Agent tooling

Agent tooling is the set of small tools I use to make AI coding agents more reliable, cheaper to run and easier to supervise. wikillm is a TypeScript CLI and plugin bundle: an npx wizard scaffolds a vault with raw sources, compiled wiki articles, Obsidian config and host-specific CLAUDE.md and AGENTS.md schemas, and the plugin skills handle ingest, query, lint, Obsidian CLI, Marp and onboarding for Claude Code and Codex. babysit-pr (shown as PR Toolkit) carries a GitHub pull request or GitLab merge request through failing checks and review feedback, then records a short demo of the change working. berkays-mods is a set of Claude Code mods for orchestrator and worker sessions. claude-config is the cost-first setup those sessions run on: one Opus session plans and reviews, Haiku reads, Sonnet builds, each at an explicit effort level. htmlpub is a private workspace for publishing self-contained HTML reports and reviewing them with text-anchored comments, and yt-scribe turns a YouTube link into a transcript and then into readable notes through Codex or OpenCode.

• –wikillm vault scaffold: npx wikillm asks for personal or project mode, host targets, features, domain and optional automation guidance before it creates the vault structure
• –wikillm host-aware schemas: CLAUDE.md and AGENTS.md are generated from the same TypeScript renderer, with command prefixes and install/update copy kept in host profiles
• –wikillm compiled wiki: ingest skills turn raw files into one-concept markdown articles with YAML frontmatter, [[wikilinks]], provenance, indices and git commits; query reads the compiled articles first and lint checks links, frontmatter, orphans, contradictions, stale claims, duplicates and graph clusters
• –wikillm Obsidian and Marp integration: skills use the Obsidian CLI for search and backlinks when it is available and can generate slide decks from wiki content through Marp
• –babysit-pr: two skills for Claude Code and Codex on GitHub and GitLab. Babysit PR makes scoped fixes, validates and pushes, then reports readiness tied to the current PR head; PR Demo records a short video of the change working (Playwright for web, VHS for terminal apps) and attaches it to the PR with the recorded commit
• –berkays-mods: nine Claude Code mods (two adapted from other MIT projects), among them limit-resume (continues a session when the 5-hour limit resets), workers (a pane with every worker session), proof-gate (asks a worker that reports done without screenshots to send them), effort-gate (blocks Opus at xhigh/max in workers unless approved) and route-ledger (records every model and effort choice, with /routing-review)
• –claude-config: shared rules, three agents (scout on Haiku, builder on Sonnet, reviewer on Opus), an orchestrate skill that runs a feature as parallel background sessions in git worktrees, and installers for macOS, Linux and Windows that merge the shared settings into each device
• –htmlpub: Clerk-authenticated Next.js dashboard, immutable version history, text-anchored owner comments, a blocking htmlpub review wait so an agent can publish and continue once the review is decided, revocable share links, and an isolated cookie-free renderer for the published HTML
• –yt-scribe: fetches public YouTube captions without downloading video or audio and polishes them into notes, summaries, cleaned text or article-style prose; stable JSON output, an optional MCP server, and agent skills installed with yt-scribe setup

• –Keeping Claude Code and Codex first-class without duplicating the whole workflow
• –Creating vault instructions without clobbering a project's own root CLAUDE.md or AGENTS.md
• –Making a markdown wiki behave like a maintained knowledge base instead of a loose notes folder
• –Being honest about automation: wikillm can record commands and prompts, but the host app owns scheduling, and a monitoring skill does not outlive the session unless the host provides an automation facility
• –Treating a green check as not enough: babysit-pr ties readiness to the current PR head and names a specific blocker (pending checks, missing approvals, conflicts, access) instead
• –Guardrails that only see text: effort-gate reads the literal command a session runs, so it denies what it cannot parse but cannot see through variables or scripts written and then run

Each tool is small and keeps its limits explicit. wikillm separates pure vault planning from disk writes, centralizes host differences in profiles and renders schema files in-process so setup is deterministic; the skills own the long-running behavior, with ingest compiling the wiki, query reading compiled articles first and lint keeping graph health visible. babysit-pr and htmlpub make the agent's work reviewable: a readiness report or a recorded demo for a PR, a versioned page with anchored comments for a report. berkays-mods and claude-config move the supervision into the setup itself, with explicit model and effort per worker, a gate on the expensive choices and a ledger to review them afterwards. yt-scribe keeps the human path to one command and gives agents the same CLI with stable JSON.

Six public repositories under github.com/Berkay2002, each installable on its own: wikillm as an npm CLI plus a Claude Code and Codex plugin bundle, babysit-pr as a plugin for both hosts, berkays-mods and claude-config for Claude Code, htmlpub as a CLI and plugin on top of a web app you deploy yourself, and yt-scribe as a Python CLI. Together they give agents a durable knowledge workflow, a way to show a PR working, and a setup where model choice, effort and review are explicit.

• wikillm ↗

  CLI and agent plugin that compiles raw sources into a cross-linked Obsidian markdown wiki, with no vector store or RAG pipeline. Ingest, query, lint, Obsidian CLI, Marp and onboarding skills for Claude Code and Codex.

  TypeScript · Node.js · Obsidian · Marp

• babysit-pr ↗

  PR Toolkit: Babysit PR fixes scoped CI and review feedback and reports verified readiness; PR Demo records a short video of the change and attaches it to the PR. GitHub and GitLab, Claude Code and Codex.

  Agent skills · Playwright · VHS

• berkays-mods ↗

  Claude Code mods for orchestrator and worker sessions: usage-limit resume, a workers pane, session identity, a proof gate, an effort gate, a routing ledger and config sync.

  Claude Code Plugins · TypeScript

• claude-config ↗

  A cost-first Claude Code setup: Opus plans and reviews, Haiku reads, Sonnet builds, each at an explicit effort level. Rules, agents, an orchestrate skill and installers that keep every device on the same config.

  Claude Code · Shell · PowerShell

• htmlpub ↗

  Private workspace for publishing self-contained HTML reports: immutable versions, text-anchored comments, blocking agent review waits, revocable share links and an isolated renderer. Includes a Node CLI and Codex and Claude Code plugins.

  Next.js · Neon Postgres · Drizzle · Clerk · Vercel Blob

• yt-scribe ↗

  Python CLI that fetches public YouTube captions and polishes them into notes, summaries or articles through Codex or OpenCode. Stable JSON for agents and an optional MCP server.

  Python · youtube-transcript-api · MCP