ExecutorAdapter: interface every executor implements.
createEchoExecutor: smoke-test executor that echoes the normalized command.
createAcpAgentExecutor: generic stdio ACP host for an ACP launch definition.
createBuiltInAcpExecutors: built-in launch profiles for pinned Codex, Claude, and OpenCode packages plus installed Cursor, Hermes, and OpenClaw ACP commands.
builtInAcpAgentDefinitions: the data-only built-in definitions and Registry provenance where a pinned Registry package is used.
createAcpExecutor: lower-level generic stdio ACP host for an internal integration manifest.
Git helpers such as createRunBranch, changedFiles, and branchNameForRun.
Command helpers such as nodeCommandRunner and assertCommandSucceeded.
Built-in and generic ACP executors expose an optional capability contract so setup, doctor, status, and service surfaces can report profile, cancellation, hook-completion, progress-event, approval-boundary, prompt/context trust gates, workspace isolation, ACP session-cwd conformance, secret, and completion-signal support honestly.

Example

import type { ExecutorAdapter } from "@opentag/runner";

export const executor: ExecutorAdapter = {
  id: "my-agent",
  displayName: "My Agent",
  capability: {
    id: "my-agent",
    invocation: "spawn",
    supportsProfile: false,
    supportsStreaming: false,
    supportsCancel: false,
    supportsHookCompletion: false,
    progressEvents: "audit",
    approvalMode: "opentag_policy",
    contextAccess: ["context_packet", "context_pointers", "workspace"],
    promptAssembly: "executor_adapter",
    writeAccess: "workspace",
    conversationAccess: "request",
    promptMutation: "none",
    rawContextAccess: false,
    writeActionAccess: "none",
    workspaceIsolation: "branch",
    requiredSecrets: [],
    completionSignals: [
      {
        type: "process_exit",
        required: true,
        description: "The process exits after producing the final result."
      }
    ]
  },
  async canRun() {
    return { ready: true };
  },
  async run(input, sink) {
    await sink.emit({
      type: "executor.started",
      message: `Running ${input.command.rawText}`,
      at: new Date().toISOString()
    });

    return {
      conclusion: "success",
      summary: "Handled by my-agent"
    };
  },
  async cancel() {
    return;
  }
};

Safety Notes

Codex, Claude Code, Cursor, OpenCode, Hermes, and OpenClaw run through the same generic ACP host. Codex, Claude, and OpenCode use exact package versions through npx; Cursor uses its installed native ACP command; Hermes uses its configured local command and profile; OpenClaw uses its installed Gateway ACP bridge. OpenCode launches in pure mode so external plugins cannot pollute the strict ACP stdout stream. The generic ACP host passes the attempt workspace as the ACP session cwd, scrubs the child environment, and terminates the adapter process group when a run is cancelled. OpenClaw declares supportsCancel: false: OpenTag requests cancellation and stops its local bridge, but does not attest termination of Gateway-owned tool subprocesses.