hermes-team-dashboard
Skill0 starsChanged 20 days agoMITView on GitHub
Use when building a live Hermes worker status panel.
In this repository86 other files it gives its agents.
- ai-assistant-editor-setupSkill
- api-and-interface-designSkill
- banner-designSkill
- brainstormingSkill
- caveman-exploreSkill
- caveman-reviewSkill
- chat-silence-diagnosisSkill
- ci-cd-and-automationSkill
- code-auditSkill
- codebase-designSkill
- koordinator-team-dispatchSkill
- constraint-driven-developmentSkill
- context-engineeringSkill
- cpanel-shared-hosting-deploySkill
- cross-profile-cron-opsSkill
- deprecation-and-migrationSkill
- documentation-and-adrsSkill
- domain-modelingSkill
- doubt-driven-developmentSkill
- finishing-a-development-branchSkill
- frontend-designSkill
- git-workflow-and-versioningSkill
- graphic-design-laneSkill
- hermes-gateway-web-uiSkill
- hermes-profile-hygieneSkill
- hermes-profile-provisioningSkill
- hermes-service-persistenceSkill
- hermes-stream-stale-diagnosisSkill
- idea-refineSkill
- incremental-implementationSkill
- indonesian-official-lettersSkill
- internal-commsSkill
- interview-meSkill
- investigate-firstSkill
- kanban-rate-limit-resilienceSkill
- karpathy-guidelinesSkill
- large-project-gated-deliverySkill
- lean-buildSkill
- live-agent-office-operationsSkill
- live-react-runtime-qaSkill
- live-web-release-verificationSkill
- markdown-pdf-pipelineSkill
- marketing-strategySkill
- meta-social-publishingSkill
- meta-threads-oauth-lifecycleSkill
- migrationSkill
- multi-profile-project-orchestrationSkill
- network-operationsSkill
- observability-and-instrumentationSkill
- office-skill-distributionSkill
- organization-site-menu-researchSkill
- performance-optimizationSkill
- php-upload-size-limitsSkill
- planning-and-task-breakdownSkill
- ponytail-auditSkill
- ponytail-debtSkill
- ponytail-reviewSkill
- ponytailSkill
- portable-app-handoffSkill
- receiving-code-reviewSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 3 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 17
Use when the user asks to see who on the team is working and what they are doing — a web panel, status board, "kantor virtual", or …
- Distinctive: line 67
When the user points at a repo whose look they like ("saya suka UI/UX ini, berikan untuk kantor kita"):
- Distinctive: line 90
Before saying a visual change is live, verify the rendered result, not just the response: hard-reload a browser holding the old build, …
- Very specific: line 152
Bound service restarts inside foreground cells: Run systemctl --user restart X under a short timeout followed by is-active in the same call.
- Very specific: line 159
Verify all three layers each fix: API JSON (active_profiles/feed populated when work is happening), frontend counts (call load() manually …
≈ “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.
Files
Hermes team status panel
When to Use
Use when the user asks to see who on the team is working and what they are doing — a web panel, status board, "kantor virtual", or monitoring view of Hermes profiles and kanban activity. Also when fixing an existing panel whose status/activity display has stopped updating.
Build a read-only web panel over Hermes internal data (worker roster, live activity, recent events, optional 3D/denah view). Never expose write endpoints, command execution, or anything that can touch a worker or the router.
Procedure
- Collect data (read-only), in this order:
- Workers: glob
~/.hermes/profiles/*/profile.yaml→display_name, model;hermes profile listfor gateway/model state. - Cards:
hermes kanban list --json— parse withjson.loads(s, strict=False)(card bodies contain raw control chars), or sqlite~/.hermes/kanban/boards/*/kanban.db(tablestasks,task_runs,task_events,task_comments). - Live activity (the load-bearing source):
task_eventsrows (kind=heartbeat|claimed|completed|commented,created_atepoch) joined totask_runs.profile, PLUS each profile'sstate.db→sessions.last_activity_at. Rule: heartbeat ≤5 min or session activity ≤10 min ⇒ "sedang bekerja". Return them as SEPARATE counts (kanban_busy,session_busy,total_busy) and tag each desk with its source — never one unlabeled number (see Pitfalls).
