work-suite

Keep an explicitly delegated long-horizon task moving through audited durable state, terminal delivery, and the next ready item without returning routine control.

The research-shaped sibling of `autopilot`: drive a deep investigation to completion in one go, never handing control back mid-dive. Invoke when the principal says "deep dive on X", "investigate X", "research X", "get to the bottom of X", "go to work on this", or "finish the investigation" — any mandate to investigate a topic by reading real sources (code, docs, issues, telemetry) and producing grounded, captured findings. Also fires when the agent catches itself about to hand control back with a readable-but-unread thread, findings only in chat, or a "want me to keep going? / loose ends:" handoff — mode violations here. Terminal state is the sweater test: pull every thread until there's no more sweater AND every thread is routed to its durable home. Two narrow exceptions, as autopilot (a human-only credential/access wall; an irreversible destructive op); everything else is keep pulling. Do NOT invoke for a quick single lookup or an explicitly interactive "let's think this through" session.

Open-ended codebase improvement loop. Start with one concrete seed issue, fix it, log side observations to a bundle task backlog, pick the next highest-value discovery, repeat, and close the backlog when done. Each fix is its own PR.

Keep long-running CI, deploy, merge, smoke, and batch work in the current turn by choosing the host's native notification path or a bounded foreground fallback.

Run this whenever work changes UI, rendered documents, CSS/layout/typography, visual assets, native app surfaces, screenshots, or any user-reported visual weirdness. Forces screenshot-backed dogfooding and adversarial visual review so metric-only checks cannot miss obviously broken UX.

Invoke when the operator activates a watchdog mandate -- phrases like "keep an eye on X", "watchdog this", "monitor Y across silences", "keep watching", "let me know if anything smells off". Watchdog mode is autonomous monitoring + repair across operator silences, with a strict scope of what worker takes action on vs. what worker surfaces. The skill encodes the cadence, the scope of acting-vs-surfacing, the carve-outs that keep watchdog from overreaching, and the exit signal.

Execute an authorized engineering task with strict TDD, complete changed-production coverage, and the smallest implementation that reaches a tested branch.

Explore an ambiguous engineering problem just far enough to expose the simplest viable shape. Use when the request has a real product, architecture, or scope fork; skip when the task is already clear.

Drive an authorized finished branch through current-target sync, exact-head checks, verified remote outcome, release or install, smoke, cleanup, and continuation.

Turn coordinated or risky engineering work into the smallest executable planning contract. Skip trivial work, reuse existing planning artifacts, and review in proportion to risk.