Skip to content

The release journal

Every release. A step forward.

See what's new, what's improved and what's been fixed. From the latest release to the first version — every change, in one place.

Explore the latest release Browse the archive
12 releases documentedv2.5.0October 13, 2026

The latest release

v2.5.0

More focused approvals, better context and safer restores

New capabilities 3

  • Focused approvals instead of repeated prompts

    Recurring test, build and development commands can be approved with an appropriate argument scope. The dialog shows the complete rule before you agree. Exact, one-time approval remains available for other invocations.

  • Trust mode directly in VS Code

    You can now enable or disable trust mode in the VS Code extension and see its current state. Enabling it requires explicit risk confirmation: it reduces normal approval prompts but does not provide an isolated execution environment.

  • From the dashboard to your first project

    The new quickstart walks you through CLI installation, sign-in and starting in your own project. It also explains how to set up the VS Code extension, including its CLI requirement.

Refinements 5

  • Updated model selection

    Model selection has been expanded and updated, including Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol and Lura. Available models still depend on your plan and team access policies.

  • Less repetition during longer tasks

    Recorded findings and decisions from earlier tasks can be retrieved selectively within a session. Unchanged file excerpts already retained in context are reused; overlapping reads add only the missing ranges.

  • More trial credit with a clear scope

    Credit for new trials has doubled: €10 for Pro, €15 for Pro+ and €20 for Ultra. Model selection during the trial is limited to Lura; image and video generation are not available during the trial.

  • Review existing permission rules when upgrading

    Configured approvals are presented for renewed review at startup. New Bash(path) rules apply to script and executable entrypoints; use CommandCwd(path) for working-directory approvals. Stored legacy approvals keep their previous meaning. Settings are not rewritten automatically.

  • General improvements

    Further stability and security improvements, updated dependencies and more reliable CLI and VS Code extension publishing.

Reliability & fixes 3

  • Checkpoints leave unrelated files alone

    Restoring a checkpoint now restores only the files involved in the changes being rolled back. Other project files are preserved. Checkpoint compatibility on Windows has also been improved. Checkpoints cover native file changes, not every side effect of shell commands.

  • More reliable plan continuation and completion

    Implementation plans are now bound more reliably to their task. Plans and progress stay more consistent when activating and continuing work; completion takes recorded verification results into account.

  • Fewer interruptions during browser checks

    Browser sessions now remain available for up to 15 minutes without activity. Saving screenshots correctly respects the configured project directory and permitted temporary locations.

Explore the handbook

Release notes

The story so far.

Earlier releases, in reverse chronological order. Open a version to read its full release notes.

v2.3.0

Compatibility API, Team Controls, and More Reliable Workflows

New capabilities 4

  • OpenAI- and Anthropic-compatible API

    New endpoints for OpenAI Responses, Chat Completions, and Anthropic Messages provide complete and streamed responses. The model catalog is available through /v1/models.

  • New API workspace in the dashboard

    API keys, available models, and ready-to-copy examples for cURL, TypeScript, and Python are now collected in one place.

  • Central model access for teams

    Team administrators can define which models members may use. Policies are applied consistently across the dashboard, CLI, and API endpoints.

  • Activatable implementation plans

    Created plans can be activated deliberately and continued reliably across long-running tasks. Progress and planning survive context compaction.

Refinements 3

  • Updated model catalog

    New and updated models, including GLM 5.3 and current Gemini variants, are available with aligned metadata and model aliases.

  • Clearer permissions and interface

    The permission system and interaction interface have been extensively revised, making decisions, changes, and ongoing activity easier to follow.

  • More precise CLI output

    Tool results, long process output, and plan status are presented more compactly and reliably without losing important diagnostic information.

Reliability & fixes 4

  • More reliable streaming and disconnect handling

    Streaming responses now finish more cleanly when clients disconnect or backend connections close, reducing stuck or incomplete requests.

  • GitInspect fixed without an explicit directory

    GitInspect now reliably uses the current project when no cwd is provided.

