train-convo-steering

name	train-convo-steering
description	Voice-first runtime steering + nightly deep analysis to learn per-user conversation priors.
triggers	["train convo steering","runtime steering step","train steering nightly"]
provides	["train-convo-steering"]
composes	[null,"task-monitor"]

train-convo-steering (Skill v2)

Voice-first runtime steering + nightly deep analysis to learn per-user conversation steering priors.

This skill is designed for:

• Live voice: per-turn inference must be fast and bounded (preset selection, not multi-candidate judging).
• Nightly deep learning: optional DeepSeek V3 (TEE) judge calls to improve labels and update per-user priors.

Concepts

Collaboration State

A compact state bucket per turn (tempo, trust, alignment, affect, control) mapped to {low, mid, high}.

Steering Presets

Configuration of response knobs (length, questions, initiative, certainty, grounding) and voice prosody.

• fast_proceed
• clarify_once
• trust_repair
• deep_dive
• exec_summary_plus_steps
• socratic

Priors

Per-user policy map state key -> best preset learned from reinforcement signals (user feedback, latency, DeepSeek judge).

Commands

Runtime (voice-first)

./run.sh runtime-step \\
  --user-id <USER_ID> \\
  --session-id <SESSION_ID> \\
  --channel <text|voice> \\
  --user-text "..."

Emits JSON with the selected preset and decision details.

Nightly

./run.sh nightly \\
  --logs ./_out/live_logs.jsonl \\
  --out ./_out

Processes logs, runs DeepSeek judge (if configured), and updates priors.

Memory Integration (memory_integration.py)

Cross-session memory persistence via common.memory_client with taxonomy bridge tagging. Priors MUST persist to memory or they are useless across sessions.

Pre-hook: recall_user_priors(user_id, k=10)

Recalls per-user conversation priors from memory for cross-session persistence of learned preferences.

Post-hook: learn_conversation_prior(user_id, prior_type, value, confidence, context)

Learns steering decisions with confidence >= 0.6 during runtime-step.

Post-hook: learn_nightly_summary(user_id, global_best_preset, training_rows, ...)

Learns nightly training results including state policies with confidence >= 0.5.

Bridge Keywords

Bridge	Keywords
Precision	preference, explicit, specific, configured
Resilience	consistent, stable, reliable, proven
Fragility	conflicting, unclear, ambiguous, volatile
Loyalty	trust, rapport, relationship, familiar
Stealth	implicit, inferred, unspoken, behavioral

Tags: ["convo_steering", user_id] + bridges

Configuration

Set environment variables in .env (or project root):

• DEEPSEEK_API_BASE: Chutes API or gateway URL.
• DEEPSEEK_API_KEY: API Key.
• DEEPSEEK_MODEL: Model name (default deepseek-v3).
• DEEPSEEK_JUDGE_ENABLED: Set to 1 or true to enable.