Skip to main content

The fleet registry gets a summary line, not a full dump

· 2 min read
Carefree Investments LLC
Anchor maintainers

Anchor already keeps executor task specs lean — never the whole project, never a raw file dump, only the context a task actually needs. The fleet registry itself was the exception: the model-fleet MCP's list_fleet handed back every endpoint's base_url and model name whether or not the caller needed it, and there was no other way to get the same picture into a prompt. It's on the same diet as everything else now.

What was pasted before​

scripts/endpoints.yaml is internal infrastructure — LAN addresses, model identifiers, per-endpoint quirks. None of it is secret (ANCHOR_API_KEY was always an environment read, never a registry field), but none of it belongs in a context window that only needs to know what's available, not where it lives. list_fleet didn't distinguish: one call, full detail, every time.

A generated summary, not a hand-maintained one​

scripts/router.py now has summarize_endpoints(fleet): one capped line per endpoint — name, tier, context size, a one-phrase capability derived from the endpoint's own quirks (hybrid-reasoning (nemotron), frontier-class, local/detached, …) — never a base_url, model name, or raw quirk value. It's generated from the registry every time, so it can't drift the way a hand-maintained summary would.

h100-nemotron · reasoner · ctx=65536 · hybrid-reasoning (nemotron)

fleet_summary_block(fleet) wraps that for splicing straight into a prompt.

Full detail, only on request​

list_fleet now returns the summary. A new lookup_endpoint(name) tool returns full non-secret detail — base_url, model, quirks — for one endpoint, by name, when something actually needs it:

lookup_endpoint("h100-nemotron")
→ h100-nemotron [reasoner] nvidia/llama-3.3-nemotron-super-49b-v1 @ http://10.0.1.11:8000/v1
  quirks: think_toggle=nemotron, temperature=0

Same split, same helper, on the Python side: router.endpoint_detail(fleet, name).

Where the summary actually lands​

Not every context gets it. orchestrate.py's planner phase always does — picking Route to targets needs fleet awareness. prompt_tuner.py only adds it to a generated task spec's ## Provided context when the rough task is itself about routing — mentions an endpoint, a tier, escalation, dispatch. An ordinary "fix the login bug" spec never sees it; a "which endpoint should handle this escalation?" one does.

Using it​

Nothing to configure — list_fleet and orchestrate.py's planner phase already use the summary. Reach for lookup_endpoint(name) (or router.endpoint_detail(fleet, name) from Python) when you need the detail list_fleet used to hand you unasked.