25 Apr 16:23

ef6928d

1.5.2-beta Latest

Latest

Meridian 1.5.2-beta

Meridian 1.5.2-beta is a focused pre-release update that expands subscription-provider capabilities, adds OpenAI Codex image generation, refreshes OpenCode Go model support, and fixes provider tool and upload edge cases.

Highlights

OpenAI Codex Image Generation

OpenAI Codex can now be used as a provider-backed image generation runtime.

Added a dedicated Codex Image Generation model entry derived from the connected OpenAI Codex catalog.
Enabled Codex's native image generation tool for image requests, including prompt suffixing for aspect ratio and target resolution.
Added support for text and image inputs in Codex image generation prompts.
Added generated-image extraction from saved file paths, data URLs, and base64 payloads.
Routed provider image generation through connected OpenAI Codex credentials instead of rejecting Codex models.

OpenCode Go Model Updates

The OpenCode Go catalog has been updated for newer model coverage and multimodal metadata.

Added DeepSeek V4 Pro and DeepSeek V4 Flash.
Added MiMo-V2.5-Pro and MiMo-V2.5.
Updated MiMo context lengths and model definitions.
Marked image-capable OpenCode Go models so Meridian can accept image inputs only where supported.
Added image input support for Kimi K2.5, Kimi K2.6, MiMo-V2-Omni, MiMo-V2.5, and Qwen3.6 Plus.

Tool Use and Reasoning Stability

Provider tool execution is more reliable for newer OpenCode Go models.

Preserved reasoning_content across OpenAI-compatible tool-call round trips for models that require it.
Fixed tool usage with the new DeepSeek V4 models in OpenCode Go.
Fixed tool usage with Kimi K2.x models in OpenCode Go.
Improved Anthropic SSE event parsing so events without explicit event: lines can still be handled from their JSON payload type.
Increased OpenAI Codex stdio buffering to handle larger runtime output.

File Upload Fix

Fixed device file upload in file prompt nodes by stabilizing the dropped file list before async folder and upload work begins.

Self-Hosting and Upgrade Notes

Run normal deployment updates for the backend and frontend.
No database migration is required for 1.5.2-beta.
OpenAI Codex image generation requires connected OpenAI Codex credentials and the existing Codex runtime dependencies from the 1.5.0-beta line.

Assets 2

20 Apr 20:50

MathisVerstrepen

1.5.1-beta

dd759fa

1.5.1-beta

What's Changed

feat: add Kimi K2.6 support to OpenCode Go

Assets 2

19 Apr 20:38

MathisVerstrepen

1.5.0-beta

1999b26

1.5.0-beta

Meridian 1.5.0-beta

Meridian 1.5.0-beta is a large pre-release focused on subscription-backed model access, a rebuilt model selection flow, richer usage visibility, and better runtime support for provider-native coding agents.

Highlights

Subscription Inference Providers

Meridian can now work with multiple subscription-backed provider runtimes alongside metered OpenRouter models.

Added support for Claude Agent, GitHub Copilot, Z.AI Coding Plan, Gemini CLI, OpenAI Codex, and OpenCode Go.
Added a new Account Providers settings tab for connecting and disconnecting provider credentials from the UI.
Subscription models now carry richer metadata, including billing type, connection requirements, and Meridian tool compatibility.
Available subscription models refresh automatically after provider credentials are added or removed.

Model Selector Redesign

The model dropdown has been redesigned to work well with a much broader catalog.

Split the dropdown into Pinned Models, per-provider subscription sections, and All Models for metered entries.
Added section metadata, jump navigation, and virtualization for smoother browsing across large model lists.
Added better pinned-model handling, including keyboard shortcut support.
Added warning labels when a model does not support native JSON output.

Usage Data and Cost Visibility

Usage reporting is now much more detailed, especially for multi-request and tool-heavy runs.

Added normalized usage aggregation with request-level breakdowns.
Added cost_details support so upstream inference cost components can be shown explicitly.
Added per-tool usage tracking for subscription providers.
Redesigned the usage popover to show totals, individual request passes, and parallelization model breakdowns in a clearer way.
Improved finish-reason and tool-name visibility inside usage details.

Tooling Across Providers

Subscription providers now integrate more deeply with Meridian's tool system.

Added Meridian tool support across subscription-backed providers for web search, link extraction, code execution, image generation, Visualise, and Ask User.
Added a provider-aware image generation service path instead of routing everything through the older shared flow.
Improved tool runtime registration and summary rendering for richer provider-native tool call output.
Added tool call duration tracking and chat display for faster debugging and better user feedback.

Ask User and Chat Resume Improvements

User-interruptible tool flows are more resilient and easier to continue.

