Agent Agentic Os

Trigger: "set up agentic OS", "initialize agent harness", "init my project for AI agents", "retrofit repository", "upgrade project for evolution", "sync instruction files", "where do I put CLAUDE.md", "create my agent environment", "set up persistent memory". Guides users through discovery, initializes/retrofits 3-Layer Memory, mirrors multi-tool instruction files (CLAUDE/GEMINI/Copilot/AGENTS), and guides plugin installation.

Trigger with "explain agentic os", "how do I set up a persistent agent environment", "what is the CLAUDE.md hierarchy", "explain the context folder structure", "how does session memory work", "what is soul.md or user.md", "explain auto-memory or MEMORY.md", "what is a loop scheduler or heartbeat", or when the user asks for the canonical guide.

Audit a file for TODO comments, pending work items, or technical debt markers. Useful for checking code readiness before a commit or reviewing task status. Trigger with "check for todos", "audit for debt", "list pending work", or "scan for TODOs".

SME-facing front-door skill for Agentic OS ecosystem evolution. Invokes the os-architect interview flow: classifies intent, audits existing capabilities, proposes evolution path (orchestrate / update / create), and dispatches work. Use when evolving plugins, skills, or agents — whether applying a new pattern, setting up an improvement lab, filling a capability gap, or coordinating multiple loops.

Converts temperatures between Celsius and Fahrenheit for the evolution end-to-end smoke test harness.

CRITICAL INTAKE GATEWAY: Use at the very start of ANY non-trivial engineering task, feature request, architectural refactor, or multi-file bugfix before entering plan mode or modifying code. Enforces read-only discovery, Socratic Defaulting (1-3 structured questions with recommended defaults), control plane registration in control_plane.db, and compilation of the immutable 4-Pillar TASK_SPEC.md.

Safely removes all agent lock files from the context/.locks/ directory to resolve deadlocks caused by crashed agents leaving stale locks behind. Use when the user says "/os-clean-locks", "clear all locks", "reset agent locks", or when an agent is deadlocked and cannot acquire a lock because a previous agent crashed and left a stale lock behind in context/.locks/. Verifies lock existence, discovers and removes stale lock directories, updates OS state via kernel.py, and emits event bus notifications. Requires Python 3.8+ standard library only.

Stateless evaluation engine that scores and gates skill improvement iterations using headless Python evaluation scripts. Use when the user says "evaluate this skill", "run autoresearch loop on", "optimize this skill", "run the eval loop", or when another agent proposes a change and needs validation.

Deterministic graph-planned self-evolution engine. Enforces 6-node state transitions, worktree isolation, verifier sovereignty, and asymmetric Layer 2 knowledge persistence.

Trigger with "run health check", "check os metrics", "system monitor", or when the user wants to review the Agentic OS liveness metrics across the Event Bus, locks, and memory arrays. Scans context/events.jsonl, os-state.json, and context/memory.md deterministically via kernel.py — no conversational judgment required. Migrated from the former os-health-check agent (2026-09-05): deterministic Bash+Read diagnostic, no interview, no adversarial judgment — fits the skill archetype, not the agent archetype.

Conducts a full-system adversarial audit of agent plugins, skills, specifications, and orchestration against enforced runtime contracts using deep reasoning.

Reviews a completed os-eval-runner lab run and backports approved changes to master plugin sources. Trigger with "backport the eval results", "review the lab run", "apply eval improvements to master", "check what the eval agent changed".

Selects confirmed native worktree facilities or the governed portable repository worktree fallback, validates placement, and reports cleanup guidance.

Bootstraps a skill evaluation lab repo for an autoresearch improvement run. Trigger with "set up an eval lab", "bootstrap the eval repo", "prepare the test repo for skill evaluation", "create an eval environment for this skill", "set up the lab space for this skill", or when starting a new skill optimization run that needs a standalone test environment.

Maintains a persistent, folder-based log of all agentic-os experiment runs. Each run writes one dated file to context/experiment-log/ and updates index.md. Supports five source types: verifier (qualitative), tester (qualitative), orchestrator (numeric), planner (qualitative), survey (mixed). Handles both numeric results (eval scores, KEEP/DISCARD, delta) and qualitative results (PASS/FAIL/PARTIAL, gap analysis). Use after any experiment run to persist findings before temp/ is cleared.

