Changelog

What’s new in Rayu Code

User-facing changes to the CLI and the VS Code extension, newest first.

Rayu-CLI Changelog#

All notable user-facing changes to Rayu-CLI are documented here, newest first.

3.0.02026-11-21#

RayuCode v3 is a rewrite of the CLI as a single native binary. It is not an update to v2: the flags, config format, and several commands changed. Read Installation and CLI reference if you are coming from v2.

Highlights#

  • One native binary. rayu is a Rust executable for Linux, macOS and Windows on x86_64 and aarch64. No Node, npm, Bun or Python runtime, and no sudo: the installer writes to ~/.local/bin.
  • Verified downloads. Every release asset is published with SHA256SUMS, and both the installer and rayu update verify the checksum before installing.
  • Automatic migration from v2. The installer puts v3 in place first, renames the old Node CLI to rayu-legacy, keeps ~/.rayu (settings, API keys, sessions), and offers to remove the legacy binary.

CLI changes#

  • One-shot mode is rayu run <message>. -p now means --provider. --print, -c/--continue, --output-format, --input-format, --permission-mode, --dangerously-skip-permissions, --add-dir, --system-prompt, --settings, --max-turns, --json-schema, --mcp-config, --plugin-dir, --bare, --worktree, --tmux, --ide are gone.
  • Removed subcommands: rayu doctor, rayu install, rayu uninstall, rayu completion, rayu agents, rayu plugin, and rayu mcp add|add-json|get|remove|serve.
  • New and reshaped commands: rayu run (--json, --ndjson), rayu login, rayu account, rayu model, rayu provider, rayu usage, rayu session, rayu server, rayu restart, rayu permissions, rayu ambient, rayu memory, rayu cloud, rayu mcp test|list|login|logout, rayu telegram link|status|disconnect|uninstall, rayu auth|auth-test, rayu provider-doctor, rayu browser, rayu companion, rayu pair, rayu transcript, rayu replay, rayu acp, rayu api-bridge.
  • Settings are TOML at ~/.rayu/config.toml. The v2 settings.json chain is read for compatibility only.

Safety#

  • No per-tool approval prompts. Safety is a destructive-bash gate, an optional pre_tool hook that can block with exit 2, and narrowable tool exposure through [tools] / --tools.
  • Plan mode is advisory. /plan produces a plan card; it does not restrict what the agent runs afterwards.
  • Ambient permission requests are reviewed with rayu permissions, or approved from Telegram.

New in v3#

  • Swarm coordination with todo-lane workers, per-role models (swarm, review, judge, memory, ambient), and named agent profiles.
  • Hooks in TOML ([hooks]) for turn_start, turn_end, session_start, session_end, pre_tool, post_tool, post_tool_feedback, and pre_tool_transform, with a documented RAYU_HOOK_* environment contract.
  • Headless JSON and streaming through rayu run --json and rayu run --ndjson, including credit reporting.
  • Rayu-only release builds via ENABLE_OTHER_PROVIDER=false. Public builds ship the Rayu provider; bring-your-own-provider builds set the flag to true.
  • Transcript inspection and search, per-tool transcripts, and Bash telemetry with per-row metrics.
  • Telegram remote control with session and model pickers, chat approvals, and a single machine-wide link.
  • A VS Code baseline sharing the agent engine, plus an SDK and a shared UI package.

Fixed#

  • Pinned task and turn status ordering in the TUI.
  • Lost turn telemetry and transcript details across turns and cancelled batches.
  • Telegram pairing and tool approvals drifting from the terminal session.
  • Agent profiles being withheld or desynced across remote TUI sessions.
  • list_models routes are capped and filtered so profiles are never dropped.
  • Missing tool-row metrics, native search result counts, and Bash previews.
  • The fullscreen composer position and theme previews on short terminals.
  • Draft stash losing complete composer state.

