osouthgate

agent-plus | Trigger doctrine for Claude Code — when to proactively probe for and offer agent-plus framework upgrades. Pure markdown skill; the runtime is the existing `agent-plus-meta upgrade-check` / `agent-plus-meta upgrade` subcommands (frozen v0.13.5 contract) — snooze ladder, `update_check`/`silent_upgrade` config, `.bak` rollback, and a post-upgrade doctor gate are already built. Probes at most once per session, only surfaces an offer on `upgrade_available` with no active snooze, and never mutates without an explicit user choice.

agent-plus | Trigger doctrine for Claude Code — when (and when NOT) to offer to install the agent-plus framework on the user's behalf. Pure markdown skill; the runtime is the existing `install.sh --unattended` one-liner that chains into `agent-plus-meta init --non-interactive --auto`. Surfaces an offer, never auto-executes; honors a session-scope decline flag; gates every offer on a `command -v agent-plus-meta` probe so installed users are never pestered.

agent-plus | The meta plugin for the agent-plus framework. Workspace bootstrap, env-var readiness, identity cache, marketplace lifecycle, extensions. Creates `.agent-plus/` (one shared dir with skill-feedback / skill-plus), reports which sibling-plugin env vars are set (names only), and caches resolved project IDs / service handles into `services.json` so subsequent calls don't re-discover them. Use at session start, when env config changes, or when an agent asks "is X configured here?".

agent-plus | One-call structured triage of a git diff. Returns per-file role classification (source/test/config/doc/generated/build/fixture/migration), risk tier (low/medium/high) with reasons, public-API touch detection, co-changed-test detection, secret-risk path flagging, and aggregate stats. Replaces the 5-20 Read calls an agent burns reading each modified file individually to figure out "what kind of change is this and how risky is it". Stateless, no network, stdlib Python only.

agent-plus | One-call cold-start orientation for an unfamiliar repo. Returns a structured map — file tree (compact per-folder summary by default, full per-file listing with --tree-mode full), language mix, framework + build-tool detection, top-level deps, entrypoints, README highlights — so you don't burn 60+ Read/Glob/Grep calls re-mining the same facts every fresh session. Stateless, no network, stdlib Python only.

agent-plus | Local-first self-assessment for Claude Code skills. After you use any skill (your own or someone else's), call `skill-feedback log <skill> --rating N --outcome ... [--friction "..."]` to append one entry to `~/.agent-plus/skill-feedback/<skill>.jsonl`. The skill author can then run `skill-feedback report` for an aggregate, or `skill-feedback submit <skill>` to bundle entries into a GitHub issue body for the skill's source repo. No telemetry leaves the machine unless the user explicitly runs `submit`.

agent-plus | Mine the Claude Code session log to find commands the user keeps typing by hand, then scaffold them into proper Claude skills under `.claude/skills/<name>/`. Read-only audit of existing skills against the framework contract via `list`. Cross-source feedback aggregator joining `skill-feedback` ratings with implicit session-mining failure signals via `feedback`. Promote project-local skills to a `<user>/agent-plus-skills` marketplace via `promote`. Stdlib Python, local-only, no network.

Reviews code for quality, bugs, security, and best practices. Use when changing code or before merge. Checks security, authentication, and validation; ensures new code follows project rules and conventions.

Read-first wrapper around the GitHub REST API. Single-call PR/branch overviews (PR state + checks + reviews + workflow runs + mergeable) for triage. Use whenever the user wants the state of a PR, a branch's CI, or a failing run — without you chaining `gh pr view`, `gh pr checks`, `gh run view`, `gh pr list --head` per investigation. Resolves PRs and issues by branch/title. `run wait` polls on a run ID or branch. One write surface, `pr comment`.

Manage Langfuse instances (cloud or self-hosted) from the CLI — export/import prompts for backup and cross-env migration, send smoke-test traces, check health across multiple instances, and inspect a specific user's recent activity (sessions, traces, errors) read-only from the public REST API. Use whenever the user wants to back up prompts, move prompts between Langfuse instances, verify that ingestion is working, check whether a Langfuse deployment is up, or debug what a given Langfuse user has been doing without clicking through the UI.

