workbench

Build a reliable browser-automation script for a website by first recording the flow with Playwright codegen, then translating it into a replayable script using agent-browser (preferred) or falling back to puppeteer. Use whenever the user wants to "automate this site", "write a scraper for X", "build a fetcher / downloader for Y", "log in to Z and grab N", "turn this manual web flow into a script", or asks for a new finance fetcher / bank statement downloader / web scraper / form filler / login automation - even if they don't explicitly mention Playwright, puppeteer, or codegen. Especially relevant when the target site has logins, iframes, OTP, downloads, or anti-bot defenses that make naive automation flaky.

Refactor logging and console output to follow structured logging standards — color scheme, log levels, prefix pattern, and output flow. Use when the user asks to clean up logging, standardize output, add structured logging, fix console messages, or when reviewing code that has inconsistent print/log/echo statements. Also applies when creating new CLI scripts or commands that produce user-facing output.

Set up, extend, or debug a `varlock` schema in a Node.js/TypeScript project - schema-driven env with validation, type generation, and 1Password secret resolution. Use this skill whenever the user asks to adopt varlock, migrate off dotenv, set up `.env.schema`, add a new var to an existing varlock schema, wire up the `@varlock/1password-plugin` (`op(...)` refs, `@initOp`), diagnose "Unknown resolver function" / "No 1Password plugin instances found" / varlock parse errors, or investigate why a varlock-loaded script exits silently. Also use when the user mentions "AI-safe .env", schema-first env, or wants to commit env shape without leaking secrets.

Examine uncommitted changes and split them into logical, well-scoped git commits. Handles untracked files by either staging them or adding them to .gitignore. If the repo is a submodule, also records the pointer bump in the parent repo. Runs on haiku in a forked subagent; reports back what it did, reminds the user to push, and flags any open questions.

Merge the current branch into the repo's default branch locally with a --no-ff merge commit. If run from a linked git worktree, it also commits pending changes, merges the worktree branch into the default branch in the main checkout, then removes the worktree and deletes the branch. Use when the user says "merge this locally", "merge no-ff and clean up the worktree", "finish/land this worktree", "merge my branch into master locally", or similar. Does not push and does not build.

Event-driven monitoring of another herdr pane from an agent: run the bundled debounced watcher as a BACKGROUND task so the harness re-invokes you only when the watched pane's content settles after a real change (resize/scroll-immune, no wasted turns). Use when running inside herdr (HERDR_ENV=1) and you need to watch a sibling pane's output and react step-by-step - guiding a user through a setup wizard or installer they're driving, waiting on a long build or streaming logs, or following an interactive CLI - especially when you want to comment ONLY when something actually changes. Triggers: 'watch this pane', 'tell me when the pane changes / is done', 'monitor the wizard/installer/logs in the other pane', 'ping me only if it changes', 'guide me through X in a pane and react as it advances', 'loop every N seconds but only speak on change'. Optionally tracks progress and action items in a user-provided markdown file (e.g. a project wiki page) and updates it as steps complete.

Investigate high CPU (and related resource) usage on the local machine, including tracing it into Docker containers. Use when the user says the machine/fans are running hot, something is "eating CPU", the system is slow, or asks what's using CPU/Docker CPU. macOS-focused.

Land the current branch and clean up after it - picks the right route automatically (GitHub PR self-merge, or a local no-ff merge when there is no GitHub or pushing straight to the default branch is allowed), then removes the worktree it ran from. Use whenever the user wants a finished branch landed - "merge my PR", "self-merge", "merge this branch", "land this", "ship it", "finalize and clean up", "wrap up this worktree". Creates the PR when none exists, enables auto-merge, polls until it merges or CI fails, then removes the worktree and deletes the branch. Refuses to run from the default branch or a detached HEAD.

