DeepSeek Harness Plugins · Memory
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
Bundle a DSH session—logs, subagents, media, and workspace files—into a single ZIP for portable export and full restore on any other instance.
Cross-session knowledge service for DeepSeek Harness backed by SQLite FTS5 trigram retrieval and L1 query expansion, enabling CJK substring search.
Automatic context curation plugin for DeepSeek Harness: tiered cleanup of tool outputs and dead logs, pinning critical user requirements, and archiving to memory before compression.
A MySQL storage backend for DeepSeek Harness that persists KV units with InnoDB ACID transactions, serving as a drop-in replacement for the built-in JSON backend with crash-safety and cross-process visibility.
Aggregates token usage from all local session logs, presenting overview metrics, trend charts, activity heatmaps, and model breakdowns in the settings page with a persistent daily-usage sidebar widget.
DSH-native persistent memory for DeepSeek Harness with multi-scope storage, session-aware retrieval, rule-based auto-capture, retention cleanup, and built-in secret scanning.
Idle-triggered context compaction for DeepSeek Harness that condenses stale session history after a quiet window, cutting wasted tokens on repeated requests.
Scans DSH session logs for corrupted records, losslessly repairs shape drift with automatic backups, and provides real-time watch to catch new violations before they break history loading.
Adds a floating per-session todo panel to DSH with server-side cross-device storage, drag-to-reorder, and a badge on sessions with unfinished tasks.
A time memory engine for agents that records state and action events, then performs trend, anomaly, and causal analysis to answer what happened and why.
Imports Codex conversation history into DeepSeek Harness sessions and creates or reuses workspaces from historical working directories.
Renders per-reply badges showing triggered skills and tools used or failed, derived from engine logs via projection with no model self-reporting.