Skip to content

Releases

Release notes and changelog, populated going forward.

  • The assistant can offer to report a failure. After a server-side failure of one of your tools (a 5xx, a network error or timeout, an open circuit), the assistant drafts a report and opens a consensus dialog. The user edits the text, switches off any attachment (conversation excerpt, page path, failed tool calls) and approves or rejects it. Only an approved report is stored, exactly as previewed. It is offered once per problem per conversation. Denials, invalid input, feature requests and the assistant’s own mistakes get no offer. See Problem reports.
  • /problem. Users can start a report themselves from the composer’s / menu. The assistant explains first when the failure does not look like a fault of the application, and files anyway when the user insists.
  • A Problems page. Reports are grouped by assistant, server, tool and error category, with filters for status, source and the assistant’s assessment, and Resolve, Dismiss and Reopen. Resolved and dismissed problems’ reports are deleted after 30 days (DIOSC_PROBLEM_REPORTS_PURGE_DAYS). See Problems.
  • Webhook sinks. Forward report.created, problem.created, problem.status_changed and report.erased to a URL, signed with HMAC-SHA256 (X-Diosc-Signature), with optional JSON templates, up to 6 attempts and a delivery log that never holds the payload. There is no default sink: nothing leaves DioscHub until you add one. Sink URLs follow the MCP egress policy; secrets need MCP_AUTH_MASTER_KEY.
  • Metrics. diosc_problem_reports_total, diosc_problems_open, diosc_problem_webhook_deliveries_total and diosc_problem_webhook_delivery_failures_total. See Observability.
  • Privacy. Reports are exported with a subject (problem-reports.json), deleted by user and session erasure (the problem keeps its count; sinks get report.erased), and survive the retention window’s deletion of their conversation. Free-text personal data is not detected by redaction; the user’s consent is the control. Privacy-notice wording is in Data handling.
  • New permissions: problems:read (Viewer and up) and problems:manage (Operator and up). New license capability: problemReports, in Pro and Enterprise.

After upgrading, problem reports are on for every assistant on a Pro or Enterprise license, existing assistants included. The first admin who can read assistants sees a one-time notice with Review settings, Turn off for all assistants and Dismiss. To upgrade with reports off, set PROBLEM_REPORTS_DEFAULT=off before the first start of the new version; each assistant then stays off until an admin turns it on. Five database migrations run automatically. See System settings.

  • Configuration transfer carries the per-assistant setting and the deployment default. An export from an older version leaves the target’s value unchanged. See Problem reports travel.
  • Do not name an MCP tool report_problem (for example instance report, tool problem). DioscHub withholds the built-in for that assistant and logs a warning until the tool is renamed. See the naming collision.
  • Known limits. Whether the assistant offers a report is the model’s judgment: smaller models can miss an offer after a server error, or offer in plain text, which does nothing. /problem always works. Webhook events are in-process: an instance that stops between storing a report and queueing its deliveries loses those deliveries.

Data subject rights and a retention window on every tier

Section titled “Data subject rights and a retention window on every tier”
  • Subject access export. An Owner or DPO can export everything DioscHub holds about one person within one assistant as a zip of JSON files (conversations, files, analytics, feedback, approvals, usage, error records, preferences, audit and profile) and their original uploads, with a README of what is left out and why. It answers access (Art. 15) and portability (Art. 20) requests. The download needs the admin session, lasts 24 hours, and is deleted when the person is erased. See Subject access export.
  • Retention window. Conversations and users’ files can now be deleted automatically after 30 to 3650 days without activity. It is off by default, set on the Data Privacy page, and a change that deletes sooner waits 7 days and needs a typed confirmation. Pinned conversations are kept unless you choose otherwise, and where session history is licensed the widget warns users 7 days before a conversation goes. Anonymous visitors’ conversations expire by the same rule. Approval and error logs never outlive the window. See How long conversations are kept.
  • Waiting for BI. Where Business Intelligence is licensed, the window can hold conversations BI hasn’t classified yet for up to 30 more days.
  • Every tier. Export, erasure and the retention window need no licence capability, Free included.
  • Changed: subject lookup by user id on the Erase a subject tab refuses anonymous visitor ids (anonymous, anon_…). The retention window is not carried by configuration transfer.

Spreadsheets the assistant can read, outline and query

Section titled “Spreadsheets the assistant can read, outline and query”
  • Three table tools. describe_workbook shows a workbook’s sheets, table regions, merged and hidden cells, pictures, charts and a preview; load_table loads one A1 area (or a whole CSV) as a typed table; query_table runs a read-only SQL SELECT over it. They replace list_tables and describe_table, which are gone. They work on .xlsx, CSV and TSV files in a user’s own library, not on knowledge-base files. See Spreadsheets and the table tool limits.
  • Pictures in workbooks. Up to 20 pictures per Excel workbook are extracted when it is indexed, and the assistant can look at them with read_file. They are deleted with the workbook. See Pictures in workbooks.
  • Better spreadsheet search. file_search and kb_search hits from a spreadsheet carry the A1 range, and start with a Columns: line naming the headers. The header names are also embedded with each chunk.
  • A new Excel reader. Workbooks are read by DioscHub’s own streaming reader: formulas as their cached values, dates as ISO dates, and the same result whatever order the workbook’s parts are stored in, so a workbook whose parts are stored in an unusual order no longer ends failed / corrupt.
  • Table tool output is framed as file data, like read_file; see Framed as data, not instructions.
  • Quota. Cached tables and workbook pictures count toward a user’s storage quota and are checked against it when stored, and an upload over the quota first makes room by deleting cached tables. See How it is enforced.
  • Removed: background-model structure detection. Spreadsheet structure is now found when the file is indexed, and the background-model refiner is gone. The TABULAR_LLM_STRUCTURE_DETECTION environment variable has no effect. A configuration export that still carries the tabularStructureDetection field imports as before, and the field is dropped. Usage recorded by the refiner keeps its Spreadsheet detection label.

After upgrading, spreadsheets indexed before this release have no outline and no extracted pictures: describe_workbook tells the assistant to re-index them. Run a bulk Re-index of chat files of kind table, or let a user re-index one file with POST /api/files/:id/reindex, which now accepts a ready spreadsheet or CSV without an outline. The widget’s Re-index button doesn’t cover that case yet.

Removed images are no longer sent to the model. An image already shown in a conversation, a workbook picture included, used to be sent to the model again on later turns after its file was deleted or redacted. Each model call now checks the file first and sends a short “no longer available” note instead. See Read file content stays in the conversation.