Changelog

All notable changes to this project are documented below.

Unreleased

Fixed — policy guardrails

• calculateMaxDiscount now reads the policy it is asked about. It previously ignored the policy entirely and answered from hardcoded default-policy thresholds: against a policy with a 30% margin floor and an 8% cap it reported 15% with limiting factor volume_tier, naming a rule that policy does not contain. New extractPolicyThresholds recovers a policy's limits from its own rule conditions and is now the single source of truth, replacing four separate copies of those constants.
• A policy stating no recognizable limit now reports max_discount: 0 with limiting_factor: 'undetermined' rather than guessing.
• Insight checks registered under two ids each emitted every finding twice and double-counted summary.total. Each id now maps to exactly one check.
• sim-004 through sim-007 were exported as insights but never registered, so four of the eight simulation insights could never fire. All eight are wired up.
• Frequency checks divided per-round violation counts by the per-session order count, reporting impossible figures like "Margin floor triggered in 138.0% of evaluations (69 of 50)". Metrics now carry totalEvaluations as the matching denominator.
• The policy-review and pre-deployment checklists defined no isComplete handlers and always reported 0%. Verifiable items are now scored; the rest are surfaced as manual.
• detectEdgeCases hardcoded default-policy boundaries; it now reads the policy under test.
• evaluate() treated every fired rule as a violation, so a policy could not express an advisory rule. Only type: 'violation' blocks approval now, other events land in a new notices field, and applied_rules means "rules evaluated" rather than duplicating the violations.
• The abandoned session outcome is now produced: a buyer walks away once even its own floor cannot clear the policy. marginImpact reports revenue conceded, not the discount rate.

Added

• MCP 2026-07-28 protocol support on SDK v2. Every tool declares an outputSchema and returns structuredContent; none did before. 2025-era clients are still served.
• UCP 2026-04-08: CATALOG_CAPABILITY, SUPPORTED_UCP_VERSIONS, DEFAULT_PRODUCT_MARGIN, and GUARDRAIL_UCP_PROFILE.
• run_simulation, analyze_simulation, and the five UCP checkout tools are now documented.
• Rule Flow tab in the playground, wiring up a diagram component that shipped unused.
• ADR-004: MCP 2026-07-28 migration and SDK v2.

Changed

• Breaking: the three ui:// MCP Apps resources are removed — ext-apps has no SDK v2-compatible release, and the panels were already non-functional. See ADR-004.
• Breaking: UCP discount entries in totals[] are now negative per 2026-04-08, and the checkout store emits them at all — a discounted checkout previously reported total === subtotal. extends widens to string | string[]; identity linking is renamed to dev.ucp.common.identity_linking.
• MCP server now exposes 12 tools (v0.0.1 shipped 3).
• Unknown tools return JSON-RPC -32602 instead of a custom UNKNOWN_TOOL string.
• Toolchain moved to TypeScript 6, ESLint 10, Node 24 LTS, Next 16 and Fumadocs 16.
• Documentation corrected against the running server: tool counts, applied_rules (was triggeredRules), policy_id (default, not default-b2b-policy), and the simulate_checkout_discount response shape, which documented a rejected[] array that does not exist.

Security

• pnpm audit --audit-level=high went from 17 high advisories to zero. Dropping MCP SDK v1 removed the hono/ajv/express tree outright; the rest are pinned via overrides.

v0.2.1 / v1.0.0 (ucp-types) — earlier 2026

Added

• UCP alignment with Universal Commerce Protocol
• @guardrail-sim/ucp-types package with discount and checkout types
• @guardrail-sim/insights package with policy health checks
• @guardrail-sim/simulation package with adversarial buyer personas
• UCP-aligned MCP tools: validate_discount_code, simulate_checkout_discount
• Standard UCP checkout tools and /.well-known/ucp profile discovery
• Interactive policy playground in documentation site
• ADR-002: UCP Protocol Alignment; ADR-003: UCP Specification Review

v0.0.1 (December 2025)

Added

• Initial policy engine implementation with json-rules-engine
• MCP server with evaluate_policy tool
• Default policy with rules:
  • margin_floor - 15% minimum margin
  • max_discount - 25% maximum discount
  • volume_tier - Volume-based limits
• get_policy_summary and get_max_discount MCP tools
• Documentation website with Fumadocs
• ADR-001: Policy Engine Library selection

Technical

• pnpm monorepo structure
• TypeScript strict mode throughout
• ESLint + Prettier configuration
• Husky pre-commit hooks
• Conventional commits with commitlint
• 76 passing tests across all packages (177 as of the current Unreleased entry)