2.5.32026-11-18 (Beta)#

  • Telegram: pick your model by tapping — /model with no argument opens an inline, paginated card (ten models a page, Prev/Next) and a tap switches the model. It reads the same catalog and uses the same selection path as the typed /model <id> <provider>, which still works, and a card that has gone stale re-resolves its model instead of selecting the wrong one
  • Telegram: switch sessions by tapping — /sessions and a bare /switch open a tap-to-switch session card with Refresh and Close. A session that has closed in the meantime is refused instead of becoming a dangling pointer, and switching to the session the chat already drives is a no-op, so it never detaches a running turn
  • One Telegram link for the whole machine — the CLI and the editor now share a single Telegram link even when they are signed in to different Rayu accounts: the owning account is recorded, and a surface that does not own the link can show it and route through it but cannot overwrite or unlink it. While the hosted bridge is live the link cache heals itself from the backend, a link left behind by a closed session is adopted on the next engine start, and the editor's Integrations card now separates "connected" from "linked — not listening" and offers "Drive this conversation"
  • Fixed: pasting several images kept only the last one — a multi-image paste fired its pill insertions in one synchronous loop, so each pill overwrote the previous one and every image without a surviving reference was pruned; pasting 3 still attached 1. Insertions now accumulate, so every pasted image attaches. On Windows the clipboard's full file list is also read, because Windows Terminal forwards only the first copied file, so a multi-file Explorer copy now attaches all of them
  • Fixed: Khmer text overflowed and wrapped wrong — Khmer combining marks are zero-width in terminals but were counted as cells, so the measured width was too large and text overflowed its area, mis-wrapped, or was cut mid-cluster. Khmer now measures its spacing code points with the marks at zero width, wraps grapheme-atomically, splits clusters wider than two cells to match the renderer's cell model, and slices at real cell boundaries
  • Fixed: resuming a conversation could fail with "No conversation found" — the engine now resolves a transcript the same way the editor host does: the shared session-file resolver, then a scan by the globally unique session id, so a transcript under a sibling git worktree, a cloud-synced path (iCloud/OneDrive/Dropbox) or a differently resolved symlink is found. The resumed session's writes are anchored to the directory the transcript was actually loaded from, instead of a second copy being started elsewhere

2.5.02026-11-07 (Beta)#

  • Updates install themselves in the background — a new release is fetched and installed through the rayucode.com installer without interrupting the session you are in, so updating no longer means stopping work to run it by hand.
  • Fixed: clipboard images paste on Windows — pasting an image from the clipboard now attaches it to the prompt on Windows instead of silently doing nothing.
  • Subagents run on Rayu-hosted models by default — a subagent now prefers a Rayu-hosted model, and the Rayu provider is given priority, so a subagent no longer fails because its model resolved to nothing.
  • Fixed: the Orchestrator no longer rejects work that was done — verification was hardened so a run that completed its packets is not turned away on a contract problem that isn't real, and a contract result now names the project root so workers and checks resolve paths from the right place.
  • Steadier background runs and releases — session routing was hardened and the release pipeline tightened, so long-running Orchestrator work and the builds that ship it are less likely to break.

2.4.542026-11-06#

  • Fixed: an expired Rayu session no longer needs /logout before /login — a hosted request rejected with 401 now renews the token once and replays the request; if the refresh token itself is dead, the stale session is cleared so /login opens instead of replying "already signed in". The login gate now treats an unrenewable session as signed out, which is what forced /logout → /login after an update or a deploy
  • Choose your models in Settings — the editor's Settings tab now has a Models section for the subagent model (one default plus an optional override per agent type), the image generation model, the video generation model and the WebFetch model. Changing the subagent or WebFetch model restarts idle conversations so the change takes effect; image and video choices apply to the next generation with no restart
  • Nine commands that used to do nothing now open real editor UI — /permissions opens a rule manager to view, add and remove allow/deny/ask rules, /memory picks a memory file and opens or creates it, /doctor, /status and /hooks show read-only reports, /stats adds a per-tool usage tab to the Runtime Center, /add-dir adds a working folder and resumes the conversation, /export saves the conversation as Markdown, and /help opens the command reference
  • Fixed: the permission rule editor saved rules the engine rejects — Bash() silently became a blanket allow of every Bash command, and a malformed rule was reported as saved yet never took effect. Adding a rule now validates it first; removing one stays lenient, so an existing rule can always be taken out

