while all other services use digest-pinned images. Fix: Pin to SHA digest. Effort: 0.5 hour
• G11. Set up GCP Cloud Monitoring dashboards — Zero monitoring dashboards configured on GCP. All observability relies on default Cloud Run console metrics. Fix: Create dashboards for request rate, latency, error rate, instance count across all 5 services. Effort: 2 hours

Cross-Property Copy Remediation (Audit 2026-04-04)

Adversarial copy audit across 4 web properties (docs-site, portal, developers site, main website). 8 critical, 9 high, 8 medium findings. Source: adversarial compliance review plan.

Immediate (no blockers — copy changes only)

• CR1. Complete C12 compliance language sweep — "compliance-ready" → "compliance-aligned" across 3 repos. aegis-governance (4 files), portal (3 files) completed 2026-04-04. undercurrent-core had no instances. Completed: 2026-04-04
• CR2. Add qualifier to trust bars — "Aligned with:" prefix + FedRAMP → "FedRAMP High" across docs-site (3 files), portal (2 files), developers site (2 files). Completed: 2026-04-04
• CR3. Fix quantum-resistant superlative — "the only" → "the first" + Ed25519/ML-DSA hybrid clarification. Market research confirms defensible. Completed: 2026-04-04
• CR4. Fix LIBERTAS "live from day one" — → "built in from day one" across 5 locations (deep-dive, product-cards, capabilities, layout FAQ, libertas page meta). Completed: 2026-04-04
• CR5. Remove unsourced statistics — Removed "5-10×", "~87%", "months to weeks" from LIBERTAS copy. Replaced with qualified language ("measurably fewer failures", "higher deployment capacity", "compress compliance timelines"). AFA stats ("~60% of issues", "23% engineering time") remain — tracked separately. Completed: 2026-04-04
• CR6. Fix EU AI Act deadline — "enforcement deadline" → "full enforcement" (partial enforcement already active). Completed: 2026-04-04
• CR7. Drop "Operational" from SOC 2 description — "Runbook templates aligned to trust criteria" across docs-site + developers site. Completed: 2026-04-04

Short-term (this week)

• CR8. Verify AFA "9 gates" claim — VERIFIED: AFA genuinely has 9 gates (AEGIS 6 + Entropy + Supply Chain + KPI Alignment). Source: AFA/src/afa/gates/engine.py. No copy changes needed. Completed: 2026-04-04
• CR9. Fix portal AFA "priority support" claim — Developer tier upgrade banner claims support not in tier spec. File: portal afa/dashboard/page.tsx Effort: 15 min
• CR10. Remove regulatory_reporting feature references — Removed from portal (lib/types.ts, lib/tier-data.ts, upgrade-form.tsx). Now consistent with aegis-governance tiers.py. Completed: 2026-04-04
• CR11. Soften "100% logging coverage" claim — Changed to "comprehensive logging coverage for all evaluation paths." Completed: 2026-04-04
• CR12. Agora IV on About page — Remove mention or add "(coming soon)". Repo: undercurrent-core Effort: 15 min

Medium-term (this sprint)

• CR13. Add inline compliance disclaimers — Added "Aligned with:" qualifier and "Not a certification claim" disclaimer to 3 developer-facing pages: aegis-landing-page.tsx (badge section), aegis-overview-section.tsx (compliance table intro), libertas-capabilities-section.tsx (SOC2/PCI-DSS/HIPAA). Portal pricing page and docs-site compliance page already had proper qualifiers. Repo: undercurrent-core (brand/website submodule) Completed: 2026-04-04
• CR14. Verify team credentials — Gather supporting evidence for specific claims on team page: BMW dealer rankings (internal BMW reporting data), Claude Code token count (Anthropic dashboard), product count, QuarryX backtest methodology. Most claims are verifiable with internal data — assemble the evidence, then decide if copy needs clarification or just citation. File: brand/website/src/components/sections/team-section.tsx Repo: undercurrent-core Effort: 2 hours
• CR15. Rebuild llms-full.txt — Regenerated from updated MDX content (roadmap DX phases, D4 completion, README quickstart update). File: docs-site/public/llms-full.txt Completed: 2026-04-04
• CR16. Update compliance page for FedRAMP High — Note FedRAMP High as target baseline, FIPS 140-3 PQC module gap (industry-wide, earliest Q2 2027), classical FIPS crypto path planned. File: docs-site/content/docs/compliance/index.mdx Effort: 1 hour

FedRAMP High Readiness (Long-term)

Technological alignment for FedRAMP High authorization. Formal 3PAO engagement deferred until organizational readiness (C18: team expansion). FIPS 140-3 validated PQC modules unavailable industry-wide until Q2 2027+ (CIQ NSS module target). Budget: $250K-$3M+ for authorization process. Only 48 cloud services hold FedRAMP High authorization as of March 2026.

