agent-studio

Audits, plans, and orchestrates a phased Postgres Row-Level Security (RLS) rollout for the agent-studio multi-tenant database (61 Prisma models). Produces SQL migration drafts, cross-tenant test suites, and rollout runbooks. NEVER auto-applies migrations — human approval gates every write operation. Triggers: "rls audit", "rls rollout", "rls migration", "enable rls", "tenant isolation", "row level security", "row-level security", "rls phase 1", "rls phase 2", "rls plan", "rls verify", "rls audit produkcije", "uradi rls audit", "krenimo sa rls", "rls staging". Do NOT use for: applying migrations without review (skill never does this), designing tenancy model from scratch (assumes existing organizationId schema), Redis cache audit (separate workstream), threat modeling (human task).

Scaffolds a NEW AgentStack agent with a DETERMINISTIC quality gate and anti-hallucination input-guard baked in by default — not the prompt-only guardrails that LLMs ignore. Builds the full flow (kb_search → processor → function-validator → condition-gate → pass/error emitters → terminal message nodes), wires the KB, and smoke-tests BOTH the happy path and a deliberately bad input to prove the gate fails closed. Use this whenever the user wants to "create an agent", "scaffold an agent", "build a new agent", "napravi/kreiraj agenta", "dodaj agenta u pipeline", "set up an AgentStack agent", or wants an agent that "won't hallucinate", "blocks bad output", "has a quality gate", or "is production-safe" — even if they don't say "deterministic" or "gate". Prefer this over plain scaffolding when safety, anti-hallucination, or fail-closed behavior matters. Do NOT use for debugging an existing agent (use a debugger skill) or editing one field (use as_patch_node_field directly).

Consolidated RLS rollout status dashboard across 4 systems: git, GitHub CI, Railway, and PostgreSQL. Produces a phase-tracking table with drift alerts. Read-only — never modifies files or DB. Triggers: "/rls-status-checker", "rls status", "check rls rollout", "rls health", "koliko smo daleko sa rls", "rls dashboard", "proveri rls", "rls phase status", "gde smo sa rls"

Runs a structured health check across the Agent Studio AgentStack system and produces a scored report with prioritized fixes. Covers 5 dimensions: broken flows (all agents), model availability, memory wiring (kb_search nodes), KB embedding status, and duplicate agent names. Uses 3 global MCP calls + per-agent calls only for production-scoped agents. Triggers: "health check", "system status", "check agents", "što je pokvareno", "šta ne radi", "agent health", "provjeri agente", "provjeri sistem", "status sistema", "health report", "before I run pipeline", "nešto ne radi u pipeline-u", "pipeline fails", "pipeline ne radi", "da li je sve ok", "is everything ok", "system check", "provjeri sve agente". Do NOT use for: running agents, fixing a specific known issue (use as_patch_node_field directly), adding KB documents, or scaffolding new agents.

Fully scaffolds a new AgentStack agent from spec to live deployment using SOMA-standard architecture. Automates: agent creation, standard flow building (kb_search → processor → extractor → optional web_search and call_agent nodes), Obsidian vault initialization (DESIGN_SPEC.md, agent-card.md, instincts.md, evo-log.md), KB seeding with 3 documents, and smoke testing. Triggers: "create a new agent", "scaffold an agent", "napravi novog agenta", "add agent to pipeline", "dodaj agenta u pipeline", "set up new AgentStack agent", "build a new agent from scratch", "I want a new agent that does X", "new agent", "kreiraj agenta", "napravi agenta", "hocu novog agenta", "napravi mi agenta", "build agent", "create agent". Do NOT use for modifying existing agents (use as_patch_node_field directly) or debugging flows.

Empirically verifies the 27 key claims from the Agent Studio V1 architectural audit by running real commands — typecheck, vitest, lint, knip, npm audit, license-checker. Produces a structured V2 verification report that confirms or refutes each claim, surfaces new findings, and provides patch instructions for V1 corrections. Zero side effects on the codebase. Triggers: "verify audit", "audit verify", "verifikuj audit", "potvrdi audit", "v2 audit", "validate audit findings", "audit verifikacija", "ponovo proveri audit", "phase 2 audit", "audit accuracy check", "potvrdi nalaze", "test the audit", "is the audit accurate". Do NOT use for: executing audit recommendations (future skill), running E2E or load tests (out of scope), threat modeling (human task), live request tracing (manual analysis).