- Workers: glob
- Serve: single-file stdlib server (
ThreadingHTTPServer), token in?token=enforced on/and/api/*(401 without); inject the token into the HTML as a__TOKEN__replacement so the frontend fetch authenticates. Cache payload ~4 s, frontend refresh ~8 s. - Frontend: one
index.html(inline CSS/JS) with stat cards, per-worker desks (status dot + current task), event feed. Optional 3D view as a lazy ES module so WebGL only exists while its tab is open. - Verify BEFORE reporting done:
curlAPI with token → 200 and JSON showsactive_profiles/feedpopulated when someone actually works; without token → 401.node --checkevery JS file, then in the browser call the render/load function via JS and assert the counts — a parse error in an inline script leaves the page silently empty while curl still returns 200.systemctl --user is-active <unit>.
- Persist: user unit in
~/.config/systemd/user/<name>.service(WantedBy=default.target,Restart=always), thensystemctl --user daemon-reload && enable && start.
Adopting an existing upstream panel instead of building
For frontends that speak the Hermes JSON-RPC gateway directly (Studio/
Hermes3D-class), see references/hermes-native-frontend-gateway.md for the
verified backend-start recipe (hermes serve --isolated), token pinning, and
the 403-vs-101 WS probe.
When the user points at a repo whose look they like ("saya suka UI/UX ini, berikan untuk kantor kita"):
- License first: clone to scratch, read LICENSE — MIT/Apache lets you deploy and modify; refuse copyleft silently. Prefer zero-dependency stacks (stdlib Python + local JS) so nothing needs installing.
- Read the data contract before running it: how agents are fed (file
poll / SSE webhook / API), where its bridge assumes
HERMES_HOMEand the kanban boards live, how many agent slots andhomepositions the frontend maps support. - Add the auth gate BEFORE it ever listens publicly — upstream
dashboards ship with none. Token via env var; read the secret from
token.txtin a tiny launcher script so no unit file or repo file ever contains it; enforce on/, the office page AND/api/*. - Test on a side port, then promote to the production port so the
user's saved link and token stay identical. Move the old panel aside as
<dir>-v1-archive(rollback target), never delete it; keep two systemd units (server + data emitter) bothenabled. - Verify in-browser before reporting: page loads with token / 401
without, agent count == roster size,
workingcount == live cards, then screenshot — numbers equal to the kanban, not "it rendered". - Replacing a UI framework, preserve the user's link and secure data path. Stage the candidate in a separate directory/port; inspect LICENSE, install dependencies, identify auth and backend path, adapt the emitter to the actual board DB/profile roster, then test before cutover. If the old panel is token-gated, the replacement must stay gated before any public bind. Use a dedicated gateway session token (never overwrite
HERMES_DASHBOARD_SESSION_TOKEN), bindhermes serveto loopback, and publish only over the user's authenticated/restricted path. Cut over the same public port/token only after browser+API verification; archive old files for rollback, then remove old units/processes after the new service is proven. - Live backend data must outrank all demo motion. In
json_file/API mode, animation may move avatars between poses but must never randomize agent status, task text, meeting counts, or chat bubbles. Update every periodic callback, not just the initial data loader; sample status counters repeatedly over 10+ seconds and ensure they remain identical to the API payload. - Before saying a visual change is live, verify the rendered result, not just the response: hard-reload a browser holding the old build, inspect console errors, and assert roster count + status counts + notification text match the API payload. A 200 alone is not proof the new UI loaded.
Pitfalls
- A live payload can still be overruled by the demo behavior loop — the panel looks "error" while every server check is green. Replacing the agent array is only half the fix: upstream dashboards run a periodic simulation (
updateBehavior()-style timer) that re-randomizes everystate.status(meeting/working/walking), overwrites the real task string with canned phrases, spawns random chat bubbles that bury the floor, and writes hardcoded notification text ("Gathering in meeting room… x/5 attending"). Symptom the owner reports as "terlihat seperti error": counters jumping (4 working · 2 meeting · 4 walking against 2 real cards) and bubble spam. Gate the ENTIRE simulation on the data mode: live mode may only move positions (status/task come from the payload untouched), and the bubble spawner + every notification string/title get a live variant that states the truth ("N bekerja · M standby — data Kanban live"). Verify by sampling the counters 3× over ~10 s: they must be identical. - Patch one canonical file, sync it, then checksum before restarting. Copying a stale source over the live directory silently reverts a fix that just landed (copy-back killed a notification patch and restored the old counters in one move). Pick the deploy path as canonical, edit it, push to any mirror, and compare md5 of both copies as the last step before
systemctl --user restart. - Serve HTML/JS with