Critical path: F1-F3 (FIPS crypto + policies, no blockers) ─┐ C3-C5 (exercises) → C18 (team expansion) ─┴→ F4 (SSP) → F5 (3PAO)

Medium-term (Q2-Q3 2026 — no external dependencies)

• F1. Add FIPS-approved classical crypto provider — ECDSA P-256 signatures + ECDH P-384 key agreement using OpenSSL FIPS module (via Python cryptography library with FIPS backend). This becomes the FedRAMP-compliant crypto path; PQC hybrid remains as optional/additional. NIST-recommended "crypto agility" approach for the gap period. New files: src/crypto/ecdsa_fips_provider.py, src/crypto/ecdh_fips_provider.py Effort: 3-5 days
• F2. FIPS 199 security categorization — Formal categorization of AEGIS information types per NIST SP 800-60. Required for FedRAMP SSP. New file: docs/compliance/FIPS-199-categorization.md Effort: 4 hours
• F3. Create MA and PS policies — Maintenance (MA) and Personnel Security (PS) are complete gaps in the FedRAMP control family assessment. No code changes, documentation only. New files: docs/compliance/MAINTENANCE-POLICY.md, docs/compliance/PERSONNEL-SECURITY.md Effort: 6 hours

Long-term (Q3-Q4 2026 — depends on C18, F1)

• F4. Draft System Security Plan (SSP) — FedRAMP SSP template with control implementation narratives for all 421 High baseline controls. docs/compliance/system-description.md provides ~70% of system description. Effort: 4-8 weeks Blocked by: F1, F2, F3, C18
• F5. Engage 3PAO for readiness assessment — Pre-assessment gap analysis before formal authorization pursuit. Budget: $50K-$100K for readiness assessment alone. Effort: External Blocked by: F4
• F6. Monitor FIPS 140-3 PQC module availability — Track CIQ NSS module (target Q2 2027), wolfSSL wolfCrypt PQC, and CMVP Modules in Process list. When first validated PQC module ships, trigger migration from F1 classical crypto to PQC+classical hybrid under FIPS. Effort: Ongoing monitoring

Feature Accuracy & Tier Enforcement (Audit 2026-04-04)

Rigor audit of all customer-facing feature claims across docs-site, portal, developers site, and README. 50+ claims audited against implementation code. Source: adversarial feature-accuracy review with code-documentation reconciliation.

Audit results: 35 VERIFIED | 10 QUALIFIED (limitations disclosed) | 5 OPEN (copy remediation) | 0 FALSE | 1 UNVERIFIED-DEFENSIBLE ("first" PQC governance platform)

CRITICAL: Tier Feature Enforcement (Pre-Revenue Blocker)

src/aegis_governance/tiers.py defines 4 tiers with feature flags (4/9/15/18 features), but zero feature enforcement exists in the codebase — feature flags are metadata only. Any bearer token can use Enterprise/Financial Services features (shadow_mode, crypto_overrides, RBAC, PQ crypto). Only rate limits (Unkey) and API key count limits are enforced. This is by-design for Phase 1 (visibility only), but must be resolved before publishing per-tier feature differentiation on pricing pages.

• T1. Implement tier feature gating in API server — Hybrid model (research-backed): core evaluation universal, 2 features gated. shadow_mode=true requires Professional+ (natural CI upgrade trigger). CSV compliance export requires Financial Services (regulatory need). Added has_feature() and min_tier_for_feature() helpers to tiers.py. Inline checks in api_server.py (2 insertion points, ~12 lines). 403 responses include tier name + upgrade URL. Files: src/aegis_governance/tiers.py, src/api_server.py Completed: 2026-04-05
• T2. Add tier enforcement tests — 21 tests: 11 tier helper tests, 5 shadow_mode gating tests, 5 compliance export gating tests. Per-tier Unkey mock fixtures. Files: tests/test_tier_enforcement.py (new, 21 tests) Completed: 2026-04-05
• T3. Sync portal tier data with tiers.py — Portal (undercurrentai/portal) has its own tier definitions that must match. Automate or document the sync process. Repo: undercurrentai/portal Effort: 2 hours

Remaining Copy Remediation (from Audit 2026-04-04)

