Changelog

​

Added

• Added a curated llms.txt agent entry point with task routing, product boundaries, Jev/NSR examples, and operational verification guidance. Added complete per-topic indexes generated from navigation, with link and freshness checks in CI.

​

Added

• Added Jev + NSR industry-pack explanations and a complete Decisions API walkthrough, with seven downloadable packs, 106 scenarios, and an offline request builder that excludes fixture model answers.

​

Changed

• Linked industry examples from the Jev concepts, workspace, and Decisions API guides; documented deployment-dependent public-demo idempotency and execution lookup.

​

Documentation reliability

• Expanded nested request fields, enum choices, constraints, alternatives, and idempotency headers in generated references. Added schema checks for 242 published JSON request examples across Commerce, ResponseCX, Sandbox, and Computer Use, correcting invalid bounds and example values.
• Unified rate-limit guidance around service-specific contracts and replaced the broken queue retry sample with an executable, tested GET retry helper.
• Corrected Sandbox API-key authentication from controller source, rebuilt its create/execute/teardown walkthrough, and filled missing file-operation request schemas. Removed spurious JSON bodies from 23 Temporal control endpoints after checking their handlers.
• Corrected EDI request references and response media/envelopes, Computer Use SSE media types, NSR request guidance, and Cart.com order validation and response wrapping from service source.
• Added deployment prerequisites to contract-only and legacy endpoint pages; reconciled the README, Commerce overview, reporting guide, and unsupported legacy wallet/network examples with working entry points.
• Fixed a site-wide browser error by keeping the Node maintenance script in a .cjs file. Correctness findings and missing MDX dependencies now fail checks; unavailable example hosts and unpublished package commands also block validation.

​

Added

• Added source-backed getting-started walkthroughs for Desktop, web Console, and mobile Console, with installation paths, first-task checkpoints, MCP/tool setup, and troubleshooting.

​

Changed

• Linked app onboarding from the navigation and product references, refreshed the Console overview, and corrected mobile tab labels and entry points.

​

Added

• Added Jev + NSR architecture, public first-decision, and organization setup guides based on the latest engine and console source. Covered credential boundaries, evidence-aware Playground runs, fresh policy tests, deployment capabilities, and provider usage/retry semantics without claiming live-provider verification.

​

Added

• Added a Chat Desk first-shift walkthrough with website and managed-email exercises, reply-versus-macro recovery, browser-draft limits, snoozing, and an operator handoff template.
• Added source-backed Chat Desk, hosted-widget handoff, and channel coverage guides, including managed-email delivery, operator ownership, internal notes, and current SMS/WhatsApp/Voice/ChatGPT boundaries. Distinguished the hosted loader from the standalone widget integration.

​

Changed

• Reworked the onboarding hub into outcome cards and a five-step signup walkthrough with verification checkpoints; added direct troubleshooting routes and highlighted recovery criteria.
• Extended walkthrough summaries and success checkpoints to evaluations, configuration changes, and durable agents. Made support test cases easier to scan and prioritized two next actions at the end of key guides.
• Streamlined the homepage, shortened onboarding navigation labels, moved secondary community links to the footer, and added consistent guide summaries and result checkpoints.
• Unified theme accents, added a brighter dark-mode logo, and refined responsive layouts, table spacing, footer contrast, narrow-screen wrapping, and keyboard focus.

​

Added

• Added a workflow configuration-change walkthrough covering full candidate preservation, evaluation versus apply-time gates, optimistic concurrency, independent read-back, and reviewed restoration.
• Added a durable-agent MCP operating walkthrough covering governed release prerequisites, launch, status, steering, cancellation, and evidence handoff.
• Added a first agent-evaluation walkthrough with MCP argument examples, duplicate-case checks, idempotent creation, baseline/retest evidence, and explicit handling of ungraded or partial suites.
• Added an onboarding troubleshooting guide for workspace access, MCP authentication and scopes, agent tests, accelerator jobs, and missing replies, with recovery criteria and a read-only diagnostic prompt.
• Added a capability guide distinguishing the Rust commerce and durable workflow engines, with operational examples, long-running agents, schedules, and MCP task patterns.
• Added a first-week operating guide for reviewed live rollout, source-backed MCP reporting, correction-to-evaluation workflows, partial/ungraded test handling, and measurable improvement. Linked it throughout platform onboarding.
• Added source-checked platform onboarding from signup through a first support-agent answer, a shadow workflow trial, and MCP management. Included UI steps, evidence-based time-to-value milestones, accelerator recovery, connection checks, and bounded agent task prompts.

