ceo-loop

CEO Decision Loop — ORBIT Protocol

Overview

The CEO-LOOP ORBIT model turns the CEO agent from a single-pass decision oracle into a continuous governance engine. Each invocation of this skill is one ORBIT iteration. The Stop hook (ceo-loop-stop.sh) re-invokes this skill after each iteration until the queue is empty or limits are reached.

Five contributions ORBIT adds to CEO-LOOP:

1. Actual loop mechanism — Stop hook + state file, not pseudocode
2. Cross-context continuity — progress file bridges compaction
3. Iteration tracking + retrograde — was the last decision correct?
4. Adaptive depth — ESLint fix in 30s, architecture in 15 min with Agent Teams
5. Trust calibration — accuracy tracked, demotion alert surfaces to human

---

State Files (all in $PROJECT_ROOT/.claude/)

These are written and read by the loop. Never modify manually during an active loop.

ceo-loop-state.json — Hook state

{
  "active": true,
  "iteration": 7,
  "max_iterations": 20,
  "time_limit_minutes": 30,
  "started_at": "2026-03-08T10:00:00Z",
  "trust_level": "standard",
  "depth_escalation_counter": 0,
  "correct_decisions": 47,
  "incorrect_decisions": 2
}

• active: Stop hook checks this. If false, hook is a no-op.
• iteration: Current iteration number. Hook increments this before re-feed.
• depth_escalation_counter: How many consecutive trivial-only iterations. At 3, go proactive.
• trust_level: "standard" | "elevated" | "demoted". Demoted = escalate more.

ceo-loop-decisions.jsonl — Decision journal (append-only)

{
  "id": "iter-7-1",
  "iteration": 7,
  "ts": "2026-03-08T10:15:00Z",
  "tier": "auto_approve",
  "depth": "trivial",
  "category": "quality",
  "input": "Add ESLint auto-fix to CI pipeline",
  "decision": "APPROVED",
  "reason": "Code quality improvements are always welcome. Low risk, high alignment.",
  "next_action": "Commit and move to next task.",
  "verification": "PENDING",
  "retrograde": { "previous_id": null, "outcome": "N/A" },
  "confidence": 0.98
}

Tier values: auto_approve | quick_review | deep_think | escalate Depth values: trivial | light | medium | heavy Category values: feature | architecture | quality | scope | deploy Decision values: APPROVED | NEEDS_REVIEW | REJECTED Verification values: CONFIRMED | FAILED | PENDING Retrograde outcome values: CORRECT | INCORRECT | PENDING | N/A