Skip to content
  • There are no suggestions because the search field is empty.

Release Notes - 3.14.0

Rhino 3.14.0 delivers browser automation for customer web apps, a self-aware Copilot that answers product questions from protected built-in guides, Copilot decision points in Trace Explorer, encrypted solution packs, routines that run on any Knowledge Reserve content, OIDC claim-based admission, keyless credentials for OpenAI-compatible providers, and the retirement of Document Studio and legacy UAN code.

 

Highlights
  • Browser Automation (Web Apps): Copilot and Tasks can open a customer's web app in an isolated, hardened Playwright runner to map its screens, compare as-is and to-be apps, make and verify changes through the UI, and upload Knowledge Reserve files. Every run leaves evidence and recordings. Off by default.
  • Self-Aware Copilot: Copilot recognizes "how do I" and product questions and answers them from protected built-in platform guides. It knows who is asking and which features are on, and can explain why an External MCP tool isn't available.
  • Copilot Decision Points in Trace Explorer: Every question Copilot asks a person is shown with the answer and the work that led to it, and can be downloaded as questions.zip. Individual skill answers can be pinned and unpinned.
  • Encrypted Solution Packs: Packs can be sealed and imported with the customer's license key. Unencrypted packs install at the scope of the screen that imported them.
  • Routines on Any Knowledge Reserve Content: Routine runs can target any project KR folder or file (docs, generated, config), not only app assets.
  • Knowledge Reserve Images and Long Documents: Copilot, tasks, and routine agents can view images stored in the KR, and people can preview them in the KR viewer. Heading-dense long documents are indexed to the end.
  • Enterprise Identity and Model Authentication: OIDC claim-based admission lets SSO users in on first login. OpenAI-compatible providers can get credentials from Bedrock IRSA or Azure Managed Identity instead of a stored key.
  • Platform Retirement and Hardening: Document Studio and legacy UAN code are removed, a critical folder-deletion authorization flaw is closed, and a broad set of dependency vulnerabilities is patched.

 

Rhino 3.14.0 introduces Web Apps, a new admin capability alongside API Connections and External MCP. A Web App is a customer application that Copilot and Tasks can operate through a real browser.

  • Isolated runner: Browsers run in a separate, hardened browser-runner container, not in the backend. An egress fence blocks private, link-local, and cloud metadata addresses, and only operator-allowed hosts can be reached.
  • Configuration and sign-in: A Web App configuration page registers apps with site-level credentials, including SSO/MFA sign-in.
  • Extraction: Copilot maps an app's screens and fields. A discovery map guides follow-up questions, and browser sessions feed the map.
  • Parity testing: Compare an as-is app with its transformed to-be app.
  • Live changes: Configure an app through its UI and verify the result, including pages drawn on a canvas or as images.
  • Screenshot fallback: When the accessibility tree isn't usable, Copilot navigates from screenshots.
  • Knowledge Reserve uploads: Upload KR files into an app through the browser.
  • Run visibility and evidence: A run card shows connected apps and progress; every step is recorded with screenshots and recordings.
  • Proven on Appian: A Copilot Task configured and verified an Appian app end to end through the UI.

Why it matters: Teams can work with applications that have no API, verify transformed apps against the originals, and keep an auditable record of what Copilot did in the customer's app.