​

Changed

• Added consistent onboarding readiness and finish-line checkpoints, a copyable progress record, and an MCP resume prompt for handoffs between teammates and AI assistants.
• Refreshed the documentation design with a guided homepage, clearer onboarding navigation, coordinated light and dark colors, and improved card, table, and keyboard-focus styling.
• Corrected the active-horizon reference: validation creates a release record, recipe launches require approval, and accepted controls must be verified through status reads.
• Rebuilt the MCP directory around tasks, connection verification, and service boundaries instead of static tool counts. Corrected the workflow overview to distinguish durable execution from external-call idempotency.
• Made hosted platform onboarding prominent on the documentation homepage and starting navigation. Linked existing REST, provisioning, agent, and MCP references to the new walkthroughs and clarified Response MCP authentication and dynamic tool discovery.

​

Changed

• Rebuilt manufacturing onboarding around a tested BOM and two-batch work-order example. Clarified additive completion quantities, unchanged inventory, transition-policy limits, production recovery, and agent permissions; added it to automated example checks.
• Replaced broken warranty snippets with a runnable claim approval, completion, and denial workflow. Documented coverage-date arithmetic, claim read-back limits, separate replacement fulfillment, and agent decision boundaries; added the program to automated checks.
• Rebuilt fulfillment guidance with a runnable shipment example, application transition checks, verified order/inventory boundaries, warehouse binding coverage, and carrier-event recovery guidance. Added it to automated example checks and linked it from order onboarding.
• Replaced legacy supplier snippets with a runnable purchase-order approval workflow, rejected-transition checks, supplier-delivery boundaries, and agent authorization guidance. Added the program to automated example checks and linked it to receiving onboarding.
• Rebuilt warehouse onboarding around a runnable inbound-receiving workflow with partial receipts, over-receipt and cancellation guards, and an explicit inventory check. Added it to the example runner and clarified receiving, stock posting, and put-away boundaries.
• Added a guide for connecting real business operations, with data-ownership and identity mapping, service-specific readiness requests, agent task briefs, and completion evidence. Updated the API directory to lead commerce newcomers to a working local path.
• Added a shared first-operation walkthrough for developers and AI agents, with a persisted order/stock fixture, generated MCP configuration and task prompt, and independent verification. Linked it from both starting navigation paths and added its programs to CI.
• Rebuilt agent onboarding around task briefs, actual tool schemas, service selection, and verified operational outcomes; removed stale tool counts and blanket retry advice.
• Replaced the legacy comprehensive-testing guide with an integration test plan and an automated payment-replay test across process restarts, plus concrete guidance for hosted, asynchronous, and performance checks.
• Rebuilt the error-handling guide with executable stock-rejection tests, service-specific error fields, bounded retry guidance, and reconciliation steps for uncertain writes. Added its program to the documentation example checks.

​

Fixed

• Corrected cross-engine advice that treated conflicts as successful replays and recommended retrying all throttled or failed writes without checking endpoint support.