Added local draft persistence for pending tool questions so answers survive refreshes and accidental navigation.
Improved resume behavior after submitting tool responses.
Added better handling for pending tool call IDs across backend services and frontend components.

Canvas and Navigation

Added a quick-create flow that places new graphs directly inside the currently selected folder or workspace from the home page.
Improved recent-canvas and sidebar history interactions for folder-heavy workspaces.
Fixed routing node execution issues.
Improved parallelization and text-to-text node behavior around cancellation and stream state handling.

Self-Hosting and Upgrade Notes

Self-hosted deployments should treat 1.5.0-beta as an upgrade that adds new runtime requirements.

Run alembic upgrade head. This release adds a migration for tool call duration tracking.
Backend environments now need Node.js and npm available for the bundled Gemini CLI and OpenAI Codex bridge runtimes.
Local installs should also install the runtime package dependencies in api/app/gemini_cli_runtime and api/app/openai_codex_runtime.
Users who want subscription-backed model access must connect provider credentials from Settings -> Account -> Providers.

Assets 2

09 Apr 20:58

MathisVerstrepen

1.4.1-beta

c420d6a

1.4.1-beta

Hotfix 1.4.1-beta

Fix OpenRouter structured-output compatibility by sanitizing unsupported integer JSON Schema constraints generated from Pydantic models.
Fix visualisation structured requests so they use the same OpenRouter-compatible schema formatting.
Fix prompt improver history labels so run entries reflect the actual selected optimizer model.

Assets 2

09 Apr 19:36

MathisVerstrepen

1.4.0-beta

339868f

1.4.0-beta

Meridian 1.4.0-beta

Meridian 1.4.0-beta is a feature-heavy update centered on better prompt authoring, a more capable tool runtime, stronger chat rendering, and a broad round of reliability and security hardening across the stack.

Highlights

Prompt Improver

Meridian now includes a dedicated Prompt Improver workflow for prompt nodes.

Audit an existing prompt and score it across structured improvement dimensions.
Select recommended dimensions or choose your own optimization focus areas.
Let the optimizer ask clarification questions before rewriting when more context is needed.
Review suggested changes one by one, accept or reject them, then apply the improved prompt back to the graph.
Re-open past improvement runs from node history.
Override the optimizer model globally in settings or select a different optimizer model directly from the Prompt Improver dialog.

New Tool Runtime

The tool system has been expanded substantially.

Added sandboxed Python code execution with bounded resources, artifact persistence, and inline chat rendering for generated files.
Added the Visualise tool for Mermaid, SVG, and HTML outputs.
Added structured Ask User tool support so a model can pause, ask a clarifying question, and continue after the user replies.
Added persistent tool call storage and richer formatted tool call detail views in chat.
Added per-tool and per-node configuration for visual generation behavior and model selection.

Chat, Markdown, and Mermaid

Rendering and chat ergonomics received a large pass.

Markdown processing is more resilient around tool output, artifact links, and mixed structured content.
Mermaid rendering now supports validation and optional retry flows when generated syntax is invalid.
Mermaid runtime loading was streamlined to reduce unnecessary upfront cost.
Assistant message copy now excludes tool call noise.

Canvas and Navigation

Added graph loading and error states for a safer graph-opening flow.
Added custom 401, 403, 404, and 500 error pages.
Improved edge compatibility and snapping on the graph canvas.
Added pagination for graph history and recent-canvas loading, which improves responsiveness on larger workspaces.
Continued UI cleanup, including canonical Tailwind class migration and removal of obsolete graph UI pieces.

Fixes and Hardening

Security and Provider Safety

Enforced repository path restrictions under the clone root.
Enforced SSH host key verification for GitLab clones.
Stopped leaking Git credentials through URLs and telemetry.
Fixed OpenRouter auth header reuse across requests.
Improved shared HTTP client behavior and non-git HTTP/2 handling.

Reliability and Correctness

File uploads no longer materialize the full payload in memory before processing.
Account deletion now removes on-disk data correctly.
Rate limiting was rewired correctly.
Startup and shutdown lifecycle handling now tracks cron tasks correctly.
Canvas persistence now updates PostgreSQL and Neo4j more consistently.
Fixed inconsistent GitLab behavior, including multi-instance handling.
Fixed canvas backup import when workspace_id is missing.
Fixed file upload node refresh behavior in the graph after attachment changes.
Fixed several chat resume and user-awaiting state edge cases.
Tightened validation around file-view query parameters.

Self-Hosting and Upgrade Notes

Self-hosted deployments should treat 1.4.0-beta as an upgrade that requires both config and infrastructure updates.

Update your Docker configuration. 1.4.0-beta introduces a dedicated sandbox_manager service and a new [sandbox] config section.
Code execution and generated artifact workflows depend on the sandbox manager and its worker image.

Assets 2