Prodgy Overview

Latest Version v0.0.14 Stable

Changelog v0.0.14 History

PBIs

Unified Prodgy Assistant meeting-join notice with fixed links

New The privacy/LGPD consent notice the Prodgy Assistant posts to the chat when it joins a meeting in Microsoft Teams is now the same message whether the bot was summoned by email invite or by the @Prodgy join command. The Privacy Terms and Terms of Use links now derive from the same URL base, fixing the wrong links that appeared on the @Prodgy join path. The message language still follows the language of whoever triggered the assistant.

Sorting on the Knowledge Base documents table

New The Knowledge Base documents table is now sorted by Updated at from newest to oldest by default, and you can sort by any column with a single click on its header (Name, Updated at, File size, Status and Quality) — click again to toggle the direction. Sorting is applied on the server, so it covers the entire collection and not just the current page.

Workspace settings available to regular users (Users and Security tabs restricted to admins)

New Users with the User role can now open Workspace Settings from the gear in the workspace selector and use the General, API Tokens and Connected Apps tabs. The Users and Security tabs remain admin-only — and that restriction is now enforced on the backend as well (member management and workspace deletion require an administrator role), not just hidden in the UI.

Force the chunking strategy on Knowledge Base nodes

New Knowledge Base workflow nodes (Save / Upsert in storage) now expose an optional Chunk strategy field to override the automatic content-based chunking — for example forcing Single block to keep a document whole. It defaults to Automatic, so existing agents and workflows are unaffected, and it is not exposed in the manual upload screen. The Meeting Scriber now stores the meeting summary as a single block.

Stop a workflow when a critical node fails (halt_on_failure)

New Workflows can now mark a node with halt_on_failure so a hard failure stops the run immediately, preventing downstream steps from acting on missing data (for example, posting an empty review or casting a default vote). The existing continue_on_failure (soft-fail) still takes precedence.

Configurable Self-Learning block for the Code Reviewer

New The workspace context block where the Code Reviewer stores the coding standards it learns is now configurable per install. The field appears only when Self-Learning is enabled and defaults to the previous behavior, so existing installs are unaffected.

Reliable answers from spreadsheets uploaded to the knowledge base

New Spreadsheets (xlsx/csv) uploaded to the knowledge base now produce exact answers to analytical questions — totals, sums, averages, counts and breakdowns by column (for example, sales by segment and country). Instead of guessing from text, the assistant runs a real query over the spreadsheet data. It works in the Playground, in communicators (Slack/Teams), in workflows and for MCP clients. When there is more than one spreadsheet, the system picks the right one or returns the result per spreadsheet, without summing data of different natures (distinct currencies/definitions).

Robust support for complex spreadsheets (merged cells, multi-level headers)

New Spreadsheets formatted for human reading — with merged cells, multi-level headers, section-title rows or unnamed columns — are now handled safely. The system tries to reconstruct the data automatically (including with your organization’s AI) and, when it cannot do so safely, indexes the sheet in descriptive mode (searchable by meaning) instead of risking wrong numbers. No rows are lost in the process. When a sheet falls into descriptive mode, the document’s Quality tab in the Knowledge Base shows a notice explaining what it means and when it happens.

Meeting Scriber posts the summary to the meeting chat

New Besides email, the Meeting Scriber now posts the meeting summary back to the source chat (for example, Microsoft Teams). The message is laid out for chat reading (overview, highlights, next actions with the responsible person when mentioned) and written in the meeting’s language.

Meeting summaries use the workspace context

New Meeting summaries are now generated with the workspace context (product, audience and conventions), producing more accurate overviews, action items and requests directed to Prodgy.

Meetings managed as documents in the knowledge base

New Processed meetings (summary and transcript) are now stored and managed like any other document in the knowledge base — searchable, viewable and removable from the Knowledge screen.

Self-learning Code Reviewer

New The Code Reviewer can now learn from its own reviews: after a review with relevant findings it distills generalizable coding standards and saves them to a workspace guidelines block, so future changes — and AI coding assistants — follow them.

Code Reviewer: product context, documentation and work-item feedback

New The Code Reviewer now understands the workspace’s product context, can require documentation updates for business or core changes (documentation gate), and can comment on the related work item about scope beyond the acceptance criteria and operational requirements. Each check is optional.