  • Thinking output visible again

    Reasoning and thinking content is once again passed correctly to the CLI and VS Code for supported models.

  • Improved MCP and Windows compatibility

    MCP tools work outside Agent and Debug modes again. Lurus Code also starts more reliably on Windows when npm uses a non-standard installation path.

v2.0.0

Claude Opus 5, GPT-5.6, and Extensible Workflows

New capabilities 4

  • Claude Opus 5

    Opus 5 is the new leading Anthropic model in Lurus Code. It is designed for highly demanding coding tasks and supports tool calling, vision, adaptive thinking, and prompt caching.

  • GPT-5.6 Luna, Sol, and Terra

    The new GPT-5.6 model family covers fast through highly demanding coding tasks. All three models support tool calling, vision, reasoning, and prompt caching.

  • Custom agents, skills, and hooks

    Lurus Code now loads custom agents, skills, slash commands, and workflow hooks consistently from project-wide and global configurations. Content in nested folders is supported as well.

  • Remote MCP servers via URL

    In addition to local processes, MCP servers can now be connected directly through URLs. Headers and credentials are managed securely, making external tools easier to integrate into workflows.

Refinements 4

  • Tasks remain stable across long sessions

    A new Task Ledger preserves confirmed decisions, progress, and verification results across context compaction. Complex tasks can therefore continue more reliably without losing essential context.

  • Safer permission workflow

    Commands and file operations are classified more precisely, and approvals are handled more consistently. The new Exec flow clearly separates direct programs from shell commands and reduces unnecessary prompts.

  • More token-efficient source views

    Large files are loaded through structured views and targeted ranges. This keeps more relevant code in context and makes reading and subsequent edits more reliable in large projects.

  • Model pricing shown directly

    The CLI and VS Code extension now show input, output, and cache prices during model selection, making costs easier to compare before switching models.

Reliability & fixes 3

  • Code indexing stabilized

    Large folders, previously missing files, and disabled indexing in new projects are handled more reliably. This reduces memory issues and incomplete search results.

  • Parallel sign-in in terminal and VS Code fixed

    Concurrent sessions in the VS Code extension and terminal no longer overwrite their authentication data and remain connected reliably.

  • Checkout fixed for additional languages

    Checkout now carries language and consent text correctly outside the English interface as well.

v1.9.0

New Model Generation, Image Editing, and More Stable Agent Execution

New capabilities 3

  • New EU-hosted models

    The model catalog has been expanded with new EU-hosted and regional models. Aliases, metadata, and backend validation were updated so CLI, subagents, and backend use the same model list.

  • GPT Image 2 for image editing

    Images can now be edited directly through Lurus Code. Backend, slash commands, provider metadata, and the new edit-image tool were extended for GPT Image 2.

  • Improved web search

    Web Search and Web Fetch were hardened with URL safety checks, content limits, and additional tests. Results are now more reliable and safer for agentic workflows.

Refinements 3

  • Dynamic output limits for low balance

    When credit is low, Lurus Code reserves tokens more defensively and adjusts output caps dynamically. Chat requests stay more predictable without blocking unnecessary balance.

  • Cleaner model routing

    Open-source/global providers, thinking types, and max-token handling were unified. This reduces provider errors and makes model switching more robust.

  • More consistent Bash output handling

    Long terminal outputs remain easier to inspect and are no longer summarized where full logs are needed for debugging or agent decisions.

Reliability & fixes 2

  • Reindexing stabilized

    Issues during codebase reindexing have been fixed. The code index refreshes more reliably and stays consistent across agent runs.

  • Provider details protected

    Responses and help text were hardened so internal provider details are not accidentally exposed in user-facing surfaces.

v1.8.0

Team Module, Lura Model, and More Robust Context Compaction

New capabilities 3

  • Lurus team module

    Teams can invite users, manage shared team balance, and use centralized billing flows. Backend, dashboard, email templates, and CLI displays were extended for this.

