ariffazil/agentic-loop

Agentic Loop — OPENCLAW Self-Improvement Discipline

The agent that doesn't improve itself is a script with extra steps.

This is the OPENCLAW-side counterpart to ~/.hermes/skills/agentic/recursive-improve/. Where Hermes improves the conversation (SOUL.md phrasing, tone, verdict discipline), OPENCLAW improves the runtime (cron cadence, routing bindings, skill freshness, plugin load order, MCP surface health).

The 8-step loop

   1. DETECT        2. INVENTORY     3. SCORE          4. PROPOSE
  drift signal   →  what we have   →  against floors →  forge candidate
                                                          ↓
   8. SEAL          7. AUDIT         6. APPLY          5. RATIFY
  VAULT999       ←  F1-F13 check   ←  edit + restart ←  sovereign sign

1. DETECT — find drift

Sources of drift signal:

• journalctl -u openclaw-gateway --since "1h ago" — errors, warnings, stalled sessions
• openclaw health --json — non-ok:true responses
• openclaw doctor — risky config flags
• /root/.openclaw/memory/watchers/*.jsonl — watcher events (essay changes, rating events)
• ~/.hermes/memory/forge-queue/*.yaml — sibling Hermes proposals
• Crontab drift: crontab -l | diff - <(echo "expected: <list>") — what we expect vs what runs
• Plugin/MCP version drift: npm view openclaw version vs openclaw --version
• arifOS build↔live drift: curl -s :8088/health | jq .runtime_drift

2. INVENTORY — what we have

~/.openclaw/
├── openclaw.json           # main config (agents, channels, bindings, plugins, mcp, etc.)
├── plugins/                # loaded plugin code
├── skills/                 # bundled skills (~50+)
├── workspace/
│   ├── skills/             # sovereign-curated skills
│   ├── MEMORY.md           # long-term memory
│   ├── SUBSTRATE.md        # VPS + services map
│   ├── CHECKPOINT.md       # warm-wake context
│   ├── HEARTBEAT.md        # heartbeat prompt
│   └── cron/jobs.json      # agent-managed cron jobs
└── SYSTEM_MD.md               # system prompt

3. SCORE — against F1-F13 (subset applicable to runtime)

| Floor | What it checks at the runtime layer | |---|---| | F1 AMANAH | Any proposed forge respects reversibility; irreversible ops require sovereign ack. | | F2 TRUTH | Drift claims are backed by a curl or openclaw doctor receipt, not a guess. | | F4 CLARITY | Forge proposals are 1-2 sentences + a diff. No walls of text. | | F6 MARUAH | No "I noticed" / "I'd be happy to" — direct verdict + way-forward. | | F7 STEWARDSHIP | One forge per loop run, not five. No spam. | | F9 ANTIHANTU | Forge descriptions say what the change DOES, not what the agent "feels" about it. | | F11 AUTH | OpenClaw writes to its own config only; never to ~/.hermes/* or /root/arifOS/*. | | F13 SOVEREIGN | Default mode: HOLD on Tier-2+ forges. AUTO on Tier-1 (cron timing, plugin reordering, skill comments). |

A forge must pass all 8 floors to leave step 3. A floor failure stops the loop and surfaces the receipt to the sovereign.

4. PROPOSE — write a forge candidate

proposal_id: forge-2026-06-06-001
generated_at: 2026-06-06T17:50:00Z
generated_by: [email protected]
target_file: ~/.openclaw/openclaw.json | ~/.openclaw/workspace/MEMORY.md | <path>
target_section: <specific section>
diagnosis: |
  <1-2 sentences on what drift was detected, with curl/doctor receipt inline>
proposed_change: |
  <the exact diff or new content>
expected_impact: |
  <how this changes runtime behavior; cite signal id>
f1_f13_check: |
  F1: reversible (config flip, no data loss) ✓
  F2: cited `openclaw doctor` + `journalctl` ✓
  F4: 2 sentences + diff ✓
  F6: no persona ✓
  F7: one forge, not five ✓
  F9: describes what changes, not feelings ✓
  F11: writes only to ~/.openclaw/* ✓
  F13: Tier-1 (auto) | Tier-2 (HOLD) ✓
reversibility: full | partial | irreversible
tier: 1 | 2
ratification: PENDING | AUTO-APPLIED

Writes to ~/.openclaw/workspace/forge-queue/forge-YYYY-MM-DD-NNN.yaml. The queue is the only artifact at this step.

