Skip to main content
Version: V0.4.100

​
March 2026

​
Overview

This release expands how SIYA connects to your tools and files: new Microsoft 365 and class-rules integrations, and a clearer file viewer for XML. Context compaction stays accurate across providers, and security is tightened end to end—with fewer false alarms in guardrails.

​
Improvements & new features

  1. Microsoft 365 — SharePoint and OneDrive are available as MCP connections so SIYA can work with your Microsoft content where your admins allow it.
  2. Class rules and steadier MCP — A class-rules MCP server is available, and MCP connection handling is unified so connections behave more consistently across tools.
  3. File viewer — XML files include a tree view so structure is easy to scan without leaving the viewer.
  4. Context and compaction — Token counts respect each AI provider’s rules, so automatic context compaction stays accurate as conversations grow.

​
Bug fixes

  1. Sending messages — You can no longer send multiple messages at once in a way that cancelled in-flight tasks.
  1. File management — Tabs in the file management popup respect the current view mode.
Siya
Siya
  1. MCP stability — A race condition in MCP client lookup is fixed by using a map instead of a plain list.
  2. Prompts and caching — Prompt token counts are more accurate, tool order is stable for caching, and prompt caching is refactored with clearer provider checks, shared logic, and Langfuse visibility.
  3. Live actions — Jump to live appears only when you are reviewing past actions, not during live activity.
Siya
Siya

​
Security

  1. Secrets in answers — Production system prompts include defense-in-depth rules so credentials and secrets are less likely to appear in model output.
  2. Sensitive configuration — Sensitive configuration is not exposed through chat or the REST API in ways users could misuse.
  3. Secret coverage and scanning — More environment variables are treated as secrets; connection-string redaction is fixed; AWS-style keys are included in credential scanning.
  4. Sanitization — Deep sanitization handles non-plain objects (for example dates, buffers, and regular expressions) more safely.
  5. Guardrails and MCP — Guardrails are tuned to reduce false positives, and an MCP prompt-injection issue is fixed.

​
Support

For questions or issues related to this release, please contact your support team or refer to the project documentation.