Add typed between-turn guidance to delegated agent runs #74

Outcome

Add a typed policy layer between agent turns that recommends the next action, reason, and confidence while deterministic guardrails retain authority.

Current behavior

The orchestrator can continue or fan out work, but it has no explicit between-turn contract for rate limits, new evidence, verification gaps, executor changes, or a correct stop. A generic continue loop can waste capacity or conceal why a run changed direction.

Source: docs/episode-triage.md, episodes 206 and 245.

Contract

• Produce a typed recommendation such as continue, steer, wait, change admitted executor, request input, verify, checkpoint, or stop.
• Include reason, confidence, evidence references, and the policy version.
• Keep the output advisory. Deterministic authority, budget, consent, and safety guardrails can refuse or replace it.
• Record requested, recommended, effective, and refused actions separately.
• Evaluate guidance against accepted outcomes, not token volume or uninterrupted runtime.
• Support a no-action result when waiting or stopping is correct.
• Keep model and executor substitution visible.

Acceptance criteria

• Each delegated turn records the recommendation and the effective guarded action.
• A low-confidence or unsupported recommendation cannot widen authority or spend.
• Rate-limit, verifier-failure, user-input, and completion scenarios select distinct typed outcomes.
• Operators can explain why a run continued, changed, waited, or stopped without exposing private reasoning text.
• Offline evaluation compares policy versions against a fixed accepted-outcome corpus.

Verification

Test every action, guardrail override, confidence boundary, stale evidence, privacy projection, and replay. Run an offline evaluation and the relevant release gates.

Dependencies

Depends on #41 and the ACP executor-contract issue in project 7.