Take over and drive a specific orchestration task's in-pane gates to completion, on the user's explicit request. Use inside a herdr orchestration session (see `/workbench:orchestrator-init`) when the user names one or more tracked tasks and tells you to push them through - phrases like "drive it to completion", "drive Dev-connect and Composer", "push the composer task through", "answer its gates and keep it moving". This is the deliberate exception to the standing "don't proxy in-pane gates" rule, scoped to exactly the named task(s). If the user has not explicitly asked you to drive a named task, do NOT use this skill - leave gates for the user to answer in the pane.

Initialize an orchestration session's live status document (the "target") and keep it current. Use at the START of any orchestration/multi-task session - whenever the user says things like "start an orchestration session", "orchestrate these tasks", "here's the status doc / target", "track everything in this doc", or "we'll open several tasks and monitor them". Also use whenever the user hands over a markdown file to serve as the session's status target. If a session is about to run several tasks in parallel and no status target is established yet, invoke this skill.

Schedules a work meeting end-to-end: resolves attendee names to emails, finds a mutually-free timeslot honoring working-hours and clustering preferences, and opens a prefilled Outlook invite ready to send. Use whenever the user wants to "schedule/set up/book a meeting", "find a time with <person>", "invite <person> to a meeting", "put something on my and X's calendar", or gives a subject + people + rough time. Handles optional people, date, time, and duration (default 30 min). Resolves names via Slack/Outlook and disambiguates using the vault people registry. Distinct from /workbench:task-obsidian (personal to-dos) and /workbench:task-jira (work tickets).

Delegate a dev task by spawning it as a real `claude` agent in its own tracked herdr tab, instead of an Agent-tool subagent. Use when running inside herdr (HERDR_ENV=1) and the user wants a task run in a visible, tracked tab they can watch or talk to - phrases like "spawn this in a herdr tab", "run it as a tracked tab", "do this the herdr way", or when running several such agents in parallel and monitoring them together. If HERDR_ENV is not 1, fall back to normal Agent-tool subagents (see `/workbench:orchestrate-agents`).

Reads, creates, edits descriptions, comments on, transitions, and searches JIRA issues using the `tasks` CLI (a Python single-file script from the ai-enablement repo, callable from anywhere once installed). Use this whenever the user asks to do anything JIRA-related: look up a ticket ("what is AIE-282", "show me FOO-123"), create or update an issue, change/append to a description, post a comment, move a status, list issues under an epic, fetch custom fields, or run a raw Atlassian REST call. Auto-triggers on bare KEY mentions like AIE-NNN / DRV-NNN, and on words such as "JIRA", "Atlassian", "issue", "ticket", "epic", "sprint", "transition", "status". Works in any folder; does not require being inside the ai-enablement repo.

Lists, opens, and creates personal tasks in the user's Obsidian vault. Tasks are plain `- [ ]` checkbox lines living in the notes that explain them, found with ripgrep - there is no task database and no CLI. Use this whenever the user asks about their personal tasks / to-dos / "what's on my list" / "what am I working on" / "add a task" — anything that lives in the Obsidian vault rather than in JIRA. Auto-triggers on phrases like "my tasks", "my todos", "my to-do list", "obsidian task", "vault task", "add to my list", "what should I do next". Distinct from `/workbench:task-jira`, which handles work tickets (AIE-NNN / DRV-NNN).

Read from and write to the user's personal Obsidian knowledge vault at `$OBSIDIAN_VAULT` (default `~/homebot/obsidian`) - wiki pages, contacts, calendar events, finance/receipts, kids-money, "tell someone" items, day logs, movies, storage, and other structured data tables. Use whenever the user wants to capture, look up, route, or update personal knowledge in their vault: "add this to my wiki/notes", "save this link", "remember that...", "note that...", "log this", "what do my notes say about X", "look up X in my vault", "add a contact", "create a calendar event", "tell P / tell someone about X", "journal this", receipts / bank exports / manual payments, or any free-form note that belongs in the vault. The vault's own `CLAUDE.md` is the routing source-of-truth; this skill just sends you there. Distinct from `/workbench:task-obsidian` (personal to-dos / tasks) and `/workbench:task-jira` (work tickets) - reach for those when the input is an actionable task or a JIRA ticket rather than knowledge to file.