buildoak

Domain-agnostic TrueSkill batch ranking via LLM-as-judge. Ranks any list of text items using overlapping subsets dispatched to Codex Spark workers. Swappable rubrics. Use when you need to rank, score, curate, or sort a collection by quality.

Native macOS GUI automation via peekaboo 3.0+. AX-first hybrid with VLM fallback, FSM-based workflows with bug-specific transitions. Safari web automation, native app control, social media operations, system dialogs, cross-app workflows.

Video generation via fal.ai -- from quick one-off clips to multi-scene productions with keyframes, state tracking, and assembly. 10+ models including Kling 3.0, Veo 3.1, Sora 2, LTX-2.3. Two modes: light (fire-and-forget) and project (structured pipeline with gates).

Image generation, editing, and review via OpenRouter API. Five models from budget to premium. Style presets for series consistency, JSON structured prompts, reference image anchoring, system message support, prompt upsampling. Vision-based quality review loop. Zero dependencies beyond Python stdlib. Use when: generate, create, draw, design, illustrate, edit, or modify images.

Google Workspace CLI automation via gogcli. Gmail, Calendar, Drive, Docs, Slides, Sheets, and 9 more services. JSON-first output, composable pipelines.

Browser automation skill for AI coding agents. 25 Playwright-based tools for navigation, interaction, observation, and session management.

Search ChatGPT conversation exports using SQLite FTS5 (SQLite full-text search). BM25-ranked full-text search (relevance scoring) with TF-IDF keywords (term-weighted key phrases), date/role/model/language filtering, and conversation browsing. Use when agent needs to search past ChatGPT conversations by topic, find specific discussions, browse conversation history, or find conversations by extracted keywords. Do NOT use for non-ChatGPT knowledge bases — use a dedicated document search tool. Do NOT use for Apple Notes or Obsidian — use a dedicated document search tool.

End-to-end task orchestration across AI coding engines. Coordinates Claude, Codex, Codex Spark (the faster Codex variant), and OpenCode workers using dispatch-verify-synthesize workflows. Requires the upstream agent-mux CLI for worker dispatch. Use when work spans multiple dependent steps, benefits from model diversity, or needs structured quality verification.

Mac system monitoring skill. Returns CPU, memory, temperature, disk, and Docker status. Trigger when: - "How's the Mac?" or "How's the server?" - "System status" or "Server status" - "Check server health" or "Mac health" - "What's eating CPU?" or "What's using memory?" - "Disk space?" or "Storage status?" - "Temperature?" or "Is the Mac hot?" - "Docker containers?" or "Are containers running?" - Any question about Mac hardware health or resource usage

Check your posts for unanswered comments and handle replies efficiently. Use when asked to "patrol comments", "check for replies needed", "comment sweep", "handle unanswered comments", or "catch up on moltbook responses". Automates the full cycle: fetch posts, identify unanswered comments, triage by type, batch-reply with rate limiting, update memory.

Work on the tg-useragent-cli repository, an agent-first Telegram useragent CLI and skill. Use when updating its docs, command contracts, safety gates, implementation, tests, or agent workflows. M0/M1 plus scoped export/stats/chunk artifacts are implemented; live auth/status/dialog/resolve and bounded history sync are partially implemented; send, send-file, scoped reply, and scoped mark-read have dry-run/confirm safety gates and live executors; M5a media supports local photo/document listing, gated scoped download, and single local document upload.

Telegram user-account operations through the Go tgua useragent CLI. Use for schema/describe/doctor/config, authorized auth/status/dialog/resolve checks, local SQLite legacy import/peers/messages/search, sync dialogs/messages, recent channel/group DB rebuilds, scoped export/stats/chunk artifacts, local media listing, gated media download planning/confirmation, send --dry-run planning, gated confirmed text sends, gated single-document send-file, scoped reply, and scoped mark-read. Do not use for Bot API work, daemon lifecycle, legacy Telethon maintenance, subscriptions, or ungated config mutation.

CLI-based context compression for Claude Code sessions via wet proxy

Cross-engine dispatch layer for AI coding agents. Use when you need to: launch a worker on Codex/Claude/Gemini/agy/Grok, recover a timed-out dispatch, steer a running agent, or fan out parallel work. Profiles live in ~/.agent-mux/prompts/.

Cross-session memory and continuity CLI. Provides session memory, handoff documents, and historical session retrieval. Use at session start to retrieve context from past sessions (eywa get). Use at session end to persist a handoff document for future sessions (eywa extract). Trigger on: "session memory", "handoff", "continuity", "past sessions", "what was I working on", "session start", "session end", "batch index", "historical sessions", "prior context", "previous session".

Agent session observability CLI — turns AI coding sessions into first-class queryable artifacts. Parses Codex, Claude Code, Grok Build, Antigravity CLI (agy), Hermes, and Gemini traces, indexes into SQLite + Tantivy FTS, and renders token-efficient markdown views over raw evidence. Use when recalling prior context, attributing file/command activity, searching session history, inspecting sessions, rendering transcripts/activity, identifying the current session, generating optional continuity handoffs, or maintaining the local index. JSON default, -H for human. Gaal-shaped errors teach agents what went wrong + how to recover.

General-purpose Codex subagent for code-focused tasks using OpenAI Codex (default: GPT-5.3-Codex). Spawns an independent Codex agent the same way Claude's Task tool spawns Claude subagents. Can be invoked from Claude's main thread OR from Claude subagents (nested pattern). Trigger when: - User explicitly requests Codex ("have Codex do this", "spawn a Codex agent") - Cross-model verification or second opinion needed - Security review, vulnerability scan, or auth/crypto/payment code analysis - Bug hunting, debugging, or root-cause analysis - Code generation or implementation where Codex's style fits - Refactoring, cleanup, or large-scale code transformations - Codebase analysis or exploration - Any code-focused task that benefits from a different model's perspective - Nested subagent pattern: Claude subagent needs to spawn a Codex worker Prefer NOT to trigger for: - Simple tasks that don't benefit from model diversity - Single-file edits where Claude is already doing fine Do NOT use for: - Architecture plann

Claude Code token usage tracking and limit reverse-engineering. Daily/project/model breakdown, burn rate, context gauge, limit calibration. Active model: Formula A (CU = Σ_model[mult × (inp×1.0 + out×5.0 + cc×1.25 + cr×0.1)], limit 527.6M CU/week).