Social sign-on when accepting an organization invite

New Users invited to an organization can now accept the invite and sign in with social login (for example, Google/Microsoft), without having to create a separate password.

Configurable feedback popup

New The in-app feedback popup is now configurable: it can be turned on or off globally and triggered after a chosen number of sessions, so feedback is requested at the right time without being intrusive.

Real-time execution status on the executions screen

New The executions screen now shows the status of each node in real time for agents running on the new engine, so you can follow a run as it happens instead of waiting for it to finish.

Re-running a meeting no longer duplicates data

New Re-processing the same meeting is now idempotent — it updates the existing summary and transcript instead of creating duplicates — and the Execution Details screen gives clearer diagnostics when a run needs attention.

Bugs

Spreadsheet questions in the knowledge base chat returned only a sample (or nothing)

Bugs Fixed analytical questions about spreadsheets uploaded to the knowledge base not being answered over the full data once deployed — the assistant would only return the spreadsheet’s column overview and a sample of the first rows, come back empty, or appear to hang, instead of running the real query across all rows. Questions like “which projects use technology X” now return complete answers in the Playground, in communicators (Slack/Teams), in workflows and for MCP clients. Spreadsheet questions are also answered regardless of the question’s language — for example, asking in Portuguese about a spreadsheet whose columns and values are in English. This restores the Reliable answers from spreadsheets uploaded to the knowledge base capability listed above.

Code Reviewer did not remove its webhook when uninstalled on non-Azure providers

Bugs Fixed uninstalling a Code Reviewer agent not removing the webhook it had registered when the agent was connected to a non-Azure DevOps provider (for example, Bitbucket). The uninstall always ran the Azure DevOps cleanup routine, which never matched the repository’s webhook and left it active. The uninstall now picks the cleanup routine that matches the agent’s integration, so the webhook is removed on the correct provider.

Magic Fill auto-fill now works with any file type

Bugs Fixed the document import Magic Fill auto-fill only accepting PDF files and failing with an “Only PDF files are allowed” error for spreadsheets, CSVs and other formats. Magic Fill now accepts any file type — for non-PDF documents it suggests the form fields (name, description, facets, metadata) from the filename and the values already filled in, while PDF content extraction keeps working as before.

Knowledge Base overview generation ignored existing content and the user's prompt

Bugs Fixed the “magic” overview generation (Sparkles button) on the Knowledge Base Context screen producing content from scratch — it silently skipped the RAG retrieval (failing on credentials it didn’t need and swallowing the error), so existing Knowledge Base content was never used. The retrieval now reuses the platform Core LLM (no agent/workflow integration credentials), failures are logged instead of silenced, and the prompt the user types in the generation dialog now drives the Knowledge Base search, so retrieved documents reflect what the user actually asked for.

Rescheduled meeting dropped while the Assistant was in the call

Bugs Fixed the Prodgy Assistant dropping a rescheduled meeting when it had already joined the original call. Because Teams reuses the same join link, the scheduler matched the new occurrence to the call already in progress (by URL only) and discarded it, so the meeting never reappeared in the Queues. The scheduler now also considers the start time: a reschedule to a later time gets its own slot in the queue (and the Assistant rejoins at the new time), while a duplicate invite for the meeting currently in progress is still merged into the active call.

Recurring meeting occurrences scheduled at the wrong time for non-UTC organizers

Bugs Fixed recurring meetings being scheduled (and shown in the Queues) shifted by the organizer’s UTC offset — e.g. a 10:30 PM (GMT-3) series landed 3 hours off. The recurrence expansion read the meeting’s UTC clock as if it were the organizer’s local time. It now converts the start instant into the organizer’s timezone before computing each occurrence’s local day and time. Single (non-recurring) meetings and the webhook payload’s local timestamps were already correct and are unchanged.

Cancelling one meeting could remove a different meeting at the same time

Bugs Fixed a meeting cancellation removing the wrong meeting from the Prodgy Assistant queue. When a cancellation could not be matched to its own calendar event, it fell back to matching only by start time and sender — so a different meeting scheduled for the same time (even a freshly scheduled one) could be removed and disappear from the Queues. The cancellation now also checks the calendar event id and the meeting title, so it only removes the meeting it actually refers to.

