메뉴
Weekly plain-language briefing of what the Zouroboros loop did on its own — autonomous changes, model health, and the short list of items that actually need the operator's judgment. Read-only.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Local-only multi-agent orchestrator with token optimization, 6-signal composite routing, and persistent memory. Routes all tasks through local executor bridges (Claude Code, Hermes, Gemini, Codex) with DAG dependencies, auto-episode creation, and cognitive profiles.
Zouroboros self-healing watchdog that monitors scheduled agent models and automatically switches failing agents to backup models, restoring originals when they recover.
Hybrid SQLite + Vector persona memory system for Zo Computer. Episodic memory with temporal queries, graph-boosted search, BFS path finding, knowledge gap analysis, auto-capture pipeline, and per-workload model routing. Default generation workloads use OpenAI `gpt-4o-mini`; embeddings use OpenAI `text-embedding-3-small`.
Security governance gate for AI agent skills, backed by NVIDIA SkillSpector. Statically scans a skill (local path, directory, zip, or Git URL) for 65 vulnerability patterns across 16 categories (prompt injection, data exfiltration, privilege escalation, supply chain, excessive agency, tool/MCP misuse, dangerous code via AST, YARA signatures) and returns a 0-100 risk score with a SAFE / REVIEW / DO_NOT_INSTALL recommendation. Use BEFORE adopting any third-party skill, and for the periodic baseline scan of the Skills/ directory.
Detector for model classifier SOFT-blocks — refusals, filtered-to-empty, truncated-with-disclaimer, content-policy flags — that masquerade as successful empty output to existing infrastructure (agent-model-healer only catches HARD failures: HTTP 4xx/5xx, timeouts, genuine empties). Classifies an output as soft_block / genuine_empty / ok using a per-provider signal catalog, then routes a per-domain fallback (security → broader-tolerance model, bio/chem → human review, distillation → refuse-by-design, never routed around). Every block + fallback is appended to an audit ledger. A quarterly system-card-diff agent flags expanded classifier scope so the catalog stays current.
Rendered-output verification station for the swarm post-flight eval. After a maker subagent produces a visual deliverable (UI/route/site), the station captures a full-page screenshot via agent-browser, then an independent verifier model (not equal to author, per aiewf P0-2) reads the screenshot against seed acceptance criteria + project DESIGN.md tokens + prior screenshots. Match marks the task complete; mismatch emits a structured visual diff for the maker's next iteration. The verifier is the exit condition, not the maker. Recurring visual failure modes graduate to instincts.yaml via the extract-patterns gate (ZOU-452).
| name | operator-digest |
| description | Weekly plain-language briefing of what the Zouroboros loop did on its own — autonomous changes, model health, and the short list of items that actually need the operator's judgment. Read-only. |
| compatibility | Created for Zo Computer |
| metadata | {"author":"marlandoj.zo.computer","version":"1.0.0","category":"zouroboros-infrastructure"} |
A weekly, human-facing summary of the Zouroboros self-improvement loop's autonomous activity.
The loop runs unattended: self-heal (introspect → prescribe → evolve), the model healer, the consensus gate, the curiosity explorer. Each writes structured JSON to the data dir. On their own those files are operator-hostile blobs. As the operator steps back and lets loops close without supervision, the risk is comprehension debt — the system keeps changing itself and the operator loses the thread of what changed, why, and where their judgment is still required.
digest.ts reads the loop's on-disk state for a time window (read-only — it never mutates loop
state) and renders ONE briefing: what changed and why, what the verification layer is watching,
and the short list of things that genuinely need a human. It emits a markdown file, a styled
HTML email body, a PDF attachment, and a JSON manifest on stdout for the scheduled agent.
# Default: last 7 days, data dir from $ZOUROBOROS_DATA_DIR or /root/.zouroboros
bun Skills/operator-digest/scripts/digest.ts
# Custom window / paths
bun Skills/operator-digest/scripts/digest.ts --days 14 --out /tmp/digest --no-pdf
bun Skills/operator-digest/scripts/digest.ts --since 2026-06-01 --data-dir /root/.zouroboros
| File | What it contributes |
|---|---|
evolution-history.json | autonomous interventions: playbook, metric, delta, kept/reverted |
intervention-ledger.json | per-intervention state deltas + anti-Goodhart drift flags |
chronicle.json + chronicle-votes.json | curiosity proposals; accepted votes = escalations |
healer-state.json | model failovers + per-model health probes |
reputation.json | consensus voter reputation (agreement / vendor-error rates) |
agent-model-healer/assets/fallback-chain.json | friendly model labels + retired-model list |
The subject line and accent are driven by a deliberate three-level scheme. The point is to
reserve the alarm for the loop actually misbehaving, so the operator can trust a ❌ to mean
"look now" rather than crying wolf on routine self-correction.
❌ action — an anti-Goodhart drift flag tripped: a change improved its visible metric
while the hidden held-out set moved against it (possible metric-gaming). This is the only
case where the loop itself is suspected of misbehaving.⚠️ review — items need human judgment but the loop is behaving: curiosity proposals the
operator accepted (awaiting escalation) and live models failing their last health probe.✅ clear — nothing waiting.Two deliberate exclusions keep the signal honest:
❌._removed* block in fallback-chain.json) is expected fallout
of the removal, not an operator concern. Unhealthy lines also carry the probe age so a stale
probe can't silently overstate urgency.{ "status": "action|review|clear", "subject": "…", "summary": "…", "windowLabel": "…",
"mdPath": "…", "htmlPath": "…", "pdfPath": "… or null", "changedCount": 0,
"needsJudgment": 0, "driftDetected": false, "reverts": 0, "healthOk": true }
scripts/digest.ts — the generator (read-only; no dependencies beyond Bun + a PDF renderer)<data-dir>/operator-digest/operator-digest-<YYYY-MM-DD>.{md,html,pdf}A weekly [ZBR] Zo agent runs the script, parses the manifest, and emails the operator the
HTML body with the PDF attached (subject = manifest.subject, which carries the status emoji).
The script does the work; the agent only renders + delivers. PDF rendering falls back through
wkhtmltopdf → weasyprint → chromium; if all fail, the email still sends with the HTML body.