Trigger with "remember this", "update memory", "what should we record from this session", "capture learnings", "write a session log", or when closing a session. Guides agents on managing memory hygiene across sessions, deciding what to write to dated memory logs, what to promote to long-term memory.md, and when to archive.

Pattern 5: Concurrent Event-Driven Multi-Agent Loop. Coordinates multiple Claude sessions as OS threads sharing a common event bus and memory address space. Every loop cycle is a full improvement cycle: execute, eval against benchmark (KEEP/DISCARD), emit friction events, and close with surveys, metrics, memory persistence, and Triple-Loop triggers.

Discovers and persists the user's available AI environments (Claude, Copilot CLI, Agy CLI, Cursor, etc.) to context/memory/environment.md. Run once after OS setup or whenever the environment changes. os-architect and os-evolution-planner read this file to select the right delegation backend and cheapest brainstorm model automatically. Invoked by os-architect on first run if environment.md is absent.

Codifies the plan-and-delegate workflow for evolving plugins, skills, and agents. Given a target (plugin/skill/agent name) and an evolution goal, this skill first brainstorms 2-3 approach options using the cheapest available model, presents them for selection, then writes a structured task plan and Copilot CLI delegation prompt for the chosen approach. Called by os-architect for Path B (update) and Path C (create) executions. Can also be invoked standalone.

Continuously improves an existing agent skill based on eval results using the RED-GREEN-REFACTOR cycle. Apply when a skill's routing accuracy is low, trigger descriptions need sharpening, or os-eval-runner scores are below target. (1) run a RED baseline to observe the failure mode, (2) apply a focused patch and verify with os-eval-runner (GREEN), (3) refactor to close loopholes until score meets threshold. Integrates with os-eval-runner as the objective eval gate. NOT for scaffolding new skills — use create-skill (agent-scaffolders) for that.

Verifies that os-architect actually causes evolution — not just words. Dispatches os-architect in single-shot simulation mode for a given test scenario, then checks for real artifact presence (new files, HANDOFF_BLOCK, plan files). Reports PASS / FAIL with grep evidence. Accumulates results into a test report. Use after any changes to os-architect, os-evolution-planner, or improvement-intake-agent.

Trigger with "show me the improvement chart", "how are we improving", "progress report", "graph the eval scores", "show cycle of improvement", "what's the trend", "are we getting better". Produces a visual/text summary of how the agentic loop is improving across cycles. Do NOT use this to run the learning loop or evaluate a specific skill change.

Consumes friction_cluster_agent hotspot reports and synthesizes systemic refactoring proposals for human review, for Tier 3 architecture friction. Trigger with "synthesize a refactoring proposal from the friction hotspots", "what's the systemic fix for this friction cluster", or when os-architect/self-evolution escalates a Tier 3 (Regression / Architecture) friction event per github-issue-logging-policy.md. Migrated from the former repository-improvement-agent (2026-09-05): deterministic report-synthesis task, no interview, no self-directed git/PR execution — fits the skill archetype, not the agent archetype. Never creates branches, commits, or PRs itself — see "Human Gate" below.

Audits closed GitHub issues to verify whether root causes were genuinely resolved, if follow-on execution friction appeared, or if systemic improvements were retained. Trigger with "audit closed issues", "review issue resolution quality", "was this issue actually fixed", or as a post-closure quality gate on issues labeled resolution:fixed or resolution:superseded. Migrated from the former issue-resolution-reviewer agent (2026-09-05): deterministic audit against fixed criteria, no interview, no code writes — fits the skill archetype, not the agent archetype.

Audits and rewrites AI agent instruction files (AGENTS.md, CLAUDE.md, GEMINI.md, .github/copilot-instructions.md) in any repo. Strips stale or foreign content, applies Karpathy's four behavioral principles, ensures platform-specific sections, and makes each file authoritative rather than a copy of another. Trigger when the user says "optimize my CLAUDE.md", "audit agent instructions", "improve my AGENTS.md", "apply Karpathy principles to my agent files", "clean up my copilot instructions", "review my GEMINI.md", or "update my AI instruction files".