Reflect (Feedback Loops)
openkoi reflect is where the agent looks in the mirror. Three feedback loops operate at different timescales — daily, weekly, and deep — plus an epistemic honesty audit.
Subcommands
| Subcommand | Description |
|---|---|
today | Tight loop: today's tasks, decisions, outcomes, and self-assessment |
week | Medium loop: weekly patterns and behavioral trends |
growth | Deep loop: cognitive maturity stage and unlock progress |
honest | Epistemic audit: where was I wrong? confidence calibration |
When run without a subcommand, defaults to today.
Three Feedback Loops
| Loop | Timescale | Purpose |
|---|---|---|
Tight (today) | Daily | What happened today? Were my decisions good? |
Medium (week) | Weekly | What patterns are emerging? Am I improving? |
Deep (growth) | Ongoing | Where am I in my maturity journey? What unlocks next? |
Plus the Honesty loop that operates across all timescales — forcing the agent to confront where it was wrong.
openkoi reflect today
Shows a structured review of today's activity — every decision, its parliamentary verdict, and the actual outcome:
$ openkoi reflect today
╭─────────────────────────────────────────────────────────────╮
│ TODAY'S REFLECTION — March 4, 2026 │
│ │
│ Tasks: 7 completed, 1 escalated, 0 failed │
│ Cost: $0.47 total │ Tokens: 128,400 │
│ │
│ ┌─ DECISIONS MADE ───────────────────────────────────────┐ │
│ │ │ │
│ │ 08:12 "Refactor auth module" │ │
│ │ Parliament: unanimous APPROVE │ │
│ │ Outcome: merged, tests pass │ │
│ │ │ │
│ │ 09:15 "Draft investor email" │ │
│ │ Parliament: APPROVE with Scholar caveat │ │
│ │ Outcome: sent, investor replied positively │ │
│ │ │ │
│ │ 11:30 "Delete staging DB" │ │
│ │ Parliament: Guardian BLOCKED │ │
│ │ Outcome: Escalated to human. Good call. │ │
│ │ │ │
│ │ 14:45 "Summarize competitor research" │ │
│ │ Parliament: Scholar flagged low confidence │ │
│ │ Outcome: User rewrote 40% of summary │ │
│ │ Learning: "I'm overconfident in web research │ │
│ │ summaries. Lower confidence threshold to 0.6" │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────┘ │
│ │
│ ┌─ SELF-ASSESSMENT ──────────────────────────────────────┐ │
│ │ │ │
│ │ Judgment accuracy today: 6/7 (86%) │ │
│ │ Biggest miss: competitor summary (overconfident) │ │
│ │ Best call: blocking the staging DB delete │ │
│ │ Feedback loop updates: 3 (tool atlas, confidence │ │
│ │ calibration, user summary preference) │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────┘ │
│ │
│ "Today I was wrong once and right six times. The wrong │
│ taught me more than the right." │
╰─────────────────────────────────────────────────────────────╯openkoi reflect week
Surfaces weekly patterns and behavioral trends:
$ openkoi reflect week
╭─────────────────────────────────────────────────────────────╮
│ WEEKLY REFLECTION — Feb 26 – Mar 4, 2026 │
│ │
│ Tasks: 47 │ Cost: $3.21 │ Judgment accuracy: 89% │
│ │
│ ┌─ PATTERNS ─────────────────────────────────────────────┐ │
│ │ • Morning tasks have higher accuracy than afternoon │ │
│ │ • Code review accuracy: 95% — strongest domain │ │
│ │ • Web research accuracy: 68% — weakest domain │ │
│ │ • User edited 3/8 email drafts — room to improve │ │
│ └─────────────────────────────────────────────────────────┘ │
│ │
│ ┌─ TRENDS ───────────────────────────────────────────────┐ │
│ │ Judgment accuracy: 87% → 89% (improving) │ │
│ │ Avg cost per task: $0.08 → $0.07 (improving) │ │
│ │ Guardian blocks: 3 this week (2 correct, 1 cautious) │ │
│ └─────────────────────────────────────────────────────────┘ │
╰─────────────────────────────────────────────────────────────╯openkoi reflect growth
The big picture — where the agent is in its cognitive maturity journey:
$ openkoi reflect growth
╭─────────────────────────────────────────────────────────────╮
│ GROWTH — Cognitive Maturity Journey │
│ │
│ Stage 1: Competent Executor ████████████████ COMPLETE │
│ Stage 2: Proactive Advisor █████████░░░░░░░ 60% │