• CR9. Fix portal AFA "priority support" claim — Removed "and priority support" from AFA Community plan upgrade banner. Feature does not exist in tier spec. Repo: undercurrentai/portal, File: afa/dashboard/page.tsx Completed: 2026-04-04
• CR12. Agora IV on About page — Remove mention or add "(coming soon)". Repo: undercurrent-core Effort: 15 min
• CR13. Add inline compliance disclaimers — Added "Aligned with:" qualifier and "Not a certification claim" disclaimer to 3 developer-facing pages: aegis-landing-page.tsx (badge section), aegis-overview-section.tsx (compliance table intro), libertas-capabilities-section.tsx (SOC2/PCI-DSS/HIPAA). Portal pricing page and docs-site compliance page already had proper qualifiers. Repo: undercurrent-core (brand/website submodule) Completed: 2026-04-04
• CR14. Verify team credentials — Gather supporting evidence for specific claims on team page: BMW dealer rankings, Claude Code token count, product count, QuarryX backtest methodology. Repo: undercurrent-core, File: brand/website/src/components/sections/team-section.tsx Effort: 2 hours
• CR16. Update compliance page for FedRAMP High — Note FedRAMP High as target baseline, FIPS 140-3 PQC module gap (industry-wide, earliest Q2 2027), classical FIPS crypto path planned. File: docs-site/content/docs/compliance/index.mdx Effort: 1 hour

Feature Claim Verification Summary

All core technical claims verified against implementation:

Developer Experience & CI Hygiene (Discovery 2026-04-04)

Comprehensive 6-agent codebase discovery (TODO scanner, code quality audit, ROADMAP audit, dependency/CI analysis, docs-code contradiction finder, feature completeness analysis). Project health: A+ — 0 TODOs, 4874 tests, ~94.89% coverage, all features production-deployed, zero deferred bugs.

Sequencing principle: DX items first (unblock contributors and reduce onboarding friction), then operational hardening.

DX Phase 1: CI & Build Reliability (no blockers, immediate)

All items are independent — can be done in parallel. Fixes supply chain risk, install correctness, and build reproducibility.

• DX1. Update outdated GitHub Actions — 5 actions across 3 workflows behind by 1-2 major versions. Supply chain risk (unpinned google actions) + missing security patches. sdk-publish.yml: checkout@v4→v6, setup-python@v5→v6. docs-deploy.yml: checkout@v5→v6, setup-node@v4→v5. aegis-deploy.yml: pin google-github-actions/auth@v2 and setup-gcloud@v2 with commit hashes (only unpinned actions in the repo). Files: .github/workflows/sdk-publish.yml, .github/workflows/docs-deploy.yml, .github/workflows/aegis-deploy.yml Effort: 1 hour
• DX2. Add numpy<2.0 to pyproject.toml [engine] — Dockerfile.lambda:20 pins numpy{'<'}2.0 but pyproject.toml [engine] group does not. pip install -e ".[engine]" outside Docker may pull numpy 2.x, breaking scipy on older platforms. File: pyproject.toml Effort: 5 min
• DX3. Pin Lambda runtime version — Dockerfile.lambda:5 uses public.ecr.aws/lambda/python:3.11 without patch pin — non-reproducible builds. Fix: Pin to python:3.11.11 (or latest 3.11.x). File: Dockerfile.lambda Effort: 5 min
• DX4. Narrow dependency version ranges — cryptography>=41.0.8,{'<'}48.0.0 too permissive (narrow to <43.0.0); coincurve>=18.0.0,{'<'}22.0.0 too wide (narrow to <20.0.0). Extends D15 (btclib <2025.0.0). File: pyproject.toml Effort: 15 min

DX Phase 2: Documentation & Onboarding (no blockers, short-term)

Reduce contributor friction. All items independent.

