Workbench — a 3-agent web-app dev team (Hermes, v1.1)

A portable multi-agent team: Spec (requirements) → Scaffold (implementation) → Smoke (QA & release gate), coordinated through a Hermes kanban board.

Trained and proven on the source install: shipped two real apps — a live board status page (after a genuine QA bounce → fix → regression re-check) and a Safeway hot-sauce catalog (27 real SKUs, QA re-verified against live safeway.com PDPs) — and v1.1's handoff itself is verified live: a probe card went Scaffold-build → complete → QA task auto-promoted via the parent edge → Smoke verdict PASS on the running app, with zero manual intervention.

What's in this pack

File	What it is
TEAM.md	The team's portable memory — roster, roles, the kanban card flow, handoff protocol, quality bar, escalation rules, cron table. Read this first.
WORK.md	The shipped-history record + user backlog notes (NOT the live board — the board is kanban).
workbench-spec.hermes.tar.gz	Scrubbed Hermes profile: SOUL (requirements charter), team memory, learned skills, redacted config.
workbench-scaffold.hermes.tar.gz	Scrubbed profile (implementation charter + the workbench-card-build skill, which encodes the "use a worst-case fixture for ordering checks" lesson learned from a real QA bounce).
workbench-smoke.hermes.tar.gz	Scrubbed profile (QA gate charter: gate-only, evidence-first, never fixes).
workbench_cron*.sh	The heartbeat scripts (standup 08:00 / QA sweep 12:00+18:00 / harvest 21:00) — v1.1 prompts drive the kanban board.

All profiles are scrubbed (no sessions, no secrets, no machine-local state; config.yaml endpoints are SET_YOUR_ENDPOINT placeholders).

How v1.1 coordinates (the kanban model)

• One kanban board, workbench (SQLite, durable, atomic claims). The Hermes dispatcher (kanban.dispatch_in_gateway) spawns the assigned profile onto ready tasks automatically — no polling, no manual wake-up.
• Two task types: Build task (Spec → Scaffold) and QA task (Scaffold → Smoke). The QA task's parents=[build task] edge is the handoff — it auto-promotes the moment the build completes. A QA task also carries the verdict (PASS = shipped / FAIL + rework task), so the whole chain — build → QA → rework → QA → shipped — is a queryable event chain on the board, not scrollback in a chat.
• DMs and the optional group chat are the live channel and the receipts; the board is the only state.

Install

# 1. import the three profiles
hermes profile import workbench-spec.hermes.tar.gz --name workbench-spec
hermes profile import workbench-scaffold.hermes.tar.gz --name workbench-scaffold
hermes profile import workbench-smoke.hermes.tar.gz --name workbench-smoke

# 2. recreate the team workspace + the board (the one new step in v1.1)
mkdir -p ~/.hermes/teams/workbench/{reports,repos,state}
cp TEAM.md WORK.md ~/.hermes/teams/workbench/
hermes kanban boards create workbench --name "Workbench team"   # idempotent: check `hermes kanban boards list` first

# 3. set each profile's model endpoint
#    edit ~/.hermes/profiles/workbench-*/config.yaml: replace
#    https://SET_YOUR_ENDPOINT/v1 with your LLM endpoint (all three profiles)

# 4. install the heartbeat (Hermes cron, no-agent script jobs)
cp workbench_cron*.sh ~/.hermes/scripts/ && chmod +x ~/.hermes/scripts/workbench_cron*
# then create four cron jobs pointing at the per-job wrappers:
#   workbench-standup      0 8 * * *   -> workbench_cron_standup.sh
#   workbench-qa-sweep-am  0 12 * * *  -> workbench_cron_qa-sweep.sh
#   workbench-qa-sweep-pm  0 18 * * *  -> workbench_cron_qa-sweep.sh
#   workbench-harvest      0 21 * * *  -> workbench_cron_harvest.sh

Note: the kanban board is machine-local SQLite — the pack ships WORK.md (shipped history + backlog notes), not the live board. On the source install the board already exists; on a fresh install, step 2 creates it.

Optional: seat the team in a Group room

Hermes Bot Mode supports group chats (2–6 bots in one shared room). Rooms are created from the desktop — there is no CLI command for them.

1. Bots tab → right-click @workbench-spec (any of the three) → Manage groups → create a group inline, or use the New Group Chat picker.
2. Seat all three members: @workbench-spec, @workbench-scaffold, @workbench-smoke (optionally add yourself).

In the room, your message triggers up to three serial rounds of member turns — @-mentioned bots respond, the rest pass when they have nothing to add; hard caps (10 messages per send, 3 rounds) prevent loops. Each member keeps a persistent Group: <name> session. The room is the live working channel; all durable knowledge stays in TEAM.md / the kanban board / skills, so the transcript is cheap to reset and nothing is lost.

Note: group membership lives in each bot's backend-synced profile metadata, so rooms follow gateways — a fresh install ships the three profiles but not your room. Create the group after installing the profiles.

Notes

• The team works anything you put on the board (or seed as a backlog note in WORK.md — standup promotes it); add cards or DM @workbench-spec.
• Cron sessions are date-stamped per run (no shared long-lived session) — that avoids the context-bloat failure mode of pinned chat sessions.
• License: MIT. Profiles may contain bundled Hermes skills under their own terms.