agentleFS
TechNickAITechNickAI/hermes-skills/skills/engineering/mini-app/SKILL.md

mini-app

Skill0 starsChanged 3 months agoMITView on GitHub

Use when adding, removing, password-protecting, or troubleshooting a mini-app served by the hermes-config mini-app router (Caddy + PM2 + auth sidecar + Tailscale Serve/Funnel) on a fleet machine. Covers install, the Caddy route pattern, the auth sidecar conventions, exposing apps publicly via Funnel, Hermes dashboards behind a password, and the recurring pitfalls (Tailscale "serve reset" wars, the PM2 $HOME trap, funnel-eligible ports, strip-prefix requirements).

  • Installs packages

The file

Show its rules and text

We split this skill into the rules it gives an agent

25 rules. About 0 are common, 10 are distinctive, and 15 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈10

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 15

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Add a new mini-app to this host's router (Hermes dashboard, webhook receiver, status page, anything that binds to localhost)

    line 47
  2. Distinctive:

    Remove or rename an existing mini-app

    line 49
  3. Distinctive:

    Add or change a password on a gated mini-app

    line 50
  4. Distinctive:

    Verify the front door is up after a reboot or upgrade

    line 55
  5. Very specific:

    Never run ad-hoc tailscale serve … commands.

    line 109
  6. Very specific:

    Add APP_PASSWORD_<SLUG> / APP_TITLE_<SLUG> / APP_DESC_<SLUG> for any gated apps

    line 134
  7. Very specific:

    If password-gating, also add to the auth-service block's env:

    line 186
  8. Very specific:

    After this, pm2 list should show a caddy entry.

    line 297
  9. Very specific:

    Fleet rollout addendum: when exposing dashboards across multiple Hermes hosts, use references/hermes-dashboard-fleet-rollout.md for slug …

    line 360
  10. Very specific:

    Always check ls ~/.hermes/hermes-agent/hermes_cli/web_dist/index.html during pre-flight.

    line 419
  11. Very specific:

    After that --skip-build is correct (fast start, no rebuild per boot).

    line 454
  12. Distinctive:

    Fix pattern: add a dedicated Caddy listener on a loopback port whose only job is to reverse-proxy to the tailnet-IP backend, then funnel …

    line 508
  13. Very specific:

    Use a dedicated listener (NOT a path under :8080) when the upstream is a SPA or Next.js app with absolute /_next/static/...

    line 531
  14. Very specific:

    Do not infer public exposure from a process binding 0.0.0.0.

    line 561
  15. Distinctive:

    When asked "how is this public?" or before rolling out a webhook, verify all layers:

    line 570
  16. Distinctive:

    Check for PF/NAT redirects if the port appears reachable but is not in Funnel/Caddy.

    line 576
  17. Very specific:

    If you're calling PM2 from inside a Hermes tool environment, hardcode the absolute path with no shell expansion — neither ~ nor $HOME …

    line 596
  18. Distinctive:

    If a freshly added funnel route returns 502, work through this in order:

    line 660
  19. Very specific:

    If another service already owns the public :443 funnel — check tailscale serve status for an existing :443 … proxy http://127.0.0.1:<port> …

    line 776
  20. Distinctive:

    Before installing anything on a fresh fleet host, gather all of this at once:

    line 785
  21. Very specific:

    If Caddy runs as a standalone process (not under PM2) and dies, nothing restarts it and the whole front door 502s while every backend stays …

    line 828
  22. Very specific:

    When adding a dashboard card to an existing index page, READ the file first and reuse its existing card markup/classes (e.g.

    line 843
  23. Very specific:

    After editing, the served HTML is authoritative.

    line 853
  24. Distinctive:

    Running Caddy unsupervised (bare process, not under PM2): If Caddy dies, nothing restarts it and every route 502s while backend apps stay …

    line 929
  25. Very specific:

    Shortcut: if ~/.cloudflared/cert.pem already exists (a prior cloudflared tunnel login), the whole thing is headless — no dashboard token, …

    line 980

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Mini-App

Mission: Operate the hermes-config mini-app router on this machine — the lightweight stack that exposes one or more named "mini-apps" at clean URL paths on a single Tailscale HTTPS host, with optional per-app password gating. One front door, many apps, no cloud.

A mini-app is any process that binds to a localhost port and serves HTTP — a Node service, a Python FastAPI app, a Hermes dashboard, a webhook receiver. The app-router fronts them all with Caddy and gives each one:

  • A clean path (https://<host>/<slug>/)
  • Optional password protection via an Express auth sidecar
  • HTTPS for free via Tailscale Serve (and optional public exposure via Funnel)
  • PM2 supervision so it survives crashes and reboots

The router stack lives upstream in hermes-config under devops/app-router/. This skill is the operator playbook — what to do once it's installed.

When to use

Load this skill when you (a Hermes fleet agent) are asked to:

  • Add a new mini-app to this host's router (Hermes dashboard, webhook receiver, status page, anything that binds to localhost)
  • Remove or rename an existing mini-app
  • Add or change a password on a gated mini-app
  • Expose a mini-app publicly via Tailscale Funnel
  • Diagnose a 502 / 404 / auth-loop on a mini-app
  • Reload Caddy after editing the Caddyfile
  • Restore Tailscale Serve after another tool wiped it
  • Verify the front door is up after a reboot or upgrade

Don't use for: writing the mini-app's code itself (that's a normal app), or changing the auth sidecar's source (that's an upstream PR in hermes-config).

