v0.1.33 — Know your context — Recursive Releases

See how full the model's context window is — and what's eating it. A ring in the files bar fills as the active agent reports how much of its window is in use. Tap to expand a breakdown: rules, tools, skills, conversation history, attachments, and the rest. If the agent can't report a real number, the ring stays hidden instead of guessing. Switch models mid-session and the meter starts fresh for the new one.

Review changes turn by turn, not just everything at once. The file viewer and Changes tab now share a scope picker: all uncommitted work, everything from one session, the last turn, or any numbered turn with a preview of what you asked. Commit suggestions follow the same boundaries — grouped by what you actually said, with follow-up turns folded together when they clearly continued the same fix.

Set how much freedom the agent gets before you send. Supervised asks before edits and commands. Approve for me lets it work inside the project and only checks in when stepping outside. Full access runs without prompts. Pick a preset from the composer; the menu tells you when your chosen agent treats two options the same. If a plugin connection needs sign-in or won't start, a card in the feed offers a fix instead of hiding the problem in tool output.

One project, multiple folders. Attach sibling repos or supporting directories at any point — including mid-session — mark one as primary, and the workspace remembers the full set. Git and branch operations stay on the primary folder; the others are there when the agent needs to read or write across repos.

160+ new inference providers — plus any OpenAI-compatible endpoint you already have. Browse provider logos in the model picker, or add a custom base URL and API key in Settings — Together, Groq, vLLM, LM Studio, and internal gateways are discovered from the endpoint without a restart. DeepSeek ships bundled with session-bound tools and reasoning support.

300+ plugins from the tools you already use. Browse integrations from Stripe, 1Password, Figma, Sentry, and dozens more — installed with one click and checked for compatibility before anything lands on disk. Featured picks sit alongside your installed and local plugins. Search by provider name finds the integration you want, not just the catalog that contains it. Broken plugin connections get a guided fix flow instead of failing quietly.

Rich attachments in your inbox. Agents can send HTML reports, inline images, and downloadable files with each message. Full-bleed layouts replace the markdown body when the whole message is the attachment; everything else renders inline with chips you can preview or open in the workspace file viewer. Chat attachments follow the same path — no more downloading to inspect what the agent sent.

Set spending guardrails before the agent runs away. Cost Management in Agents settings caps iterations, tokens, spend, and runtime — with a repeat guard that stops the same tool call from looping. OpenAI-compatible adapters that can spiral on tool loops are the main audience, but the limits apply workspace-wide when you turn them on.

Each session gets its own rules and setup. Project rules, plugin connections, and workspace config apply only to the session that asked for them — switching projects or resuming a different chat won't drag one workspace's settings into another.

Also in this release:

Turn rail — jump between turns from a table of contents beside the feed instead of scrolling the full transcript.
Working parents — when a side chat or subagent is active, its parent session moves into Working so orchestration stays visible.
Open a folder without a project — pick any directory and start chatting; the composer treats it as a valid workspace.
Unified attachment cards — files, folders, text snippets, and design annotations share one card layout in the composer rail.
Plugin-branded chips — @-mentions and slash commands show plugin icons and tints in the composer.
Native browser fallback — sites that block iframe embedding open in the full desktop browser instead of a dead preview.
New Tab search — Task, Automation, and Plan tabs opened empty land on a search box, not a dead placeholder.
Grouped New Tab menu — Code, Source Control, Planning, and other tab types organize into sections.
Faster workspace load — lazy plugin and provider catalogs, lazy slash commands, slimmer session previews, and a faster session list on large workspaces.
Browser tabs stay out of the way — agent browser windows open in the background and close when work is done.
Stop stays visible — in-process provider sessions keep Stop through the full model and tool loop.
Artifacts survive compaction — tool results and session artifacts stay on the owning session after compaction.
Design annotation stability — browser preview annotations and embedded navigation are more reliable.
Crash recovery prompt — the desktop app asks before reloading a crashed window so you don't lose unsaved state.
Terminal polish — refreshed styling; smoother wheel scroll in nested feeds; macros open at the top in the order you wrote them.
Clearer accents — permission presets and colored chips are easier to read across light and dark themes.