Swarm DAO herdr Adapter

Run each Swarm DAO deliberation agent as a real coding agent inside a herdr workspace — "the runtime your coding agents live on". herdr owns the agent's terminal and tracks its lifecycle (working / idle / blocked / done); the operator can attach to any agent pane live with herdr and watch it think.

herdr workspace: swarm-dao-p1-strategist   ← real pi/claude/codex/grok… agent
herdr workspace: swarm-dao-p1-critic       ← another agent, another workspace

How it works

For every agent, the adapter:

1. Creates an isolated herdr workspace (one root pane, --no-focus — it never touches your existing layout).
2. Starts the agent: herdr agent start <name> --kind <kind> — blocks until herdr detects the agent is ready for input. A freshly created pane may not be at its shell prompt yet (slower shell init under load); herdr then reports agent_pane_busy and the adapter retries on the same pane until the readiness budget (startTimeoutMs) is spent.
3. Submits the deliberation prompt: herdr agent prompt <name> '<prompt>' --wait — settles on idle, done, or blocked. A stalled prompt (herdr enforces a 5 s observed-state-change window) is recovered: the adapter grace-polls the agent state and, if the prompt took effect, waits for settle; only a swallowed submission is re-prompted, exactly once.
4. Harvests the ANSI-stripped terminal transcript: herdr agent read <name> --source recent-unwrapped.
5. Closes the workspace (unless keepPanes).

A blocked agent (approval/question UI) surfaces as an error output, never as a vote. Outputs feed the same deterministic tally as every other host — no AI authority, no proposal-state mutation.

Prerequisites

• herdr installed (brew install herdr) and the server running (run herdr once in any terminal)
• The chosen kind's executable installed and authenticated (pi, claude, codex, gemini, cursor, grok, opencode, …)

Usage

import { createHerdrHostAdapter } from "@guyghost/swarm-dao-herdr-adapter";
import { loadConfig, handleDaoDeliberate } from "@guyghost/swarm-dao-core";

const adapter = createHerdrHostAdapter({
  workDir: process.cwd(),
  kind: "pi", // or "claude", "codex", "grok", …
  // agentArgs: ["-m", "gpt-5.4"],       // passed to the agent executable
  // timeoutMs: 300_000,                 // per-agent prompt timeout
  // keepPanes: true,                    // keep workspaces for inspection
});

const result = await handleDaoDeliberate(
  { adapter, workDir: process.cwd(), deliberationMode: "auto", controlToolName: "dao_check" },
  1,
);

Boundary

Deliberation is read-only analysis: agents share the repository checkout for context (each workspace's cwd is the repo root). Execution-side isolation stays with the delivery layer's GitWorkspace (execution.isolation).

Testing

bun test packages/herdr-adapter          # unit (fake runner) + real-server error path
HERDR_IT=1 bun test packages/herdr-adapter   # + full round-trip with a real agent
                                            # (HERDR_KIND, HERDR_KEEP to tune)