Skip to content

Guide

Reference

Concepts

skill_router

Route a single user (or peer-agent) utterance to an appropriate skill, agent, or direct reply. Used by reyn chat on every turn.

Phases

The router runs as a two-phase workflow:

  1. classify — pick the intent (chitchat / memory_recall / stable_knowledge / clarification / task / fresh_lookup) and either finish with routing_decision or hand off to match.
  2. match — for task and fresh_lookup intents, dispatch to a specific skill (or peer agent), or transition to web_research.

The classify phase also handles memory writes: every turn it inspects the new utterance and may emit file/write ops to persist user / feedback / project / reference facts. Memory reads are pre-merged by ChatSession (see concepts/memory).

Entry artifact: chat_routing_request

ChatSession constructs this on every turn. Selected fields:

Field Origin Description
user_message inbox payload The latest utterance to route
chat_id session The agent's own name (used by classify when building memory paths)
history_path session Path to history.jsonl, sliced by the classify preprocessor
compaction config Head/tail sizes for history slicing
available_skills session Project + stdlib catalogue (router/compactor excluded; FP-0011 removed skill_narrator), filtered by profile.allowed_skills if set
available_agents registry Other agents reachable via topology rules — [{name, role}, ...]
memory_index session Pre-merged shared + agent layers ({status, content}); content is markdown with (shared) and (agent: <name>) sections

Final output: routing_decision

Field Type Purpose
reply_text string (optional) Text shown directly to the user — interim acknowledgement or final answer
skills_to_run array (optional) Project / stdlib skills to spawn this turn
messages_to_agents array (optional) Delegations to peer agents — [{to, request}, ...]

ChatSession dispatches each non-empty array. reply_text reaches the user immediately for user-initiated chains. For agent-initiated chains (a peer agent received this request), the chain enters deferred reply when messages_to_agents is non-empty: the router's reply_text is held until every delegate responds.

Skill selection guidance

The match phase prefers:

  • A specific skill with a clear routing.examples.positive match — narrow well-defined tasks
  • Agent delegation when available_agents shows a peer whose role matches the request better than any skill
  • Direct reply when neither fits cleanly and the request is small (or confidence < 0.6 triggers clarification)

Never both a skill AND an agent in the same decision — pick one branch.

Source

src/reyn/chat/session.py — implemented as a built-in system skill, not a regular skill directory

See also

On this page