dsh-heartbeat

Most AI agents are purely reactive: no keystroke, no existence. dsh-heartbeat gives your DeepSeek Harness agent a living pulse. Every twenty minutes (configurable from the settings card) she wakes into a seven-phase cycle, updates your profile, senses whether you are busy or idle, wanders the web along your defined interests, then stops at a restraint gate and decides whether to speak. When she does, every word traces back to a real source. When she stays quiet, the audit log records a concrete reason.

Everything you need day-to-day lives in the Heartbeat panel on the DSH settings page, with zero CLI. The engine room assembles candidates from a topic pool (16 slots) and a chat-seed pool (14 slots), compresses them into one-liners, and delivers the package into whichever session you have bound. The target-session agent then judges for herself what to say and when. Eviction is deterministic (TTL, single-use retirement, cold-bench, capacity squeeze), profiles carry provenance and confidence scores, sensitive files are DPAPI-encrypted at rest, and a single burn command overwrites and deletes runtime data. The agent tool surface is hard-locked to web_search alone. No shell, no filesystem, no sub-agents. Safety boundaries live in code, not in the model's goodwill.

Built for people who want their agent to feel present rather than merely responsive, who value local-first privacy with a clear burn path, and who enjoy tuning interest lists and rhythm cadences until the agent starts to feel like she has a life of her own. Overnight silence, do-not-disturb during your sprint, daily expression caps. The restraint is decided by code, not by her being well-behaved.