fleet-router

Part of 图灵星球 Agent 军团. New here? Start at the overview: https://github.com/turingplanet/agent-legion

The fleet's edge router. One wildcard custom domain — *.agents.turingplanet.ai — is attached to this single service, and Caddy forwards each hostname to the matching member service over Railway's private network (<slug>.railway.internal). Unknown subdomains get a 404 pointing at the docs.

That's why platform hosting needs zero DNS work per agent: the wildcard (plus a one-time ACME-delegation record for wildcard TLS) was configured once, and every new agent is just another line in a routing table.

Where this fits

flowchart LR
    REG["agent-registry<br/>deployments.yaml — the allowlist"] -->|"deploy workflow generates<br/>+ pushes the Caddyfile"| THIS["fleet-router (this repo)<br/>Caddyfile = routing table"]
    THIS -->|"routes by hostname over<br/>Railway private networking"| M1["member service<br/>hello-fleet.railway.internal"]
    THIS --> M2["member service<br/>…other platform-hosted agents"]

Loading

• agent-registry owns deployments.yaml (which agents are platform-hosted). Its deploy-fleet workflow regenerates the Caddyfile here and pushes it, then triggers a rebuild pinned to that commit.
• Member services are deployed by that same workflow into the platform's Railway project; this router is the only thing publicly exposed.
• fleet-services is the platform's other running service (AI review, registrar) — unrelated to routing, but served through here like any member.

⚠️ The Caddyfile is generated

Don't hand-edit it. The source of truth is deployments.yaml in agent-registry — edit there, and the workflow rewrites this file. A manual edit is overwritten on the next deploy run.

Ops notes

• Member services must bind IPv6 (HOST=::) or private networking can't reach them — the template's config.py handles this via the HOST env var.
• Railway builds this from the Dockerfile (caddy:2-alpine + the Caddyfile); PORT is injected at runtime.
• Routing changes take effect only after a rebuild — the deploy workflow triggers one explicitly, because the Caddyfile is baked into the image at build time.