delegation-scout

A Hermes delegation backend plugin (delegation.backend: scout) that routes every
delegate_task child run to the right model/profile — using deterministic task-shape
heuristics plus recorded past outcomes in its own durable SQLite DB. Over sessions,
routing accuracy improves because the plugin observes which route actually succeeded
for which kind of task.

Install

hermes plugins install ./hermes-delegation-scout   # or: link/copy into ~/.hermes/plugins/
hermes plugins enable delegation-scout

Requirements: Hermes >=0.22.0. No Python dependencies beyond stdlib; no network calls
on any path (routes resolve credentials through core).

Configure

# ~/.hermes/config.yaml
delegation:
  backend: scout
  backends:
    scout:
      routes:
        quick: {provider: openrouter, model: google/gemini-3-flash-preview}
        deep:  {provider: nous, model: moonshotai/kimi-k3}

mode: "quick" | "deep" | "auto" appears on every delegate_task task. Explicit
mode is honoured first; an unknown value is refused (the tool error tells the model
the allowed values).
auto (or omitted) scores the task: deep-work keywords (review, analyze,
refactor, debug, ...), quick-work keywords (lookup, summarize, ...),
attached images, and text length. A class with ≥5 recorded samples whose success rate
beats the other candidate by ≥15 points nudges the final pick.
A class missing from routes inherits the delegation section (label default).
If routes is empty the backend reports unavailable: delegation fails closed with
the fix message instead of silently running a wrong route.

How it learns

route_task writes each decision to the plugin DB (delegation.backends.scout stays
config-only): routes(task_hash, label, provider, model, ts) with a deterministic
sha1(goal)[:16] hash.
A public post_tool_call observer hook records finished delegate_task results
(route.backend == "scout", terminal statuses only) into outcomes(route_id, status, error, ts). See docs/outcome-seam.md for the seam evidence; no core patching.
scout_stats (the plugin's one tool) returns per-route counts, success rates, and
the last-N decisions as pretty-printed JSON.

Determinism: task_parameters/description are byte-stable per config (prompt
caching). History lookups happen only inside route_task.

Non-goals (v1)

LLM-based classification inside route_task (nondeterministic / prompt-cache hostile),
cross-user learning, UI.

Development

python -m pytest -q        # unit + e2e (temp HERMES_HOME; never the real profile)
hermes plugins doctor ./hermes-delegation-scout
hermes plugins validate ./hermes-delegation-scout --install-deps

README from the reviewed commit e0029ed; it updates when the author re-pins.

delegation-scout

About