heimdall

Suggest whether the machine needs a DISK or MEMORY cleanup — and catch the case where Heimdall ITSELF is the hog (runaway orphaned python from the presence keeper). Use when the machine is slow / swapping / laggy, disk or "System Data" is full, memory is exhausted, at the start of a long session, or when the user asks to free up space or RAM. Detection + suggestion only; hands off deep disk cleanup to the mac-deep-clean skill.

Autonomous superskill manager — the main orchestrator agent for complex development tasks. Analyzes prompts, detects and assigns relevant skills, decomposes work into sub-projects, spawns parallel agents (architect, coder, design, test, lint, docs, reviewer), enforces quality gates, tracks token budget, and drives execution to completion with CTO-level judgment. Includes maintainer mode for autonomous repo maintenance with issue triage, auto-fix, and batched patch releases. Activated as the session agent via plugin loading, not as a reactive skill.

The deliberate self-improvement loop for the maintainer — after every Nth maintenance cycle (or on /hmd:self-improve), step back from fixing issues and improve Heimdall's OWN capability. Ports karpathy/autoresearch to Heimdall's routing/planning: collect evidence from metrics.jsonl + queue dead/done stats → form testable hypotheses → run a bounded routing experiment → evaluate against a measured delta → keep the improvement ONLY if it beats a baseline on enough samples, else roll it back. Falsifiability over vibes. Use when a maintainer run has completed N cycles, when the user asks Heimdall to "get better over time", "improve itself", "tune routing", "learn from past runs", or invokes /hmd:self-improve.

Use when matching a React Native screen to a Claude Design HTML canonical at ≥95% visual parity on real Android/iOS hardware — closes the loop with a VQA stub mode, Playwright canonical renderer at 1080×2444, and a dual-metric pixelmatch + SSIM diff harness producing composite triptychs.