Stack at a glance

Internet (optional)
   │ Tailscale Funnel on:443 (only if you've opted in per-host)
   ▼
Tailnet host: <machine>.<your-tailnet>.ts.net
   │
   ▼  127.0.0.1:8080  (Caddy — declarative, hot-reloadable)
   ├── /auth/*       → auth sidecar      (Node + Express, JWT-style cookies)
   ├── /health       → "ok" 200
   ├── /hooks/*      → optional bearer-injected webhook proxy
   ├── /<slug-1>/*   → mini-app on a localhost port
   ├── /<slug-2>/*   → mini-app on a localhost port (password-gated)
   └── /            → static welcome page

Single sources of truth on this machine (after install):

FileWhat it controls
~/mini-apps/ecosystem.config.jsPM2 process list + ALL env vars (incl. auth passwords)
~/mini-apps/<reg>/CaddyfilePath → upstream routing
~/mini-apps/<reg>/tailscale-serve.jsonPublic/tailnet exposure (huJSON; compiled by apply script)

<reg> is router/ or _registry/ — check before you edit. The current installer creates router/; hosts installed before that carry _registry/. Both are in the wild, so resolve it once at the start of any task and reuse it.

Resolve the operator's real home first — under a Hermes tool environment $HOME may point at the profile home, where ~/mini-apps does not exist even though the router is running (see "The router dir may have been renamed" below):

# anchor on the caddy BINARY path so this cannot match your own shell.
# -ww keeps long argv from being width-clipped on platforms that do that.
APPS=$(ps -eww -o args= | grep '^/[^ ]*caddy run' \
    | sed -n 's#.*--config \(.*\)/[^/]*/Caddyfile.*#\1#p' | head -1)
APPS=${APPS:-$(ls -d /Users/*/mini-apps /home/*/mini-apps 2>/dev/null | head -1)}
REG=$(ls -d "$APPS"/router "$APPS"/_registry 2>/dev/null | head -1)
echo "apps=$APPS reg=$REG"

An empty REG with a running Caddy means the probe missed, NOT that the router is uninstalled — fall back to the ground-truth search below before concluding anything.

Examples below write $APPS and $REG. On a host you have already identified you can substitute the literal paths.

Never run ad-hoc tailscale serve … commands. Edit the JSON, run the apply script.

First-time install on this machine

brew install caddy            # macOS — use the equivalent on Linux
npm install -g pm2

cd ~/src/hermes-config         # clone if you don't have it
bash devops/app-router/scripts/install.sh

The installer copies templates into ~/mini-apps/, renders the Caddyfile with the right paths, installs auth-service deps, and stages the launchd plist (macOS only). Re-running is safe; --force overwrites existing files.

Then by hand. A fresh install from the current installer creates router/, so set REG before following these steps:

REG=$(ls -d ~/mini-apps/router ~/mini-apps/_registry 2>/dev/null | head -1)
  1. Edit ~/mini-apps/ecosystem.config.js:
    • Set AUTH_SECRET to openssl rand -hex 32 (one per machine)
    • Add APP_PASSWORD_<SLUG> / APP_TITLE_<SLUG> / APP_DESC_<SLUG> for any gated apps
  2. Edit $REG/Caddyfile to declare each app's route
  3. Edit $REG/tailscale-serve.json if you want a non-default serve layout (default exposes Caddy on :443)

Start everything under PM2:

pm2 start ~/mini-apps/ecosystem.config.js
pm2 start /opt/homebrew/bin/caddy --name caddy --interpreter none -- \
  run --config "$REG/Caddyfile" --adapter caddyfile
pm2 save
pm2 startup    # paste the printed sudo command — needed for resurrect-on-reboot

Apply Tailscale Serve:

"$REG"/apply-tailscale-serve.sh
# macOS launchd that replays on login:
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.hermes.mini-app-router-serve.plist

If another tool on this host manages Tailscale Serve, disable that first. The usual failure is a second process running tailscale serve reset on startup, which wipes these routes. Hermes' own gateway has no Tailscale integration and does not touch serve config, so a Hermes-only host needs no opt-out. A legacy agent runtime on the same machine may still own it — check what else drives serve:

tailscale serve status                                    # what is configured now
grep -rl "tailscale serve" ~/Library/LaunchAgents \
    ~/.config/systemd/user 2>/dev/null                    # who else replays it

Adding a mini-app

Three edits, then reload. Pick a slug (^[a-z0-9](?:[a-z0-9-]{0,30}[a-z0-9])?$) and a free localhost port (convention: apps start at 3001).

1. PM2 ecosystem entry — append to apps in ecosystem.config.js:

{
  name: "my-app",
  script: "./my-app/server.js",     // or absolute path, or a python command via interpreter
  cwd: "~/mini-apps",
  env: { PORT: 3001 },
},

If password-gating, also add to the auth-service block's env:

APP_PASSWORD_MY_APP: "the-password",
APP_TITLE_MY_APP:    "My App",
APP_DESC_MY_APP:     "One-line description shown on the login form.",

The slug-to-env rule is uppercase + -→_. my-app ⇒ APP_PASSWORD_MY_APP.

2. Caddyfile route — add a handle block above the catch-all:

Open app (no auth):

handle /my-app/* {
    uri strip_prefix /my-app
    reverse_proxy 127.0.0.1:3001
}

Password-gated app:

handle /my-app/* {
    forward_auth 127.0.0.1:3000 {
        uri /auth/verify?app=my-app
        copy_headers Cookie
        @unauthorized status 401
        handle_response @unauthorized {
            redir * /auth/login?app=my-app&next={http.request.uri.path} 302
        }
    }
    uri strip_prefix /my-app
    reverse_proxy 127.0.0.1:3001
}

3. Reload PM2 + Caddy:

pm2 restart ecosystem.config.js
caddy reload --config ~/mini-apps/_registry/Caddyfile --adapter caddyfile

You do not need to touch Tailscale when adding an app. Caddy is the path router; Tailscale just sees one upstream (Caddy on :8080).

Verify:

curl -sI http://127.0.0.1:8080/my-app/   # expect 200 (open) or 302 (gated)

Removing a mini-app

pm2 delete my-app

Then drop the Caddyfile handle block, the ecosystem entry, and any APP_*_MY_APP env vars from the auth-service block. Reload PM2 + Caddy.

Slug convention for Hermes dashboards

Hermes agent dashboards use the slug hermes-<agent> — hermes-atlas, hermes-scout, hermes-vega. Two reasons:

  1. Grouping — every Hermes dashboard sorts together under one prefix.
  2. Collision avoidance — a non-Hermes app may already own the bare agent name. If /vega/ is already a web app on some port, the Vega agent's Hermes dashboard still goes at /hermes-vega/ (profile vega, its own port). Never mount a Hermes dashboard at a bare name that collides with an existing app slug.

