zouroboros

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).

Zouroboros self-healing watchdog that monitors scheduled agent models and automatically switches failing agents to backup models, restoring originals when they recover.

Change-gated progress monitor for long-running builds. Watches one canonical progress file (markdown task list) and only texts the operator on a real event — milestone completed, blocker appeared, build finished, or a stall — instead of a dumb clock that pings every N minutes. Use to babysit a long build/task without notification noise, or to replace a FREQ=MINUTELY check-in automation.

Evolution engine for the Zouroboros self-enhancement pipeline. Takes a prescription (seed + playbook) and executes the improvement: either via autoloop (file-targeting playbooks) or via script execution (procedural playbooks). Measures before/after, keeps improvements, reverts regressions, and stores results as episodes.

Self-diagnostic health scorecard for the Zouroboros system. Measures memory recall quality, knowledge graph connectivity, swarm routing accuracy, eval calibration, procedure freshness, and autoloop efficiency. Outputs a ranked scorecard with improvement opportunities. Run daily or on-demand to feed the self-enhancement pipeline.

Self-prescription engine for the Zouroboros system. Takes an introspection scorecard, identifies the weakest subsystem, maps it to a known improvement pattern, and generates a seed YAML + autoloop program.md for autonomous improvement. Includes a governor that flags high-risk prescriptions for human approval.

Autonomous repo drift remediation — runs quality gates, clusters uncommitted files by branch scope, auto-commits in-scope clusters, pushes, and creates a draft GitHub PR. Called by the [SYS] Audit Repo Drift automation after drift thresholds are exceeded. Read-only when --dry-run is passed.

Socratic interview and seed specification generator. Use before any major build task to clarify requirements, expose hidden assumptions, and produce an immutable spec ("seed") with acceptance criteria and an ontology. Prevents wasted cycles on ambiguous work. Adapted from Q00/ouroboros.

Progressive 3-stage verification pipeline for evaluating artifacts against specifications. Stage 1: mechanical checks (lint/test/$0). Stage 2: semantic evaluation against acceptance criteria. Stage 3: multi-perspective consensus for high-stakes decisions. Adapted from Q00/ouroboros.

Agent fleet health utilities — path preflight checker and auto-remediating weekly doctor with apply mode.

Change-gated progress monitor for long-running builds. Watches one canonical progress file (markdown task list) and only texts the operator on a real event — milestone completed, blocker appeared, build finished, or a stall — instead of a dumb clock that pings every N minutes. Use to babysit a long build/task without notification noise, or to replace a FREQ=MINUTELY check-in automation.

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. Gives personas persistent memory with semantic search (text-embedding-3-small), HyDE query expansion (gpt-4o-mini), LLM-powered memory gate, 5-tier decay, swarm integration, context budget tracking, cross-persona sharing, conflict resolution, and multi-hop retrieval. Requires OpenAI API key.

Socratic interview and seed specification generator. Use before any major build task to clarify requirements, expose hidden assumptions, and produce an immutable spec ("seed") with acceptance criteria and an ontology. Prevents wasted cycles on ambiguous work. Adapted from Q00/ouroboros.

Weekly self-audit of the Zo skill/identity/script-health surface. Walks Skills/*/scripts/, probes each script's --help under a bounded retry budget, checks identity files, and emits findings plus an atomic machine-status file. Use when running or maintaining the "[SYS] Audit Skills & Personas" scheduled agent, when investigating introspection findings, or when changing the SKIP_HEALTH skip list.

CLI toolkit for creating AI characters — generate identity profiles, avatar images, voice configs, and content workflows, with an interactive zo.space tutorial page.

YouTube transcript ingestion pipeline and RAG query interface for the AI Engineer persona. All 763 @aiDotEngineer channel videos indexed in Qdrant. Run weekly to stay current.

Inject cinematic AI-generated b-roll into a talking-head ("spine") video for a visually richer presentation. Reads the presenter transcript (SRT or plain text), picks N cutaway moments, generates one silent clip per moment via the fal-ai-media skill (veo3.1-fast text-to-video, or nano-banana-2 stills turned into Ken Burns clips), and intercuts them over the spine with a deterministic, proven ffmpeg formula — full-frame cutaway or picture-in-picture. Use when the user wants to turn a HeyGen/Mimir/Alaric talking head into a graphics-rich cut, "add b-roll", or "make this video less of a static talking head". Resumable, fail-loud, file-based DAG: plan -> generate -> compose. A --dry-run path proves the whole pipeline with ZERO fal.ai spend by fabricating labeled placeholder clips.

Generate and edit images and videos via the fal.ai API using models including nano-banana-2, gpt-image-2, kling-v3-std, and veo3.1-fast, routing through the user's FAL_KEY to minimize Zo credit usage.

