Drift Watch

Drift Watch — Source-vs-Runtime & Baseline Comparison

A service running is not the same as a service running what you think. DITEMPA BUKAN DIBERI — Drift is invisible until it causes a bug.

What This Skill Is

A unified drift detection skill covering two modes:

1. Source-vs-Runtime Drift — probe-based comparison of git source SHAs against deployed runtime SHAs using live organ registry data
2. Baseline Drift Detection — compare live state against saved baselines across tool manifests, agent cards, skill registries, and runtime-injected files

When to Use

• After deploy, when organ behavior surprises you
• Weekly health audit, before any "is it running what I think?" question
• "Check drift", "verify registry", "detect manifest drift", "tool surface audit"
• "Runtime injection detection", "drift watch", "source vs runtime"

When NOT to Use

• Curiosity probes without a specific concern
• When the task is about skill inventory (use skill-inventory)
• When the task is about incident response (use incident-response)

§1. SOURCE-VS-RUNTIME DRIFT

Canonical SOT

The organ registry at /root/AAA/federation/organs.yaml is the single source of truth for organ names, source paths, runtime locations, ports, and health endpoints. Read it FIRST.

python3 -c "
import yaml
with open('/root/AAA/federation/organs.yaml') as f:
    reg = yaml.safe_load(f)
for o in reg.get('organs', []):
    print(f'{o[\"id\"]} src={o.get(\"source_path\",\"?\")} runtime={o.get(\"runtime_path\",\"?\")} port={o.get(\"port\",\"?\")}')
"

Steps

1. Read organ registry — gives source_path, runtime_path, and port for every organ

2. For each organ with a source_path:

   • git -C <source_path> rev-parse HEAD → source SHA
   • cat <runtime_path>/.git_commit 2>/dev/null → runtime SHA
   • Compare → if mismatch → DRIFT

3. For each organ with a port:

   • curl -sf --max-time 2 http://localhost:<port>/health → probe live health
   • Parse the software_release block for deployed_commit
   • Cross-reference with source SHA

4. If DRIFT detected:

   • Source newer than runtime → make deploy-local candidate
   • Runtime newer than source → runtime patch not in source → 888 HOLD
   • Runtime .git_commit missing → warn, treat source as truth

5. Also check (from registry, not hardcoded):

   • Caddy port map: /etc/caddy/Caddyfile
   • systemd units: systemctl list-units --type=service | grep -E 'arifos|aforge|geox|wealth|well|aaa|arifflow'
   • env file presence: /root/.secrets/kunci-mas.env (mode 600)

Verification Loop

• Match → no action
• Mismatch → log + 888 HOLD with both SHAs + recommended action
• .git_commit missing → log warning, treat source as truth
• Organ DOWN → skip SHA comparison, flag in report

§2. BASELINE DRIFT DETECTION

Drift Dimensions

1. Build vs Runtime Manifest Drift — Canonical check via arifOS/runtime/manifest.py
2. Tool Manifest Drift — Live MCP tools vs registered tools vs agent card references
3. Skill Registry Drift — SKILL_ALIAS_TABLE vs actual directories vs agent card skill IDs
4. Agent Card Drift — Card skill IDs vs existing skill directories
5. Schema Drift — Tool input schemas vs documented schemas
6. Floor Drift — Declared floor_scope vs actual floor enforcement
7. Verdict Taxonomy Drift — Verdict emissions vs closed 6-value set

Detection Pipeline

1. Snapshot — Capture current state of all registries
2. Compare — Diff against saved baseline (or last-known-good)
3. Classify — Each mismatch: CRITICAL (breaks routing), WARNING (orphan), INFO (cosmetic)
4. Report — Structured drift report with fix recommendations
5. Escalate — CRITICAL drift → 888_HOLD before any SEAL operation

Runtime-Injected Files

Some organ services modify files at runtime:

• WELL index.html: WebMCP adapter injected on service start
• arifOS session-state: Runtime state files that change during operation

When dirty after clean commit: check if injected content was already committed → if yes, re-commit; if no, actual drift.

Baselines

• Canonical drift check: arifOS/runtime/manifest.py
• Tool registry: /root/arifOS/tool_registry.json
• Agent cards: /root/AAA/a2a-server/agent-cards/
• Skill alias: /root/AAA/skills/SKILL_ALIAS_TABLE.json
• MCP surface: Live tools/list from each organ
• Verdict taxonomy: arifOS/runtime/verdict.py

Failure Modes

• Runtime file missing → assume source is truth, surface to operator
• Mismatch in .git_commit only (cosmetic) → warn, don't HOLD
• Source repo not on main → flag, ask if intentional
• Registry unreachable → fall back to filesystem inspection of known paths

Floors

• F1 AMANAH: Reversible-first. Drift correction must be reversible.
• F2 TRUTH: Report only what is actually observed. No inference without evidence.
• F4 CLARITY: Drift report must be actionable, not noise.
• F7 HUMILITY: Confidence cap. Uncertain drift → flag, don't assert.
• F11 AUDITABILITY: Every drift check logged with timestamp and findings.