​

Changed

• Replaced the customer guide’s legacy subscription snippets with a tested customer, address, and order workflow, and rebuilt the integration-testing guide around isolated native-engine tests.
• Added a CI check that runs nine JavaScript programs extracted directly from eight embedded onboarding pages, including the cross-process persistence example.
• Rebuilt the inventory quickstart as a runnable stock-reservation example and added a payment/refund quickstart with retry checks, balance guards, and explicit provider-integration boundaries.
• Replaced the SDK setup guide with versioned Node/Python installation checks and a clear distinction between embedded libraries and hosted REST clients.
• Rebuilt the order and return quickstarts as complete local programs with assertions for stock rejection, reservation release, return retries and lifecycle state.
• Added a tested embedded-commerce quickstart using the published npm package: create a customer and order with decimal-string prices, then verify persistence from a second process.
• Updated iCommerce MCP onboarding with an explicit server executable, absolute database paths, tool profiles, connection checks and write-configuration guidance.
• Reorganized the documentation landing page around tasks and interfaces, with direct links to working integration guides and the API directory.
• Replaced the developer guide’s incomplete SDK examples with standalone JavaScript, Python and curl requests, expected results and service-specific error handling.
• Reworked the ResponseCX quickstart to capture the returned agent ID, verify configuration, explain activation and deactivation, and diagnose validation and idempotency errors.
• Added a reproducible API support checklist.

​

Fixed

• Corrected inventory method signatures and reservation-confirmation behavior; documented pending-refund balance usage, separate order payment state, and the Node 1.35.1 refund-completion limitation.
• Corrected the Node order and return lifecycle signatures and documented the 1.35.1 return-disposition limitation and the boundary between state changes and external fulfillment or refunds.
• Corrected generated ResponseCX examples to preserve fixed schema values and show realistic agent creation and activation requests. Added generator regression checks for explicit examples and shell quoting.
• Fixed package-name validation for version-pinned install commands, with a regression check that still rejects unknown packages.
• Corrected the AI-agent getting-started page’s service URLs and retry-key example.
• Converted the favicon to PNG to match its filename and allow local preview favicon generation.

​

Changed

• Updated the documentation logo in light and dark themes to match the main StateSet app navigation.

​

Added

• A real 404 page (instead of a silent redirect home) that routes lost readers to the API router, host status, and Getting Help; last-modified timestamps on every page; a search prompt; dark branded social-card thumbnails.
• Launch a brand that runs itself — the capstone guide: onboard across six services, connect the commerce stack, put agents on every channel, gate what moves money, measure outcomes, and close the improvement loop. Six phases, every one linking to a walkthrough verified against the running service, with a plain table of what “autonomous” means and where the human controls sit.
• UPSTREAM-FINDINGS.md — every verified server-side issue from the grounding pass, one paste-ready section per repo.

​

Added

• Sandbox API surface tripled — the tab now documents the 49 operations the live controller serves (executions, pause/resume, fork/promote, the full file toolkit, desktop control, secrets, API keys, usage, webhooks), from the server’s own /openapi.json merged with six routes it serves but does not publish (verified live). Was 17. The tab gains an overview, and the host-status and API-history pages now refresh themselves weekly via CI.

​

Fixed

• Every documented SDK call now resolves against the shipped SDKs — JavaScript 415/415 (was 47%), Python 32/32 (was 57%), stateset-agents imports 49/49. Samples for unshipped surface are rewritten to the real method name or to the REST call, with a note where SDK support has not shipped.
• Voice and Sandbox end-to-end guides — Put an agent on the phone (web call from the browser first, then a real number, with the DNC/quiet-hours gates explained) and Give your agent an isolated runtime. Every engine now has an end-to-end guide.
• Getting Help — where to ask, what to check first, and the dispute process.
• The API router’s key table gains a “where the key comes from” column — the mint endpoint where one exists (Voice, Temporal, EDI, Computer Use), the app or onboarding flow where it doesn’t — and a deprecation-policy bullet (the Deprecation header, and per-API history pages as the migration diff).

​

Fixed

• Zero D-grade reference pages remain. Voice’s 17 untyped responses are transcribed from the handlers (four list endpoints the spec declared as bare arrays actually return {ok, tenant, count, …} envelopes; the two WebSocket routes now describe their real frame protocols). The final ten stragglers across NSR, Computer Use, EDI, Temporal and Agents are typed from source — including NSR’s model export documented as the honest 501 it is, and Computer Use’s transcript/export as the text/markdown and text/csv they return.
• Pages whose response is a non-JSON scalar (a transcript, a CSV export, TwiML, an event stream) state the content type instead of warning that “the server returns JSON”.
• Every per-API history entry links to the reference page it added.