Cache-Control: no-storeor the fix never reaches the owner. Browsers keep the previousoffice.htmlafter a server-side fix, so the owner still screenshots the old bug. Sendno-store, must-revalidatefor.html/.js/.css(and/) in the response override, and tell the owner to hard-refresh once (Ctrl+F5). - A broken-panel report arrives as a video: diagnose the browser state, don't argue from server logs. Service
active, HTTP 200,NRestarts=0and an empty JS error array can all be true while the page is wrong. Pull the report (Drive link when Telegram rejects >20 MB), extract frames withffmpeg -vf fps=1/3, vision-analyze them, then sample the live page via JS (agentStatescounts, notification text,bubbles.length) and diff against the API payload — that separates a browser-side demo-state problem from a data problem. - Answering
/?token=with a JSON short-body breaks the page (the handler returns{"ok":true}instead of serving HTML). Plant the cookie inside an overriddensend_response()so the normal static-serving path keeps running; then verify HTML is actually returned and that a later fetch (no query string) authorizes via the cookie. - Live data must REPLACE, not merge. Frontends that push incoming
agents into a built-in demo array keep the demo avatars forever (16 shown
vs 11 roster). Patch the apply function to delete ids absent from the
payload — from
AGENTS,agentStatesandagentStats— before applying. - From a profile session
HERMES_HOMEpoints at the profile dir, so a board glob under it finds nothing and everyone reports Standby. Glob the global home too (~/.hermes/kanban/boards/*/kanban.db), dedupe, and prove with one worker showingworkingwhile its card isrunning. - New roster members need a home position:
getHomePositionstyle maps default to center, stacking every unmapped avatar at (0,0). Extend the map for every roster id and keep a round-robin fallback. - Emitter pushes on change only — diff the payload, POST the webhook when it changed, fall back to writing the served JSON file when the webhook is down; a 10 s blind push loop is wasted traffic.
- Card status alone makes the panel lie. Once every card is
donewhile workers keep chatting/running cron, a status field derived from cards freezes at 0. Activity must come from heartbeats +sessions.last_activity_at— Chat and cron runs emit no card-status change. If showing chat/session activity too, include it explicitly as a separate source and separate count; never silently include the coordinator's own current conversation in a number labeled as workers working. Treatsessions.last_activity_atas a short-lived active-session signal (not proof of task progress) and label it accordingly. - User units have no shell PATH. Without
Environment=PATH=~/.local/bin:...in the unit, asubprocess.run(["hermes", ...])fails silently and the payload degrades tomodel: "?"/gateway: "unknown". Add PATH, then re-check the API JSON. - Bare ES-module specifiers need an import map (
import ... from 'three'dies with "Failed to resolve module specifier" unless the served HTML has<script type="importmap">). - Restart the unit after editing server-side Python — files served from disk (HTML/JS) hot-reload, Python does not.
- Get exact
old_stringfrom the file before patching a minified one-line HTML (grep it out, assert it is present); a guessed match silently does nothing while you believe the edit landed. - Serve token-gated pages over LAN IP only after binding
0.0.0.0and confirm the public URL responds — user opens the panel from other devices. - Never collapse session and card activity into one unlabeled "busy"
number. Expose
kanban_busy/session_busy/total_busy, tag each desk with its source (sesi aktifvs kartu), and state the rule in the stat subtext ("kerja kartu ≤5 mnt · sesi ≤10 mnt"). Counting chat sessions silently made the board claim "2 sedang bekerja" while zero cards ran (the coordinator's own chat counted); deleting sessions then showed 0 while the user was actively working — either way the number was unexplainable. - Prove an animation/3D loop with counters, not pixels.
canvas.toDataURL()reads a cleared drawing buffer and reports "static" while rendering is healthy, and full-page screenshot hashes can match too. Instrument the frame function withwindow.__frame/window.__renderedcounters, sleep, assert they advanced andwindow.__frameErris 0; screenshots only confirm composition, not motion. - Bound service restarts inside foreground cells. Run
systemctl --user restart Xunder a short timeout followed byis-activein the same call; a restart cell that hangs past the tool timeout kills the kernel state mid-turn and leaves you unable to reply — after any such timeout, re-check the unit state first thing, before reporting anything else. - Prove panel "live-ness" against the label, not just against emptiness.
A connected panel whose numbers don't move is still broken to the user.
Verify all three layers each fix: API JSON (
active_profiles/feedpopulated when work is happening), frontend counts (callload()manually via JS and assertstatBusy/feedchildren render), and desk classes (.desk.busymatchesactive_profiles) — a missing frontend patch shows correct API numbers while desks stay grey, and the user reports "tidak berjalan real time". - Feed "terbaru" must show relative time. A feed built from card history
alone reads as stale-but-latest during idle hours; emit
rel("7 jam lalu"/"baru saja") next to the timestamp so the user can tell idle from broken without asking.
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
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.

