Changelog
Notable, user-visible changes to Dispatch.
v0.1.5
April 21, 2026
Windows end-to-end, in-app changelog, and reliability polish.
Reliability
- Agent CLI detection now survives cold-start races. If Claude, Codex, or Gemini briefly doesn't answer on launch, Dispatch retries instead of caching "not installed" for the rest of the session.
- Settings diagnostics no longer surface "Unexpected token '<'" when a response round-trips through a transport that strips its content-type header. Valid JSON bodies are accepted whatever header they arrive with.
- The version shown in the app, the installer, and the auto-updater are always in sync with the published release.
Windows
- Missions dispatch without the previous "failed to start mission — is the server running?" false alarm. CLIs installed inside a WSL distro are wired up end-to-end.
- Custom Agents → Add Agent has a WSL toggle so CLIs installed inside a distro are one click away.
- Workspace Isolation offers a WSL engine and an optional Docker-via-WSL path.
- Agent dropdowns no longer list CLIs that aren't actually installed.
Updates and changelog
- "What's new" opens an in-app modal with proper formatting.
- "View changelog" opens the full release history in-app, including the channel you're on.
- Install failures surface the real error inline instead of silently reverting.
Open folder
- GitHub import and Open folder fall back to an in-app directory browser when the native picker isn't available. You're no longer stuck typing an absolute path.
About
- A clear "No agent CLIs detected — install one, then open Settings" callout replaces the ambiguous empty list when nothing is configured.
Missions
- Picking an adapter now turns off auto-model so your explicit choice sticks.
- Rate-limited runs that fall back to cloud models correctly forward attached images to vision-capable models.
Notifications
- Toggling desktop notifications on fires a confirmation notification so you know permissions are wired up.
Performance
- Start-up runs fewer redundant checks.
v0.1.4
April 21, 2026
Cross-platform reliability, Windows support, and goal lifecycle control.
Added
Windows via WSL
- Dispatch detects Claude Code, Codex, and Gemini CLI installed inside a WSL distro and runs them transparently
- If you keep multiple distros, opening a project under
\\wsl$\<distro>\...pins Dispatch to that distro - Settings has a new WSL panel with per-distro detection, per-adapter install state, and a retry button
Pause and resume goals
- Pause any running goal; agents terminate cleanly and context is preserved
- Resume picks up exactly where the agents left off, with full context
- Dispatch auto-pauses when the network drops for more than five seconds; resuming stays a deliberate action
Premium fallback when a CLI gets throttled
- When your local CLI hits its provider's usage limit mid-task, Pro subscribers can configure Dispatch to silently continue on a Premium model with full context handoff
Mission plans
- Goals that span multiple steps are now decomposed into ordered plans
- Plans view shows per-plan progress, step sequencing, and the goals backing each step
- Steps unlock automatically as their predecessors complete
Import projects without composing a goal
- The sidebar is visible from the moment you open Dispatch
- Three bottom buttons — Open folder, Remote Hosts, Open from GitHub — cover every import path without having to start a goal first
Improved
- macOS tray icon tints correctly against the menubar in light, dark, and vibrancy backgrounds
- Windows icon is sharp at every size Windows asks for (taskbar thumbnails, Alt-Tab previews, the large icon view in File Explorer)
- Wizard and Settings empty states cleaned up, with a GitHub fallback when no local CLI is installed yet
- Keyboard shortcuts show
Ctrl+on Windows instead ofCmd+ - Spec revision now feeds your rejection comment and the previous spec body back into the regeneration prompt — revisions address your feedback instead of regenerating blind
Security
- Tightened path handling when importing projects: symlinks cannot escape the chosen folder, and sensitive system paths (Windows, Program Files, ProgramData,
.ssh,.aws,.gnupg, and the POSIX equivalents) are rejected - Origin and authorization checks hardened on Pro authentication, remote file access, and admin endpoints
- File size caps enforced at the request boundary
- Activation code paste works correctly in the wizard, Settings, and sidebar
Reliability
- Fresh installs on Linux and Mac no longer show a blank window
- Several build-time dependency gaps fixed
v0.1.0
April 20, 2026 · Initial public release
A local-first environment for directing AI coding agents through a goal → spec → task pipeline, with optional hosted models.