Read-first wrapper around the Linear GraphQL API. Single-call issue context (comments + relations + state), name-resolved teams/projects/states/labels, `issues create --from-markdown` turns a design doc into a Linear issue without hand-stitching flags. Use whenever the user wants to read, triage, or write Linear issues — creating an issue from a markdown doc, fetching issue context (including comments and relations) without chaining 5+ MCP calls, listing project work by state, or moving an issue and waiting for webhook-driven transitions. Wraps a static personal API key, so it sidesteps the MCP OAuth browser-auth wall entirely.

Read-first wrapper around the Railway CLI. Single-call env overviews (services, deploy status, recent errors/warnings, env var NAMES-only) for fast incident triage. Use whenever the user wants to see the state of a Railway environment — what's running, what's broken, which env vars exist on a service — without you having to chain `railway list`, `railway service status`, `railway logs`, `railway variables` per service.

Manage a Supabase project from the CLI — list projects, run SQL files (with post-run row-count verification), run inline queries, audit RLS coverage, and generate TypeScript types. Wraps the Supabase Management API + the local `supabase` CLI so agents don't have to hand-parse the untrusted-data envelope.

Read-first wrapper around the Vercel REST API. Single-call project overviews (deployments, domains, env NAMES-only, warnings) for incident triage. Use whenever the user wants the state of a Vercel project — what deployed, what's failing, which env vars exist, which domains are verified — without you chaining `vercel list`, `vercel inspect`, `vercel env ls`, `vercel domains ls` per project. Also covers deployment logs via `logs` and async deploys via Deploy Hooks with `--wait`.

Check OpenRouter credit balance, aggregate per-key usage stats, search and filter the 350+ model catalogue by price / context / capability, and manage API keys via the provisioning endpoint. Designed for both interactive use and cron jobs (balance alerts, usage summaries).

Manage a remote Hermes Agent deployment from the CLI — list/create/pause/trigger cron jobs, check status, inspect env vars, swap models, and send one-shot chat calls to the OpenAI-compatible /v1/chat/completions endpoint. Use this skill whenever you need to control a Hermes instance that lives on another machine (VPS, Coolify app, etc.), since the upstream `hermes` CLI only talks to a local gateway.

Plan and autonomously build a software task end-to-end. Triggered by `/supergoal`, "plan and ship X", "supercharged plan", "autonomous build", "plan it out and don't stop until it's done", "I don't want to babysit this", or any non-trivial feature/refactor/redesign the user wants driven to completion. Strongly prefer over a plain plan when the user signals "every aspect", "fully", "perfectly", "until done", or wants depth + autonomous follow-through. Recons the codebase, applies preloaded memory, researches best practices with whatever tools are available, decomposes into the right number of phases, gets one confirmation, then prepares a single ready-to-paste `/goal` command — one paste between you and done — that drives the entire chain to completion with built-in retry, fix-spec recovery, and per-phase memory writeback. Works on Claude Code and Codex.

Configure Claude Code for a long unattended run — set the operating posture so Claude keeps working autonomously for hours or days without babysitting. Triggered by `/autonomous-run`, "run this overnight", "run Opus autonomously", "set up a long autonomous run", "I want to close my laptop and let it work", "keep going until it's done without asking me", or right before dispatching a big `/goal`/supergoal. This is a setup/posture skill, not a planner — it wires up the five levers that make unattended runs reliable (auto permissions, dynamic-workflow fan-out, /goal or /loop persistence, cloud execution, end-to-end self-verification), detects which are available this session, and hands off cleanly. Pairs with the `supergoal` skill (which does the planning + phase execution); run this to harden the run before or at supergoal's dispatch.

Extract transcripts from YouTube videos, local audio/video files, and summarize transcript files. Use when the user asks to download, fetch, extract, transcribe, caption, or summarize a YouTube video, a YouTube URL, a video ID, a local media file such as .mp4/.mov/.mp3/.m4a/.wav, or an existing transcript/text file. Supports captions-only, chunked local media transcription through OpenAI, YouTube audio fallback through OpenAI, and summaries through OpenAI or opt-in Ollama.