Published on npm: https://www.npmjs.com/package/occ-orchestrator/v/0.5.0

npm install -g occ-orchestrator@0.5.0

Release notes follow the npm package version (occ-orchestrator).
33 commits since v0.4.0. Roll-up of every change since the last
release; per-feature detail lives in the section below + the older
2.x notes that document the earlier platform additions still kept on
disk for reference.

Headline features

• Interactive canvas overhaul — Obsidian Note kind, Terminal Agent
  with model selector + connected-context propagation (Portal URL,
  File content, Sticky text, Obsidian markdown all flow into the
  agent's system prompt), Portal/Terminal overlays render at native
  size with a single transform: scale(zoom) so SPAs stop reflowing
  at low zoom, File Viewer file picker, right-click no longer arms a
  phantom box-select / node-drag, custom-fontFamily CSS tokens with
  emoji/symbol fallbacks (no more "tofu" boxes for ⏸ ⚡ ↺ ▶).
• Server platform — codex CLI runner (mirrors claude-runner,
  no API key needed), customizable system prompts (/system-prompts
  GET/PUT/reset, hot-reloaded JSON), generic /agent-chat endpoint
  that routes through any configured provider, execution time-travel
  via GET /executions/:id/timeline (per-step checkpoint history).
• Pipelines from canvas — canvasToPipelineChains decomposes a
  multi-stage canvas into PipelineChainRef[]; SaveChainModal
  detects pipeline canvases via isPipelineCanvas and routes through
  savePipeline. New summarize_output field on PipelineChainRef
  to control how a stage's output flows downstream.
• Canvas-only chains — chains with zero steps and one or more
  non-step canvas items (Portal / Sticky / Terminal / File / Link /
  Free Text / Obsidian) save and reload exactly via the new
  canvas_items?: CanvasItemSerialized[] side-car. Loader's Zod
  schema accepts empty steps; executor refuses to run them with a
  clear "open in canvas editor" error.
• Floor color slot palette — replaces hardcoded color: "#hex"
  on FloorData with a colorSlot: 0..6 index into a design-space-
  derived palette resolved at render time, so floor colors stay in
  sync with the active theme. Persistence layer auto-migrates legacy
  hex records via inferSlotFromHex.
• Settings UI refactor — collapsible sections (state persisted in
  localStorage), system prompts editor (SystemPromptsSection),
  provider section gains the codex provider type, MCP/schedule
  sections light tweaks.
• Workflow chat planner — concrete bug fixes: planner respects
  existing canvas (no more disconnected parallel chains on
  "améliore"), MODIFY mode patches preTools / advanced /
  outputVar / depends_on rewires atomically, named-input
  detection drives the RUN modal and the [ACTION:RUN] JSON inputs
  block, floor-aware canvas context.
• Test coverage push +1075→1233 (frontend 27.6% → 49.4%
  statements) across 12 new test files: canvasContext,
  preToolsToData, detectInputs type-inference matrix, tools/pre-tools
  serialization, pretoolFields schema integrity, canvasRenderer
  helpers + main render, connectionHit (100%), useCanvasInteractions
  (regression for the right-click bug + drag/keyboard branches —
  86%), workflowChat applyPlanToCanvas + buildCanvasContext,
  workflowChat sessions / persistence (90%+), floors store,
  physarumSim (72%), blobRenderer (75%), workflowChat actions
  ([ACTION:ANALYZE/MODIFY/RUN]).
• Misc dev quality — frontend deps bumped to current wanted
  range (vite, react, vitest, storybook, react-query...), mcp-server
  deps bumped (@modelcontextprotocol/sdk 1.27→1.29, better-sqlite3,
  eslint...), index.html served Cache-Control: no-store so the
  browser stops holding stale asset hashes after every rebuild,
  .gitignore excludes system-prompts.json runtime artifact.

Removed

• Interactive Portal mode (Playwright + WebSocket screencast) —
  added in 2.2.0 / commit 96675f2, removed before npm release.
  See the explanatory section in the older 2.x notes below for the
  full rationale (anti-bot detection at the binary level kept finding
  new signals; macOS Dock-icon UX was hard to make invisible). Portal
  nodes keep only the static iframe proxy via GET /portal?url=....
  Surface dropped: mcp-server/src/portal-sessions.ts,
  mcp-server/src/portal-ws.ts, frontend-react/src/api/portal.ts,
  InteractivePortalOverlayItem, the async live-snapshot path in
  canvasContext.ts, REST endpoints (POST /portal/session,
  GET /portal/sessions, POST /portal/session/:id/navigate,
  DELETE /portal/session/:id, GET /portal/snapshot), deps ws
  • @types/ws, types portalMode / portalPersistKey, vite proxy
    ws: true on /portal. Net diff: −1473 / +35 lines.

Notes

• Versioning: this is the first OCC release that's published from a
  CI-validated main. Every commit since v0.4.0 was tested on the
  full matrix (ubuntu/macos/windows ×ばつ Node 20/22).
• Historical 2.x section below is preserved as a reference for the
  earlier platform additions; future releases will use 0.x semver
  matching the npm package version.

What's New

One Server, One Port, One Process

The React dashboard is now served directly by Express. No more running two servers.

Before (v0.4.0):

cd mcp-server && npm run rest # Backend :4242
cd frontend-react && npm run dev # Frontend :5173 (proxy to :4242)

After (v0.4.1):

occ start # Backend + Dashboard at http://localhost:4242

How it works

• npm run build:all builds the backend (TypeScript) and frontend (React) together
• Express serves the built React app as static files with SPA fallback
• API endpoints work as before on the same port
• Dev mode with Vite hot reload still works unchanged for development

npm package now includes the dashboard

npm install -g occ-orchestrator now gives you the full experience: CLI + API + Dashboard.

Full Changelog: v0.4.0...v0.4.1

What's New

Zero-Friction Onboarding

5 commands from install to first chain execution:

npm install -g occ-orchestrator
occ init my-project && cd my-project
occ doctor
occ start
occ run hello-world -i topic="quantum computing"

New CLI Commands

occ init [name] — Create a new OCC project with:

• chains/ directory with 3 example chains (hello-world, web-analyzer, parallel-pros-cons)
• pipelines/ directory ready for multi-chain workflows
• .env with sensible defaults
• .gitignore for databases and secrets

occ start — Start the OCC REST server automatically. Finds the entry point whether installed via npm, from source, or from the repo root.

occ doctor — Check all prerequisites with actionable fix hints:

• Node.js version
• Claude CLI installed + authenticated
• Ollama running (optional)
• Docker available (optional)
• Chains directory
• OCC server status

Improved Error Messages

When the server is not running, the CLI now suggests occ start and occ doctor instead of a cryptic error.

Help Reorganized

CLI help grouped into: Setup, Chain execution, Monitoring, Human-in-the-loop, Generation, System.

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

What's New

Prerequisites Check Modal

On first launch, the dashboard now shows a Setup Check modal that verifies all dependencies before you start:

Required checks:

• OCC Backend running
• Claude CLI installed
• Claude CLI authenticated
• SQLite database

Optional checks:

• Chains directory
• LLM Providers configured
• Ollama (local models)
• Docker (sandboxed execution)

Features:

• Action buttons per item — copy install commands, open download pages, navigate to Settings
• Click-to-copy on hint codes with "Copied!" feedback
• Animated re-check with per-line spinners
• Auto-refresh every 5s when required items fail (detects when user fixes an issue)
• "Don't show again" checkbox (persisted in localStorage)
• Glass morphism design matching the dashboard theme

Getting Started improvements

• README now includes Claude CLI install instructions
• Setup Check modal mentioned in Quick Start section

Full Changelog: v0.3.0...v0.3.1

What's New

Real Benchmark Suite

• Economy of scale analysis: 10-step, 4-wave strategic analysis workflow
• OCC vs naive: 70% faster, 70% cheaper (0ドル.18 vs 0ドル.60/run)
• OCC vs smart manual: 29% faster (69s vs 97s)
• Model routing alone: 80% cost reduction
• 3 reproducible scripts: run-benchmark.sh, run-benchmark-raw.py, run-benchmark-scale.py
• Full methodology in BENCHMARKS.md

Token Tracking Fixes

• Claude: now counts cache_read_input_tokens + cache_creation_input_tokens (was showing 9 instead of 26K+)
• Ollama/HuggingFace: added stream_options: { include_usage: true } for streaming

Typed Chain Inputs

• 9 input types: string, number, boolean, enum, file, image, json, url, text
• Validation at loader, executor, and REST layers
• RunModal with specialized widgets per type (file picker, toggle, range, preview)

Image Generation

• New image_gen step type and image_generate pre-tool
• Providers: OpenAI (DALL-E 3), HuggingFace (FLUX), Stability AI
• Outputs served via /images/:filename endpoint

HuggingFace Provider

• Full integration via router.huggingface.co/v1
• 63 curated models (free + PRO tiers) with live search
• Settings UI with tier filter, tag filters, rate limits panel

Workflow Chat Improvements

• Full markdown renderer (bold, italic, code, headers, blockquotes, images)
• Canvas context injection in chat stage
• Agentic actions: RUN, STOP, DEBUG, ANALYZE, DRY_RUN, MODIFY
• Triple regex fallback for action detection

Other

• Universal output renderer (JSON, CSV, HTML, images, file downloads)
• Windows CI fixes (EBUSY SQLite, path traversal, Docker)
• Path resolution for directories with spaces
• Ollama exact model name matching

Stats

• 2344 tests across 59 files (was 1814/50)
• Coverage: Statements 82%, Branches 75%, Functions 79%, Lines 84%
• 102 REST endpoints, 28 MCP tools, 30 pre-tool types, 6 LLM providers

Install

npm install -g occ-orchestrator

Full Changelog: v0.2.0...v0.3.0