• DX5. Consolidate environment variable documentation — api_server.py:33-44 docstring lists 10 env vars but ~42 exist across src/ (AEGIS_KMS_KEY_ID, AEGIS_HSM_PIN, AEGIS_HSM_PKCS11_LIB, AEGIS_HSM_TOKEN_LABEL, AEGIS_HSM_WRAPPING_KEY_LABEL, AEGIS_HSM_WRAPPED_PRIVATE_KEY, AEGIS_KMS_WRAPPED_PRIVATE_KEY, AEGIS_CUSTOMER_ID, AEGIS_MCP_ALLOWED_HOSTS, AEGIS_AUDIT_BUCKET, AEGIS_MASTER_KEK_PUBLIC, DATABASE_URL, K_REVISION, AWS_DEFAULT_REGION, Secrets Manager ARN variants, etc.). Fix: Update docstring with all variables grouped by required/optional/Cloud Run/Lambda/legacy. File: src/api_server.py Effort: 2 hours
• DX6. Fix CLAUDE.md CI validator count — §10 references "9 compliance validators" but tools/ci/ contains 11 (includes aegis_shadow_backfill.py and generate_precommit_config.py). File: CLAUDE.md Effort: 5 min
• DX7. Fix CLAUDE.md changelog timeline — v4.6.117 entry shows "4737 → 4768" but its input baseline should match v4.6.116's output (4768). The v4.6.116 entry then shows "4769 → 4788" (removing 2 DynamoDB tests). Verify which baseline is correct via git history. Historical record only — current count (4788) is correct everywhere else. File: CLAUDE.md Effort: 15 min

DX Phase 3: Code Quality & Consistency (short-term)

Improve codebase consistency for contributors. DX8-DX10 independent; DX11 is a standalone migration.

• DX8. Standardize logger formatting — Converted 88 f-string logger calls to %-style lazy formatting across 20 files in src/. Enabled ruff G rules (flake8-logging) to enforce going forward; G201 ignored as separate DX item. Completed: 2026-04-07
• DX9. Add @abstractmethod to SerializableMixin — src/workflows/serialization.py:45-67 has 5 methods raising NotImplementedError but lacks @abstractmethod decorators — missing implementations caught at runtime instead of class definition time. File: src/workflows/serialization.py Effort: 30 min
• DX10. Narrow exception breadth in schema_validator — src/telemetry/schema_validator.py:173 catches (ImportError, Exception) — the Exception makes the ImportError redundant and swallows unexpected errors. Fix: Narrow to (ImportError, ModuleNotFoundError). File: src/telemetry/schema_validator.py Effort: 15 min
• DX11. Upgrade Dockerfiles to Python 3.12-slim — All 3 Dockerfiles use python:3.11-slim. Python 3.11 enters security-only mode May 2026 (next month). CI matrix already tests 3.9-3.12. Files: Dockerfile, Dockerfile.api, Dockerfile.lambda Effort: 2-4 hours

DX Phase 4: Runtime Safety (medium-term)

• DX12. Add drift observation window cleanup — src/integration/pcw_decide.py:60-67 — _drift_observation_windows dict keyed by id(drift_monitor) can leak memory in long-running processes if monitors are created/destroyed without explicit reset_drift_window(). Fix: Use WeakValueDictionary or TTL-based eviction. File: src/integration/pcw_decide.py Effort: 2 hours

Cross-Section Execution Sequence (2026-04-04)

Logical ordering across all open items, incorporating dependency chains and DX-first prioritization.

Week 1 — COMPLETE (DX1-DX4: CI actions, numpy, Lambda pin, dep ranges)

Week 2 — COMPLETE (DX5-DX7: env var docs, validator count, changelog fix)
  CR9, CR12 (quick copy fixes, 30 min total — remaining from Week 2)

Week 3 (Code quality + Dockerfile migration):
  DX8  (logger formatting)  ─┐
  DX9  (abstractmethod)      │ All independent, ~2 hours total
  DX10 (exception breadth)  ─┘
  DX11 (Python 3.12-slim)       ← 2-4 hours, test against CI matrix
  G10  (agora-iv container tag, 30 min)

Week 4 (Vendor risk assessments — parallelizable):
  G3 (GCP)    ─┐
  G4 (Neon)    │ 5 docs + dashboards, ~15 hours total
  G5 (Unkey)   │ Follow AWS template
  G7 (Stripe)  │
  G8 (GitHub) ─┘
  G11 (GCP Cloud Monitoring dashboards)

Week 5-6 (Compliance drills — sequential):
  C3  (access review, 4h)
  C10 (break-glass procedures, 3h)
  C4  (BCP/DRP test, 8h)           ← depends on C10
  C5  (IRP tabletop, 4h)
  C11 (DSR templates, 4h)

Week 6+ (Medium-term operational — no blockers):
  DX12 (drift window cleanup)
  N6   (distribute sandbox rate limit)
  C6   (decision accuracy metrics, 10h)
  S8   (stripe_customer_id on Customer model)
  G9   (AWS resource cleanup — includes DynamoDB decommission)
  CR9, CR12 (quick copy fixes — portal + undercurrent-core, 30 min)
  CR13 (compliance disclaimers), CR14 (credentials), CR16 (FedRAMP page)
  D10  (SNS alarm subscriptions)
  F1   (FIPS classical crypto, 3-5 days)  ─┐ No blockers — start
  F2   (FIPS 199 categorization, 4h)       │ early to unblock F4
  F3   (MA/PS policies, 6h)               ─┘