2.4.32026-11-02 (Beta)#

  • The Orchestrator run, laid out in Rayucode — the background-work panel now gives each run its own section: every packet with its gate, a status badge on each worker row (pending, running, merging, completed), and the selected packet's problems, worktree and verdict
  • A slash-command menu in the editor — typing / in the Rayucode composer opens a menu built from the same command set the terminal uses. A command the editor cannot run is hidden rather than offered, aliases and argument hints come from the engine, and the menu works before the first session snapshot arrives
  • Drag files into Rayucode without holding Shift — a new, collapsed "Rayucode Context" Explorer pane accepts an ordinary drag and inserts the same @-mention every other context route produces. The chat panel itself still needs Shift, because VS Code blanks a webview's pointer events while a file is being dragged
  • /clear, /reset and /new work in the editor — the extension restarts the tab's engine so the composer behaves like the terminal. Only the exact command matches, so /clear now stays an ordinary prompt
  • Rewind a conversation's file changes — file checkpointing is now enabled for the editor's engine, so the edits a conversation made can be rewound
  • Re-linking Telegram restarts the conversation safely — a newly connected bot is picked up by restarting the session, but only after an in-flight turn ends and its reply has been flushed to disk; nothing is interrupted mid-turn
  • Generated images and video play in the editor — media results render as media instead of a raw tool payload, and generated clips play inline
  • Fixed: the composer showed three focus outlines — the card border, a shadow ring and the textarea's own outline stacked on focus, and high-contrast themes added a second border. Focus is a single line now
  • Fixed: effort and thinking switched off for dotted model ids — claude-opus-5.5 (Rayu-hosted) and claude-sonnet-4-6 (first-party) are the same model, but capability checks matched the dashed spelling only, so a dotted id was treated as legacy: effort and adaptive thinking were silently disabled and the effort control disappeared from Rayucode. Both spellings now resolve identically
  • The whole request's credits, not one turn's — background agents a request launches, and the turns their completion notices start, now count toward that request's total. The credit window stays open while such an agent runs, has finished but not yet queued its notice (up to 30 s), or has a notice turn queued; a line printed meanwhile says Credits so far: X. Rayucode's "Completed in" now covers the whole wait
  • Credit totals read at scale — two decimals below 1,000, then k/M (2.5k, 1.2M); a real charge under 0.005 reads <0.01; the exact figure is in the VS Code tooltip
  • Fixed: image-heavy sessions no longer fail with "request too large" — compaction now also triggers on accumulated image bytes (12 MB by default), not just the token estimate, which is blind to how large images are on the wire. The 413 message now separates one oversized upload in the current turn from images accumulated across the conversation, and points at /compact or /clear
  • Fixed: != broke commands in the Bash tool — a command containing both quote styles (node -e "a!==b", jq, awk) had every ! rewritten to \! before it ran; commands now reach the shell unchanged
  • The Orchestrator works in projects that are not git repositories — contract paths are relative to a project root (the git root, or the directory the session works in), absolute paths inside it are converted instead of refused, and changes are tracked by file size and modification time when there is no git. Before, every file a worker wrote was reported as "owned by no packet" and a verified run could never pass the gate. Worktree packets still need git and are refused up front
  • Fewer wasted turns in an Orchestrator run — a packet worker started in the foreground now runs in the background instead of being refused; check commands may be up to 2,000 characters (was 500); and the contract result names the project root and flags dependencies that its shared interfaces make unnecessary, so more packets run at once
  • Fixed: the verification gate no longer gets stuck — it forgot that it had given its last word, so reminders started over and the status printed twice; rejections that were not the project's fault (the verifier did not finish, the state of the files could not be recorded) no longer count as failed repairs; repair rounds restart with each contract version; and the accept hint now shows a real problem id
  • Telegram /connect is one flow for every provider — the inline model-picker keyboard is gone. Connect saves the provider's preset default model, warms the live catalog in the background, and points at /model, which prints a copyable catalog instead of a paginated keyboard
  • Kiro as a Rayu-hosted provider — Kiro models can now be used on your Rayu plan with API-key sign-in and the usual plan and credit limits, including tool use and thinking output
  • Optional diagnostics log — Rayu can append redacted warnings and errors (upstream failures, access errors) to a separate JSON-lines file alongside its normal log. An unusable path is reported once and never stops Rayu from starting
  • Console and pricing page restyled — every signed-in screen (overview, usage, credits, plan, invoices, API keys, models, referrals, team, settings) now shares the site's updated visual design: hairline surfaces, 8px cards, 6px buttons, no glow. The plans comparison follows the same theme and animates in as it scrolls into view
  • Landing testimonials — the Rayu landing page now shows customer testimonials with an author, role and rating, in display order
  • Faster, sturdier installs — the installer now provisions ripgrep alongside the CLI and retries downloads with backoff (five attempts, plus an HTTP/1.1 fallback), so an install survives a flaky network
  • An AI support assistant on rayucode.com — a fox-launched chat on every marketing page and a full "Ask Rayu" page answer questions about API keys, installation, the CLI, the VS Code extension, providers, models and plans. Answers are grounded in these docs and cite the section they came from, stream token by token, and reply in the visitor's language

