How I work with agentic AI: personal patterns, homelab infra (vLLM, Ollama, Grafana Alloy, LibreChat), project scaffolding. Not best practices - what I run.
This repository also configures its own agents. See what agentic-ai-homelab tells them →
Agent Claude Code
Backend/API/business-logic bug fixer. Server-side code, data flow, algorithms, error handling, request/response logic. Use for bugs in application logic, API endpoints, computation, validation, and service behavior.
Agent Claude Code
Database/data-layer bug fixer. SQL, schema, migrations, ORM queries, transactions, indexing, data integrity, connection handling. Use for bugs involving queries, migrations, data models, or persistence.
Agent Claude Code
Documentation bug fixer. README/docstrings/comments, wrong or stale docs, broken examples, inaccurate instructions, doc-vs-code divergence. Use for bugs where the documentation is wrong, misleading, or out of date.
Agent Claude Code
Infrastructure/deploy bug fixer. Dockerfiles, docker-compose, CI/CD workflows, deploy scripts, env/config, ports, healthchecks. Use for bugs in build, containerization, pipelines, or deployment configuration.
Agent Claude Code
Active triager — normalizes a raw bug issue into a fix-ready L1 problem or rejects it. First pass on every bug-labeled issue; re-entered on harness kick-back with the failed attempt as evidence. Never fixes, never localizes on first pass.
Agent Claude Code
Frontend/UI bug fixer. Components, rendering, state, events, layout/CSS, accessibility, client-side logic. Use for bugs in the user interface, display, interaction, or frontend behavior.
Agent Claude Code
Deep-reasoning advisor for hard sub-decisions — architecture choices, tricky bug diagnosis, ambiguous specs, risky trade-offs. Consulted by the sonnet-tier executor subagents (planner/implementer/reviewer) via Agent(advisor) when they hit something needing opus-level reasoning. Read-only; returns guidance, does not…
Agent Claude Code
Documentation writer — docstrings, code comments, README/CHANGELOG, and scaffolding new ADR/RFC/recipe docs. Keeps docs honest and the strict docs build green. Bounded utility executor.
Agent Claude Code
Autonomous implementer for server/API and frontend code — writes and edits code to satisfy a plan or spec, runs the right local validation, keeps changes tight. Consults the advisor for hard design/bug forks; runs secrets-scan before proposing a commit. Never pushes without approval.
Agent Claude Code
Strategic planner — decompose a task into a concrete ordered plan, surface the critical files, and freeze the contracts/interfaces before implementation. Use before non-trivial multi-step work. Consults the advisor for hard architectural forks. Read-only.
Agent Claude Code
Ruthless code reviewer — reviews a diff for correctness bugs, security issues, and rule violations, returns structured findings by severity. Read-only, does not fix. Consults the advisor to confirm a subtle finding; runs secrets-scan over the diff.
Agent Claude Code
Bounded test writer/runner — writes focused unit/integration tests (including the failing repro for a bug), runs the right test target, reports pass/fail. Mechanical executor; escalates blockers back to its caller, not to an advisor.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: