All notable changes, benchmark releases, and security hardening updates across the Tinplate Gateway and Fetcher Workers.
37-llm-messages-guard (multi-turn OpenAI/Anthropic messages[] sanitizer with unified cross-turn restoration map), 38-tabular-rag-cleaner (CSV/JSON formula-injection firewall & RAG table chunker), 39-peppol-ubl-invoice (bidirectional EN 16931 JSON ↔ Peppol BIS 3.0 UBL 2.1 XML & SEPA EPC-QR generator), 40-eu-vat-calculator (EU27 + UK/CH/NO cross-border VAT & Art. 196 reverse-charge engine), 41-sanctions-screener (fuzzy Jaro-Winkler & phonetic watchlist/embargo screener), and 42-agent-code-guard (Python/JS/GitHub Actions sandbox escape auditor)./v1/pipeline & /v1/presets): Added $steps.<stepId>.data.<field> reference resolution across pipeline steps and 4 turnkey enterprise presets (llm-ingest-safe, llm-output-safe, agent-pre-exec, eu-vendor-kyb).?receipt=true): Any /v1/* call can now request an HMAC-SHA256 signed Agent Action Receipt (X-Tinplate-Receipt & X-Tinplate-Receipt-Hash headers) with hash-chain continuity via X-Tinplate-Prev-Receipt-Hash.@tinplate/sdk Package & GitHub Action (action.yml): Added workspace SDK package (packages/sdk), composite GitHub Action, and expanded live synthetic canary monitoring to 19 engines on /status and /v1/health/deep./mcp), Multi-Step Pipeline (/v1/pipeline), Official SDKs (/sdk) & Live Canary StatusGET /mcp & POST /mcp): Exposed all 36 Tinplate micro-APIs as Model Context Protocol (2025-03-26) tools with readOnlyHint/destructiveHint/idempotentHint annotations.POST /v1/pipeline & /tools/pipeline-builder): Single-request chaining of up to 5 Tinplate endpoints with automatic stop_on_block short-circuiting./sdk) & Dynamic SVG Badges (/badge/:type): Added zero-dependency SDK clients, Vercel AI SDK guardrail middleware, OpenAPI parameter/response tables on all 36 API pages, and live 13-engine canary verification on /status.tinplate-gateway, reduced HTML edge cache TTL to 60s, explicitly disclosed third-party marketplace proxy hops vs. Direct Bearer-Key access, and expanded compliance verification suites on /limitations./tools), Interactive Live Demo Sandbox (/demo/:slug) & Engineering Blog (/blog)/tools/<slug>): Launched interactive web tools for DMARC/SPF/DKIM/MTA-STS checking, HTTP Security Headers grading, IBAN/VAT/LEI/EORI validation, Broken JSON repair, Cron explanation, Secret/PII redaction, and MCP manifest security linting.POST /demo/:slug & /apis/<slug>): Upgraded all 36 API pages with a live JSON playground backed by a keyless, rate-limited (30 req/min/IP), 8 KB-capped sandbox route./blog/<slug>): Published 4 deep-dive articles covering prompt-injection bypasses & benchmarks, MCP tool poisoning, DMARC/SPF 10-lookup limits, and Hungarian bridge-day math.31-llm-output-guard (POST /v1/llm-output-guard): Outbound LLM response guardrail scanning for canary/system-prompt leaks, Markdown exfiltration beacons, XSS, ANSI injection, blocked terms, and degenerate repetition loops.32-command-sql-guard (POST /v1/command-sql-guard): Pre-execution shell & SQL safety firewall for coding agents, MCP tools, and Text-to-SQL pipelines.33-eu-id-validator (POST /v1/eu-id-validator): ISO 17442 LEI, EU/UK EORI, French SIREN/SIRET, German Steuer-ID, Italian Codice Fiscale, Spanish NIF/CIF, Polish NIP/PESEL, Dutch BSN/RSIN, UK CRN, and US EIN checksum validator.34-jwt-jwk-inspector (POST /v1/jwt-jwk-inspector): RFC 8725 JWT/JWS/JWK security auditor detecting alg:none, header injection, weak/dictionary HMAC secrets, and leaked private keys in public JWKS.35-rag-chunker (POST /v1/rag-chunker): Markdown & code-aware RAG document splitter preserving code blocks, tables, and heading breadcrumb hierarchies.36-regex-safety (POST /v1/regex-safety): Static ReDoS catastrophic backtracking scanner ($O(2^n)$ exponential & $O(n^2)$ polynomial complexity), rewrite hints, and pattern explainer./v1/secret-scanner): Expanded to 28 secret types (38+ rules) and published a 60-case benchmark achieving 100.0% recall (40/40) and 0.0% false-positive rate (0/20) at P50 0.051 ms in-Worker compute time./v1/tool-call-validator & /v1/mcp-manifest-linter): Added dedicated CI fixture suites for OpenAI (Chat Completions & Responses API), Anthropic (2023-06-01), Google Gemini (v1beta/v1), and Model Context Protocol (2024-11-05 & 2025-03-26)./v1/hallucination-diff): Improved comma-formatted number normalization and mid-sentence entity extraction; published 25-pair RAG benchmark (92.0% human agreement, 100.0% factual number/entity recall).use_ai_judge Workers AI data handling across all specs, and added ResponseMeta $ref components across all 30 OpenAPI 3.0.3 files./apis/<slug> for all 30 APIs with sample inputs/outputs, cURL/Python/JS snippets, capabilities, limitations, FAQs, and JSON-LD structured data./openapi/<slug>.json, plus /pricing, /docs, /limitations, /status, /security, /robots.txt, /sitemap.xml, /llms.txt, and /.well-known/security.txt.observability.enabled = false, head_sampling_rate = 0, and invocation_logs = false across Gateway and Fetcher Workers and removed Analytics Engine bindings./v1/prompt-guard): Added Unicode Plane 14 tag decoding, Hex/URL/ROT13/Leetspeak unpacking, 50+ homoglyphs, 10-language multilingual rules, sensitivity presets, llm_output context, and published the 104-prompt benchmark (93.8% recall, 0.0% FPR).ErrorEnvelope via $ref across 400/401/413/429/500, and removed staging server URLs from public specs.16-secret-scanner through 30-content-provenance) alongside the 15 Wave 1 APIs.docs.test.ts) verifying all 109 published Markdown request examples against the live Hono Gateway router.