0.7.0 Released! - Codex authentication, Thinking, Caching, Oh my! And there is no more flicker! #907
acoliver
announced in
Announcements
0.7.0 Released! - Codex authentication, Thinking, Caching, Oh my! And there is no more flicker!
#907
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Release v0.7.0
This is a gigantic release. For high end users, have your GPT-5.2 (or GPT-5.2 codex) agent have its Opus 4.5 subagents do the work and if you hit your quota it can automatically failover to the next without stopping (assuming you pre-auth). For cheapskates, create a load balanced "alloy model" of every freetier known to you and round robin between them. For everyone cached token support across the board and no more dang UI flicker! And that's note even all of the "headline" features.
Installation
Install or upgrade LLxprt Code using npm:
Or use directly with npx:
The Bottom Line: 297 commits of pure feature density. If you've ever hit a rate limit, run out of API quota mid-task, or struggled with SSH-based OAuth flows - this release transforms your experience. The multi-account failover alone is a game-changer for power users.
Headlines (The Big Stuff)
Multi-Account OAuth with Automatic Load Balancing & Failover
The crown jewel of 0.7.0. Authenticate multiple accounts per provider (e.g.,
claude-work,claude-personal,claude-backup) and LLxprt automatically switches between them when you hit rate limits. No more downtime when one account gets throttled - it seamlessly fails over to your next authenticated account. Works across Claude, OpenAI, Gemini, and all OAuth-supported providers.OpenAI OAuth Support (Browser & Browserless)
Full OAuth integration for OpenAI (Codex). Works both in regular browser-based flows AND in SSH sessions/headless environments via device authorization - you get a code in your terminal that you can enter at https://auth.openai.com/codex/device from any device with a browser. No more API key management. (note that 5.2-codex came out right after the release so it isn't in the menu but if you type /model gpt-5.2-codex or start with llxprt --provider codex --model 5.2-codex then it works just fine)
Claude Opus 4.5 & Sonnet 4.5 Support
Native support for Anthropic's latest models including Claude Opus 4.5 and Sonnet 4.5 with full OAuth integration.
Advanced Reasoning Model Support
<think>blocks showing the model's internal reasoning process if you /set reasoning.includeInResponse truePrompt Caching Support & Optimization
Added full prompt caching support for both Anthropic and OpenAI providers with cache hit/miss metrics and cost savings tracking. Improved cache hit rates through optimized prompt construction settings (
include-folder-structure,enable-tool-prompts, and configurable cache TTL). See your token savings and estimated cost savings in real-time.Finally Fixed the Terminal Scrollback Spam
Completely rewrote the terminal UI rendering to eliminate the infamous scrollback spam/flicker that would fill your terminal buffer with repeated output during long sessions. Your terminal now stays clean and stable no matter how long you run or how many tools the model calls.
Opening Acts (Really Cool Stuff)
Kimi AI Provider Support
Native support for Kimi AI including their k2-thinking reasoning models with proper API key authentication. Full integration with the load balancing system.
OpenAI Vercel AI SDK Provider
Alternative OpenAI-compatible provider using Vercel's AI SDK. Experimental option that may become the default if it proves more stable/compatible than our current implementation. Supports reasoning models and various OpenAI-compatible gateways.
Project-Specific Memory
Use
/memory add "fact" --scope projectto save context specific to your current workspace, separate from your global memory.Smarter Stream Recovery
Automatic retry with exponential backoff when API streams fail or terminate unexpectedly. Much more reliable under spotty network conditions.
Honorable Mentions
Better Command Autocomplete with Fuzzy Matching
Improved autocomplete for
/set,/profile, and/authcommands with fuzzy matching support, including bucket suggestions for multi-account setups.Universal Cache Metrics
See prompt cache hit/miss stats across all OpenAI-compatible providers, not just first-party APIs. Track your token savings and estimated cost reductions.
Subagent Output Now Readable
Fixed subagent/task output to display properly across multiple lines instead of overwriting itself on one line.
Security Patches
Fixed multiple security vulnerabilities including HMAC signature verification and DNS rebinding protection in the MCP integration.
Better Remote Deployment
Fixed various issues when running over SSH or in cloud environments (logo rendering, OAuth flows, etc.).
Context Dumping
New
/dumpcontextcommand to export your full conversation context in native SDK formats for debugging or external analysis.What's New
Thanks
Huge thanks to the following contributors for their pull requests in this release:
All Changes
Full Changelog: v0.6.2...v0.7.0
All reactions