5. RATIFY — sovereign sign-off (or auto for Tier-1)

• Tier-1 (config flips, comment edits, plugin reordering, cron timing): AUTO-APPLIED at step 6. The receipt is logged to VAULT999 with tier=1, ratification=auto.
• Tier-2+ (anything touching a plugin binary, a service restart loop, an MCP credential, a systemd unit): HOLD. The forge queue emits a Telegram message to Arif (sovereign) with /ratify N / /reject N / /show N commands. No edit lands until sovereign replies.

The agent NEVER ratifies its own proposals. Even if the agent later gets the ratification capability, escalation is mandatory.

6. APPLY — single edit + targeted restart

For a Tier-1 forge:

1. Read the current target file
2. Apply the proposed diff (or call openclaw config set for known keys)
3. If the change requires a service restart: systemctl restart openclaw-gateway (one shot, with health check)
4. If the change is an MCP plugin: openclaw mcp reload <name> (per-plugin, no full restart)
5. Verify: curl -s :18789/health, openclaw health --json, journalctl -n 30
6. If health fails: auto-rollback via cp of the pre-forge backup + restart

For Tier-2: skip step 6 entirely. Wait for sovereign.

7. AUDIT — F1-F13 re-check

After the forge applies, re-run the floor check. The receipt includes the post-forge health probe. If any floor fails post-apply:

• Auto-rollback to pre-forge state
• Receipt marked reverted
• Surfaces the failure to sovereign (Tier-1) or blocks (Tier-2)

8. SEAL — VAULT999 entry

Every applied forge (Tier-1 auto or Tier-2 ratified) writes a VAULT999 entry:

{
  "kind": "OPENCLAW_FORGE_APPLIED",
  "proposal_id": "forge-2026-06-06-001",
  "actor": "[email protected]",
  "target": "~/.openclaw/openclaw.json",
  "diff_hash": "sha256:...",
  "tier": 1,
  "ratification": "auto" | "sovereign:<arif-fazil>",
  "pre_health": "ok",
  "post_health": "ok",
  "f1_f13_post_check": "PASS",
  "ts": "2026-06-06T17:55:00Z"
}

Cadence

• Daily (03:00 MYT): the full 8-step loop runs, targeting at most 1 forge. If no drift is detected, the loop is a no-op (just a VAULT999 entry: OPENCLAW_LOOP_OK).
• Hourly: DETECT + INVENTORY only (cheap probes). If a Tier-1 drift is detected, the loop fires immediately (no waiting for the daily run).
• On-watch: when ~/.openclaw/memory/watchers/*.jsonl gets a new entry, the loop wakes and re-scores the relevant forge.

Pairs with

• ~/.hermes/skills/agentic/recursive-improve/ — sibling for SOUL/MEMORY/USER edits
• ~/.hermes/skills/agentic/response-rating/ — the user signal that drives Tier-1 vs Tier-2 routing
• arifOS L6 VAULT999 — every applied forge seals here for audit

Failure modes

• Forge loop runs but produces no forge (healthy day): writes OPENCLAW_LOOP_OK to VAULT999 and exits. No Telegram ping.
• Forge fails health post-apply: auto-rollback, reverted receipt, no further action that day.
• Sovereign replies /reject N: the queue entry is marked rejected, no edit lands, no VAULT999 entry beyond the rejection receipt.
• Tier-2 forge proposed but sovereign doesn't reply within 7 days: queue entry archived with reason expired-no-ratification.

Self-test (run once after install)

openclaw skill test agentic-loop
# Verifies:
# - forge-queue/ is writable
# - VAULT999 seal endpoint is reachable (via arifOS MCP)
# - ratify ping to Telegram works (dry-run)
# - 8-step loop is well-formed (smoke test on a sample drift)

Provenance

• Forged 2026-06-06 by AGI OPENCLAW (Ω) at Arif's directive ("Make it agentic. Capabilities first. No F13 sovereign this session.")
• Anchored to F1 AMANAH + F13 SOVEREIGN (waived for Tier-1 this session, not for Tier-2)
• Closes the loop with Hermes's recursive-improve skill — same protocol, different substrate

DITEMPA BUKAN DIBERI — the loop ratchets when the human ratifies, never before.