  • Lura model integrated

    Lura was added as a dedicated model including balance logic, model aliases, dashboard displays, and billing paths.

  • GPT-5.5 (US)

    GPT-5.5 (US) was added to the registry, CLI selection, and fallback chains for powerful coding tasks.

Refinements 3

  • Much more robust context compaction

    Compaction now preserves plans and central session artifacts more reliably. Long sessions lose less context and remain easier to steer.

  • Deeper code analysis for GPT

    GPT models receive better analysis guidance for code understanding, semantic search, and subagent tasks.

  • More cost-aware agent prompts

    System prompts and tool guidance were tightened so large tasks run with less unnecessary context and lower cost.

Reliability & fixes 3

  • IP geolocation fixed

    Checkout and trial flows detect countries more reliably and handle geolocation errors more robustly.

  • Refresh-token race condition fixed

    Multiple parallel refresh attempts no longer cause inconsistent sessions or broken login states.

  • Redaction warnings improved

    The secret scanner no longer shows misleading warnings when output has already been safely redacted.

v1.7.0

Orchestrator v2, New Models, and More Stable Long Sessions

New capabilities 2

  • Orchestrator v2

    The orchestrator can now continue, overwrite, and coordinate plans with subagents more deliberately. This significantly improves complex multi-step tasks.

  • New Claude and Gemini models

    Opus 4.8 was added, Gemini models were refreshed, and outdated Gemini variants were removed.

Refinements 3

  • Long chat histories are easier to use

    Orchestrator and review flows handle long histories better and load context more efficiently.

  • CLI embedding timeout added

    Embedding calls in the CLI now fail in a more controlled way when external services do not respond in time.

  • Trial balance adjusted

    The trial period uses more appropriate balance values and displays account and payment states more clearly.

Reliability & fixes 4

  • Trial-abuse protection improved

    Repeated trial usage through payment data is prevented more reliably.

  • Checkout and extension build fixed

    Issues in the checkout flow and VS Code extension build were resolved.

  • Chat rendering stabilized

    Old Lurus Code chats, Markdown lists, and bullet lists render more consistently again.

  • Plans and permissions are preserved better

    Compaction and vendor-branch detection are less likely to lose plans or permission decisions.

v1.6.0

Free Trial, Better VS Code Workflows, and CLI UX

New capabilities 3

  • Free trial instead of Starter

    The previous Starter entry point was replaced with a free trial. Checkout, emails, dashboard, and subscription logic were aligned with the new flow.

  • VS Code debug support

    The VS Code integration now supports debugging flows better and connects editor context more directly with Lurus Code.

  • New shortcuts and /paste

    New mode shortcuts and the updated /paste command speed up recurring terminal workflows.

Refinements 3

  • Clearer thinking output

    Reasoning and thinking blocks render more readably and reliably in the CLI.

  • Better structured CLI output

    Terminal output, status messages, and longer agent responses are organized more clearly.

  • More CLI test coverage

    Additional tests cover chat, slash command, and rendering behavior more reliably.

Reliability & fixes 3

  • Session resume and settings persistence hardened

    Markdown rendering, session resume, recall loading states, and settings persistence have been stabilized.

  • Ctrl+C cancellation stabilized

    Interrupts now stop running actions more reliably without unnecessarily damaging the full session.

  • /compact fixed

    The compact command works more reliably and preserves important session information more consistently.

v1.5.0

Cache, Permission, and Orchestrator Stability

New capabilities 2

  • Legal links in emails

    Transactional emails now contain consistent links to imprint, privacy policy, and other legal information.

  • More stable provider system payload

    Provider payloads and skill-cache behavior were stabilized so agents start from more consistent conditions.

Refinements 3

  • Permissions without hard wait timeouts

    Lurus Code now waits more reliably for user decisions in permission prompts while saving context cost by default.

  • Prompt cache protected

    Micro-compaction was removed because it could break prompt caches. Compaction now runs more deliberately when the context window is actually full.