Self-Aware Copilot
  • Copilot recognizes platform guidance questions ("how do I", "what should I use", setup and configuration) and loads a protected built-in guide before answering.
  • Guides cover setting up External MCP, API Connections, and GitHub versioning, and are kept current as features change. Guide screenshots are captured by script and label-checked nightly.
  • Copilot knows the user's role and which features are enabled on the site, so it doesn't recommend features the user can't use.
  • Copilot reports live External MCP tool availability and explains why a tool can't be used.
  • Answers can show app-owned images; outside image addresses are blocked.
Copilot Decision Points and Learnings
  • Decision points in Trace Explorer: Each question Copilot asked shows the prompt, context, options, and response, followed by the work since the previous question: tool actions and Knowledge Reserve documents touched. Inputs and results stay behind a click; protected content is marked unavailable with a locator.
  • Download questions: questions.zip contains one folder per run or chat with a questions.csv and one JSON file per question.
  • Decision points are always on; the copilot_decision_points feature flag is removed.
  • Pin individual answers: Skill answers (learnings) can be pinned and unpinned one at a time from the Trace Explorer artifact and from Copilot, so answers a client doesn't want reused can be excluded.
Copilot Steering and Reliability
  • Follow-ups queued while Copilot is answering now keep their image attachments, and the images reach the model when the message runs.
  • A claimed follow-up is settled on the server when its stream starts; the recovery handoff no longer runs the next question unexpectedly.
  • A follow-up queue 404 no longer bounces the whole app to the global "page not found" screen.
  • Copilot asks for External MCP authentication only when a tool call needs it.
  • External MCP tool names always carry their connection prefix, so two connections that expose the same tool name each reach the right server (also shipped in 3.13.1).
  • External MCP OAuth uses a name-based callback URL, shows a copyable redirect URI, and falls back when discovery fails.
  • The multi-question form card no longer overflows the panel and hides its submit buttons.
  • Fixed unexpected conversation endings, event-loop blocking in the backend, and Copilot failing to answer basic questions about a project's routines.
Solution Packs
  • Encrypted packs: Support-authored packs can be sealed and imported with the customer's license key, so a pack file handed to a customer is no longer readable plaintext. Sealed packs are accepted in the install dialog and installed through the existing installer.
  • Install at the importing scope: An unencrypted pack installs at the scope of the screen that imported it (project or System Reserve), regardless of the scope written in the file.
  • No cross-scope overwrites: A System Reserve install never overwrites content written by a project install, and the reverse.
  • Installing a system pack and then importing Dynamic Objects no longer creates disconnected plugin and project definitions.
  • Promoting an artifact type to System Reserve now brings the Dynamic Object types it binds to, so it renders with data in other projects.
  • A Knowledge Reserve search benchmark pack (judge routine, Benchmark Report artifact, task template) is available for administrators to install. MCP clients can read a finished routine run's Knowledge Reserve usage with read_routine_kr_usage.
Routines and Graph
  • Any KR target: Routine run paths (Run panel, REST launch, Copilot execute_routine) can target any project Knowledge Reserve folder or file, including source/docs/, generated/, and config. Applies to script and DSL routines.
  • Deleting a supporting document, steering file, skill, or generated file or folder cancels in-flight routine runs that targeted it.
  • search_graph with latest_only=True no longer silently drops rows from other writers when two routines write the same Dynamic Object label.
  • search_graph(latest_only=False) on labels with 5,000+ distinct items gives callers a way to recover (also shipped in 3.13.1).
  • The routine run summary reports the real number of links created instead of always 0.
  • Script-path map_agentic no longer injects the entity block twice into each agent prompt.
  • A NUL byte in an extracted entity no longer fails the whole routine run.
  • The Business Schema Consolidator no longer flags valid cross-asset citations as missing.
  • Fixed bulk deletion of large Dynamic Object selections running as a long, spinner-only request.
  • Artifact edits can create or extend Dynamic Object types, so changes to what an artifact shows are backed by the graph rather than artifact source.
Knowledge Reserve
  • Images: Copilot, Tasks, and routine agents can view image files stored in the Knowledge Reserve. The KR viewer previews png, jpg, jpeg, gif, and webp files instead of showing "no text to display".
  • Long documents: Cross-heading passage packing and a passage cap raised from 400 to 1000 mean facts late in heading-dense documents (registers, catalogs) are found. Repeated sibling headings stay attached to their body.
  • Bulk delete: Admins can bulk-delete Supporting Documents, Steering, Skills, and Generated items from the selection menu. System-scoped items are skipped.
  • Relative links in rendered markdown (for example a skill's SKILL.md) open the linked file instead of a 404.
  • Folders inside an uploaded app archive list their contents again.
  • Fixed missing bottom padding on KR buckets.
GitHub Project Versioning (follow-ups)
  • Copilot no longer asks again for the repository and branch a project is already tied to.
  • Commits include only what the project changed since the stored commit, so a file a load can't write no longer blocks commits.
  • Loading no longer refuses changed DSL routines.
  • Mixed connections can no longer read as one account and write as another.
  • Admin message for a blocked app, commit SHA in the reply, no empty commits, and organization repositories in the lookup.
Identity, Connections, and Models
  • OIDC claim-based admission: Sites can configure an OIDC claim and required value; matching users are admitted on first login without an admin enabling them. Denied users see a "not authorized" screen. Behavior is unchanged when not configured.
  • Keyless credentials for OpenAI-compatible providers: Authentication is a configurable delegate. Supported options: a short-lived Bedrock bearer minted from IRSA, an Azure AD bearer from Managed Identity, or an API key. Credentials are refreshed automatically and kept out of logs, errors, and the API catalog.
  • API Connections per-user credentials: Connections can use shared or per-user credentials, mirroring External MCP.
  • GPT-5.6 Luna is added to the model registry and can be configured on a site.
Artifacts

  • Button and color chart primitives in the component library were audited.
Platform, Retirement, and Security
  • Document Studio retired: The Document Studio frontend, api/documents.py, and its three MCP tools are deleted. Staged documents remain readable in the Knowledge Reserve; source tables are kept.
  • Legacy UAN retired: UniversalApplicationNotation readers and the legacy project export route are removed. Stored UAN and UAN JSON Import are kept.