0.10.0 RELEASED

[acoliver]

Release v0.10.0

The biggest release since 0.5. Four months of work, 1,100+ commits, and two changes that touch everyone: we moved off ~/.llxprt to OS-standard directories, and the CLI now runs on Bun instead of compiled Node bundles. Both are handled for you automatically.

Installation

Install or upgrade LLxprt Code using npm:

npm install -g @vybestack/llxprt-code

Or use directly with npx:

npx @vybestack/llxprt-code

The Headlines

We moved out of ~/.llxprt and into your OS's standard directories. Your config, data, cache, and logs are now separated the way your platform expects — XDG paths on Linux, ~/Library/... on macOS, %APPDATA%/%LOCALAPPDATA% on Windows. On first launch we copy (not move) your existing global ~/.llxprt over, so nothing is destroyed and you can roll back. Per-project <project>/.llxprt directories are untouched. If you have scripts pointing at ~/.llxprt, they need updating — see Breaking Changes.

LLxprt runs on Bun now, and there's no build step. The npm package ships TypeScript source plus its own pinned Bun, and a native launcher execs it directly. There's no Node relay process, no bundle/llxprt.js, no dist/. From your side nothing changes: npm install -g, npx, and Homebrew all work exactly as before, and you do not need to install Bun yourself. Windows users get the biggest win — the old multi-process startup chain is gone.

Agent Skills. Entirely new. Drop a SKILL.md bundle into <project>/.llxprt/skills, your user skills directory, or ship one in an extension, and the model can pull it in on demand. Only the name and description go into the prompt up front; the full instructions and bundled assets are injected only when the model actually activates the skill, and only after you consent. Manage them with /skills list|enable|disable|reload in-session or llxprt skills list|install|uninstall|enable|disable from the shell — install from a Git repo, a local directory, or a .skill archive. The status bar shows your enabled/disabled counts.

Fresh installs no longer silently pick a provider for you. Previously an unconfigured LLxprt would quietly fall through to Gemini. Now an unconfigured session points you at /setup to pick a hosted provider, a local model, a custom endpoint, or a saved profile — and non-interactive runs fail with real setup guidance instead of sending a request somewhere you didn't choose. Existing configs are unaffected.

A much bigger model roster. Claude Opus 5 (now the Anthropic default), Sonnet 5, and Fable 5. GPT-5.6 Sol / Terra / Luna, with Sol as the Codex default. Kimi K3. Correct GLM 4.x/5.x context sizing. LiteLLM as a first-class provider. Ollama Cloud and Makora aliases. Configurable reasoning-field names so thinking output stops disappearing on OpenAI-compatible endpoints.

Slow models are no longer mistaken for dead ones. This was the single most-reported class of annoyance. The stream watchdog now understands provider-level liveness instead of assuming silence means death, the aggressive inter-chunk timeout is off by default, ESC and timeout aborts no longer poison your next prompt, and first-chunk connection failures actually get retried. Fireworks/Kimi buffering hangs and never-finishing turns are fixed too.

Long sessions survive things that used to kill them. Load-balancer pools exhausted by 429s now reset instead of staying pinned to a dead backend. An all-429 pool stays retryable instead of dumping you back to the prompt. Real quota errors survive to the surface instead of being overwritten by a generic timeout message. Near-limit turns attempt compression before giving up. Revoked or concurrently-refreshed OAuth tokens no longer spin in reauth loops or let a second LLxprt instance revoke your first one's session.

Zed/ACP and the VS Code companion got serious attention. The Zed integration now runs on the same Agent engine as the terminal CLI instead of a hand-rolled chat loop, so retries, tool continuation, provider activation, auth, and subagents all behave identically wherever you drive it from — and the task tool crash and todo_pause continuation-loop bug in Zed are fixed. The VS Code companion no longer drops your open files on first connect or wedges on reconnect, and write_file/ast_edit open the native diff view again.

Synced through gemini-cli 0.26.0, up from 0.23.0 in the 0.9 line.