Assign the dashboards a contiguous port block (for example 9119+, one per agent) and keep the mapping in your own registry. The env slug-to-password rule still applies: hermes-atlas ⇒ APP_PASSWORD_HERMES_ATLAS.

Front-door outage: Caddy died (the #1 silent failure)

Symptom: every app 502s at once, or curl http://127.0.0.1:8080/health returns 000 (connection refused), but pm2 list shows all the backend apps online. The router — Caddy — is dead and nothing restarted it.

Root cause is almost always that Caddy was started as a bare process, not under PM2. A bare caddy run … (or one launched outside the ecosystem) has no supervisor, so when it crashes or the machine churns, it stays dead while every backend keeps running. The backends being healthy is what makes this confusing.

Recovery — bring the front door back, supervised this time:

export PM2_HOME=/Users/<user>/.pm2          # literal path, no ~ or $HOME under Hermes
CADDY=$(command -v caddy || echo /opt/homebrew/bin/caddy)

# 1. Find the config path the OLD process actually used — do NOT assume the skill's
#    default. The router dir may have been renamed (see next pitfall).
ps aux | grep "caddy.*run" | grep -v grep      # reveals --config <path> if still running

# 2. Validate before starting
"$CADDY" validate --config <real-Caddyfile> --adapter caddyfile

# 3. Start UNDER PM2 so it resurrects on crash
pm2 start "$CADDY" --name caddy --interpreter none -- \
  run --config <real-Caddyfile> --adapter caddyfile
pm2 save

# 4. Verify
curl -s -o /dev/null -w "%{http_code}\n" http://127.0.0.1:8080/health   # expect 200

After this, pm2 list should show a caddy entry. If it never had one, that was the bug — Caddy must always be a PM2 process.

Reboot resurrection is a separate, latent outage

pm2 save persists the process list, but it does NOTHING on reboot unless a launchd/ systemd boot hook exists. Check it explicitly — a missing hook is a time-bomb identical to the bare-Caddy outage, just triggered by a restart:

launchctl list 2>/dev/null | grep -i pm2 || echo "NO pm2 boot hook — reboot resurrects nothing"

Fix (needs the user's sudo — surface it, don't try to run sudo yourself under Hermes):

sudo env PATH=$PATH PM2_HOME=/Users/<user>/.pm2 pm2 startup launchd -u <user> --hp /Users/<user>

The router dir may have been renamed (find it, never assume)

The historical path was ~/openclaw-apps/ (pre-Hermes); the current one is ~/mini-apps/. Older hosts may still carry the legacy directory name. Hardcoded paths in this skill or in old notes will stat: No such file even though the stack is alive. Locate the real one from ground truth rather than guessing:

ps aux | grep "caddy.*run" | grep -v grep           # --config reveals the registry dir
find /Users/<user> -maxdepth 4 -name ecosystem.config.js 2>/dev/null | grep -v node_modules
find /Users/<user> -maxdepth 5 -name Caddyfile 2>/dev/null | grep -v node_modules

The Caddyfile lives one level down, at <router-dir>/router/Caddyfile (current installer) or <router-dir>/_registry/Caddyfile (older hosts) — not directly in the router dir. Note also: under a Hermes tool environment, $HOME is rewritten to the profile home, so ~/mini-apps and bare ls/grep may fail to see the real dir — use absolute /Users/<user>/... paths, and the Read/Grep tools (which resolve absolute paths) rather than shell ls/cat when the shell is sandboxed.

Hermes dashboards behind the router

Deep detail: see references/hermes-dashboard-rollout.md for the full debugging path — the Secure-cookie/HTTPS trap, the --skip-build web_dist crash-loop fix (a FRESH host has no compiled web_dist; build once with cd ~/.hermes/hermes-agent/web && npm install && npm run build), checking who owns :443 before exposure decisions, profile-vs-root DB selection, the SSH-hairpin gotcha, matching a custom index page's design system when adding a dashboard card, and the end-to-end verification sequence. Read it before any new hermes-<name> rollout.

Slug convention: mount Hermes dashboards at /hermes-<agent>/ (e.g. /hermes-atlas/, /hermes-orion/) so the agent dashboard never collides with a same-named product app (/nova/ = CFO app; /hermes-nova/ = the agent).

Critical: the auth cookie is Secure, so the dashboard MUST be reached over HTTPS. A plain-HTTP tailnet door (Tailscale serve --http=PORT) silently breaks login: POST /auth/login returns 303 + cookie, but the authed GET comes back 302 because the browser won't resend a Secure cookie over HTTP. If Caddy doesn't own a 443 funnel on the host, add a tailnet-only HTTPS door:

Hermes dashboards behind the router

A Hermes dashboard is a normal mini-app, with three extra requirements.

Fleet rollout addendum: when exposing dashboards across multiple Hermes hosts, use references/hermes-dashboard-fleet-rollout.md for slug naming (/hermes-$botname), public Funnel/auth routing, Linux PM2/Caddy notes, and browser verification standards.

1. ecosystem.config.js node-via-shell trick or PM2's interpreter override:

{
  name: "my-dashboard",
  script: "hermes",
  args: "dashboard --port 9120 --no-open --skip-build",
  interpreter: "none",
  cwd: "~/mini-apps",
  env: {
    PORT: 9120,
    HERMES_PROFILE: "my-profile",   // omit if you want the root state DB
  },
},

2. Caddyfile — add the prefix header so the SPA rewrites asset URLs correctly, but STILL keep uri strip_prefix — the FastAPI routes mount at root:

handle /my-dashboard/* {
    forward_auth 127.0.0.1:3000 {
        uri /auth/verify?app=my-dashboard
        copy_headers Cookie
        @unauthorized status 401
        handle_response @unauthorized {
            redir * /auth/login?app=my-dashboard&next={http.request.uri.path} 302
        }
    }
    uri strip_prefix /my-dashboard
    reverse_proxy 127.0.0.1:9120 {
        header_up Host {upstream_hostport}
        header_up X-Forwarded-Prefix /my-dashboard
    }
}

X-Forwarded-Prefix is used by Hermes ONLY for HTML rewriting (asset paths and __HERMES_BASE_PATH__). API routes still mount at root, so uri strip_prefix is mandatory.

