Changelog & Roadmap

Programmatic tool calling, real document generation, auto-update, and a capability system that stops smaller models denying what they can do.

• Programmatic tool calling: the agent can write a short script that calls its own tools inline, in one turn, instead of a round-trip per call. Chaining several reads before a write, or looping over a list of files, now costs one step instead of ten.
• Document generation: it writes real PDF, Word, Excel, and PowerPoint files you can open in the corresponding app. Spreadsheets keep numbers numeric, so you can actually sum them.
• Auto-update, with release notes shown in-app after an upgrade, and an announcements channel for things worth saying between releases.
• A capability system built from the live tool list rather than a hand-written blurb. Smaller models were reading a stale description and concluding they had no browser tools while holding nine of them; the description can no longer drift from reality, and a router names the right tools for a request at the moment it is asked.
• The right sidebar and bottom bar are now tabbed and can host any view, so switching to the file tree no longer throws away what you were looking at.
• A usage screen with lifetime totals, streaks, and a year-long activity heatmap, all from real recorded token counts.
• The file browser previews anything: images, SVGs, and PDFs render properly instead of failing as invalid text, and you can paste or drag any file straight into the composer.
• Fixed: provider errors (out of credits, rate limits) ending a turn silently with the reason visible only in the log; the browser never auto-starting on Windows; image input being dropped on models that could have read it; model choice not surviving a restart; and file mentions only ever seeing the first workspace folder.

A real persistent terminal, vision with an automatic OCR fallback, goal tracking, and a stalling bug fixed for good.

• A full performance pass, front to back: the panel resize drag, diff rendering, and message list no longer do unnecessary work on a long thread; the provider request context is no longer deep-cloned repeatedly on retries; token-budget checks no longer re-walk the whole transcript every turn; and the app now shuts down cleanly instead of leaving terminals, dev servers, and local models running as orphaned processes.
• Persistent terminal tools backed by the same PTY sessions the app's own terminal tabs use: open one, keep it alive across steps (an activated venv, a running dev server), type into it, read back what happened.
• Real vision for images, with an automatic OCR fallback when the selected model can't take image input at all, so a text-only model still gets the content.
• Goal tracking: a durable objective for a long task, separate from the moment-to-moment checklist, restated every turn so it survives compaction.
• Honest partial scans (a capped search says so, instead of quietly looking exhaustive), line-numbered file reads, and a list_tools call to see exactly what's registered right now.
• Fixed the bug behind every turn's chance of a 120-second stall, a 111MB read on every turn, New Chat not actually persisting, and the Review panel's diff view failing on real changes.
• Gesture control moved out into its own standalone app; the dead Rust orchestrator module removed; bash and powershell always stream live output now.

AI Elements across the board, a real browser for web preview and search, and OpenAI's Responses API.

• Sources, Plan, Sandbox, Snippet, Web Preview, Queue, Checkpoints, and Confirmations: the full AI Elements set, wired into the agent's own tools.
• Web preview, web search, and web_fetch now route through a real (fingerprint-spoofed) browser process instead of an iframe or plain HTML scrape. It renders sites that block embedding, and launches itself automatically the first time it's needed.
• Direct OpenAI calls now speak the Responses API instead of the legacy Chat Completions endpoint, with true incremental streaming and reasoning summaries.
• One-click Docker install for the sandbox tool, and a fixed cron scheduler that was silently treating every automation's local time as UTC.
• A refreshed welcome screen with the real Vivianne mark, replacing the placeholder CLI command box.

Initial release: the coding agent, the memory pipeline, and the multi-provider harness.

• A single tool-use loop: propose, run, verify, with a hard step ceiling.
• Three-tier memory pipeline: auto-capture, L1 fact extraction, L3 persona synthesis.
• Bounded delegation, one level deep, always traceable in the session log.
• One harness across model providers, with automatic context compaction.
• Runs on cloud model providers, with the SSM-MoE local architecture in development.