Plugins

Self-built DSH memory plugin: SQLite persistent storage + database-level audit guard (enforceAudit) + audit log. Auto-captures knowledge after conversations, agents recall via memory_recall, supports Claude Code/Codex/Cursor multi-agent sharing. Zero external dependencies.

Filters tool outputs, prunes stale context and discloses skills on demand for DSH 0.2.1-alpha.1.

Context engineering and token optimization plugin for DeepSeek Harness (DSH): tool output filtering, context pruning and progressive skill disclosure for LLM coding agents

Five tools for reading a video as raw material rather than as something to watch. video_outline maps a whole file as text and draws nothing at all: 4-60 equal windows, each with its cut count, speech share, intensity bar and tags, then the windows worth opening and the exact start/end for the call that follows -- zero image tokens. video_probe returns metadata and a reading plan without decoding frames. video_storyboard turns a window into one contact sheet plus a text index binding every cell to its timecode, cut count and frame time; cells whose text would be illegible at that scale are marked zoom, and dedupe global folds a picture that comes back after a cut. video_zoom re-renders 1-9 cells from a short window at a larger scale. video_read decodes only keyframes, OCRs each screen, and returns an outline plus \[mm:ss] text -- the path for lectures, slide decks and screen recordings, where the words are the content. It also reads a file with sound and no picture as a transcript (subtitle track, then DSH_ASR_CMD, then a transcript file you pass; long audio is cut at 45 s onto silence), composes six changed frames into one sheet when a file has no readable text at all, and can write the result out as SRT, WebVTT or text. Cuts, dissolves and one-frame flashes are found in a single ffmpeg pass; from ten minutes up a file is analysed on its keyframes only, 2.7x faster on a two-hour lecture; and a cell is never shrunk below a fifth of the source, so a 1440p sheet spends the same tokens on fewer, bigger cells. OCR is the engine Windows ships, tesseract with line boxes, or DSH_OCR_CMD. It takes a B station link (bilibili.com/video/BV..., a bare BV id, an av id or a b23.tv short link) wherever it takes a file: up to 300 seconds is read from the CDN in place with nothing written to disk, longer videos are downloaded once and cached within a 4 GB / 24 video budget, play URLs are cached for ten minutes, and a link that is not a single video says what it is. Needs ffmpeg, uses ffprobe when present, and has no other runtime dependencies. It is a wire-compatible reimplementation of SAOS-3624/dsh-video-storyboard: it registers the same video_probe / video_storyboard / video_zoom tool names, so the two cannot be enabled together.

Forgetting-curve in-place dialogue compression for DeepSeek Harness: recent turns stay verbatim, middle turns are structurally dehydrated (zero-LLM), ancient turns fold into a bounded summary pair — unbounded conversations on a bounded memory budget.

On-demand memory for DeepSeek Harness: session logs are distilled into layered markdown (a summary, an index and per-topic files), and a mem_query tool returns only the matched entries rather than the whole index. Nothing is injected per turn, there are no dependencies, and every entry keeps a source pointer back to the conversation it came from.

One plugin brings the rtk token-saving setup from Claude Code to DeepSeek Harness: commands rtk can rewrite are rewritten, and builds, tests and logs that rewriting cannot reach are compressed too; common commands such as git, cargo, npm, pytest and docker are covered (rtk itself supports 100+ commands). After compression the model is told what was elided and how to recover it — when the sandbox cannot reach the rtk store, the original is saved separately and its path is given — and `/rtk stats` shows how much was saved. When a rewrite is unreliable the original command runs as usual: if rtk cannot run, or the command rtk returns would drop arguments, nothing breaks. The rtk binary ships with the package, so mainstream platforms need no separate rtk install, unlike most other rtk plugins.

Context curation for DeepSeek Harness: content-aware tool-output slimming, dead-log sweep, pinned user requirements and in-flight work, archive-then-compact with an audit trail.

File-state cache skill for coding agents: one explore call replaces ls+grep+read, reads remember they already happened, and edits return a diff plus the exact changed line range instead of a full re-read. State is a JSON cache on disk, so it survives across tool calls and restarts. Requires Python 3.8+ (standard library only).

Multi-agent frugality defense plugin: read-ledger dedup metrics, compaction-immune rules, completion gate, cheap-model review lane.