New Models & Providers

• Claude Opus 5 — in the model menu and profile wizard, and now the Anthropic default. 200K context / 32K output on subscription. opus-latest resolves here.
• Claude Sonnet 5 — 200K context, 128K output, adaptive thinking, correct sonnet-latest handling and dated-snapshot token limits.
• Claude Fable 5 — available on Anthropic/Claude Code OAuth. Its HTTP-200 safety-classifier refusals now show as a visible notice instead of the session appearing to just stop.
• GPT-5.6 Sol, Terra, and Luna — 262K context, max reasoning effort, Responses API path. Sol is the new Codex default.
• Kimi K3 — 1M context, 131K output, max effort.
• GLM catalog — GLM 5.x correctly recognized as 200K, 4.x as 128K. The unknown-model fallback dropped from a dangerous 1M to a conservative 200K so compression fires before the backend rejects you.
• LiteLLM provider — /provider litellm points at the OpenAI-compatible proxy on 127.0.0.1:4000 and reads LITELLM_API_KEY. Thanks @RheagalFire.
• Ollama Cloud and Makora aliases with their own endpoints, defaults, and OLLAMA_API_KEY / MAKORA_API_KEY.
• Refreshed Codex menu — GPT-5.6 family, 5.5, 5.4, 5.4-mini, 5.3-codex-spark; stale entries removed.
• reasoning.fieldName for Ollama and OpenAI-compatible streaming, with sensible auto-detection of reasoning_content then reasoning.

Notable Fixes

Streaming and cancellation

• Slow-but-alive models aren't killed for not emitting visible text yet; timeout errors now say which watchdog fired.
• The inter-chunk idle watchdog is off by default — 10 minutes still wasn't enough for frontier reasoning models. Configure it if you want it.
• ESC cancellation and idle aborts close the stream iterator properly, so the next prompt doesn't hang or abort until you restart.
• The first stream read happens inside the retry boundary, so connection-time 429s and 5xxs are retried instead of escaping instantly.
• Responses stream progressively again rather than buffering the whole thing before display.
• Fireworks/Kimi no longer withholds ordinary text waiting for Kimi markers, and terminal stop reasons are preserved so turns actually finish.

Load balancing, routing, and identity

• A 429-exhausted failover rotation resets for the next attempt instead of staying stuck on the quota-dead backend.
• An all-429 aggregate stays retryable rather than fatally dropping to the prompt after one cycle.
• Real provider quota errors are no longer replaced by a generic five-minute idle-timeout message.
• Profile load fails loudly on an unavailable provider instead of silently falling back to gemini; z.ai/GLM profiles stop getting routed to gemini-2.5-pro; Codex stays labeled Codex instead of turning into gemini:gpt-5.6-sol.
• Load-balancer runtime state can no longer be accidentally saved as an ordinary profile.
• The footer, response header, and diagnostics show the actual effective profile-qualified model.

Compression and context

• Preflight tries compression before cancelling a send; the hard-limit path makes a final attempt before producing diagnostics.
• An empty or masked summary no longer aborts your whole turn.
• /compress distinguishes "nothing to do", "recently compressed", and actual failure.
• Media blocks become text placeholders for providers that can't summarize them; overload/rate-limit failures during compression are retried; your most recent request is preserved so the model doesn't wake up with amnesia.
• A 0.5% margin stops turns failing by a handful of tokens.
• HTTP 413 on oversized tool calls now feeds the model guidance to retry smaller instead of ending the loop.
• Context limits apply on the first profile load, and provider switches re-estimate with the new tokenizer instead of double-counting cache and showing you 348,860/262,144.

Auth and keychain

• 401/403 revocations invalidate caches and force-refresh instead of retrying a locally-"valid" but rejected token.
• Cross-process refresh re-reads from disk after lock contention, so a second LLxprt instance can't replay Anthropic's single-use refresh token and revoke your session.
• Reauth is bounded instead of hanging a turn forever, and remaining buckets are authenticated together so failover doesn't pop another browser mid-turn.
• Third-party Anthropic-compatible endpoints (z.ai and friends) no longer trigger Anthropic OAuth when the profile has its own key and base URL.
• macOS: /key autocomplete no longer hits the Keychain on every keystroke and spams you with authorization prompts.
• Windows OAuth no longer launches a browser that hides or strands your terminal.

Subagents

• Launch crashes, post-scheduler-failure hangs, and stale timeouts fixed.
• Subagents inherit usable tool defaults and the parent's turn budget; disabled tools are hidden rather than throwing; the creation wizard no longer crashes on getModelBehavior is not a function.
• ESC actually cancels foreground subagent work, and a cancelled task no longer produces a bogus "concurrent executions" error on your next prompt.
• Streaming output accumulates into readable prose instead of being replaced token-by-token, and whitespace-only chunks are preserved.
• ACP/Zed: task no longer crashes on an unbound scheduler method, and todo_pause actually breaks the continuation loop.

Windows and terminal

• Storage EINVAL crashes from : in credential filenames, and /dev/tty crashes on a platform that has no /dev/tty. Thanks @HumanBot000.
• The launcher resolves native Bun correctly and preserves multiword and metacharacter arguments.
• CJK/IME input survives intact, and ripgrep results with C: drive-letter paths are no longer truncated.
• /copy doesn't crash; Windows Terminal uses OSC-52; an exited PTY no longer crashes on resize.

