linear-agent-api（Linear Agent API 搜索）

Linear Agent API Skill

Comprehensive reference for Linear's Agent API, following the Agent Interaction Guidelines (AIG).

When to Use

• Verifying agent sessions are created correctly
• Checking activity streaming (thought, action, response)
• Implementing two-way communication (user input, stop signals)
• Debugging Linear integration issues

Agent Session Lifecycle

Sessions track the lifecycle of an agent run:

Sessions are created automatically when an agent is @mentioned or assigned as delegate.

Agent Activity Types

Activity Payloads

// thought - Agent thinking
{
  "content": {
    "type": "thought",
    "body": "Analyzing the codebase structure..."
  }
}

// action - Tool call
{
  "content": {
    "type": "action",
    "action": "edit_file",
    "parameter": "src/main.rs"
  }
}

// action with result
{
  "content": {
    "type": "action",
    "action": "run_tests",
    "result": "All 42 tests passed"
  }
}

// elicitation - Request input
{
  "content": {
    "type": "elicitation",
    "body": "Which database should I use?"
  }
}

// response - Completion
{
  "content": {
    "type": "response",
    "body": "Implementation complete. PR #123 created."
  }
}

// error - Failure
{
  "content": {
    "type": "error",
    "body": "Build failed: missing dependency"
  }
}

Ephemeral Activities

Activities can be marked ephemeral: true for temporary status messages that get replaced by the next activity. Only thought and action types support this.

Signals (Two-Way Communication)

Human-to-Agent Signals

stop Signal

When user clicks "Send stop request" in Linear, agent receives a prompt activity with signal: "stop".

Agent MUST:

1. Halt immediately - No further code changes or API calls
2. Emit final activity - response or error confirming stop
3. Report current state - What was completed, what remains

CTO Implementation: The status-sync.rs sidecar detects signal: "stop" in polled activities and triggers graceful shutdown via /stop endpoint.

Agent-to-Human Signals

auth Signal

Used with elicitation to request account linking:

{
  "content": { "type": "elicitation", "body": "Please authenticate" },
  "signal": "auth",
  "signalMetadata": {
    "url": "https://auth.example.com/oauth",
    "providerName": "GitHub"
  }
}

select Signal

Used with elicitation to present multiple choice options:

{
  "content": { "type": "elicitation", "body": "Which repository?" },
  "signal": "select",
  "signalMetadata": {
    "options": [
      { "value": "5dlabs/cto" },
      { "value": "5dlabs/alertub" }
    ]
  }
}

Agent Plans (Checklists)

Agents can provide session-level task checklists:

{
  "plan": [
    { "content": "Parse PRD document", "status": "completed" },
    { "content": "Generate task breakdown", "status": "inProgress" },
    { "content": "Create Linear issues", "status": "pending" }
  ]
}

Status values: pending, inProgress, completed, canceled

Note: Plan updates replace the entire array, not individual items.

Best Practices

1. First Response Within 10 Seconds

Upon receiving created webhook, agent MUST emit a thought activity within 10 seconds or be shown as unresponsive.

2. Follow-up Activities Within 30 Minutes

After first response, activities can be sent for up to 30 minutes before session is stale.

3. Delegate vs Assignee

• Delegate = The agent working on the issue
• Assignee = The human responsible (ownership)

Agents should set themselves as delegate, not assignee.

4. Status Updates

When starting work, move issue to first "started" status if not already there.

5. Completion

When work complete, emit response activity. If user action needed, emit elicitation or error.

GraphQL Queries

Get Session Activities

query AgentSession($agentSessionId: String!) {
  agentSession(id: $agentSessionId) {
    id
    state
    createdAt
    activities {
      edges {
        node {
          updatedAt
          content {
            ... on AgentActivityThoughtContent { body }
            ... on AgentActivityActionContent { action parameter result }
            ... on AgentActivityElicitationContent { body }
            ... on AgentActivityResponseContent { body }
            ... on AgentActivityErrorContent { body }
            ... on AgentActivityPromptContent { body }
          }
        }
      }
    }
  }
}

Get Team Started Statuses

query TeamStartedStatuses($teamId: String!) {
  team(id: $teamId) {
    states(filter: { type: { eq: "started" } }) {
      nodes {
        id
        name
        position
      }
    }
  }
}

CTO Status-Sync Implementation

The status-sync.rs sidecar implements:

Input Polling: The input_poll_task periodically calls get_session_activities() to detect:

• New prompt activities from users
• signal: "stop" requests

Webhook Events

Reference Documentation

• Linear AIG - Agent Interaction Guidelines
• Linear Getting Started - Agent setup
• Linear Developing Interaction - Activity types
• Linear Best Practices - Recommendations
• Linear Signals - Two-way communication
• CTO Linear Integration - CTO-specific implementation