​

Fixed

• Install commands for @stateset/icommerce-skills now lead with the repository (the package is not yet on npm), and every page whose example imports an unpublished SDK says so up front instead of letting the install fail.

​

Added

• Brand Onboarding — the one flow that provisions a brand across six services (WorkOS, Knowledge Base, ResponseCX, Temporal, NSR, Voice), documented for all four ways in: the live self-service portal at onboarding.stateset.com, the operator CLI, the provisioning API’s draft → approve → launch → verify sequence, and the MCP server that lets an agent onboard a brand itself. Linked from the quickstart, the API router, and the For AI Agents pages.

​

Added

• Two cross-engine journey guides — the “close the operation, not just the conversation” thesis as running requests: From “where is my order?” to a closed operation (ResponseCX → Sync → Temporal → NSR) and A retailer’s 850 to cash (EDI → Commerce Engine → Temporal → EDI → AR). API calls are drawn solid, integration seams dashed and named as your code.
• Errors across the engines — every engine’s error envelope side by side, grounded in each server’s source, with the five error families and what an agent should do with each.
• API Reference router — the old two-surface intro is now a task → API router: what you want to do, which of the twelve surfaces does it, the host, the auth header, and the env var its examples use.
• Host status page, generated by probing every documented host (scripts/build-hosts-page.mjs) — a 503 a reader hits is now something the docs told them to expect.
• Per-API history pages (scripts/api-history.mjs) — operations added or removed in each API surface, derived from the spec’s own revision history, listed in every tab’s Overview group.
• CONTRIBUTING.md — how the generator, overlays, exclusions and checks fit together.

​

Changed

• Every generated curl now uses a per-engine env var ($RESPONSECX_API_KEY, $STATESET_SYNC_API_KEY, …) instead of YOUR_API_KEY — export once, paste everything on that tab.
• Set L2 and Sequencer pages carry an Early preview banner.
• The Documentation group “StateSet One” is now “Using the APIs”.

​

Added

• Five end-to-end guides, one per newer engine, each grounded in the server’s spec and source: Sync a Shopify store with NetSuite, Your first EDI 850 → 855 → 856 → 810 cycle, Gate a refund with a verified decision (NSR), Run a durable order workflow end to end (Temporal), Automate a back-office tool that has no API (Computer Use).

​

Fixed

• Every documented stateset CLI invocation is now one the CLI accepts (63 rejected flag/subcommand pairs rewritten to the real grammar or to the REST call).
• EDI overview: the outbound 810 example showed a JSON response; the server returns the wire-ready X12 interchange.
• Temporal: GET /v1/workflows/order-fulfillment/{id}/status documents the bare "<status>:<phase>" string the handler actually returns.

​

Removed

• 349 legacy StateSet One API pages (api-reference/<resource>/*, response-api-reference/*) that targeted a host with no deployment. Each URL now redirects to the iCommerce or ResponseCX page for the same endpoint where one exists, otherwise to that tab’s overview. The api-reference/v1/* pages are kept.

​

Changed

• Realistic examples on every generated page. Request and response examples now use values a reader can recognise — emails, SKUs, addresses, plausible amounts — chosen from the field name, with the spec’s own examples taking precedence.

​

Fixed

• Temporal Engine (137), Sync Server (74) and ResponseCX (17) responses typed. Each server’s OpenAPI left these as untyped objects or referenced schemas it did not export; the fields are now transcribed from the handler that returns them (spec/overlays/*.json, merged at generation time so a fresh export never wipes them). Pages whose handler answers 201/202/204 no longer list a spurious 200. Reference pages grading D went from 231 to 12.

