Data-room intelligence

• Roddick searches data-room content directly — combined keyword and semantic retrieval over an indexed store — instead of relying on file summaries alone.
• Answers and drafts cite their sources; citations click through to the source file preview at the cited page.
• Key financials (revenue, EBITDA, debt schedules, cap tables) are extracted from data-room files into a structured store and looked up exactly, instead of re-read from summaries.
• Numeric claims in drafts are checked against extracted facts and source passages; figures that can't be verified are flagged inline.
• Excel models are read natively — workbook structure, cell ranges, formula tracing — rather than as flattened text.

Documents

• Version history: browse earlier versions of a document, review what changed, and restore. Snapshots are captured automatically as documents are edited and before AI drafts, restores, and finalization.
• Document lifecycle: Draft / In review / Final status on every document. Finalizing freezes editing while keeping export, history, and download available.

Notifications & activity

• In-app notifications with a sidebar bell and a notifications page; comment activity on a document notifies the people involved in it.
• A project Activity tab: document creation and finalization, comments, data-room uploads, completed generations, and client views in one feed.
• Documents show when a client last viewed them.

Platform

• Every assistant run records a full trace — tool calls, timing, tokens, cost per step — with an admin viewer to replay runs.
• Cost accounting now covers prompt caching and web search, attributed per run and per step.
• Long conversations manage their context automatically, so extended sessions stay coherent instead of degrading.
• A document eval harness grades generated drafts against firm rubrics — the quality gate for prompt and pipeline changes.

Fixes

• Accepted AI proposals always apply; cases where an apply was silently blocked are fixed.
• Deleting a conversation, note, or a member's access now asks for confirmation.
• Double-clicking or double-pressing Enter no longer creates duplicate clients, projects, or comments.
• Menus, dialogs, and the formatting toolbar stay on-screen at small window sizes.
• Failed document loads and file previews show a clear error instead of an empty page or endless spinner.
• Switching conversations quickly no longer shows the previous conversation's messages.
• Streaming responses render more smoothly, and the app loads faster on first visit.
• "Load older" in the activity feed and run list no longer skips entries.
• Finalizing a document reliably disconnects live editing sessions.
• Tightened access checks across data-room reads, project tabs, collaboration, and streaming.
• Unexpected assistant errors show a friendly message instead of internal details.
• Truncated names reveal in full on hover; dates display consistently.
• Partially failed project-access saves can now be retried.

Charts

• Charts render reliably in environments with a strict content-security policy.
• Chart nodes mount into a dedicated full-width container, fixing sizing on first insert.

Other

• Roddick gate image as the login page background.
• Ask attaches the active selection range as message context, matching Edit.

Charts

• Chart generation in chat responses and documents, rendered inline with automatic light/dark theming.
• Charts included as an editable element in the agent edit flow, with appearance and underlying data revised independently.
• Bar charts annotate every value, with negatives shown in parentheses.
• Charts rendered in PDF document exports.

Document generation flow

• Full-document generation from the project data room, parallel option to blank document.
• Long document lists collapse within the project view.

Editor

• Text alignment (left/center/right/justify) for paragraphs, headings, and table cells/headers.
• Table handling improvements.

Other

• Selection actions streamlined to Ask and Edit.
• Visual and theme refinements across the app.

Changelog begins here. Changes prior to this release are not recorded.

• Upload files to the data room from within Roddick.
• Global search across projects, documents, notes, conversations, and data-room files, in a Cmd+K palette.
• LLC Lab: a thesis workbench and MOIC attribution bridge over selected data-room files.
• Share an individual document with an external user.
• Browse SharePoint sites and drives when linking a data room, instead of pasting a URL.
• Import Word documents from the data room into the editor.
• Meetings: uploads, transcripts, and summaries — transcripts become citable sources.
• Background assistant tasks: long runs continue after you close the tab, with a task inbox and completion notifications.
• Specialist subagents that read large data rooms in parallel without crowding the main analysis.
• Block-level edit proposals: review AI edits block by block instead of as one combined replacement.
• Generation pipelines for long deliverables: research → outline → parallel section drafts → verification → review.
• Self-review: drafts are graded against firm rubrics and revised before delivery.
• Project memory that carries deal context across sessions.
• Skills as executable playbooks — instructions, templates, pipeline, and rubric in one runnable unit.
• Skill refinement suggestions learned from repeated manual edits to generated drafts.
• AI shortcuts on a selection (Challenge, Dig deeper).
• Web research with page fetching and cited sources.
• Exhibits generated from extracted financials; select and edit existing charts.
• Word export on firm templates; PDF cover page and page numbers.
• The assistant's reasoning surfaced as progress during long runs.
• Word and Excel add-ins.
• Microsoft sign-in for external users.

• Read-only fallback when live collaboration can't connect, instead of a blank document.
• Typing responsiveness in very large documents.
• Conversation history paging as the list grows.
• Repeat reads of an unchanged data-room file skip re-downloading and re-processing.
• No duplicate data-room processing runs when triggered twice in quick succession.
• Permission changes apply to already-open editor tabs without a reconnect.
• Automatic SharePoint permission re-sync, replacing the manual admin step.
• Multi-cell AI edits keep inline formatting.
• Cell edits lock only the affected cells, not the whole table.
• Re-inviting a previously removed user without manual cleanup.
• Full keyboard and screen-reader support in dialogs.
• A recovery screen instead of a blank app on rendering errors.
• External users offered only the model options their account allows.