MCP and startup

• MCP discovery is non-blocking and non-fatal — no more black screen while servers connect, and no more silently swallowed prompts and shell passthrough during startup.
• Servers declaring JSON Schema draft 2020-12 (Playwright MCP and others) validate correctly instead of failing every advertised tool.
• The -i initial prompt is actually processed instead of just appearing in history.
• A fresh checkout with no git metadata can't crash startup.

Sessions and memory

• --continue restores normal wrapped chat instead of roughly one token per line.
• Project LLXPRT.md is found when you start LLxprt from a subdirectory, not just the workspace root.
• Settings search and typing flow, Home/End keys, external-editor fallback, and large-input rendering all fixed.
• /profile list supports direct d-key deletion with confirmation, and refuses to delete profiles a load balancer still references. Thanks @Ayush7614.

Breaking Changes / Action Required

1. Global config moved out of ~/.llxprt. Update anything that hard-codes that path.

   	Linux	macOS	Windows
   Config — settings, profiles, prompts, commands, skills, policies, sandboxes, subagents, hooks, global memory, .env	~/.config/llxprt-code	~/Library/Preferences/llxprt-code	%APPDATA%\llxprt-code\Config
   Data — credentials fallback, conversations, history, todos, user extensions	~/.local/share/llxprt-code	~/Library/Application Support/llxprt-code	%LOCALAPPDATA%\llxprt-code\Data
   Cache	~/.cache/llxprt-code	~/Library/Caches/llxprt-code	%LOCALAPPDATA%\llxprt-code\Cache
   Log/state — debug, temp, checkpoints, shell history, OAuth locks	~/.local/state/llxprt-code	~/Library/Logs/llxprt-code	%LOCALAPPDATA%\llxprt-code\Log

   Migration runs once, copies rather than moves, and records itself in <data>/.migration-complete.json. Your old ~/.llxprt is left alone — delete it yourself once you're satisfied. If you set LLXPRT_CONFIG_HOME, migration is deliberately skipped; unset it for one startup if you want the import. Project-local <project>/.llxprt does not move.

2. You must now choose a provider. New and unconfigured users get pointed at /setup. Automation must configure a provider or profile explicitly — there is no implicit Gemini path anymore.

3. Gemini Code Assist "Login with Google" OAuth is gone. Use GEMINI_API_KEY / keyfile or Vertex AI credentials. /auth gemini login is no longer a route.

4. Qwen OAuth is gone — Alibaba discontinued the free OAuth tier. Qwen is API-key-only now via DashScope (DASHSCOPE_API_KEY) or OpenRouter. OAuth providers are down to three: Gemini, Anthropic, Codex.

5. google_web_search and google_web_fetch are removed. Use exa_web_search and direct_web_fetch. Update any profiles, policies, subagents, or prompts naming them.

6. --experimental-ui and @vybestack/llxprt-ui (OpenTUI) are removed. Ink is the only terminal UI; that's where the effort goes now.

7. GEMINI_YOLO_MODE is no longer honored. Use the --yolo flag (or -y, or --approval-mode yolo) instead. We recommend the flag over an environment variable generally — env vars have a habit of leaking into places you didn't intend, and auto-approving every tool call is not something you want silently inherited by a shell, a CI job, or a child process. Note that gemini-cli interoperability is fully retained: GEMINI.md, .geminiignore, gemini-extension.json, .gemini/extensions, and GEMINI_API_KEY all still work.

8. Anthropic's unpinned default is now Claude Opus 5. Pin a model or profile if you need cost and behavior held stable.

9. For integrators only: legacy TypeScript aliases were removed. Use AgentEventType, ServerAgentStreamEvent, ServerAgent*Event, AgentErrorEventValue, ServerFinishedOutcome, InformationalStreamEvent; LlxprtExtension instead of GeminiCLIExtension; LLXPRT_CONFIG_DIR instead of GEMINI_DIR. GeminiCodeRequest has no alias — use ContentBlock[]/IContent or ToolResultContent. Anything launching dist/ or bundle/llxprt.js must switch to the package's llxprt bin. Direct AuthPrecedenceResolver consumers must pass providerKeyStorage or use createAuthPrecedenceResolver().

Thanks

Huge thanks to the contributors who sent pull requests in this release:

• @Ayush7614 — direct profile deletion, CI assignment action, provider test coverage
• @RheagalFire — LiteLLM provider
• @HumanBot000 — Windows storage and /dev/tty crash fixes

And to the upstream gemini-cli contributors whose work came in through the 0.24.5, 0.25.2, and 0.26.0 syncs.

[Ayush7614]

Let’s Go @acoliver