--help skill for general domain operations

Quarterly drift watcher for marlandoj's collaboration style baseline. Scans typed user messages from zo_conversations.duckdb and staging payloads, diffs against a frozen baseline (median chars, frustration rate, cadence), and exits 0/1/2 for no-drift/drift/error. Identity files are never auto-modified — drift requires human review.

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.

Cross-Zo orchestrator → worker dispatch helper. Sends prompts and files between this Zo and a worker Zo (faunaflora) via the /zo/ask API using ZO_B_API_KEY. Supports prompt dispatch, JSON output parsing, file get/put via base64, and task-with-file-result workflows.

Whole-repo audit for over-engineering. Like ponytail-review, but scans the entire codebase instead of a diff: a ranked list of what to delete, simplify, or replace with stdlib/native equivalents. Use when the user says "audit this codebase", "audit for over-engineering", "what can I delete from this repo", "find bloat", "ponytail-audit", or "/ponytail-audit". One-shot report, does not apply fixes.

Harvest every `ponytail:` comment in the codebase into a debt ledger, so the deliberate shortcuts and deferrals ponytail leaves behind get tracked instead of rotting into "later means never". Use when the user says "ponytail debt", "/ponytail-debt", "what did ponytail defer", "list the shortcuts", "ponytail ledger", or "what did we mark to do later". One-shot report, changes nothing.

Code review focused exclusively on over-engineering. Finds what to delete: reinvented standard library, unneeded dependencies, speculative abstractions, dead flexibility. One line per finding: location, what to cut, what replaces it. Use when the user says "review for over-engineering", "what can we delete", "is this over-engineered", "simplify review", or invokes /ponytail-review. Complements correctness-focused review, this one only hunts complexity.

Programmatic HTML→video generation for the Zouroboros Academy and Alaric/Mimir explainer content, backed by HeyGen HyperFrames. Author a composition as a single HTML file (GSAP/anime.js timelines, captions, voiceovers, audio-reactive visuals) and render it to a real MP4/WebM/MOV/GIF locally with headless Chrome + ffmpeg — no manual ffmpeg zoompan pipeline, no cloud render. Use this to build title cards, faceless explainers, slideshows, product/launch videos, motion graphics, and captioned clips. The heavy multi-step orchestration skills (faceless-explainer, slideshow, etc.) live in Integrations/hyperframes/skills/.

Mechanized Definition of Done for an implementation plan. Runs four gates over a change set and emits one consolidated pass/fail report: (1) EVAL — the deterministic checks (tests/compile/dry-run), a hard gate and the real correctness guarantee; (2) GAP AUDIT — a wire-what-you-build scan for placeholder/stub markers and declared-but-unreferenced functions/exports (advisory); (3) CONSENSUS — frontier-model review per changed source file via the consensus-gate (advisory; mine the findings, not the verdict); (4) COMPLEXITY — single-axis over-engineering review per changed file via ponytail-review (advisory ONLY, never gates). Use at the end of every plan so closeout does not depend on remembering to do it.

Weekly audit + self-healing pipeline. Lints every Projects/*/01-brand/DESIGN.md against the Google Labs spec, diffs declared tokens against live site CSS, auto-heals mechanical OKLCH approximation drift (opens a PR), and surfaces judgment-call findings (WCAG, missing slots, rogue hexes) for human review via email.

Autonomous single-metric optimization loop that edits a target file, runs experiments, commits improvements, and reverts regressions until a numeric goal is reached.

Socratic interview skill that clarifies requirements before any major build task and produces an immutable seed spec with acceptance criteria and ontology.

Progressive 3-stage evaluation pipeline: mechanical checks, semantic verification against acceptance criteria, and multi-model consensus for high-stakes decisions.

Mechanical deep-research pipeline. Chains peer-reviewed literature (Consensus MCP), open web (web_research), and the internal Qdrant corpus into a single provenance-tracked synthesis, emitted as a markdown report. Use when the user asks to "research X deeply", "do a deep dive on X", "what does the research/evidence say about X", or wants a sourced report fusing what we already know with what's published. Runs as a resumable, fail-loud, file-based DAG: plan -> gather (parallel) -> fuse -> synthesize -> claim-check -> report -> persist.

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.

Multi-vendor validation gate with direct synthetic.new integration. Routes code and logic through three frontier models (GLM-5.1, Kimi-K2.6, MiniMax-M2.5) in parallel; flags disagreement before merge or evolution. Currently active in production as the Phase 1 gate for zouroboros procedure evolution.

Generalized blueprint for creating sophisticated personas on Zo Computer with safety rules, skills, MCP servers, and enhancement tools. Adaptable to any domain.