2.4.22026-10-31 (Beta)#

  • Orchestrator runs follow a contract — before launching implementation workers, the Orchestrator publishes a contract: one packet per worker with the files it owns, the shared interfaces with their exact definitions, the order packets run in, and required checks. Each worker is briefed with it, dependent packets wait for their dependencies, a packet gets one worker at a time, and at most four workers write at once by default
  • Changes are checked against the packets — a worker editing a file another packet owns, or one no packet owns, is told so in its tool result, and the Orchestrator sees the problem in the worker's completion notice and /orchestrator-status. Shell commands are checked too, and whenever no worker runs, the whole tree is compared with the snapshot taken when the run started, so a change nobody accounts for is raised instead of passing as someone's work
  • Worktree packets are merged back — a packet marked as risky works in its own git worktree from the run's snapshot and is merged 3-way into your tree when it finishes, all or nothing, and recorded for /undo and the review card. On a conflict nothing is written; the worker gets the conflict markers in its worktree to resolve
  • A verifier whose PASS needs evidence — the new verifier agent runs the contract's required checks after the workers finish. Rayu records the commands it actually ran and accepts a PASS only when every required check ran, passed and is reported, and the verifier changed nothing
  • The run cannot be reported as done while it is not — when the Orchestrator ends its turn with a packet unfinished, a problem open, or no accepted verification of the current tree, it is sent back to work with the reasons. A verification gate is on by default (a soft mode only reports). After three failed verifications it must report what is unfinished instead of claiming success. You see the gate's status in the transcript and in Rayucode
  • See the run — /orchestrator-status shows packets, worktrees, problems, the latest verification and the gate; accept <id> accepts a problem you reviewed and diff <id> shows its recorded change. The background-tasks dialog (CLI) and the background-work panel (Rayucode) show every packet, and worker rows keep updating after they finish
  • The planner and verifier are available in Rayucode and -p — they were available only in the interactive terminal before
  • Fixed: a same-size edit made within a second of git add could be missed by the Orchestrator's working-tree snapshots

2.4.12026-10-30 (Beta)#

  • Rayucode settings, account and credits, right in the editor — a new Settings tab (sidebar gear, Command Palette, or the chat's overflow menu) covers General, Account, Integrations and About. Account shows your profile, plan and renewal date, credits, usage limits and registered devices, and a credits item in the status bar warns when your remaining allowance or daily turns drop below 10%. General settings are shared with the terminal, so both surfaces always agree
  • Integrations moved into Settings — connect Telegram by scanning the shared bot's QR code or pasting your own bot token, turn on Studio Remote or pair it with a QR code, and install Rayu's GitHub Actions workflows without leaving VS Code. The Telegram connect panel no longer lives in the chat
  • Rayu-hosted image and video generation — generate images and clips on your Rayu plan with no provider key, billed in credits; bringing your own fal.ai key still works for video
  • Per-turn credit reporting — a "Credits: …" line after each turn, in the terminal and in VS Code, totals everything Rayu charged for that turn, including side queries and media generation
  • Platform VSIX packages — platform builds ship a compiled engine and bundled ripgrep (the universal VSIX keeps the plain engine bundle), plus a memory watchdog that replaces an engine at its memory limit and continues the same conversation instead of starting over
  • A stricter orchestrator — in orchestrator mode the main agent can no longer edit files or run non-read-only shell itself and must delegate to workers; changing mode warns when it would break running workers, and SendMessage is worker-only when agent teams are off
  • Shift+Tab is always bound — cycling permission modes now works with the standard key on Windows (Meta+M remains an alias)
  • Signing out of Rayucode matches /logout — the same caches are cleared, and a configured API key is kept
  • Fixed remote-server safety — a plaintext (non-HTTPS) custom API endpoint is now refused for Rayu credentials, and redirects on it are rejected instead of followed
  • Breaking: NVIDIA image/video backends removed — /generate-image and /image-video now use Rayu-hosted, Vertex or fal.ai; NVIDIA text models are unaffected. In VS Code the engine now starts in the permission mode chosen in the extension, so a default mode written for the terminal no longer overrides it

