@agent-dashboard/adapter-claude-code

The first-party adapter that lets the Agent Dashboard spawn and observe Claude Code agents.

Usage

import { claudeCodeAdapter } from '@agent-dashboard/adapter-claude-code';

claudeCodeAdapter.on('agent.tool_called', (p) => console.log(p.tool, p.arg));

const handle = await claudeCodeAdapter.startTask({
  agentId: 'agent-1',
  storyId: '1.1',
  story: { /* parsed BMAD story */ },
  workspaceDir: process.cwd(),
});

The adapter requires claude on PATH. The orchestrator (Story 4.3+) constructs an instance via createClaudeCodeAdapter({ command, args }) when it needs to point at a different binary.

Overriding the binary (tests)

Pass command to createClaudeCodeAdapter to point at any executable that consumes the prompt over stdin and emits Claude-style stdout lines. The bundled test fixture at tests/fixtures/fake-claude.mjs is the reference implementation used by the adapter's own test suite.

import { createClaudeCodeAdapter } from '@agent-dashboard/adapter-claude-code';

const adapter = createClaudeCodeAdapter({
  command: process.execPath,
  args: ['./tests/fixtures/fake-claude.mjs'],
});

Forking as a template

The adapter is intentionally small (≤300 lines under src/). Use it as the starting point for adapters targeting Cursor, Aider, Codex, or any other agent runtime:

1. Copy this package and rename it.
2. Replace parseClaudeLine with the heuristic for your runtime's stdout format.
3. Replace buildPrompt with the prompt template your runtime expects.
4. Update runtime and command defaults.
5. The shape of claudeCodeAdapter, the lifecycle (start → events → completed/crashed/killed), and the cross-platform spawn/kill helper in @agent-dashboard/core are all reusable as-is.

See @agent-dashboard/protocol for the contract every adapter implements.