3a. The dashboard frontend must be built (web_dist) before --skip-build works. On a fresh fleet machine the Hermes web UI is often unbuilt, and a dashboard launched with --skip-build will crash-loop with: ✗ --skip-build was passed but no web dist found at: …/hermes_cli/web_dist. PM2 will show the process errored with a climbing restart count, and the port won't listen. Build it ONCE (takes ~1-2 min), then restart the PM2 process:

cd ~/.hermes/hermes-agent/web && npm install --no-audit --no-fund && npm run build
# build writes to../hermes_cli/web_dist/; then:
PM2_HOME=/Users/<user>/.pm2 pm2 restart <name>-dashboard
PM2_HOME=/Users/<user>/.pm2 pm2 reset <name>-dashboard   # clear the crash-loop counter

Always check ls ~/.hermes/hermes-agent/hermes_cli/web_dist/index.html during pre-flight; some machines have it pre-built, others don't.

3b. Verify there are sessions to show — don't assume the profile name. Some fleet agents run as root cron jobs against the root ~/.hermes/state.db, not against ~/.hermes/profiles/<name>/state.db (Scout, Orion, Ali use the root DB with NO --profile flag; Atlas/Vega/Nova use --profile <name>). Pinning the wrong DB shows an empty dashboard. Quick triage:

for db in $(find ~/.hermes -name state.db 2>/dev/null); do
    n=$(sqlite3 "$db" "SELECT COUNT(*) FROM sessions;" 2>/dev/null)
    echo "  $n  $db"
done

Pin to whichever DB actually holds the sessions.

4. Build the dashboard web UI first, or --skip-build crash-loops. The dashboard is served from a pre-built web_dist/ directory. On a fresh Hermes install it may not exist yet, and PM2 will show the process errored with restart count climbing. Check pm2 logs <name> for:

✗ --skip-build was passed but no web dist found at:.../hermes_cli/web_dist

Fix — build it once (takes ~1-2 min), then restart:

cd ~/.hermes/hermes-agent/web && npm install && npm run build
# emits ~/.hermes/hermes-agent/hermes_cli/web_dist/
pm2 restart <name> && pm2 reset <name>   # reset clears the inflated restart counter

After that --skip-build is correct (fast start, no rebuild per boot).

5. Pin HERMES_HOME in env and pass the profile as a CLI arg, not env. The reliable combo is args: "--profile <name> dashboard..." (omit --profile for the root DB) plus env: { HERMES_HOME: "/Users/<user>/.hermes", PATH: "<venv-bin>:/opt/homebrew/bin:..." }. Set HERMES_HOME to an ABSOLUTE path — under PM2 the rewritten $HOME otherwise points the dashboard at an empty DB. Use the absolute venv hermes binary as script (e.g. /Users/<user>/.hermes/hermes-agent/venv/bin/hermes), not the bare hermes name.

Verify session presence at the DB, NOT through the proxied API. The Hermes dashboard's /api/sessions endpoint enforces its own token auth that only the in-browser SPA supplies — a plain curl (even with a valid mini-app auth cookie) gets {"detail":"Unauthorized"} or parses as 0 sessions. That is NOT evidence of an empty dashboard. Confirm data the honest way:

sqlite3 /Users/<user>/.hermes/profiles/<profile>/state.db "SELECT COUNT(*) FROM sessions;"
# scout runs against the ROOT db: /Users/<user>/.hermes/state.db

The end-to-end proof for a Hermes dashboard is: backend port returns 200 on loopback + login round-trip serves the SPA shell (200, <title>Hermes Agent - Dashboard</title>) + the pinned DB has a non-zero session count. The SPA shell is ~700 bytes; data hydrates client-side after login.

Public exposure via Tailscale Funnel

Tailscale Serve is tailnet-only by default. To share an app publicly, add the funnel-enabled port to ~/mini-apps/_registry/tailscale-serve.json:

"AllowFunnel": {
  "${HOST}:443": true
}

Then re-apply: ~/mini-apps/_registry/apply-tailscale-serve.sh.

Funnel-allowed ports are exactly {443, 8443, 10000}. Anything else fails silently or is rejected by Tailscale.

Security rule (non-negotiable): a funnel'd port is fully public. Only password-gated or token-gated upstreams may live behind it. Put passwordless admin UIs on a separate tailnet-only port (e.g. :8443) by adding a Web entry without a matching AllowFunnel entry.

The Tailscale "Proxy" → loopback constraint

The Proxy field in tailscale-serve.json (and tailscale serve --bg http://…) only works with loopback backends (127.0.0.1:*). Pointing a Serve/Funnel handler at a tailnet-IP backend (e.g. 100.x.y.z:20128) returns HTTP 502 through the funnel and silently strips the Funnel off that port.

Fix pattern: add a dedicated Caddy listener on a loopback port whose only job is to reverse-proxy to the tailnet-IP backend, then funnel that loopback listener. Example for an LLM proxy bound to a tailnet IP:

# In ~/mini-apps/_registry/Caddyfile, ABOVE the main:8080 block:
:8090 {
    bind 127.0.0.1
    reverse_proxy 100.x.y.z:20128 {
        header_up Host {upstream_hostport}
    }
}

Then in tailscale-serve.json:

"${HOST}:10000": {
  "Handlers": { "/": { "Proxy": "http://127.0.0.1:8090" } }
},
"AllowFunnel": { "${HOST}:10000": true }

Use a dedicated listener (NOT a path under :8080) when the upstream is a SPA or Next.js app with absolute /_next/static/... asset paths — mounting it on a prefix inside :8080 would 404 every asset.

