Apply SecondOrder AI meta-cognition patterns to Ouroboros

[kinwo]

Background

SecondOrder AI advocates exposing curated meta-cognition signals (confidence, limitations, context gaps) instead of raw chain-of-thought, and gating complex work through a planner/critic loop. Several of these patterns map cleanly onto Ouroboros's existing architecture and risk surfaces.

This issue tracks the work to adopt them. Each item below could become its own sub-issue.

Proposed adoptions

1. Task-classification gate before the ReAct loop

Where: packages/cli/src/ agent loop.
What: Add a lightweight classifier in front of ReAct. Simple Q&A bypasses tool-use; ambiguous, multi-step, or risky tasks enter the full plan→act→observe loop.
Why: Today every prompt pays full ReAct overhead. SecondOrder's framing: "answer directly when easy, switch to planning loop when ambiguous."

2. Planner + Critic pass for self-modification

Where: the 5-tier permission model for RSI.
What: Before any tier-3+ self-modification, run a structured pass that emits {goal, constraints, plan, confidence, limitations} and a critic review. Required permission tier scales with reported confidence — low confidence forces human approval even for normally auto-approved tiers.
Why: Directly hardens RSI, which is Ouroboros's riskiest surface and core pitch.

3. Trust signals as a first-class IPC notification

Where: packages/desktop/src/shared/protocol.ts and the renderer.
What: Add a meta notification type carrying {confidenceLabel, limitations[], contextGaps[]} separate from streamingText. Desktop renders a compact strip — expandable on click.
Why: Pairs with the new mid-flight steering UI — users decide whether to steer based on reported confidence rather than gut feel. Keeps presentation/agent separation clean.

4. Compact plan preview, expand on demand

Where: desktop UI, alongside the steering row.
What: When the planner pass runs, emit a one-line plan summary as a notification; full plan available on expand.
Why: Avoids dumping internal reasoning into the transcript while still letting users inspect it.

5. Event instrumentation feeding the evolution log

Where: memory's evolution-log.
What: Track classification decisions (was this task correctly routed?), confidence-vs-outcome (did high-confidence answers actually succeed?), and user steering events (where did the agent need correction?).
Why: Gives RSI real signal to learn from instead of just transcripts. SecondOrder's "event instrumentation tracking user interactions and outcomes" maps directly onto evolution-log's purpose.

6. Critic gate on memory writes

Where: durable / checkpoint memory writes.
What: Run a critic pass before persisting — does this contradict existing entries, is it sourced, does it have a stated confidence?
Why: Prevents memory rot in long-running agents. Fits the existing memory-tier model.

Suggested first slice

1. #3 — trust-signal IPC — mechanical, ships visible value alongside steering, unlocks the rest.
2. Test custom skill #2 — planner/critic on self-modification — directly addresses the highest-risk path.

Source

Patterns synthesized from https://secondorderai.kinwo.net/ — meta-cognition for LLM systems, planner+critic workflow, and trust-signal UX.