sandgardenhq/using-sgai

Using sgai from an AI Harness

sgai is a software factory system that runs AI agents in workspaces. This skill teaches you to drive it via the HTTP API or MCP tools.

Quick Start: Base URL

All examples assume BASE_URL=http://127.0.0.1:PORT where PORT is shown in the server startup log:

sgai serve listening on http://127.0.0.1:PORT

The MCP endpoint is at /mcp/external on the same server (e.g. http://127.0.0.1:PORT/mcp/external).

The Cyclical Probe/Poll/Act Loop

The core pattern for driving sgai is a continuous loop:

LOOP:
  1. PROBE  → GET /api/v1/state          # Discover all workspaces + status
  2. CHECK  → pendingQuestion != null?   # Does any workspace need human input?
  3. ACT    → based on workspace status  # Start, steer, respond, or wait
  4. WAIT   → poll again after delay     # Repeat

Step 1: Probe — Get Factory State

curl -s $BASE_URL/api/v1/state

Response shape:

{
  "workspaces": [
    {
      "name": "my-project",
      "running": false,
      "needsInput": false,
      "inProgress": false,
      "status": "agent-done",
      "currentAgent": "coordinator",
      "task": "Planning implementation",
      "pendingQuestion": null
    }
  ]
}

Key fields to check per workspace:

• running — is a session active?
• needsInput — does the agent need a human response?
• pendingQuestion — non-null when human input is required
• status — current workflow status string
• inProgress — is work actively happening?

Step 2: Check for Pending Questions

When workspace.pendingQuestion != null, the agent is blocked waiting for human input.

{
  "pendingQuestion": {
    "questionId": "abc123def456",
    "type": "free-text",
    "agentName": "coordinator",
    "message": "Which approach should we take?",
    "questions": []
  }
}

Question types:

• "free-text" — respond with a text answer
• "multi-choice" — select from provided choices
• "work-gate" — approve to proceed (select approval text)

Step 3: Act Based on Status

| Workspace State | Action | |----------------|--------| | needsInput: true | Call respond endpoint with answer | | running: false and has goal | Start session | | running: true | Monitor / optionally steer | | Session complete | Check results, start next task |

Step 4: Respond to Questions

# Free-text response
curl -s -X POST $BASE_URL/api/v1/workspaces/{name}/respond \
  -H "Content-Type: application/json" \
  -d '{"questionId": "abc123def456", "answer": "Use the microservice approach"}'

# Multi-choice response
curl -s -X POST $BASE_URL/api/v1/workspaces/{name}/respond \
  -H "Content-Type: application/json" \
  -d '{"questionId": "abc123def456", "selectedChoices": ["Option A"]}'

Sub-skills

For detailed documentation on specific operations:

• workspace-management — Create, fork, delete, rename workspaces
• session-control — Start/stop sessions, steer agents
• human-interaction — Respond to questions and work gates
• monitoring — List workspaces, get state, diffs, SVGs
• knowledge — Agents, skills, snippets
• compose — Compose wizard: state, save, preview, draft, templates
• adhoc — Ad-hoc prompt start/stop/status

MCP Interface

If using the MCP interface instead of HTTP, all tools are available at /mcp/external:

# List all 38 tools
npx mcporter list --http-url http://HOST:PORT/mcp/external --allow-http

Key MCP tools mirror the HTTP API:

• list_workspaces → GET /api/v1/state
• start_session → POST /api/v1/workspaces/{name}/start
• respond_to_question → POST /api/v1/workspaces/{name}/respond
• wait_for_question → polls + elicitation (MCP only)

Real-Time Updates via SSE

Subscribe to state changes instead of polling:

curl -s -N $BASE_URL/api/v1/signal
# Emits: event: reload\ndata: {}\n\n

When you receive a reload event, re-fetch /api/v1/state.

Common Workflow: Start a Project End-to-End

# 1. Create workspace
curl -X POST $BASE_URL/api/v1/workspaces \
  -d '{"name": "my-project"}'

# 2. Write a GOAL.md
curl -X PUT $BASE_URL/api/v1/workspaces/my-project/goal \
  -d '{"content": "# My Goal\n- [ ] Build the feature"}'

# 3. Start session in auto (self-drive) mode
curl -X POST $BASE_URL/api/v1/workspaces/my-project/start \
  -d '{"auto": true}'

# 4. Poll for completion
while true; do
  STATE=$(curl -s $BASE_URL/api/v1/state)
  NEEDS_INPUT=$(echo $STATE | jq '.workspaces[0].needsInput')
  if [ "$NEEDS_INPUT" = "true" ]; then
    # Handle question...
  fi
  sleep 5
done