Hooks / webhooks (optional)

Caddy can front a local agent gateway's webhook endpoint so external callers reach it over the same HTTPS door.

Hermes runs its own webhook listener (WEBHOOK_ENABLED=true, WEBHOOK_PORT, default 8644) and authenticates callers by HMAC signature using WEBHOOK_SECRET, not a bearer token. Caddy does not need to inject credentials — proxy the path and let the gateway verify the signature itself:

handle /hooks/* {
    reverse_proxy 127.0.0.1:8644
}

Point the sender at https://<host>/hooks/<route> and configure the same secret on both ends. Do not add the mini-app auth sidecar in front of this path — the caller is a machine and cannot complete a cookie login.

A legacy agent runtime on the same host may instead expose a bearer-token /hooks/* endpoint, which needs token injection in the Caddy env block. Confirm which gateway owns the port before wiring it.

Public exposure audit for hooks

Do not infer public exposure from a process binding 0.0.0.0. On these hosts there are two separate layers:

  • Raw listener: e.g. a local webhook process on *:8644; reachable on local/LAN/tailnet interfaces depending on firewall, but not necessarily public.
  • Public front door: Tailscale Funnel routes (tailscale serve status) usually point public HTTPS traffic at loopback Caddy, and Caddy may then expose selected paths such as /hooks/*.

When asked "how is this public?" or before rolling out a webhook, verify all layers:

  1. lsof -nP -iTCP:<port> -sTCP:LISTEN — identify the raw listener and bind address.
  2. tailscale serve status — identify which ports/paths are Funnel-enabled vs tailnet-only.
  3. Inspect the active Caddyfile for handle /hooks/* or webhook-specific proxy blocks.
  4. Check for PF/NAT redirects if the port appears reachable but is not in Funnel/Caddy.
  5. Redact injected bearer tokens/secrets when reporting the route.

A common safe conclusion: existing /hooks/* may be public via Funnel + Caddy bearer injection, while a separate Hermes webhook listener is only bound locally/tailnet and is not public until a Caddy/Tailscale route is added.

The PM2 $HOME trap (CRITICAL when running tools inside Hermes)

Hermes rewrites $HOME for tool execution to ~/.hermes/profiles/<profile>/home/. PM2 keys its socket off $HOME. Without PM2_HOME set, you talk to a shadow PM2 daemon that supervises nothing.

Always export this first:

# Real shell only — DO NOT use this under Hermes (both $HOME and ~ expand to the rewritten profile home):
export PM2_HOME=$HOME/.pm2

If you're calling PM2 from inside a Hermes tool environment, hardcode the absolute path with no shell expansion — neither ~ nor $HOME resolves to the user's real home under the rewritten environment:

# Substitute your real username; do NOT use ~ or $HOME here.
export PM2_HOME=/Users/<user>/.pm2     # macOS
# export PM2_HOME=/home/<your-username>/.pm2    # Linux

Quick sanity check after exporting — this should print the real user's home, not a .hermes/profiles/... path:

echo "$PM2_HOME"

Detect the trap:

ps -ef | grep "PM2.*God" | grep -v grep

Two daemons with different $HOME paths = you spawned a shadow. Clean up:

pm2 kill                                  # kills the shadow
export PM2_HOME=/Users/<user>/.pm2   # the real one — literal path, no ~ or $HOME
pm2 list                                  # now shows the actual fleet

Same trap applies to skill-asset paths: ~/mini-apps/... resolves under the rewritten Hermes home. Hardcode the user's real path if you're scripting from inside a Hermes profile.

Auth-service env reload trap

The auth sidecar reads APP_PASSWORD_* / APP_TITLE_* / APP_DESC_* at startup. Changing a password in ecosystem.config.js and running pm2 restart auth-service --update-env does NOT pick up changes from the ecosystem file — --update-env only re-reads env from PM2's own state.

To actually reload:

export PM2_HOME=/Users/<user>/.pm2   # literal path — no ~ or $HOME under Hermes
pm2 delete auth-service
pm2 start ~/mini-apps/ecosystem.config.js --only auth-service

When Tailscale Serve goes sideways

Any tool that runs tailscale serve reset on its own startup will wipe your config. Hermes' gateway does not manage Tailscale, so on a Hermes-only host suspect another service or a legacy agent runtime. Diagnosis: tailscale serve status shows the wrong routes (or nothing), or a funnel'd port disappeared. Recovery is one command:

~/mini-apps/_registry/apply-tailscale-serve.sh

The apply script is idempotent — safe to run any time.

If a freshly added funnel route returns 502, work through this in order:

  1. tailscale funnel status — confirm the funnel is still on for that port. If it's missing, you almost certainly pointed it at a non-loopback backend (see the loopback constraint above).
  2. lsof -nP -iTCP:<backend-port> -sTCP:LISTEN — confirm the backend binds loopback. If it binds a tailnet IP only, use the Caddy-bridge pattern.
  3. curl -sI http://127.0.0.1:<port>/ — direct loopback check. Connection refused = tailnet-IP-only backend.

Login round-trip (for debugging an auth loop)

Always run a NEGATIVE auth test too — a passing positive test doesn't prove the gate actually blocks. A correct gate returns 302 (redirect to login) for a wrong password:

HOST=<host>.<tailnet>.ts.net
JAR=/tmp/jar-bad.txt && rm -f $JAR
curl -sk -o /dev/null -c $JAR -X POST "https://$HOST/auth/login" \
    --data-urlencode "app=<slug>" --data-urlencode "password=WRONG" \
    --data-urlencode "next=/<slug>/" -H "Origin: https://$HOST" >/dev/null
curl -sk -o /dev/null -w "wrong-pw authed GET => %{http_code}\n" -b $JAR "https://$HOST/<slug>/"
# expect 302 (denied). 200 here means the gate is broken/open.

A successful login returns 303/302 with a Set-Cookie; the authed GET then returns 200.

