Skip to content

Releases: DeepMyst/Mysti

Release list

v0.4.0

Choose a tag to compare

@BahaAbuNojaim BahaAbuNojaim released this 11 Mar 19:33

What's New

4 New Providers (12 Total)

  • OpenCode: Multi-backend agent (Anthropic, OpenAI, Google, Groq) — closes #25
  • Qwen Code: Alibaba's AI coding CLI with deep reasoning
  • Ollama: Local LLM inference, privacy-first — closes #24
  • LocalAI: Self-hosted AI models — closes #24

18 Brainstorm & @-Mention Stability Fixes

  • Silence-based timeout — agents aborted after 90s of no output (no more infinite waits)
  • Auth pre-check — validates provider authentication before brainstorm starts
  • Synthesis fallback — UI shows "retrying with [agent]..." when primary synthesis fails
  • Oscillation detection — catches flip-flopping discussion positions
  • Convergence regex broadening — handles varied score phrasings in Delphi mode
  • Duplicate agent guard — prevents selecting the same agent twice
  • Cancel propagation — stops all sub-processes immediately on cancel
  • Sub-agent question timeout — auto-skips after 5 minutes if user doesn't respond
  • Max 5 mentions per message — prevents resource exhaustion
  • Refined mention parsing — tighter regex, path boundary matching, agent validation
  • File resolution warnings — user notified when file mentions can't be read
  • Retry process cleanup — cancels previous attempt before retrying

Bug Fixes

  • Windows: Fixed spawn EINVAL error with auto shell mode — closes #14
  • Codex: Brainstorm now respects mysti.codexPath setting — closes #26
  • Mentions: Full-path file mention matching (@src/utils.ts now works)
  • Mentions: Invalid agents no longer produce misleading "CLI not installed" errors
  • Brainstorm: Empty contributions no longer cause false convergence
  • Brainstorm: Text similarity no longer drops meaningful short words like "not"

Test Infrastructure

  • 360 automated tests via vitest
  • Mock provider infrastructure for brainstorm and mention testing
  • Per-provider test suites (buildCliArgs, parseStreamLine, permissions)

New Settings

Setting Default Description
mysti.opencodePath opencode Path to OpenCode CLI
mysti.opencodeModel `` Custom OpenCode model
mysti.qwenCodePath qwen Path to Qwen Code CLI
mysti.qwenCodeModel `` Custom Qwen model
mysti.ollamaPath ollama Path to Ollama CLI
mysti.ollamaModel `` Custom Ollama model
mysti.localaiPath localai Path to LocalAI CLI
mysti.localaiModel `` Custom LocalAI model
mysti.useShellForCli false Run CLIs with shell (auto-enabled on Windows)

Closes

  • #14 (spawn EINVAL on Windows)
  • #24 (Ollama + LocalAI support)
  • #25 (OpenCode support)
  • #26 (Codex codexPath in brainstorm)

Full Changelog: v0.3.1...v0.4.0

v0.3.1

Choose a tag to compare

@BahaAbuNojaim BahaAbuNojaim released this 18 Feb 13:31

Mysti v0.3.1

What's New in v0.3.0 / v0.3.1

This is the first published GitHub release for Mysti — covering all features through v0.3.1.

3 New Providers (7 total)

  • Cursor — Full integration with Cursor's headless AI agent CLI (6 models including Claude Sonnet 4, GPT-5, o3)
  • OpenClaw — Dual-transport provider with WebSocket Gateway and CLI fallback
  • Manus (Experimental) — HTTP API-based provider for Manus AI

Autonomous Mode

AI works independently with configurable safety controls — SafetyClassifier with three safety levels, MemoryManager that learns preferences, goal-based and task-queue continuation modes, and audit logging.

@-Mention System

Route tasks to specific providers with @agent mentions, attach file context with @file, with heuristic-based task generation, auto-retry, and dependency tracking.

Brainstorm Team Reasoning (5 Strategies)

  • Quick — Direct synthesis
  • Debate — Critic vs Defender
  • Red-Team — Proposer vs Challenger
  • Perspectives — Risk-Analyst vs Innovator
  • Delphi — Facilitator vs Refiner
  • Convergence detection, parallel discussion, round markers, role badges

Context Compaction

Smart conversation compaction to prevent context overflow — native CLI strategy for Claude (/compact), client-side summarization for others, per-panel token tracking with threshold-based triggering.

Slash Command System

Centralized command registry organized by sections — context, model, customize, commands, settings, support. Provider-specific commands like /compact, /thinking, /profile, /plan-act.

Per-Panel Session Isolation

Each webview panel has fully independent state, provider sessions, context, and process management.

Agent Lifecycle Management

Session lifecycle with configurable idle timeout (default 1 hour), cross-platform process tree tracking, and graceful shutdown.

v0.3.1 Fixes

  • Security: Exclude .claude/ and .env from git and vsix package
  • Contributor: Merged Patrick Fu's custom model configuration contribution

Installation

Download mysti-0.3.1.vsix and install via:

code --install-extension mysti-0.3.1.vsix

Or install from the VS Code Marketplace.

Contributors

  • @patrick-fu — Custom model configuration & provider settings