Swarm DAO tmux Adapter

The swarm-forge execution model for Swarm DAO deliberation: each agent runs as its own detached tmux pane — watchable live by the operator (tmux attach), isolated from every other agent, with completion observable through filesystem markers.

.swarm-dao-p1-strategist   ← pane running the strategist's agent CLI
.swarm-dao-p1-critic       ← pane running the critic's agent CLI
...

How it works

For every agent the adapter:

1. Writes the deliberation prompt to .dao/tmux/<proposalId>/<agentId>/prompt.md (stale completion markers from previous runs are purged first).
2. Starts a detached tmux session whose program is sh -c 'PROMPT="$(cat prompt.md)"; <command> > output.md 2> stderr.log; printf "%s" "$?" > done' — the command runs as the session program, so there is no typing race with the pane shell.
3. Polls for the done marker up to timeoutMs.
4. Harvests output.md as the agent's output (exit code and stderr.log surface as an error output), and kills the session unless keepSessions is set (which leaves the pane's scrollback alive for inspection).

The outputs feed the same deterministic tally as every other host — no AI authority, no proposal-state mutation.

Configuration

.dao/config.json:

{
  "tmux": {
    "command": "claude -p \"$PROMPT\"",
    "timeoutMs": 300000,
    "sessionPrefix": "swarm-dao",
    "keepSessions": false
  }
}

• command (required for deliberation): the agent CLI. $PROMPT carries the deliberation prompt; stdout is captured as the agent's output.
• timeoutMs: per-agent timeout (default 5 min). A timeout kills the session and records a deterministic error output.
• sessionPrefix, keepSessions: session naming and pane retention.

Usage

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

const config = await loadConfig(".dao");
const adapter = createTmuxHostAdapter({
  workDir: process.cwd(),
  command: config.tmux?.command,
});

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 and never need per-agent worktrees. Execution-side isolation stays with the delivery layer's GitWorkspace (execution.isolation).

Testing

bun test packages/tmux-adapter          # unit (fake tmux) + real tmux (skipped if absent)