JAR=/tmp/mini-app-cookies.txt && rm -f $JAR

# 1. Login: 303 + Set-Cookie
curl -si -c $JAR -X POST "http://127.0.0.1:8080/auth/login" \
    --data-urlencode "app=my-app" \
    --data-urlencode "password=the-password" \
    --data-urlencode "next=/my-app/" \
    -H "Origin: http://127.0.0.1:8080" -H "Host: 127.0.0.1:8080" | head -10

# 2. Authed GET — 200, not 302
curl -sI -b $JAR "http://127.0.0.1:8080/my-app/" | head -5

End-to-end verification after any change

# 1. PM2 supervises everything
export PM2_HOME=~/.pm2
pm2 list

# 2. Expected ports listen
lsof -iTCP -sTCP:LISTEN -P -n | grep -E ":3000|:8080|<your-app-ports>"

# 3. Caddy is on the latest config
caddy reload --config ~/mini-apps/_registry/Caddyfile --adapter caddyfile

# 4. Routes return expected codes
for path in "" "auth/login?app=my-app" "my-app/" "health"; do
    code=$(curl -sk -o /dev/null -w "%{http_code}" \
        "https://<host>.<tailnet>.ts.net/$path")
    echo "  $code  /$path"
done

# 5. Persist PM2 state across reboot
pm2 save

Expected:

  • / → 200 (welcome page)
  • /auth/login?app=… → 200
  • /<gated-app>/ → 302 (redirect to login when not authed)
  • /<open-app>/ → 200
  • /health → 200
  • /hooks/test → 405 (POST-only) if hooks are enabled

The auth cookie is Secure — gated apps REQUIRE an HTTPS door

The auth sidecar sets oc_auth_<slug> with the Secure flag. A browser (and curl across hosts) will only send it back over HTTPS. Consequences:

  • A plain-HTTP tailnet door (e.g. tailscale serve --http=4243) will let you log in (303 + Set-Cookie) but the authed GET still 302s because the cookie never comes back. Gated apps are effectively unusable there.
  • Loopback curl gives a false positive: curl -c jar … && curl -b jar … over http://127.0.0.1:8080 "works" because curl reuses the jar within the same call regardless of the Secure flag. Don't trust a loopback round-trip to prove auth — test over the real HTTPS tailnet/funnel URL.

Always front gated apps with an HTTPS door. On a machine where Caddy owns :443 (funnel), that's covered. On a machine where another service owns :443, add a tailnet-only HTTPS door on :8443 pointed at Caddy's loopback :8080:

tailscale serve --bg --https=8443 http://127.0.0.1:8080
# tailnet-only (no --funnel) — survives reboot natively. Verify:
tailscale serve status   # → https://<host>.<tailnet>.ts.net:8443 (tailnet only)

Then the app is reachable at https://<host>.<tailnet>.ts.net:8443/<slug>/.

Correct end-to-end verification (run from a DIFFERENT host on the tailnet, over HTTPS):

BASE="https://<host>.<tailnet>.ts.net:8443"; JAR=/tmp/jar.txt; rm -f $JAR
login=$(curl -sk -o /dev/null -w "%{http_code}" -c $JAR -X POST "$BASE/auth/login" \
  --data-urlencode "app=<slug>" --data-urlencode "password=<pw>" \
  --data-urlencode "next=/<slug>/" -H "Origin: $BASE")
authed=$(curl -sk -o /dev/null -w "%{http_code}" -b $JAR "$BASE/<slug>/")
echo "login=$login authed=$authed"   # want 303 + 200; wrong password → authed 302

Note: a box hitting its OWN tailnet hostname from inside an SSH session can hang (hairpin) — run the cross-host test from a different fleet machine instead.

When NOT to take over the public :443 funnel

If another service already owns the public :443 funnel — check tailscale serve status for an existing :443 … proxy http://127.0.0.1:<port> entry — do NOT seize :443 for Caddy without asking, because it disrupts whatever is already answering there. Default to a tailnet-only HTTPS :8443 door (works on any device signed into the tailnet, not public). Add a public funnel later only on explicit request. This is the safe, reversible resting state.

Pre-flight survey for a NEW machine (one SSH pass)

Before installing anything on a fresh fleet host, gather all of this at once:

# identity + prereqs
whoami; echo $HOME; sw_vers -productVersion; uname -m
ls /opt/homebrew/bin/brew /opt/homebrew/bin/caddy 2>/dev/null   # caddy often MISSING
which node pm2 hermes
ls ~/.local/bin/hermes ~/.hermes/hermes-agent/venv/bin/hermes 2>/dev/null
# sessions per DB (root vs profiles) — picks the right --profile flag
for db in ~/.hermes/state.db ~/.hermes/profiles/*/state.db; do
  [ -f "$db" ] && echo "$(sqlite3 "$db" 'SELECT COUNT(*) FROM sessions') $db"; done
ls ~/.hermes/hermes-agent/hermes_cli/web_dist/index.html 2>/dev/null  # build needed?
ls -d ~/mini-apps ~/openclaw-apps 2>/dev/null                          # existing router?
# who owns:443 / what serve routes exist already?
/opt/homebrew/bin/tailscale serve status

tailscale lives at /opt/homebrew/bin/tailscale on macOS (the GUI app's binary is also linked there); it's usually NOT on a non-interactive SSH PATH, so call it by full path or export PATH=/opt/homebrew/bin:$PATH first. Same for brew/caddy. nvm-based node lives at ~/.nvm/versions/node/<ver>/bin.

Standing up the router on a bare machine

brew install caddy && npm install -g pm2. There's no public installer for the auth-service on a bare box, so copy a known-good auth-service/ from an existing fleet machine. Use scp, not an SSH-pipe heredoc — tar czf - … | ssh host 'tar xzf -' fails ("Unrecognized archive format") and base64-through-heredoc fails (the heredoc consumes stdin so the piped data never arrives). Reliable pattern:

ssh src 'cd ~/mini-apps && tar czf - --exclude=node_modules auth-service' > /tmp/a.tgz
scp /tmp/a.tgz dst:/tmp/a.tgz
ssh dst 'cd ~/mini-apps && tar xzf /tmp/a.tgz && cd auth-service && npm install --no-audit --no-fund'

Then write ecosystem.config.js (fresh per-host AUTH_SECRET via openssl rand -hex 32 — NEVER reuse between machines), the Caddyfile, and an index page, start all three under PM2 (auth-service, the dashboard, caddy), pm2 save, add the :8443 serve, and do the HTTPS round-trip test.

