reasoning_streams

reasoning_streams

A primary agent emits its reasoning as a kind: thought stream. A mirror peer runtime subscribes, runs a smaller critic model on each thought, and delegates arcpx.mirror.critique.v1 events back into the primary's session. Bounded by a token budget — when spent, the mirror unsubscribes and the primary continues without critique.

Before ARCP

Self-critique loops drift: there's no clean way to (a) expose just the reasoning to a separate process for a second opinion, (b) cap the rounds, (c) gracefully step back when the critic budget is spent. Sidecar a "second LLM call" before each step and the indirection burns tokens whether you need it or not.

With ARCP

// mirror peer subscribes to the primary's thought stream...
await client.send(buildEnvelope({
  id: newMessageId(), type: "subscribe", timestamp: nowTimestamp(),
  payload: { filter: { session_id: [targetSessionId], types: ["stream.chunk"] } },
}));

// ...and delegates critiques back as namespaced events.
await client.send(buildEnvelope({
  id: newMessageId(), type: "agent.delegate", timestamp: nowTimestamp(),
  payload: { target: targetSessionId, task: "consume_critique", context: { critique } },
}));

The mirror is a peer runtime (agent_handoff: true, subscriptions: true, trust_level: trusted), not an Observer — it both reads and writes back into the primary's session.

ARCP primitives

• kind: thought reasoning streams — RFC §11.4.
• Subscriptions with type filter — §13.2.
• Custom extension event under arcpx.<domain>.<name>.v<n> — §21.1.
• agent.delegate for cross-runtime delivery — §14.
• tokens.used budget — §17.3.1.

File tour

• main.ts — boots primary + mirror, wires the critique queue. runPrimary emits thoughts; runMirror subscribes, critiques, delegates back.
• agents.ts — primaryStep and critiqueThought stubs.

Variations

• Multiple mirrors (security / factuality / style) subscribed in parallel; primary merges critiques by severity.
• Persist critiques to the SQLite sink in subscriptions for drift analysis.
• Replace the critic LLM with a deterministic verifier (linter, type checker) that returns severity from a rule set.