@snomiao/daemon-kit

Shared daemon helpers for codehost and agent-yes. Both projects independently grew the same Windows daemon workarounds — and the same bugs. This is their single home.

Why this exists

codehost and agent-yes each ship a serve-style command that runs a long-lived process as a managed daemon (pm2 on Windows, oxmgr elsewhere). On the same day, both repos separately fixed an identical set of Windows daemon bugs: hidden launch, restart backoff, login auto-start, and a port-clash restart storm. Duplicated fixes drift — so the daemon layer is factored out here.

Scope boundary (the co-work contract)

| Concern | Owner | | --- | --- | | pm2 process mgmt · hidden launch · exp-backoff · login auto-start · TEMP normalize · port-bind retry | daemon-kit (this package) | | WebRTC / signaling (rooms, tokens, hosts) | each project — stays separate (signal.codehost.dev vs agent-yes.com) | | Remote workspace / editor access · repo provisioning · tree/<branch> worktrees | codehost | | Agent execution · auto-yes · terminal/resize · ay ls/tail/send | agent-yes |

The two apps compose: codehost provisions & serves a worktree → agent-yes runs an agent inside it. They share this daemon layer and nothing in each other's product surface.

API

import {
  pm2Start, pm2Online, pm2Delete, pm2Available,   // pm2 daemon management
  enablePm2LoginAutostart, disablePm2LoginAutostart, // login/boot auto-start
  normalizeTempEnv,                                 // fix unexpanded %USERPROFILE% TEMP/TMP
  bindWithRetry, freeStalePort,                     // EADDRINUSE retry (+ stale-holder reclaim)
} from "@snomiao/daemon-kit";

// Repair TEMP/TMP before any native dep or child reads it (call first).
normalizeTempEnv();

// Start a restart-backed daemon (hidden launch on Windows; direct on POSIX).
pm2Start({
  name: "agent-yes",
  script: process.execPath,          // the bun runtime
  args: [entryScript, "serve", "--webrtc"],
  cwd: projectDir,
  log: "C:\\Users\\me\\.agent-yes\\logs\\daemon.log",
  expBackoffMs: 200,                 // exponential restart backoff (anti-storm)
});

enablePm2LoginAutostart("agent-yes"); // Win: HKCU Run -> hidden `pm2 resurrect`; POSIX: `pm2 startup`

// Self-heal a port clash during restart instead of crash-looping. With `reclaim`,
// also recover a LIVE stale holder (e.g. a detached child that inherited the
// listen socket and orphan-holds the port — netstat shows a dead owner pid).
const server = await bindWithRetry(() => Bun.serve(opts), {
  reclaim: { port: 7432, signature: "ay serve" }, // opt-in, destructive — your own port only
});

What each helper does

• pm2Start(spec) — replaces any same-named instance (deletes first) and starts the daemon with --exp-backoff-restart-delay so a persistent crash can't storm-restart. On Windows it launches via a hidden VBScript launcher (wscript //B, window style 0) so neither the pm2 daemon nor its fork pops a console window; on POSIX it starts pm2 directly. Returns true once pm2 reports it online.
• pm2Online / pm2Delete / pm2Available — liveness, teardown, and a resolvable-pm2 probe (resolves pm2's JS entry under the current runtime; never the bare pm2 shim).
• enablePm2LoginAutostart(name) — auto-start at login/boot. Windows: pm2 save + a hidden HKCU\…\Run entry running pm2 resurrect (no admin; works where Startup-folder writes and schtasks /create are blocked). POSIX: pm2's native pm2 save + pm2 startup (may print a one-off sudo command).
• disablePm2LoginAutostart(name) — remove the per-tool Windows HKCU Run entry. No-op off Windows by design: pm2's startup hook is global (shared by all pm2 daemons), so a single tool's uninstall must not tear it down — pm2Delete + the re-saved list already stop this daemon being resurrected.
• normalizeTempEnv() — expands a literal %USERPROFILE%\AppData\Local\Temp left in TEMP/TMP (Windows REG_EXPAND_SZ passed through unexpanded), so temp writes don't litter the cwd. No-op off Windows.
• bindWithRetry(fn, opts?) — runs a listen()/Bun.serve() with exponential backoff on EADDRINUSE, so a daemon restart racing the old instance's port release self-heals. With opt-in reclaim: { port, signature? }, after the backoff exhausts it runs one destructive reclaim then a final bind — for a LIVE holder that won't release on its own (the inherited-socket / dead-owner-pid case).
• freeStalePort(port, signature?) — the reclaim primitive (also usable standalone): phase 1 kills the port's current LISTEN owner; phase 2, if signature is given and the port is still held, kills processes whose command line matches it (the inherited-socket orphan, whose pid differs from the dead netstat owner). Cross-platform (netstat/taskkill on Windows; lsof/ss + SIGKILL on POSIX). Destructive — use only for a port you own.

Develop

bun install
bun test
bun run build      # tsc -> dist/ (+ .d.ts)

License

MIT