Re-sending an invite after cancelling a meeting did not re-add it to the queue

Bugs Fixed a recurring meeting that had been cancelled never coming back into the Prodgy Assistant queue when its invite was re-sent for the same time. The cancellation was recorded permanently, so the re-invite was silently skipped even though the recurrence pattern was registered — the meeting showed under recurring patterns but never reached the pending queue and the Assistant never joined. A fresh invite (or reschedule) for the same time now clears the prior cancellation and schedules the occurrence normally.

Agent executions list columns loaded progressively

Bugs Fixed the agent executions screen showing only some columns at first (e.g. Status) while others (Started, Completed, Agent version) stayed empty for a few seconds, even after a refresh. The list was being assembled from two overlapping data sources that returned different subsets of fields and were merged progressively. The backend now returns the start and completion timestamps in the initial response, so the whole row renders complete on the first paint.

Browser refresh keeps you on the current page

Bugs Refreshing the browser (F5) no longer bounces you back to the organization dashboard — you now stay on the page you were on (e.g. Agents, Knowledge, Playground). The issue happened because the view scope and the selected product were briefly recomputed before the organization and product data finished loading, which incorrectly triggered a redirect. The app now waits for that data to settle before deciding any navigation.

Code Reviewer approved pull requests without analyzing them when the AI provider failed

Bugs Fixed AI agents treating an error returned by the LLM provider inside an HTTP 200 response (with empty content) as a successful empty answer. This previously caused the Code Reviewer to post an empty review and fall back to an approval vote. Provider errors are now detected and reported as failures, and transient provider/network errors are automatically retried with backoff.

Code Reviewer failed on pull requests that delete files

Bugs Fixed the automated code review failing when a pull request removed a file. Deleted files are now skipped in the review loop, so reviews complete normally on pull requests that include deletions.

Playground kept showing the previous workspace's chat after switching workspaces

Bugs Fixed the Playground still showing the chat list and the open conversation from the previously selected workspace after switching workspaces. The chat list, the selected conversation and the in-memory history cache are now reset and re-scoped to the active workspace.

Facet filters on knowledge base search returned no results

Bugs Fixed facet filtering (for example, type:meeting) on knowledge base search, which returned empty even when documents had the facet. Facets set on upload and on the “Save to Knowledge Base” node are now propagated to the indexed chunks, and a backfill fixes the existing content.

Inviting a user whose email already exists

Bugs Fixed an error when inviting a user with an email address already present in the platform; the invitation now resolves to the existing account instead of failing.

Wrong characters in the Playground PDF export

Bugs Fixed garbled characters (accents and special symbols) in the PDF exported from a Playground conversation.

MCP get_my_profile returned an empty language

Bugs Fixed get_my_profile (MCP) returning an empty language for users whose locale is a 2-letter code; the user’s language is now resolved correctly.

@Prodgy join used the wrong locale and paraphrased the privacy notice

Bugs Fixed the @Prodgy meeting join using the Teams client locale and the AI paraphrasing the privacy notice; the notice is now shown verbatim in the correct language.

OpenRouter not listed when installing the Meeting Scriber

Bugs Fixed the OpenRouter integration not appearing as an option while installing the Meeting Scriber agent.

Incomplete meeting summary saved to the knowledge base

Bugs Fixed the meeting summary being stored incomplete in the knowledge base; the full summary is now saved.

Teams chat did not reference the user who asked

Bugs Fixed Prodgy responses in Microsoft Teams not correctly referencing the user who sent the prompt.

New workspace missing its default dashboards

Bugs Fixed a newly created workspace not showing the default dashboards right after creation.

Email-invite privacy notice hard-coded in Portuguese

Bugs Fixed the privacy notice posted by the Prodgy Assistant on email invitations being hard-coded in Portuguese; it now follows the recipient’s language.

Thinking indicator ignored the user's language

Bugs Fixed the “thinking” indicator in Teams chat ignoring the user’s preferred language and falling back to a heuristic.

Scheduled Meetings dashboard showed no data

Bugs Fixed the Meeting Scriber “Scheduled Meetings” dashboard showing no data.

Transcript timestamps shown in UTC

Bugs Fixed meeting transcript timestamps being delivered in UTC; they now reflect the meeting’s local time.

Node titles missing in Execution Details

Bugs Fixed node titles not appearing in Execution Details for agents running on the Prodgy executor.

Processing logs missing on the new engine

Bugs Fixed processing logs not being shown for agent executions running on the new engine.

---