Connector Model

Principle

The harness is a runtime adapter and control plane, not the coding interface.

Users already have agent surfaces they like:

• Codex application/account
• Claude Code
• Hermes
• local terminal agents

The harness connects to those surfaces, starts/observes sessions, and enforces workflow artifacts.

Adapter contract v0

Every connector should implement:

type AgentConnector = {
  id: string;
  displayName: string;
  capabilities: ConnectorCapability[];
  prepareContext(input: ProjectHarnessPack): Promise<PreparedContext>;
  startSession(input: AgentSessionStart): Promise<AgentSessionHandle>;
  streamEvents(handle: AgentSessionHandle): AsyncIterable<AgentEvent>;
  requestStop(handle: AgentSessionHandle): Promise<void>;
};

Event types

type AgentEvent =
  | { type: "session.started"; sessionId: string }
  | { type: "stage.changed"; stageId: string; status: string }
  | { type: "tool.called"; toolName: string; argsSummary: string }
  | { type: "artifact.written"; path: string; artifactType: string }
  | { type: "gate.passed"; gateId: string; artifactPath: string }
  | { type: "gate.blocked"; gateId: string; reason: string }
  | { type: "human.approval.requested"; approvalId: string; risk: string }
  | { type: "diff.changed"; files: string[] }
  | { type: "session.completed"; summary: string };

Codex connector

Initial assumption:

• Generate AGENTS.md / context pack.
• Launch Codex CLI/app session where available.
• Watch filesystem events and gate artifact paths.
• Stream logs/diffs/proofs to the UI.

Claude Code connector

Initial assumption:

• Generate CLAUDE.md and conditional instruction blocks.
• Launch Claude Code from the task worktree.
• Watch artifacts, transcript/export logs where available, and git diff.

Hermes connector

Reference implementation:

• Use Hermes Kanban board/task.
• Use Hermes tools/subagents/processes.
• Capture run packet artifacts directly.
• Works first because we control this environment.