2.2.172026-10-17 (Beta)#

  • Sessions now open as editor tabs — every RayuCode conversation opens in its own editor tab, so you can run several side by side and switch between them without interrupting a turn that is already in flight
  • The sidebar is now the agent's management hub — it shows your session list, background tasks, MCP servers, provider setup and sign-in instead of a second copy of the chat. The conversation itself — transcript and composer — lives in the session's editor tab, so it appears in exactly one place
  • Fixed the sidebar duplicating and reloading the chat — the sidebar could briefly render the open conversation and re-lay-out whenever you created or switched sessions; it now stays on the management view and updates in place

2.0.132026-09-21 (Beta)#

  • RayuCode on GitHub — you can now connect RayuCode to GitHub and put it to work on your pull requests. Install it as a GitHub App, pick the repositories you want in scope, and RayuCode reviews incoming PRs right where your team already works — leaving inline findings, summaries, and suggested fixes as review comments, no local setup required
  • Run RayuCode as a GitHub bot in your hub repo — designate a "hub" repository and RayuCode becomes a standing bot there. It listens for new PRs and issues, responds to mentions, and can act as the central review and triage agent for the whole organization — one bot, every repo that points at the hub

2.0.122026-09-20 (Beta)#

  • Orchestrator mode is now the default — the CLI now ships with the orchestrator as the built-in coordination layer. Plan, build, review, and run multi-agent workflows end-to-end from a single session, with parallel/sequential/race policies, retry/fallback, and git worktree isolation all wired in by default
  • Reusable messages in VS Code — RayuCode now supports saved, reusable prompt messages. Build a library of your most-used prompts and invoke them instantly from the IDE chat panel, with variable interpolation for project-specific context
  • Collaborator swarm migrated to orchestrator — the standalone collaborator swarm command has been retired and its capabilities fully absorbed into the orchestrator. Domain-specialized agents (backend, frontend, mobile, security, deploy) are now orchestrated through the unified /agent workflow with better task routing and shared context
  • Rayu GitHub setup command restored — /install-github-app now creates Rayu-branded @rayu and automated review workflows, defaults to securely provisioning a repository credential from the signed-in Rayu account (with manual API keys still supported), and uses trusted-actor guards and publisher-configured Rayu GitHub App and Action settings