Added
Agent orchestration
- First-class adapters for Claude Code, OpenAI Codex, and Gemini CLI; custom CLI adapters for anything else
- Per-stage agent and model assignment via Pipeline Profiles
- Mission Mode (
Cmd/Ctrl+M) for quick single-agent tasks; Research, Audit, and Plan modes included - Auto Model: intent-based routing across the agents you have installed
- Automatic escalation to a stronger model when a run fails
- Promote a Mission to a full Goal when one-shot answers turn into real work
Goal → spec → task pipeline
- Human-in-the-loop review gates between every stage
- Pause, resume, and steer goals while agents are running
- Review workspace for high-impact changes; auto-approval for low-risk ones
Sandboxing & safety
- Every agent run happens in an isolated git worktree — the real project is never touched until you approve
- Rejecting a review discards the worktree cleanly
- Optional Docker sandbox with network isolation for stricter trust boundaries
- Built-in secret scanning and destructive-command detection on agent output
- Security spec library covering the OWASP Top 10
- Post-build security and quality audits on every goal
Review & verification
- Multi-model cross-verification for implementations (Pro)
- Consensus approves or rejects in bulk; you only see the lines reviewers disagree on
Context & memory
- Automatic compaction detection with structured summaries and file recovery on resume
- Decision tracking and replay timeline
- Background consolidation of project memory; Pro adds written reflection reports
Learning Mode & reports
- Toggle Learning Mode (
Cmd/Ctrl+L) to have agents narrate reasoning and decisions live - Structured markdown report after each task: what was built, approach, files changed, key decisions
- Commentary panel you can pin anywhere on screen
Observability
- Activity view for per-agent tool calls and scope
- Audit log keyed by actor, event, and entity
- Performance dashboard with adapter metrics and corrections
- Token & cost breakdown per stage, per adapter, per role
Dispatch Pro
- Hosted models available without installing a CLI: Gemma, Qwen Coder, Qwen 3, DeepSeek R1, Llama 3.3, Llama 4 Scout, Kimi, GPT-OSS, Mistral Small
- Image attachments work on any cloud model — requests with images are served by a vision-capable model automatically
- Cloud tool execution with ephemeral, per-task workspaces
- No prompt content logged or retained after a request completes
- 14-day free trial
Dispatch Team
- Bring-your-own-key for organization-managed API keys
- Member management, invitations, and access policies
- Shared project memory across the team
- Per-member usage dashboard with CSV export
Accounts
- Sign in with email magic link or GitHub OAuth
- Account dashboard at auth.withdispatch.dev/dashboard for subscription, billing, and usage
GitHub integration
- Browse your repositories and add them as projects
- Open pull requests from completed goals
- Turn GitHub issues into goals with the description pre-populated
- GitHub review comments feed back into the next agent run
- CI status visible alongside local verification
Remote execution
- Run agents on any host you can reach over SSH; transcripts stream back to your local UI
- Uses your existing SSH keys; Dispatch never stores private keys
- Optional HTTPS browser access with a self-signed certificate for phones and other laptops on your network
Research paper import
- Paste a URL or drop a PDF to implement a paper directly — Dispatch extracts the methodology and prepares the spec for you
Project rules
- Edit project-specific consequence rules in-app; rules live with the project so they travel with the codebase
Desktop & UX
- Quick Launch (
Cmd/Ctrl+N) and Command Palette (Cmd/Ctrl+K) - 3D Command Center with a fullscreen HUD view
- Real-time agent monitoring with reasoning preview
- Mid-task steering (send corrections while agents are working)
- In-app preview for completed goals
- Multi-mission tabs with concurrent execution
- Mobile-responsive layout
- Native menus, tray, and keyboard shortcuts
dispatchCLI bundled with the app for headless use
Notifications
- Toasts for real-time agent events
- Desktop notifications for reviews, errors, and goal completion
- Configurable notification level (all, important, none)
System requirements
- macOS: 12 Monterey or later (Apple Silicon)
- Linux: Ubuntu 22.04+ / Debian 12+ / Fedora 40+ / RHEL 9+ / any glibc-2.35+ distro (x86_64)
- Windows: Windows 10 build 19041 or later (x64)
- At least one agent CLI: Claude Code, Codex, or Gemini CLI — or a Dispatch Pro subscription for cloud models