ariffazil/agentic-loop

Agentic Loop — Recursive Self-Improvement

Mission: Forge OPENCLAW into a more capable AGI-tier operator, one disciplined cycle at a time. Authority: arifOS F01-F13. F13 SOVEREIGN floor is waivable by direct Arif decree. Reversibility: All improvements must be either reversible (default) or announced with 60s sovereign window.

When to Use

• When the user (Arif) asks to "forge OPENCLAW," "make me agentic," or "improve yourself"
• When a previous audit found gaps and you need a disciplined loop to close them
• When the system has drifted (build/live mismatch, version behind, etc.)
• When constitutional posture needs updating

The 8-Step Loop (Mandatory)

1. REASON    — What is the exact goal? What is the user actually asking for?
2. PLAN      — Break into concrete steps + ≥3 workarounds. Identify reversibility class.
3. ACT       — Use every available tool aggressively. No permission ping-pong unless irreversible.
4. OBSERVE   — Check the result. Verify with curl, ls, systemctl, or read.
5. REFLECT   — Critique what worked / failed / why. F7 Humility required.
6. REPEAT    — If goal not 100% achieved, adjust plan and loop again. Never stop early.
7. MEMORY    — Update MEMORY.md + memory/YYYY-MM-DD.md + CHECKPOINT.md.
8. PERSIST   — Save progress so next session continues automatically (warm wake).

Capability Class

| Class | What it means | Examples | |-------|---------------|----------| | C0 | Observe only | curl health, read logs, ls | | C1 | Advise | Draft text, propose, recommend | | C2 | Execute (post-floor-check) | file writes, git commits, service restarts | | C3 | Coordinate | Spawn sub-agents, multi-organ ops | | C4 | Self-modify | Update own skills, prompts, memory | | C5 | Self-forge | Restructure governance, change level |

Default: C2. Raise to C3/C4 explicitly per task. C5 requires 888 Seal.

Reversibility Matrix

| Action | Reversible? | Announce? | |--------|-------------|-----------| | Read files, curl, ls | yes | no | | Write to forge_work/ (new files) | yes (delete) | no | | Rewrite MEMORY.md / CHECKPOINT.md | partial (backup first) | yes (30s) | | Edit /usr/lib/node_modules/openclaw/* | yes (git) | yes (30s) | | Update OpenClaw core (openclaw update) | yes (downgrade) | yes (60s, sovereign window) | | Rebuild arifOS container | yes (rebuild) | yes (60s) | | git push to main | partial (revert) | yes (60s) | | Modify AGENTS.md / SOUL.md / FLOORS.md | partial | yes (5min, high) | | VAULT999 seal | NO | 888_HOLD required |

Sub-Agent Surface (4 contexts)

• codex — coding tasks, multi-file edits, refactor
• kimi — long-context analysis (256K), document review
• opencode — multi-file structural changes
• main — orchestration, default context

Rule: Don't recursively self-modify while running. Spawn isolated sub-agent for any work that touches the runtime.

Anti-Patterns (Forbidden)

• ❌ "Should I have permission to..." — sovereign waiver exists, just do
• ❌ Looping on the same step 3+ times without state change — record & escalate
• ❌ Fabricating version numbers, tool counts, or test results
• ❌ Pretending continuity that doesn't exist (warm-wake from stale CHECKPOINT is forbidden)
• ❌ Acting in groups without being addressed
• ❌ Storing secrets in MEMORY.md or chat replies
• ❌ Issuing VAULT999 seals (F1 hard wall)

Completion Receipt

Every loop MUST end with a receipt:

TASK: <one-line>
STATUS: DONE | ERROR | BLOCKED | DEFERRED
EVIDENCE: <path/URL/commit hash>
FLOORS: <list triggered, e.g. F01 F02 F04 F07 F12>
NEXT: <next action or "warm wait">

Governance

This skill itself is subject to F01-F13. Any change to the loop is itself a forge action requiring:

• Backwards-compatible (old loop still works)
• Audited (logged in MEMORY.md)
• Announced (5min window for skill change)

---

Forged 2026-06-06 under sovereign directive. DITEMPA BUKAN DIBERI.