2.0.12026-09-14 (Beta)#

  • Introducing RayuCode — our new VS Code extension! Rayu's full AI coding agent engine is now available directly inside VS Code as RayuCode. Same tools, same providers, same skills — now with a native IDE chat panel, diff viewer, and webview UI. One engine, two products: terminal and IDE
  • Greatly improved model support — expanded the provider catalog with more models across Anthropic, OpenAI, Google Gemini, DeepSeek, Kimi, and OpenAI-compatible endpoints. Better model capability detection, smarter fallbacks, and more accurate context window handling
  • Pay-as-you-go billing — introduced a new pay-as-you-go pricing model alongside subscriptions. Top up your balance and pay only for what you use — no monthly commitment required. Perfect for occasional or variable usage
  • More accurate credit charging — completely reworked the credit calculation pipeline. Token counting now aligns precisely with provider billing (input/output/cache-read/cache-write tokens tracked separately), so you're charged exactly for what the model consumed — no more, no less
  • Subscription management overhaul — redesigned the subscription experience from the ground up: clearer plan comparison, seamless upgrades/downgrades, prorated billing, and a unified billing dashboard. View usage, manage your plan, and track invoices in one place
  • Team SSO (Enterprise) — added SAML/OIDC single-sign-on support for teams and organizations. Google Workspace, Okta, or any OIDC identity provider can be configured so team members sign in with their existing corporate credentials — no separate Rayu accounts needed
  • Fixed streaming interruptions — resolved an issue where long streaming responses could silently drop tokens or terminate early under high-latency network conditions
  • Fixed credit double-counting on retries — the final charge now always matches actual usage, eliminating the previous over-charge when a failed upstream call was retried
  • Fixed model fallback loops — prevented infinite fallback cycles when multiple models in the chain were unavailable; the CLI now fails fast with a clear error message
  • Fixed subscription plan changes not reflecting immediately — plan access is now refreshed in real-time after a subscription change, removing the previous delay where users had to restart the session
  • Fixed memory growth in long-running sessions — resolved a memory accumulation issue that caused slowdowns during multi-hour coding sessions
  • Fixed clipboard paste on Windows Terminal — clipboard image paste now works correctly in Windows Terminal and other Windows consoles
  • Fixed slash command autocomplete ordering — commands are now ranked by relevance and frequency instead of alphabetically, so /model, /connect, and /help appear first

1.6.02026-08-30#

  • Introduced the External Agent Orchestrator — /agent command and ExternalAgent tool to discover, launch, adopt, assign work to, and stream other agentic CLIs (Codex, Claude Code, OpenCode, and any ACP agent). Supports orchestration policies (parallel/sequential/race/retry/fallback), git worktree isolation, brokered permissions, and crash recovery
  • Overhauled the Telegram bridge — list sessions with /sessions, switch between them with /switch <n>, and get clear connection status (connected/reconnecting/disconnected) with automatic reconnection
  • Added the Rayu API Key provider (rayu) — authenticate with a rayu_sk_live_… key, and a streamlined first-run setup screen that offers Rayu sign-in or API key entry, with the full provider list one keystroke away via /connect
  • Added smarter agent behavior — per-model image capability detection, automatic repair of malformed tool arguments from weaker models, Edit/Write key coercion so non-Claude tool calls work correctly, and the active model name in the input bar and sidebar

1.5.122026-07-30#

  • Improved the model selection experience — better picker UX and more reliable model resolution when switching providers

1.5.112026-07-30#

  • Fixed a bug where Rayu could hang when running inside the default macOS Terminal app — Enter and several modified key events were not being delivered

1.5.12026-07-28#

  • Migrated the Telegram bridge for a smoother pairing flow, clearer status messages, and more reliable bot re-pairing when the linked bot changes

1.5.02026-07-25#

  • The Rayu-hosted provider now supports many more models, with real-time model list updates so newly enabled models are picked up without requiring a CLI release
  • Improved provider routing and model resolution for Rayu-hosted models

1.4.4792026-07-23#

  • Fixed the Ask User Question tool not working in full manage (full-control) permission mode

1.4.4782026-07-21#

  • Improved the Ask User Question tool for all LLM providers
  • Fixed an upstream error and a plan credit calculation bug

1.4.4772026-07-21#

  • Reliability fixes for Rayu-hosted models, improving behavior with multiple concurrent requests

1.4.4722026-07-09#

  • Fixed deferred tool loading for third-party providers — tools such as web fetch and search were called with guessed parameters and rejected; they now work correctly

1.4.4712026-07-09#

  • Fixed Kimi Code 2.7 streaming compatibility
  • Promoted Ollama to a primary provider with improved multi-key configuration
  • Fixed the update, uninstall, and install commands for cross-platform reliability

1.4.4702026-07-09#

  • Improved multi-provider compatibility and cleanup of disabled provider connections, plus native context-window support for Rayu-hosted models
  • UI polish: refined several terminal components for better readability and consistency