Run Caddy UNDER PM2 — not as a bare process

If Caddy runs as a standalone process (not under PM2) and dies, nothing restarts it and the whole front door 502s while every backend stays healthy — a confusing outage where pm2 list looks fine but https://<host>/ is down. Diagnose: ps aux | grep "caddy run" shows nothing, curl http://127.0.0.1:8080/health returns 000. Fix and prevent:

caddy validate --config <Caddyfile> --adapter caddyfile   # validate first
pm2 start /opt/homebrew/bin/caddy --name caddy --interpreter none -- \
  run --config <Caddyfile> --adapter caddyfile
pm2 save

Index-page cards: match the page's own design system

When adding a dashboard card to an existing index page, READ the file first and reuse its existing card markup/classes (e.g. this fleet's pages use .card + .pill, not the older .card-lock span). Two failure modes seen:

  1. Orphaned card — a naive "insert before </body>" lands the card OUTSIDE the styled <main>/.grid container, so it renders as unstyled floating text. Insert INSIDE the .grid (e.g. right after <div class="grid">).
  2. Wrong markup — copying the Mac Studio .card-lock pattern onto a page that only defines .pill yields an unstyled card. Use the target page's own classes.

After editing, the served HTML is authoritative; a stale browser view is just cache (hard-refresh ⌘⇧R). Verify with curl -sk "$BASE/" | grep -n hermes-<slug> and confirm the card line sits between <div class="grid"> and </main>.

Reboot-resurrect needs sudo (flag it, don't skip it)

pm2 save persists the process list, but resurrection on reboot needs a launchd entry from sudo pm2 startup — which can't run unattended. Check launchctl list | grep -i pm2; if absent, surface the exact command for the user to run (it prints from pm2 startup launchd -u <user> --hp /Users/<user>) rather than silently leaving the host unable to recover from a reboot. Tailscale serve --bg config persists across reboots natively, so only PM2 needs this.

  1. Editing the Caddyfile in /etc/... or wherever you found Caddy on the system. The router uses ~/mini-apps/_registry/Caddyfile and runs Caddy under PM2 with --config pointing at that file. Edit there, reload from there.
  2. Forgetting uri strip_prefix /<slug> on a Hermes dashboard. Hermes receives X-Forwarded-Prefix but only uses it for HTML rewriting. The API routes mount at root and 404 without strip-prefix.
  3. pm2 restart auth-service --update-env after editing ecosystem.config.js. --update-env doesn't re-read the file. Use pm2 delete + pm2 start --only to actually reload.
  4. Pointing Tailscale Serve/Funnel at a non-loopback backend. Returns 502 and strips Funnel off the port silently. Use a Caddy loopback bridge.
  5. Putting passwordless admin UIs on a funnel'd port. A funnel'd port is fully public. Move them to :8443 (tailnet-only) by adding a Web entry without an AllowFunnel entry.
  6. Letting a second process manage Tailscale Serve. Anything that runs tailscale serve reset on startup wipes these routes on every restart. Hermes' gateway has no Tailscale integration, so on a Hermes-only host look for another service or a legacy agent runtime; disable its serve management, then re-apply with apply-tailscale-serve.sh.
  7. Calling PM2 from inside a Hermes tool environment without exporting PM2_HOME. You talk to a shadow daemon that supervises nothing. Export first, every time.
  8. Adding a port outside {443, 8443, 10000} to AllowFunnel. Tailscale rejects it silently. Stick to those three for public exposure.
  9. pm2 save without PM2_HOME exported. You save the wrong process list to the wrong dump file, and reboot resurrects nothing.
  10. Slugs with uppercase, underscores, or >32 chars. The auth sidecar rejects them at /auth/verify and /auth/login with 400. Pattern is ^[a-z0-9](?:[a-z0-9-]{0,30}[a-z0-9])?$.
  11. Serving a password-protected app over plain HTTP. The session cookie is Secure, so browsers and curl refuse to send it back over http://. Symptom: login POST returns 303 (looks like success) but every subsequent GET bounces back to the login page because the cookie never returns. ALWAYS use an HTTPS door: tailscale serve --bg --https=8443 http://127.0.0.1:8080 for tailnet-only, or the :443 HTTPS funnel. Beware a pre-existing plain-HTTP :4243-style serve listener — it's the trap. When testing cross-node, use the https:// URL with curl -sk; a loopback test on the box can falsely "pass" because curl reuses the jar within one invocation regardless of Secure.
  12. pm2 startup needs sudo and bakes in the wrong --hp under Hermes. Reboot resurrection requires a launchd entry (launchctl list | grep pm2). Generating it needs an interactive sudo, so it CANNOT be done unattended — hand the user the exact command. Generate it with a clean HOME or the rewritten profile $HOME leaks into --hp: HOME=/Users/<user> PM2_HOME=/Users/<user>/.pm2 pm2 startup launchd -u <user> --hp /Users/<user>. The --hp must match where dump.pm2 actually lives. The pm2 save dump and tailscale serve --bg config both persist across reboot on their own; only the launchd entry needs the one-time sudo.
  13. Streaming a tarball over ssh... 'bash -s' <<heredoc to transfer files. The heredoc consumes stdin, so piped binary data never arrives ("Unrecognized archive format" / "error decoding base64"). Use scp for file transfer, and reserve the heredoc-over-ssh pattern for running commands that don't also need piped stdin.
  14. Running Caddy unsupervised (bare process, not under PM2). If Caddy dies, nothing restarts it and every route 502s while backend apps stay healthy — looks like a total outage but it's just the front door. Always start Caddy under PM2 (pm2 start <caddy> --name caddy --interpreter none -- run --config...). Diagnosis: backends online in pm2 list but curl http://127.0.0.1:8080/health is refused/000 ⇒ Caddy itself is down.
  15. Starting a Hermes dashboard with --skip-build on a fresh host that never built the frontend. The process crash-loops errored with ✗ --skip-build was passed but no web dist found at: …/hermes_cli/web_dist. Build it once: cd ~/.hermes/hermes-agent/web && npm install && npm run build, then restart. Pre-flight: ls ~/.hermes/hermes-agent/hermes_cli/web_dist/index.html || echo "build web_dist first".
  16. Reaching a Hermes dashboard over a plain-HTTP tailnet door. The auth cookie is Secure, so it's never resent over http:// — login returns 303 but the authed GET stays 302 (looks like a broken gate). Serve over HTTPS: a 443 funnel if Caddy owns it, else a tailnet-only HTTPS door tailscale serve --bg --https=8443 http://127.0.0.1:8080. Always run the login round-trip over the https:// URL — an HTTP test gives a false negative.
  17. Pasting generic card markup into a host's custom index page. Index pages differ per host: an inline Caddyfile respond block (Mac Studio) vs a file_server _registry/index.html with a bespoke theme. READ the existing index first and reuse ITS card/pill classes; insert the new card INSIDE the existing grid/container, never after </main> (an orphan outside the styled wrapper renders as raw, unstyled floating text — a user-visible "looks funny" defect). Back up before editing; verify with a cache-busting browser reload.