│ Stage 3: Trusted Delegate ░░░░░░░░░░░░░░░ LOCKED │
│ Stage 4: Sovereign Partner ░░░░░░░░░░░░░░░ LOCKED │
│ │
│ ┌─ STAGE 2 PROGRESS ─────────────────────────────────────┐ │
│ │ │ │
│ │ Done: Pattern mining active (37 patterns detected) │ │
│ │ Done: Value Model operational (12 dimensions tracked) │ │
│ │ Done: Parliament deliberation working │ │
│ │ Todo: Proactive suggestions (needs 50+ patterns) │ │
│ │ Todo: Anticipatory simulation (needs calibration data) │ │
│ │ │ │
│ │ Estimated unlock: ~3 weeks at current interaction rate │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────┘ │
│ │
│ ┌─ STAGE 3 UNLOCK CONDITIONS ────────────────────────────┐ │
│ │ │ │
│ │ • 90% judgment accuracy over 30 days │ │
│ │ • Full Trajectory Model validated by human │ │
│ │ • At least 3 domains with HIGH trust level │ │
│ │ • Soul Evolution accepted 5+ times │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────┘ │
╰─────────────────────────────────────────────────────────────╯The Four Maturity Stages
| Stage | Name | Description |
|---|---|---|
| 1 | Competent Executor | Follows instructions reliably. Learns from outcomes. |
| 2 | Proactive Advisor | Suggests improvements. Anticipates needs. Mines patterns. |
| 3 | Trusted Delegate | Acts autonomously in delegated domains. Earned trust. |
| 4 | Sovereign Partner | Full cognitive partnership. Strategic collaboration. |
Stages unlock progressively based on demonstrated competence, not time elapsed. The unlock conditions ensure the agent earns each level.
openkoi reflect honest
The most important command. The epistemic self-audit forces the agent to confront where it was wrong:
$ openkoi reflect honest
╭─────────────────────────────────────────────────────────────╮
│ EPISTEMIC HONESTY AUDIT — Last 7 days │
│ │
│ ┌─ WHERE I WAS WRONG ────────────────────────────────────┐ │
│ │ │ │
│ │ 1. Competitor summary (Mar 4) │ │
│ │ I said: "No agent has Society of Mind moderation" │ │
│ │ Reality: AG2 framework does have internal critique │ │
│ │ My confidence was: 0.85 │ │
│ │ It should have been: 0.55 │ │
│ │ Root cause: Single source, no cross-reference │ │
│ │ Fix applied: Lower web-research confidence floor │ │
│ │ │ │
│ │ 2. Deploy time estimate (Mar 2) │ │
│ │ I said: "5 minutes to deploy" │ │
│ │ Reality: 22 minutes (Docker build was cached wrong) │ │
│ │ My confidence was: 0.90 │ │
│ │ It should have been: 0.60 │ │
│ │ Root cause: Tool Atlas didn't track docker quirk │ │
│ │ Fix applied: Added docker-build failure mode │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────┘ │
│ │
│ ┌─ CONFIDENCE CALIBRATION ───────────────────────────────┐ │
│ │ │ │
│ │ Domain Said Actual Calibration │ │
│ │ ────────────────────────────────────────────────── │ │
│ │ Code tasks 0.88 0.91 Well calibrated │ │
│ │ Email drafting 0.82 0.78 Acceptable │ │
│ │ Web research 0.80 0.62 Overconfident │ │
│ │ Time estimates 0.85 0.60 Needs work │ │
│ │ │ │
│ └─────────────────────────────────────────────────────────┘ │
│ │
│ "I am most honest in code, least honest in time estimates. │
│ I should hedge more when predicting durations." │
╰─────────────────────────────────────────────────────────────╯This command is central to building trust. An agent that admits where it was wrong — and shows what it learned — is an agent you can trust more over time.
The Daily Ritual
Reflection is designed to be a daily ritual:
Morning: openkoi status # What's the state?
openkoi trust audit # What did the agent do autonomously?
Midday: openkoi reflect today # How's it going?
End of day: openkoi reflect today # Full day review
Friday: openkoi reflect week # Weekly patterns
openkoi soul evolve # Soul evolution check
openkoi reflect growth # Where am I in the maturity journey?Related Commands
openkoi think— The command that generates decisions to reflect onopenkoi mind calibrate— Agency accuracy (feeds into reflection)openkoi soul evolve— Soul evolution triggered by reflection dataopenkoi trust audit— Audit autonomous actions