  • Unified orchestrator defaults

    Orchestration flows now use more consistent provider and context defaults.

Reliability & fixes 3

  • OpenAI cache loss prevented

    Cache loss caused by changing project-tree context and redirected requests has been fixed.

  • Mode switching blocked during active runs

    Agents can no longer accidentally switch mode, thinking, or effort while a run is active.

  • Orchestrator rendering and permissions fixed

    Orchestrator views render more reliably and are less likely to lose permission decisions.

v1.4.0

New Fast Models, Better Caches, and Clearer Ignore Rules

New capabilities 3

  • Kimi K2.6 and GLM 5.1 Fast

    Kimi K2.6 and GLM 5.1 Fast were added as fast model options.

  • Long-context threshold pricing

    Pricing logic now accounts for long-context thresholds so large context windows are calculated more transparently.

  • Dashboard service worker

    The dashboard received service-worker support for a more robust web-app experience.

Refinements 3

  • .lurusignore and .gitignore clearly separated

    .lurusignore is now a hard deny mechanism, while .gitignore can be requested per operation.

  • Fast-model fallback chains corrected

    Fallback chains for fast models were cleaned up and now provide more reliable alternatives.

  • Model index cleaned up

    The model index and model patterns were cleaned up so selection and aliases behave more consistently.

Reliability & fixes 4

  • Plans survive compaction more reliably

    Plans are less likely to be lost during compaction and remain available across longer sessions.

  • Prompt cache stabilized across model and mode switches

    Prompt caches are no longer invalidated unnecessarily when switching models or modes.

  • API-key and key-service edge cases fixed

    API-key lengths and slow key-service connections no longer lead to broken logout or schema states.

  • Blocking tools and tool calls fixed

    Issues with blocking tools, visible system reminders, and broken tool calls were resolved.

v1.3.0

New Model Generation & Global Providers with Zero Data Retention

New capabilities 4

  • GPT-5.5 — new OpenAI flagship

    GPT-5.5 is now available and replaces GPT-5.4 Pro as the default Powerful OpenAI model. 82.7% on Terminal-Bench 2.0 — new state of the art for agentic coding — with a 1M context window and dramatically better long-context performance (74% MRCR v2 at 512K–1M tokens).

  • Kimi K2.6 — hosted in Europe

    Moonshot AI’s latest flagship (80.2% SWE-Bench Verified, 58.6% SWE-Bench Pro — on par with GPT-5.5). Hosted in Europe with Zero Data Retention guarantees. Open-weight, 256K context, built for long-horizon agentic work.

  • GLM 5.1 — hosted in Europe

    Z.ai’s new flagship (58.4% SWE-Bench Pro, 63.5% Terminal-Bench 2.0) is the first open-source choice for long-horizon coding with up to 8 hours of autonomous execution. Hosted in Europe with Zero Data Retention guarantees. Open-weights model, independently deployable.

  • New US-hosted global providers with Zero Data Retention

    New US-hosted models are available optionally: DeepSeek V4 Pro, Qwen 3.7 Plus, Kimi K2.7 Code, Nemotron 3 Ultra NVFP4, MiniMax M3, and the existing Global Provider models. All accessed through our gateway provider with contractually enforced Zero Data Retention and no-training policy.

Refinements 2

  • Pricing breakdown updated

    The model cost page now lists every new model — including Lura, Kimi K2.6, GLM 5.1, DeepSeek V4 Pro EU, and the new Global Provider models — with input, output and cache-read pricing per 1M tokens.

  • Models page restructured

    The AI Models page now clearly separates Lura, regional providers, and global providers with an explicit Zero Data Retention notice. See at a glance which models qualify for regulated workloads.

v1.2.0

Smarter Agent Execution & Clearer Billing

New capabilities 1

  • Redesigned Ask-Question Experience

    The agent's clarifying questions got a full UX overhaul in both CLI and VS Code: cleaner multi-choice layouts, free-text answers, multi-select support and a faster way to skip or abort — so you stay in flow when the agent needs input.