​

Changed

• Introduction and Why StateSet aligned to the same messaging — the competitive framing now says what the matrix says (helpdesk AI automates interactions; iPaaS is excellent orchestration; StateSet sits above and across your ERP, not instead of it), with the objection-handling answers added to the hard questions.
• StateSet Overview rewritten to the 2026 messaging matrix — “StateSet is the iCommerce company; the iCommerce Engine is the execution layer for AI commerce” — with the Engine’s parts mapped to their API tabs, the six pillars, audience messaging, and where each part stands today. Unsourced metrics are gone; the figures that remain are the scoped ones from the evidence ledger.

​

Added

• Agents API tab — the 35 routes of the stateset-agents FastAPI gateway (agents, conversations, training jobs, /v1/messages, /v1/chat/completions, observability), exported from the app’s own OpenAPI. Self-hosted; no StateSet deployment exists.
• Chat Widget API tab — the 39-route contract a backend implements for the Response chat widget (POST /api/chat with its SSE event stream, threads, catalog, cart, settings, review, analytics) plus the widget admin server’s API. Authored from the repository’s APIS.md and docs/ADMIN-API.md; customer-specific flows are left out.
• Sync Server API tab — 86 tenant endpoints from the spec api.sync.stateset.com publishes; the 39 admin operations are excluded.
• Computer Use API tab — the 90 versioned v1 endpoints, exported from the platform’s own FastAPI app. The production host answers 503 today; the pages document the contract.
• NSR API tab — 259 endpoints from api.nsr.stateset.com’s published spec, in 37 groups; the docs previously covered 13.
• EDI API tab — 116 endpoints from edi.stateset.com’s published spec.
• For AI Agents dropdown — every MCP server (15) and every skill (8) in one place. Four skills published for the first time: NSR, Computer Use, Voice, Sync Server.
• Skills overview — the .md URL convention, how to load a skill into an agent harness, and how skills pair with MCP servers.
• Sandbox: webhooks page — 24 events, subscription, payload envelope, HMAC verification in three languages.
• ResponseCX: the approvals queue — where a money-moving action waits on suggest.
• NSR: SDKs page (Node, Python, Rust, Go) and the v0.9.4 proof-verification changes.
• stateset-agents: the six additional RL objectives and the four rented-GPU commands.
• iCommerce: the 24 stable invariant error codes an agent can branch on, and the two ICP conformance profiles.

​

Removed

• Five Sync Server endpoints built for individual accounts (/sync/capacity*, /sync/luminous*) are not shown; spec/exclusions.json keeps them out of regeneration.

​

Changed

• The StateSet One API tab is no longer in navigation. All 509 of its pages target api.stateset.com, which has no deployment; the pages stay at their URLs until a host exists, and the iCommerce tab is the reference for that surface.
• Navigation is now three dropdowns — Documentation, API Reference, For AI Agents — with the tabs inside each, instead of eleven top-level tabs.
• The Resources tab is gone; its five app pages live under Documentation → Apps & Consoles.
• curl examples on NSR and Temporal Engine pages lead with the X-API-Key header each service lists first; Bearer is still accepted.
• Computer Use authenticates with X-API-Key, not Bearer; the MCP page’s host corrected to api.computer.stateset.app.
• EDI pages point at edi.stateset.com — their paths were already right, their host was not.

​

Fixed

• A generated page that did not compile: a markdown list inside a field description. Field descriptions are single-line now.
• NSR’s spec declares four query parameters as path parameters; the generator reclassifies any “path” parameter the URL cannot carry.

​

Added

• Voice API tab — 65 tenant-key endpoints from api.voice.stateset.com’s published spec.
• Sandbox API tab — 17 endpoints from the controller’s spec.
• Temporal Engine API tab — 170 endpoints from the engine’s spec at api.workstream.stateset.com; nine of its eleven workflow families had never been mentioned.
• iCommerce invariant error codes and ICP conformance profiles; sandbox webhooks; ResponseCX approvals queue; NSR SDKs and 0.9.4 changes; stateset-agents algorithms and rented-GPU commands.

​

Changed

• The sandbox host is api.sandbox.stateset.app; 92 places said .com, which does not complete a TLS handshake.
• Pages whose spec types the response as an untyped object now say so instead of showing a blank Response section (164 across five specs).