Pre-Revenue (before publishing pricing page — blocks paid tier launch):
  T1   (tier feature gating, 2-3 days) ← pricing strategy decision first
  T2   (tier enforcement tests, 1 day) ← blocked by T1
  T3   (portal tier data sync, 2h)     ← blocked by T1

Post-D12 (~2026-04-24, after 30-day KL data collection):
  D12 → D16 (shadow → enforcement transition)
  D13 (Grafana dashboards)
  N7  (PERT variance documentation)
  N8  (per-gate drift thresholds)

Strategic (Q3-Q4 2026, blocked by C18 team expansion):
  F4 (SSP, 4-8 weeks) ← blocked by F1+F2+F3+C18 → F5 (3PAO)
  C7 (external legal counsel)
  C8 (external TEVV audit)
  C9 (SOC 2 Type 2 prep) ← depends on C3+C4+C5+C18
  F6 (monitor FIPS PQC modules — ongoing)
  D14 (multi-region DR)
  D15 (btclib constraint planning)

Recently Completed

Gap Closure Sprint (PR #25) — Phase 1 Complete

Metrics (at PR #25 merge): 1997 tests, 94.47% coverage, 6 issues addressed + 13 rigor findings + 11 bug-hunt #5 + 6 bug-hunt #6 + 6 bug-hunt #8 + 8 bug-hunt #9 + 2 ultrathink + 5 QG-ultrathink-10 + shadow mode + HTTP sink + drift enforcement + MCP HTTP transport + H-1 SSRF fix + MCP hardening + TLS enforcement + parameter cookbook + QG56 ultrathink + QG57 ultrathink + BH10 (7 bugs)

AWS Deployment (ROADMAP Items 16-20) — DEPLOYED

API Endpoint: https://aegis-api-980022636831.us-central1.run.app/ Routes: POST /evaluate, POST /risk-check, GET /health

Completed (v3.26.0 — Rigor Protocol)

Metrics: 1689 tests, 94.60% coverage, 103/103 bugs fixed (100% fix rate)

Previously Completed (v3.11.0-v3.13.0)

See: Multi-Model Coherence Review for full analysis.

Recently Merged PRs

Open Issues

Recently Closed Issues

Sprint Update (PR #25): Issues #2, #5, #7, #8, #9 have Phase 1 code-side implementations complete. AWS infrastructure now DEPLOYED (4 CDK stacks live in us-west-2). Remaining work: Slack/email SNS subscriptions, multi-region replication, Locust load testing against live endpoints. Issues #6, #18, #24 closed. Milestone "Guardrail β-to-Prod" due date updated to 2026-06-30.

Release Roadmap

v1.0.1 (Patch - Pre-Release Bug Fixes) ✅ RELEASED

Released: 2026-01-31 (pre-release fixes merged before v1.0.0 SDK release) Focus: Bug fixes merged

v1.0.0 (Major - SDK Release) ✅ RELEASED

Released: 2026-02-06 (PR #23, commit cfa3783) Focus: Governance Decision SDK — public API, CLI, MCP server

v1.1.0 (Minor - Enhancements)

Target: Q1 2026 Focus: Testing improvements and new features

v1.2.0 (Minor - Features)

Target: Q2 2026 Focus: Production readiness

v2.0.0 (Major - Backlog)

Target: 2026 H2 Focus: Operational excellence

GAP Status Summary

Completed GAPs

In-Progress GAPs

Planned GAPs

Remaining Work (Post-Audit 2026-03-20)

Production Readiness Metrics

Quick Links

Integration Guides

• Parameter Reference - Complete parameter reference with derivation guidance
• Domain Templates - Worked examples for 4 domains (trading, CI/CD, content moderation, autonomous agents)

Implementation

• Implementation Plans - EPCC detailed plans
• Shadow Mode Prerequisites - Deployment guide

Research

• Research Index - All research documents
• KL Threshold Calibration - Drift detection tuning
• Market Research & Competitive Landscape - Market sizing, competitors, GTM
• MCP Security Ecosystem Review - CoSAI taxonomy + Red Hat patterns mapped to AEGIS

Analysis

• Comprehensive TODO Discovery - Historical progress
• Test Count Methodology - Test verification
• Cross-Reference Verification - Documentation accuracy

Changelog