Routing classifier + multi-agent orchestrator for Claude Code. 3-4ms Node fast-path hook, per-prompt S0-S5 tiering, parallel Sonnet workers, Oracle-gated memory with vector search, 35-test suite + eval harness. Stdlib Python + SQLite. MIT.
Skill Claude CodeCodex
Brain workbench — the model routing classifier and savings analyzer. Invoke ANY time the user says "brain scan", "brain score", "brain audit", "brain audit-skills", "brain pin", "brain apply-env", "brain test", or "/brain [subcommand]". Classifies tasks by severity (S0-S5) to route between Haiku/Sonnet/Opus while…
Skill Claude CodeCodex
Homer L5 — Three-tier memory store. Core (context-resident, 5K tokens, auto-compacted) / Recall (SQLite FTS5 searchable) / Archival (cold markdown, back-pointer protocol). Every write and edit requires a valid citation. Zero dark edits. Compaction moves entries between tiers with back-pointers in Core. Zeus and MUSES…
Skill Claude CodeCodex
Homer L3 — Epic Research Muse. Deep synthesis from web + local research sources. Zeus dispatches Calliope as a parallel subagent when a plan includes a research subtask. Calliope runs read-only, returns structured markdown with T1-T3 source citations, never writes files or modifies state.
Skill Claude CodeCodex
Homer L3 — Code Archaeology Muse. Maps existing codebases, finds call sites, builds dependency graphs, spots dead code. Zeus dispatches Clio when the plan needs "what's already in the codebase" before any new work. Read-only. Every claim cites file:line.
Skill Claude CodeCodex
Homer L3 — Measurement Muse. Pulls numeric receipts from Toke telemetry (decisions.jsonl, learnings.md, stats-cache.json, Homer VAULT, brain scans). Zeus dispatches Urania when the plan needs "how many / how much / what percentage." Read-only. Every number comes with a reproducible command.
Skill Claude CodeCodex
Homer L7 — Critic and evaluator. Scores Zeus synthesis outputs against the 13 Sacred Rules, a rubric (quality floor), and theater detection heuristics. Flags regressions vs baseline outputs. Reproduces yesterday's theater-audit pattern on any markdown text. Beat SOTA commitment.
Skill Claude CodeCodex
Homer L4 — Advisor Escalation. Invokes Anthropic's advisor20260301 API via braincli.py advise when a MUSES worker returns ROI=0 or Zeus hits an inconclusive state. Cost-capped (max 2 per session). Preconditions enforced (API key, creative-content refusal, session cap, brain reachable). Returns to Zeus for Phase 3…
Skill Claude CodeCodex
Homer L2 — Orchestrator skill. Decomposes a task, plans via extended thinking, delegates to MUSES workers in parallel (Anthropic MARS pattern), synthesizes, hands to Oracle for eval, writes checkpoint to VAULT. Use when a task would benefit from multi-agent decomposition (3+ subtasks, research burst, architecture…
Skill Claude CodeCodex
Actionable implementation blueprint skill v2.0. Use this ANY time the user says "blueprint", "make a blueprint", "create a blueprint", "deploy blueprint", or asks for an actionable implementation plan for a game system. This skill scans the existing codebase and research for relevant data, conducts deep research (10+…
Skill Claude CodeCodex
Session closure and memory persistence v2.0. Captures work completed, decisions made, and context for next session. v2.0: structured learning format, cross-skill auto-promotion, checkpoint verification, ecosystem health snapshot. Companion to init — init loads, close-session saves.