​

Fixed

• 22 broken StateSetSandbox imports — an earlier casing pass had renamed the SDK’s class.

​

Changed

• The Quickstart runs against response.stateset.com/api/v1, the live ResponseCX API. Every claim on it is read from the server’s spec. Its previous version made three requests against a host that serves nothing and one against an endpoint that never existed.
• Four long “quickstarts” (876–2,361 lines) retitled as guides and moved to their topic groups; the two iCommerce quickstarts distinguished by title and cross-linked.
• sdk-installation lost 1,643 lines describing Ruby and PHP SDKs that are published nowhere.

​

Fixed

• Twelve ResponseCX endpoint pages retired for contradicting the published spec (rules and functions are agent-scoped and have no delete; responses are read-only). The Rules concept page’s example used name/type where the schema requires rule_name/rule_type.
• Two Cloud “update” pages that were copies of their create pages.

​

Added

• ResponseCX API tab — 38 endpoints from the spec response.stateset.com publishes, with a playground on each. Every page names the scope its key needs and its per-minute rate limit.
• iCommerce API tab — the 513-page reference moved out of API Reference into its own tab.

​

Changed

• The iCommerce spec caught up with the engine: 10 new endpoints (warehouse bins, return dispositions), and monetary amount re-typed from float to exact-decimal string.

​

Fixed

• Five ResponseCX agent pages with the wrong host, prefix and method, and one endpoint that does not exist — redirected to the generated pages.
• A misspelled endpoint (/v1/attritbutes) and nine pages whose path disagreed with their own directory.

​

Fixed

• 506 Go, Java and C# strings written with JavaScript’s single quotes — compile errors in all three languages. 82 missing Go trailing commas. A Rust metrics example using lazy_static syntax the engine no longer uses, with two metric names wrong. Two Ruby blocks using capitalised locals.

​

Fixed

• 78 GraphQL blocks that did not parse — mostly a stray backtick left from a template literal; four genuinely malformed queries.
• Two mermaid diagrams that rendered as an error box (parentheses in unquoted labels).
• A section of the sandbox production guide swallowed by an unclosed YAML fence.
• 25 page titles showing { instead of {.
• Eight TypeScript blocks hiding a JSON response inside a code fence.

​

Added

• Roughly 200 MCP tools documented for the first time across six servers; every count verified against a live tools/list.
• scripts/check-mdx.mjs — every page is compiled with a real MDX parser in CI.

​

Changed

• Theme aligned to stateset.com’s design token #1A62FF; the light-mode accent had been a near-black.

​

Fixed

• 359 code samples that could not run: 341 X.method({ 'id' }) calls, 17 misplaced semicolons, one const return.
• Three pages that did not render at all.
• Numbers that had drifted: 203 tables documented as “70+”; 155k lines documented as 55k; 719 methods as “671+”.
• 64 JavaScript and 19 Python SDK calls corrected against the live clients.

​

Fixed

• Query parameters documented as request bodies; descriptions naming the wrong resource; ids documented twice.
• Eight pages the live site had served until July were 404ing; docs.json gained its first redirects.

​

Added

• A StateSet Agents section; the ResponseCX autonomy dial; Sync Server connectors and troubleshooting; NSR console surfaces; Voice translation calls.

​

Fixed

• A page told readers a shipped product did not exist, under a package name on no registry.

​

Fixed

• Every part of the request contract had defects — auth scheme, URL, method, path and query parameters, request body — and so did the response: field types contradicted by their own examples, fields documented flat that the response wraps. Each class is now enforced by a check.

​

Fixed

• An auth header the API rejects, on 297 pages. 192 curls calling an endpoint other than the one documented. Truncated request bodies, a Python SyntaxError, JSON blocks that were not JSON.

​

Fixed

• An auth header declared nowhere; curl examples calling a different endpoint than the page documents; install commands naming packages that do not exist; 572 pages whose response section named a status code and no field.