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 →
Skill Claude CodeCodex
Close a work arc — write the next docs/history/NNNN- .md continuity entry, update the history index and mkdocs nav, strike completed items in docs/wip/NEXTSTEPS.md, append any D-NNNN operational decisions to 0002-decisions.md, and suggest the version tag. Use when wrapping up a multi-step session or a milestone. Keeps…
Skill Claude CodeCodex
Pre-deploy sanity check for Docker Compose stacks. Validates each compose file with docker compose config (read-only), then flags undefined interpolation variables and missing .env before the stack is brought up. Use before docker compose up, when editing a compose file, or whenever asked to validate or check a…
Skill Claude CodeCodex
One-shot read-only health snapshot of the DGX host — GPU mode, GPU utilization and compute-app count, nvidia-smi memory and processes, and running containers. Use to answer "what is the DGX doing right now" before deploying, switching GPU mode, or debugging a stuck endpoint. Strictly read-only; never changes state.
Skill Claude CodeCodex
Validate that the docs site builds cleanly in strict mode and the changed Markdown passes the repo's own markdownlint config, before committing or pushing docs changes. Detects the project's docs toolchain (Makefile docs target, MkDocs, Sphinx, Docusaurus, mdBook) and runs its strict build, then reports PASS/FAIL with…
Skill Claude CodeCodex
Show per-subagent invocation counts and token/cost spread by model tier from the current Claude Code transcript — proves whether the model-tiered fleet (haiku/sonnet/fable) actually spreads token cost. Use to see who was invoked, how often, tokens per agent/model, and escalation rate. Local and read-only (no…
Skill Claude CodeCodex
Resolve a homelab service name to its real endpoint URL and how to call it — DGX model servers, the mini's observability and LLM gateway services, prod VPS services. Points at the operations handbook, which holds the one maintained table. Use when an agent or client needs to point at a homelab service, when writing…
Skill Claude CodeCodex
Scaffold a new documentation file — ADR, RFC, recipe, or wip note — from the repo's conventions, assign the next number/slug, wire it into the mkdocs nav, update the section index, and keep the strict build green. Use when starting an ADR/RFC/recipe/design note, or when asked to write up a decision or a runbook.
Skill Claude CodeCodex
Check or (re)boot the self-hosted homelab observability stack — VictoriaMetrics/Logs/Traces + Grafana + GlitchTip + Langfuse on the Mac mini, Alloy collectors on the mini, the DGX and prod. The check (endpoints, containers, collector loops, per-host metric freshness, alert-rule drift) is read-only and scripted…
Skill Claude CodeCodex
For a real bug, write a failing test (or fixture / matrix row) that reproduces it BEFORE fixing — the test becomes the regression guard. Use when starting on a reported bug, a production incident, or any "it's broken" investigation. Enforces reproduce-before-fix and never-invent-root-causes.
Skill Claude CodeCodex
Scaffold a per-folder AGENTS.md and wire it into the repo-root AGENTS.md scoped table. Use when adding folder-specific agent rules, splitting a monolithic root AGENTS.md into per-folder files, or when asked to create or scope an AGENTS.md for a subdirectory. Enforces that AGENTS.md is the source of truth and CLAUDE.md…
Skill Claude Code
Scan for secrets (API keys, tokens, private keys, credentials, real .env files) before committing or pushing — enforces "never commit secrets". Prefers gitleaks if installed; otherwise uses high-signal patterns. Reports findings by file:line with the value redacted. Use before any commit/push, when adding config or…
Skill Claude CodeCodex
The pre-push ritual — rebase onto the trunk, run the right validation gates, confirm green, show the diff, push only with the operator's approval, then watch CI on what was pushed until it settles. Covers feature branches with PRs and the direct-to-main path for small fixes on trunk-based repos. Enforces…
Skill Claude CodeCodex
Safely bring a local vLLM stack up. Runs the read-only preflight (gpu-mode check, compose-check, .env presence) then hands you the exact deploy command for approval. Use when deploying or restarting coder-next / autoresearch / openwebui. Never runs up/down itself — the mutating step is gated on explicit approval.
Skill Claude CodeCodex
Review one project's observability over a time window, read-only. Covers log volume, warning and error patterns (normalized and counted), GlitchTip errors (new vs. old), traces (slow operations and outbound calls), and what the triage fleet did on the project's repo, with an optional handover file. Use for "what…
Skill Claude CodeCodex
Prove a behaviour change on real history before it ships. Freeze real evidence into a corpus, replay it through the old code (from git, never re-implemented) and the new code, check the replay is faithful, mutation-test the replay, read the changed cases against explicit links, then gate it. Use for any change to…
Skill Claude CodeCodex
Weekly health and quality review of the homelab triage fleet (signal-fleet, Fleet 2) across every repo it files into (podcastscraper, orrery, agentic-ai-homelab), plus triage of the proposals it parks for the operator. Use when the operator asks "how is the fleet doing", "what did the fleet do this week", "what can we…
Skill Claude CodeCodex
Compare a branch against its trunk base by running the same make targets in both (a detached worktree at the merge base, never a stash), each tree with its own src/ first on PYTHONPATH, and print a target × {base, branch} PASS/FAIL table plus the error lines that exist only on the branch. Use when a gate fails on a…
Skill Claude CodeCodex
Run a function from the project's own TS/JS code inside a real headless Chromium (real decoder, real fetch, real CORS) and get its result as JSON — bundled with the project's bundler, served on an ephemeral port, torn down afterwards. Use when a unit test's fakes cannot answer the question, when asked to "check it in…
Skill Claude CodeCodex
Test a GitHub workflow or composite-action step locally before pushing — extract one run: step, run it under bash (never zsh) the way the runner does, with a fresh HOME/RUNNERTEMP/GITHUBENV/GITHUBOUTPUT and every ${{ … }} supplied explicitly, as named pass AND fail cases. Plus verifyunits.sh, which checks systemd unit…
Skill Claude CodeCodex
Watch every GitHub check on one commit until it settles — one line per check as it lands, the failing job's log on the first failure, cancelled checks reported as NOT RUN (never as a pass), and a GREEN / RED / INCOMPLETE verdict. Resolves short SHAs, bounded, read-only. Use after every push, when asked "is main…
Skill Claude CodeCodex
Before/after gate for a deploy's telemetry — snapshot before, then check that every service still logs with environment=prod, traces carry the deployed release, scrape jobs are up and counters have series, no GlitchTip issue appeared since the deploy (each new one listed with status), Langfuse app traces carry the…
Skill Claude CodeCodex
Run a multi-line, quoted or JSON-bearing command on a homelab host (the Mac mini, the DGX, the prod VPS) from a script file instead of inline over ssh — it copies the file, runs it with bash or python3, streams the output, returns the remote exit code and deletes the copy. Use whenever a remote command has more than…
Skill Claude Code
Run any command that may take more than 10 minutes (a full e2e or translation run, a large build, a data pipeline, a migration) detached from the tool call, with a log ending in EXIT=, a PID file, and a lock so two jobs writing the same files can't run at once. Returns the exact Monitor command to watch it. Never hits…
Skill Claude CodeCodex
Diagnose a slow or stalled endpoint or job from evidence — observability first, then a local reproduction (including a GIL-contention bench that times the request path idle, under in-process CPU load, and with the load in a child process), then a before/after proof against the old code. Use when "an endpoint hangs /…
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: