GoalOS — A Personal Goal Execution Operating System

You state the goal. The system delivers it — safely.

GoalOS is not a chatbot, not a workflow engine, not an agent framework. It is a personal operating system for goal execution. You say what you want — the system understands, plans, executes, verifies, and delivers.

中文文档

You state your goal → The system understands → You agree on "what done means" (CompletionContract) → The system executes safely → Multi-LLM verification → Verified → Delivered.

Any feature that does not directly serve this chain should be removed.

Download the binary for your platform from Releases.

# macOS/Linux
tar xzf goalos-<os>-<arch>.tar.gz
./goalos-daemon &

# Via CLI
goalos "Design a 3D rotating Rubik's Cube in HTML/CSS"

# Via HTTP API
curl -X POST http://localhost:18920/api/goals \
  -H "Content-Type: application/json" \
  -d '{"goal":"Build a CRM system"}'

# View review summary for a goal
goalos review <goal_id>

# View full review report with per-Provider reasoning
goalos review <goal_id> <action_id>

# Make a decision on FAIL/WARN verdict
goalos review <goal_id> <action_id> --retry --feedback "fix the auth issue"
goalos review <goal_id> <action_id> --accept --confirm
goalos review <goal_id> <action_id> --refine "use argon2 instead of bcrypt"

On first startup, GoalOS auto-generates ~/.goalos/config/daemon.yaml with comments. Edit it:

daemon:
  port: 18920
  autonomy_level: approve   # observe|suggest|approve|autonomous

llm:
  provider: openai
  model: qwen3.6-flash
  api_key: "sk-..."         # API key directly in config
  base_url: https://your-llm-api.com/v1
  max_tokens: 65536

# Hot-reload — no daemon restart needed
# curl -X POST http://localhost:18920/api/system/reload

Configure multiple LLM providers for parallel code review. System calls all providers in parallel → VerdictCombiner merges results with voting (R-844). Provider health checked at startup (R-861).

multi_llm:
  enabled: true
  debate_round: false        # R-860: enable cross-model debate (adds ~1.5x token cost)
  cold_review: false         # R-858: isolate verifier from builder context
  providers:
    - name: primary
      model: qwen3.6-flash
      api_key: "sk-..."
      base_url: https://your-llm-api.com/v1
      allowed_for: [L0,L1,L2,L3,L4,L5]
    - name: reviewer
      model: google/gemma-4-26b-a4b-it:free
      api_key: "your-openrouter-api-key"
      base_url: https://openrouter.ai/api/v1
      allowed_for: [L0,L1,L2]

User Goal → Agent(Align→Analyze→Plan) → MissionGraph + PlanHash
          → Governance(5-engine approval) → PipelineRunner(Check→Exec→Decide)
          → Plugin Runner(seccomp sandbox, FD3 IPC) → Artifacts(~/Goals/)
          → MultiLLM Verification(voting + debate + cold review)
          → ReviewReport → User Decision(retry/accept/refine)
          → Verifier(verification pyramid) → CompletionContract → Delivered

1. Plugin over Build — All variable capabilities are Plugins. The core never changes
2. Event over Call — Modules communicate via events. Auditable and replayable
3. File over Database — Data is files. User-owned. Zero external storage dependencies
4. Projection over State — State is derived from events. Caches are rebuildable
5. Delegate over Build — Don't rebuild what platforms already do (messaging, keychain, notifications)
6. One over Many — Use one until data proves you need more
7. Interface over Implementation — Define "what", not "how". Implementations are replaceable
8. User-Owned over System-Managed — Data belongs to the user. Summaries in file frontmatter
9. Persona Decoupled — Core produces neutral events. Persona is a replaceable rendering layer
10. Honest Feedback — System truthfully reflects state. Never fakes success. Never silently downgrades user goals. User has final decision authority

git clone https://github.com/dotnet010/GoalOS.git
cd GoalOS
go build -o goalos-daemon ./cmd/goalos/
go build -o goalos-cli ./cmd/goalos-cli/

• User Manual (Chinese)
• Architecture Meeting Minutes (Chinese)

GPL-3.0