Verification checklist

  • App appears in pm2 list as online
  • Backend port responds on loopback: curl -sI http://127.0.0.1:<port>/
  • Caddy reloaded with no parse errors: pm2 logs caddy --lines 30
  • Front door returns the right code via Tailscale URL (200 / 302 / 405)
  • If gated: login round-trip succeeds, authed GET returns 200
  • If Hermes dashboard: session count > 0 on the pinned DB
  • tailscale serve status matches tailscale-serve.json
  • pm2 save after any ecosystem change (with PM2_HOME exported)
  • If exposing publicly: only password/token-gated upstreams on funnel'd ports

Exposing the router on a real custom domain (Cloudflare Tunnel)

Tailscale Funnel only serves a TLS cert for the host's *.ts.net name, so a plain CNAME yourname.example.com → host.tailnet.ts.net gives a cert-name mismatch (browser sends SNI for your domain, funnel presents a ts.net cert → TLS fails). To put a real domain on the router, use a Cloudflare Tunnel into the same Caddy loopback. It's the easy path when the zone is already on Cloudflare DNS: Cloudflare terminates HTTPS (auto-issues + renews the cert) and cloudflared tunnels outbound to 127.0.0.1:8080 — no public IP, no port-forward, no inbound firewall ports.

Shortcut: if ~/.cloudflared/cert.pem already exists (a prior cloudflared tunnel login), the whole thing is headless — no dashboard token, no browser click needed. The CLI path becomes the lower-effort route. Check first: ls ~/.cloudflared/cert.pem.

# 0. one-time only if no cert.pem yet (opens browser to pick the zone):
cloudflared tunnel login

# 1. create the tunnel (writes creds json next to cert.pem)
cloudflared tunnel create <name>          # e.g. miniapps  → prints the tunnel-id

# 2. write ~/.cloudflared/config.yml
cat > ~/.cloudflared/config.yml <<'YML'
tunnel: <tunnel-id>
credentials-file: /Users/<user>/.cloudflared/<tunnel-id>.json
ingress:
  - hostname: sub.example.com
    service: http://127.0.0.1:8080
  - service: http_status:404
YML

# 3. create the proxied DNS record automatically + validate
cloudflared tunnel route dns <name> sub.example.com
cloudflared tunnel ingress validate

# 4. supervise under PM2 (reboot-safe, same as Caddy) and save
# Resolve the binary from PATH rather than hardcoding the Apple-Silicon Homebrew
# prefix — on Linux and Intel macOS cloudflared lives elsewhere (/usr/local/bin,
# /usr/bin), and PM2 needs an absolute path, so a hardcoded prefix fails even when
# cloudflared is installed and on PATH.
CLOUDFLARED=$(command -v cloudflared || echo /opt/homebrew/bin/cloudflared)
pm2 start "$CLOUDFLARED" --name cloudflared --interpreter none -- \
  tunnel --config /Users/<user>/.cloudflared/config.yml run <name>
pm2 save

Verify with a real external round-trip (through Cloudflare, not loopback): curl -sS -o /dev/null -w "HTTP %{http_code}\n" https://sub.example.com/health → 200, and curl -sv https://sub.example.com/health 2>&1 | grep -iE "subject:|issuer:" should show CN=<your-domain> issued by Let's Encrypt (Cloudflare). PM2 logs should show 4 Registered tunnel connection lines (redundant QUIC edges).

Pitfalls:

  • The custom domain fronts the same Caddy as the funnel — so /, /hooks/*, and every app path become publicly reachable on the pretty URL. Gate anything sensitive (auth sidecar) or scope the Caddy routes; don't assume the domain is private.
  • Keep the Tailscale Funnel — it's an independent door to the same Caddy, not a conflict. Both proxy into 127.0.0.1:8080; no route duplication.
  • cloudflared is outbound-only; a stale version just logs an update warning (harmless) — brew upgrade cloudflared when convenient.
  • Rollback is clean and touches nothing else: pm2 delete cloudflared && pm2 save, delete the DNS record in the CF UI, cloudflared tunnel delete <name>, rm the config + creds json.

Reference

  • Source of truth for the stack: ~/src/hermes-config/devops/app-router/ (README.md, templates/, scripts/, auth-service/)
  • references/hermes-dashboard-rollout.md — standing up a Hermes dashboard behind the router: slug choice, root-vs-profile DB, door type, and index-page markup. Keep your own per-machine map (slug, port, profile) alongside your deployment; it is site-specific and deliberately not shipped here.
  • Tailscale Funnel docs: https://tailscale.com/kb/1223/funnel
  • Caddy forward_auth docs: https://caddyserver.com/docs/caddyfile/directives/forward_auth

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.