Skill Claude CodeCodex
Iterative 3-cycle Blueprint refinement engine. Use this ANY time the user says "cycle", "run cycle", "deploy cycle", "use cycle on", or wants a deeply refined implementation plan that improves itself through multiple passes. Cycle runs Blueprint 3 times on a given prompt, with DevTeam fluidity analysis and Professor…
Skill Claude CodeCodex
DevTeam v4.0 — Adaptive Code Fluidity & Architecture Engine. Calibrated scoring against the 7 Laws of Code Fluidity, domain auto-detection, regression guard, and project-docs reference integration. Holy Tool status. Auto-activates on "devTeam", "deploy devTeam", "fluidity check", "architecture review" — and whenever…
Skill Claude CodeCodex
Toke maximum-execution mode. Activates on "godspeed". Ticks the audit counter, scores the incoming prompt via Brain, auto-dispatches Zeus on S3+ tasks, parallelizes subtasks, and terminates every orchestrated run with an atomic zeus gate-write so Oracle-scored synthesis lands in Mnemos.
Skill Claude CodeCodex
Supreme diagnose-research-implement-verify engine. Combines devTeam (calibrated diagnosis with regression guard) and profTeam (adaptive agents with ROI tracking) in a convergent loop. Stops when diminishing returns hit OR regression risk exceeds improvement value.
Skill Claude CodeCodex
Session initialization. Scans the current folder, detects the project type, reads CLAUDE.md / README / projectstatus / MEMORY, loads shared protocols + learnings grep-first, runs an ecosystem health check, tails the previous session's status, and prints a one-screen briefing with a project description, extracted…
Skill Claude CodeCodex
Adaptive multi-agent parallel research engine. Deploys agent configurations based on topic classification and accumulated ROI data. Cross-validates with confidence matrix. Smart Cycle compression. Failure recovery. Output: ONE master Blueprint PDF. Holy Tool status.
Skill Claude CodeCodex
Research expert skill v3.0 that acts like a team of college professors giving a scientific lecture. Use this skill ANY time the user says "use professor", "professor skill", "deploy professor", or asks for deep research on a topic with PDF output. Also trigger when the user wants comprehensive, well-sourced research…
Skill Claude CodeCodex
Build and deployment verification. Auto-detects project type from the current working directory and runs the matching checks — Python imports, pytest, npm build, CMake build, Go test, Rust cargo, or a custom health script. Returns pass/fail with a terse error summary. Use after any batch of code changes or when…
Skill Claude CodeCodex
Homer L6 — Sleep-time routing weight tuner. Mines Brain's decisions.jsonl + advisorcalls.jsonl, computes tier / model / guardrail distributions, and proposes routingmanifest.toml weight adjustments based on observed data. Aurora PROPOSES; the user decides. No auto-apply. Sacred Rule.
Skill Claude CodeCodex
Homer L6 — Sleep-time learning distillation agent. Mines every learnings.md file across skills + /.claude/shared/sharedlearnings.md + Toke/research/ + Mnemos archival. Extracts top-N patterns by composite score (ROI × confidence × confirmedcount). Writes dated best-practices KB. Absorbs the original Kiln blueprint…
Skill Claude CodeCodex
Homer L6 — Sleep-time theater auditor. Scans every /.claude/skills//SKILL.md using Oracle's theater detection, produces dated audit reports flagging PRUNE and INVESTIGATE candidates. Reads Oracle's pattern, applies it ecosystem-wide. Run on-demand via sleep run nyx or schedule nightly via cron. Output lands in…
Skill Claude CodeCodex
Toke Local LLM workbench — routes brain-classified S0/S1/S2 prompts to a local Ollama model (Qwen 2.5 14B) with softmax-entropy confidence monitoring. Critical Logic Override gate hands work back to Claude only when local confidence drops below threshold AND the user explicitly approves. Toke-native — uses Brain's…
Skill Claude CodeCodex
Cross-project status aggregator. One command, full birds-eye view of every project with a projectstatus.md file in /.claude/projects/. Flags stale or blocked projects, shows learning pipeline health, gives a prioritized "what needs attention" list.
Skill Claude CodeCodex
Fast session-start pre-flight for Toke. Loads project context, runs a health check across Brain + Homer + hooks, and delivers a one-screen briefing so godspeed routes to the current state, not stale memory.