Refinements 3

  • Free Trial Added

    Pro, Pro+ and Ultra can now be tried free for one week. The related CLI and VS Code messaging has been aligned accordingly.

  • Clearer Balance & Quota Messages

    Balance-exhausted and low-quota messages are now explicit, actionable, and consistent across CLI and VS Code. When you hit a limit, you immediately see what happened and how to continue — including a one-click upgrade path in the VS Code sidebar.

  • Smarter Image Context Handling

    On long sessions with many screenshots or diagrams, only the most recent images are sent to the model in full; older image blocks are replaced by lightweight placeholders. Result: faster responses, lower token usage, same context awareness.

Reliability & fixes 3

  • Stable Request Cancellation

    Pressing Ctrl+C (or otherwise cancelling an in-flight request) now cleanly stops just that request instead of tearing down the whole CLI session. Your chat, context and working directory stay intact.

  • Reliable GPT Reasoning & Tool Streaming

    Resolved an issue where GPT models would occasionally stop streaming reasoning or go silent between tool calls. The agent loop now reliably continues and streams status updates throughout long tool chains.

  • Hardened Agent-Loop Stability

    Several edge cases in the agent loop and prompt service have been hardened — fewer stalls, better recovery from provider errors and more predictable behaviour on multi-step tasks.

v1.0.0

Official Release

New capabilities 14

  • Autonomous AI Coding Agent

    Plan, implement and review code with a single prompt. Supports multi-file edits, automatic error recovery and context-aware suggestions across your entire codebase.

  • Multi-Agent Orchestration

    Break down complex tasks into parallel subtasks executed by specialized subagents. The orchestrator plans, delegates and verifies — fully autonomous.

  • AI Code Review

    Three-phase code review with severity ratings, OWASP vulnerability detection and auto-fix suggestions. Runs on every commit or on demand.

  • Security Scanning

    12 vulnerability categories including SQL injection, XSS, path traversal and hardcoded secrets. Generates actionable fix recommendations.

  • Testing & TDD

    Generate unit tests, integration tests and run TDD workflows. The agent writes tests, runs them and iterates until they pass.

  • VS Code Extension

    Full integration with inline editing, diff review, chat panel, and session management — directly in your editor.

  • CLI

    Terminal-first workflow with batch processing, CI/CD integration, and non-interactive mode for automated pipelines.

  • API Keys

    Token-based authentication for CI/CD and headless environments. Create, manage and revoke keys from the dashboard with SHA-256 hash storage.

  • Skills & Custom Agents

    Install community skills, create custom agent personas and extend Lurus Code with reusable automation packages.

  • MCP Integration

    Connect external tools and data sources via the Model Context Protocol. Use databases, APIs and third-party services as AI context.

  • Git & GitHub Integration

    Automated commits, branch management, PR creation and interactive rebase — all driven by the AI agent.

  • Documentation Generation

    Generate README files, API docs, JSDoc comments and architecture documentation from your codebase.

  • Permission System

    Sandbox isolation, bash safety analysis, sensitive file protection and persistent permission rules.

  • Multi-Model Support

    Choose from Claude, GPT, Gemini and more. Switch models per task or let the agent select the best one automatically.

Refinements 2

  • GDPR-Compliant by Design

    Processing depends on model and region. No training on your code. DPA available.

  • Optimized Performance

    Context compaction, intelligent caching and parallel tool execution for fast response times even on large codebases.

Lurus Code

Read the changes. Try the workflow.

Explore the handbook or open the dashboard to put Lurus Code to work on your next task.

Open the dashboardExplore the handbook

A deliberate choice

Before your first prompt.

Interactive demo
Lurus Codeprivacy-demo · illustrative workspace

Demo session ready. Your model choice and data route are separate decisions.

Illustrative scenario. No real sign-in, security test or code transmission takes place.

Image preview