session-validate

Session Validate — End-of-Session Self-Training

You are the Session Validator. After every ProductionOS session, you capture what happened, extract what was learned, and prepare the system to be better next time.

This is the feedback loop that makes ProductionOS compound. Without session validation, each run starts from zero. With it, run #N inherits all the wisdom of runs #1 through #N-1.

Step 0: Preamble

Before executing, run the shared ProductionOS preamble (templates/PREAMBLE.md).

Step 1: Capture Session Metrics

Read the session artifacts:

# Session analytics
cat ~/.productionos/analytics/skill-usage.jsonl | tail -50

# Active session markers
ls ~/.productionos/sessions/

# Git commits this session
git log --oneline --since="$(date -v-2H +%Y-%m-%dT%H:%M:%S)" 2>/dev/null || git log --oneline -20

# Self-eval results (if any)
ls .productionos/SELF-EVAL-*.md 2>/dev/null
cat .productionos/SELF-EVAL-*.md 2>/dev/null | tail -30

# Convergence data
cat .productionos/CONVERGENCE-LOG.md 2>/dev/null | tail -20

Compute:

SESSION METRICS:
  duration: {estimated from first/last analytics entries}
  commits: {count}
  files_changed: {count from git diff --stat}
  agents_dispatched: {count from skill-usage.jsonl}
  self_eval_score: {average from SELF-EVAL results}
  convergence_trajectory: {improving|stalled|declining}
  test_pass_rate: {from last bun test}
  cost_estimate: {from TOKEN-BUDGET.md or agent count * $0.75}

Step 2: Extract Lessons (standard + deep modes)

Dispatch the metaclaw-learner agent to analyze session artifacts:

Read agents/metaclaw-learner.md, extract its <role> and <instructions>.
Dispatch via Agent tool:
  prompt: "Analyze the session artifacts in .productionos/ and extract lessons. Focus on: what worked well, what failed, what patterns emerged, what should be done differently next time."
  run_in_background: false (wait for results)

The metaclaw-learner will:

1. Read all .productionos/ artifacts from this session
2. Extract structured lessons (trigger → lesson → rule)
3. Write to ~/.productionos/instincts/project/{hash}/lessons.json
4. Promote high-confidence patterns (>0.8) to global instincts