Extracts patterns from SOMA agent evo-logs, proposes new instincts, and — after human approval — appends them to the correct instincts.md files in the Obsidian vault. Use when the user says "update instincts", "ažuriraj instinkte", "instincts update", "nauči agente iz evo-logova", "analiziraj logove za instinkte", "šta su agenti naučili", "provjeri greške u logovima", "izgeneriši instinkte iz logova", "what have agents learned", "extract patterns", "ekstrakcija obrazaca", "nauči iz grešaka", "izvuci obrasce", "dodaj instinkte", or after a batch of runs when the user wants to consolidate learnings. Do NOT use for general Obsidian writes (use obsidian-knowledge-logger), for running agents (use soma-run), or for checking pipeline health (use agent-health-check). ALWAYS wait for explicit human approval before writing any instinct to the vault.

Syncs Obsidian vault files into Agent Studio knowledge bases using True-Sync (ADD new, wait READY, DELETE old). Use this skill whenever the user wants to sync vault docs to agent KBs, update agent memory, keep KB in sync after editing instincts or evo-logs, run a KB refresh, or says things like "sync the KB", "update agent memory", "push vault to KB", "KB is out of date", "sync SOMA agents", "refresh agent knowledge", "sinkronizuj KB", "azuriraj KB", "sync instincts", or "uradi kb sync". Change detection uses SHA-256 hash comparison against stored contentHash so only actually changed files are re-ingested. Handles rate limiting, ADD-before-DELETE ordering, and per-document polling automatically. Do NOT use for creating new agents or KBs from scratch (use agent-scaffolder), diagnosing broken flows (use agent-health-check), or editing agent prompts directly.

Reactive diagnostic skill for the SOMA pipeline (TI → HW → CR → Score Analyzer). Accepts a problem description, runs structured multi-dimensional checks, maps evidence to root cause via explicit IF-THEN rules, auto-applies trivial fixes, and delivers a prioritized debug report. Replaces 5–8 manual MCP calls with one structured investigation.

Audits AgentStack agents for unwired kb_search nodes (missing knowledgeBaseId), proposes a fix plan, and — after confirmation — patches each node using only live MCP data. Zero hallucination tolerance: every ID used in a patch comes from a real tool response in the current session.

End-to-end SOMA pipeline runner: validates input, runs TI → HW → CR sequentially, captures outputs at each step, writes evo-logs to Obsidian, and logs winners. One skill call replaces manual as_chat_with_agent + evo-log-writer + winners-log-logger.

Periodic enterprise quality audit and improvement of AI agents against 2026 standards. Use this skill ANY TIME the user wants to audit agents, check agent quality, evaluate agents, improve existing agents, find underperforming agents, run a quality check, do an agent review, "proveri agente", "evaluiraj agente", "unapredi agente", "koje agente treba popraviti", or says anything like "let's review all agents" or "which agents need work". Connects to Railway PostgreSQL, scores every agent on a 10-dimension rubric, identifies agents below the 8/10 enterprise threshold, and generates improved system prompts. Always use this before deploying a new version of agent-studio to production.

Enterprise AI agent creation skill following Anthropic/Google DeepMind 2026 standards. Use this skill ANY TIME the user wants to create a new agent, add an agent, build an AI agent, define a new agent's purpose, or design an agent's system prompt — even if they just say "I want an agent that does X" or "kreiraj novog agenta" or "napravi agenta za Y". Produces complete, production-ready system prompts with all required XML sections: <role>, <output_format>, <constraints>, <failure_modes>, and <example> blocks. Never lets an agent go live with a generic "You are a helpful assistant" prompt.

Session management protocol for the agent-studio project. Use this skill at the START of every session and whenever picking a new task to implement. Also use it before any file write/edit operation and when git errors occur. Triggers on: "start session", "let's work on", "pick a task", "which task", "continue working", "nastavimo", "nastavi", "šta radimo", "počnimo", as well as any time a new Cowork conversation begins on this project.