1.4.4662026-07-08#

  • Rebranded the terminal theme with the Rayu identity and a new Goose mascot, including an animated mascot and configurable banner
  • Added /banner and /mascot slash commands for displaying Rayu branding in the terminal

1.3.4622026-07-06#

  • Removed the DeepSeek Web reverse-scraping provider
  • Added cache read/write credit tracking for usage
  • Updated the Kiro provider to Sonnet 5, and improved the QR payment flow and Telegram reliability

1.3.4452026-06-29#

  • Fixed routing issues for the Rayu-hosted provider

1.3.4442026-06-29#

  • Improved prompt theming, thinking display, structured API error messages, and rendering stability

1.3.4432026-06-26#

  • Production stability improvements across sign-in, plan access, and hosted routing
  • Improved the first-launch flow and dashboard

1.3.4422026-06-24#

  • Improved interrupt handling during streaming responses
  • Added memory pressure protection and heap-limit handling for very large sessions

1.3.4412026-06-24#

  • Added QR (KHQR) payments and Telegram payment notifications, with a smoother checkout flow

1.3.4402026-06-22#

  • Improved Copilot sign-in and message handling
  • Added performance diagnostics and stability fixes for long sessions

1.3.4392026-06-21#

  • Subscriptions arrive — paid plans unlock image and video generation, and the free plan gains usage limits, with a clearer credit limit and account dashboard

1.3.4382026-06-17#

  • Added a Hugging Face (huggingface) provider preset using the Hugging Face Inference Providers endpoint (https://router.huggingface.co/v1) with HF_TOKEN
  • Added image-editor and video/image generation integrations, and improved Copilot sign-in and feature limits

1.3.4372026-06-17#

  • Fixed Kiro thinking output and context handling

1.3.4362026-06-17#

  • Renamed the Planning & Research subagent from "PA-AGENT" to "planner" and sped up codebase exploration with concurrent search subagents
  • Improved collaborator swarm memory syncing, plus agent task spinner and status displays

1.3.4352026-06-17#

  • Published a new stable CLI release with packaging improvements

1.3.4342026-06-16#

  • Fixed Kiro provider context handling and rebranded the icon from Claude's to Rayu's
  • Added /brandmark to customize the brand-mark glyph and loading-spinner style

1.3.4332026-06-15#

  • Added a new provider preset with full model discovery

1.3.4322026-06-15#

  • Fixed a thinking-output bug

1.3.4312026-06-15#

  • Fixed the Kiro context base so it matches the original 1M context

1.3.4302026-06-15#

  • The final update made the Rayu CLI free to use
  • Added Kiro sign-in through an API key and OAuth, plus a Rayu OAuth preview

1.3.4292026-06-14#

  • Improved thinking output for every provider and made responses about 10% faster

1.3.4282026-06-13#

  • Added the thinking status display to all agent subagents and the collaborator agent

1.3.4272026-06-13#

  • Completely updated the documentation to reflect Rayu as an independent CLI, removing legacy Claude Code specific references
  • Improved the thinking UI and status displays
  • Fixed Kimi thinking output and blocked thinking through the Bedrock provider

1.2.252026-06-11#

  • Extended thinking and effort now work on any provider, not just Claude — type ultrathink to trigger deep reasoning on NVIDIA, Gemini, DeepSeek, and other OpenAI-compatible models
  • Added local /ultraplan: deep multi-agent planning that explores your codebase in parallel and produces a step-by-step plan for approval — runs entirely on your own provider
  • Added local /ultrareview: a deep bug-hunt that gathers your branch diff and dispatches parallel review subagents to find and verify real bugs — runs entirely on your own provider
  • "What's new" now shows Rayu's own changelog, and an "Update available" notice appears when a newer version is published to npm

1.2.242026-06-10#

  • Fixed a spurious "Interrupted by user" error that appeared when a parallel tool timed out or a sibling command failed — interruptions are now only reported when you actually cancel
  • /ide now detects VS Code when connected through the integrated terminal
  • Clipboard image paste now works on Wayland (Linux)
  • Reworked /review_detail into an interactive diff viewer, /undo all reverts every pending file change at once, and /stickers is now /contact_me