Routing to Departments

Mechanism, not a flow

This is the plumbing under agent work: how a task becomes a particular agent's task. If you are here to see what the system does, start at Level 3 · Flows.

A task arrives with no idea who should handle it. hermes-router reads it, names one department, and writes a decree message naming that department. The department answers with its own tools and files the result in workspace/ai/.

There is one routine for all of them — hermes-dept — because a department's identity is not code. It is a profile definition in ai/hermes/profiles/<name>/profile.yml, and that is what both the router and hermes itself read.

Why route at all

A hermes profile is its own HERMES_HOME: its own config.yaml, .env, SOUL.md and skills/. With GATEWAY_MULTIPLEX_PROFILES set, every profile is served off the one listener at /p/<name>/v1, so a caller picks a profile — and therefore a toolset and an MCP server list — by URL alone.

That matters because tool schemas are the expensive axis of a hermes prompt. Measured on this stack:

Profile	Fixed prompt	Tools
default	~69 KB (~17,000 tokens)	21
research	~15 KB	4 toolsets + firecrawl
sales	~15 KB	2 toolsets
router	~2.6 KB (~650 tokens)	0

A skill costs about 76 bytes in the always-on index; a tool schema costs about 2.4 KB. So the lever is tools, not agents — a department can hold thirty agents for the price of one tool. Routing on the tool-free profile is what makes it cheap enough to run on every message.

Adding a department

A department is one directory in ai/hermes/profiles/. Create ai/hermes/profiles/legal/profile.yml:

description: Contracts, terms, licensing questions, and compliance review.
toolsets: skills, todo
mcp: openviking

Then docker compose up -d hermes-agent. That is the whole job.

Those three keys are the whole registration, and they are read by exactly two things: ai/hermes/entrypoint.sh creates the matching hermes profile on boot, and hermes-router reads the same file — over its read-only /repo mount — to build the list it chooses from. No routine to write, no registry, no router edit, and no second copy of the description to drift.

Write description as the answer to "when should work come here?". It is the only thing the router sees, so a vague one produces vague routing.

The one thing that is not automatic: hermes-dept must be enabled in services/automation/decree/config.yml (shared routines are invisible until listed). Enable it once and every department you add afterwards works with no further change.

Provisioning the profiles

Nothing to run — ai/hermes/entrypoint.sh provisions every profile in ai/hermes/profiles/ when the container starts.

That is the only place it can happen, and the reason is worth knowing before you go looking for a migration to do it with: profiles are created by the hermes CLI writing into the container's own data volume, so the host would need docker exec and decree cannot reach them at all — it has no Docker socket, and it sees the repo read-only. Config a service owns, written by that service's entrypoint, is the repo's standing answer for exactly this shape.

Idempotent and never destructive: an existing profile keeps its config.yaml. To rebuild one after changing its definition, delete volumes/hermes_agent_data/profiles/<name> and restart hermes.

Each profile gets its own API_SERVER_KEY. Secondary profiles do not borrow the default profile's credential — a profile without one returns 401 on every request, and an unknown profile name returns 404 rather than quietly falling through to the default.

To see what a profile costs:

docker exec -e HERMES_HOME=/opt/data/profiles/research hermes-agent \
  /opt/hermes/.venv/bin/hermes prompt-size

Sending work

Drop a message in automation/inbox/ (the main daemon wholesale-mounts the repo-root automation/ directory as its whole project — services/automation/decree/ is only the image build, not the running project's files):

---
routine: hermes-router
source_file: notes/competitors.md
output_name: competitor-scan
---

Find out who already sells this, and how they position against each other.

Or skip the router and address a department directly with routine: hermes-dept and profile: research — the same routine the router would have queued, no routing call. A file processor does this by writing the outbox message itself; see File Processor.

When routing fails

hermes-router never guesses. The reply is validated against the discovered department list, and anything else — an empty answer, a gateway timeout, a hallucinated name, or an honest none — routes to route-failed, which logs what happened and stops:

FAILED TO ROUTE to 'none'
  reason     : the router found no department matching this message
  candidates : research sales
  source     : test/weather.md
Nothing was run for this message.

route-failed calls no model and runs no agent. It is also decree's default_routine, so a message that arrives with no routine at all is logged and dropped rather than guessed at.

Why not OpenCode

Departments call the hermes endpoint directly — through automation/lib/hermes.sh, like every other routine — and the stack ships no AI CLI at all. Hermes is itself an agent running its own tool loop against its own MCP servers, so OpenCode was an agent wrapping an agent, and its streaming parser rejects hermes' custom event: hermes.tool.progress SSE frames outright, which fails every turn where hermes actually uses a tool.

That is also why pointing a coding agent at the gateway is not the way to work on this stack. Change the stack with a dedicated coding agent against a frontier model, on the machine that holds the repo; the gateway gets the work it is good at, on workspace/. In a routine that is hermes_chat/hermes_gate; in a terminal it is the hermes command that code-server puts on PATH.