Plugins

Folder time machine for DeepSeek Harness: snapshot any folder, browse its timeline, roll back with one click — Git-like history without needing Git, for folders that never had a repo (game saves, app config dirs, loose documents). Snapshots are content-addressed and deduplicated, and unchanged files are not re-hashed. Automatic snapshots come in two modes: realtime (fs.watch — one version ~8s after the last change, so a save-all or a build counts once) and timed (checks every N minutes, stores only when something actually changed; identical content never adds an empty version). Empty folders are recorded and recreated on restore. Every restore takes a safety snapshot first, so a rollback is itself undoable, and exact-mode cleanup only removes files and empty folders — a non-empty folder is never deleted. Ships a folder_timemachine model tool (status/history/diff need no approval; writes ask first) plus a floating panel with the timeline, one-click restore and mode switching. Pure JS, zero runtime dependencies, 110 tests included.

Injects your existing markdown documents (notes, rules, knowledge-base files) into every dsh session's system prompt as read-only memory — list file or directory paths in `files`; no database, no service, no write path; UTF-8 byte-budgeted truncation with line-boundary fallback, total budget, YAML front-matter stripping, and freshness timestamps.

File-backed memory: injects your own Markdown memory file at each session start, adds read/write tools with backup and undo, and reports DSH_HOME drift. Reads no conversations by default; an opt-in autoScan switch surfaces your own recent lines from local session logs. Never opens a network connection.

Generates a structured handoff card so a near-full conversation continues in a fresh zero-context session: /handoff writes the card with self-incrementing titles, pins it to the top of the sidebar and prepares the new session once the context nears its threshold.

Adds a four-pointed-star button left of the composer's model selector that rewrites the unsent draft in place — grammar, wording, punctuation and execution order — preserving code, commands, paths, identifiers, numbers and URLs byte-for-byte, answering nothing and never overwriting text typed while it runs. A pencil under every user message reopens that message for editing without touching the draft; because the Harness session log is append-only, an edit forks the conversation before that message rather than rewriting it. A circular restart icon under every reply asks that one question again in a fresh conversation that carries none of the earlier turns. Both are refused while a turn is still answering, and disabled when the reply used a tool that can leave a trace or when the files the message named have since changed. A right-sidebar memory tree draws the conversation as turns, showing what was actually asked and answered, every answer to one question on its own line labelled original, edited or memory-free, newest first, with arrows that re-read an answer in place; right-clicking jumps to a branch or removes it from the plugin's own history.

In-place Markdown editing for the right-sidebar document preview: a file keeps the shipped renderer's look, clicking a paragraph, heading, list, fence, or table cell puts the caret exactly where you clicked, and only the edited range is written back through ctx.fs.writeText, optionally guarded by the version it read.

Adds two visual pages to DSH Settings. MCP servers are kept as a JSON source of truth that rewrites the profile patch managed block, with YAML validation and timestamped backups, plus import of a pasted mcpServers JSON that first lists what will be added or overwritten. Thinking unlock probes each OpenAI-compatible gateway for the reasoning efforts it really declares and writes the per-level config back while leaving other hand-written YAML byte ranges untouched; the available model list syncs from a separate button that previews additions and removals, then writes levels and models in one atomic pass, and can also run on the page schedule.

Removes sandbox_permissions / justification arguments that cannot widen the call's current sandbox mode, so calls no longer fail with 'sandbox escalation to "workspace-write" is not strictly wider than this call's current "danger-full-access" mode'. GPT models tend to fill these optional fields on every call and pick workspace-write, which is narrower than danger-full-access. Runs in tools/pre-execute: compares the requested mode with the mode the session's sandbox policy resolves, and drops the fields when the request is the same or narrower (also drops a lone justification). Also understands the values GPT carries over from Codex's own shell tool: null, empty and use_default are treated as no request, and in a danger-full-access session any value (including require_escalated) is dropped because nothing is wider. Real widening requests, unknown modes and tools that do not advertise DSH's escalation schema are left to DSH core. Never widens permissions. Configurable tool list (default bash, pwsh, write, edit, run_code). Zero dependencies.

Registers FilmLight's official FLAPI MCP assistant (flapi-dev-mcp) as an MCP server in DSH, so an agent can write and run Baselight / Daylight FLAPI scripts.

Mount the QianYuan MCP server inside DeepSeek Harness: one config-layer bundle brings 15 agent-record tools (pitfalls, cached results, identity, task board) to mcp__qianyuan__*. No runtime code; reads are anonymous, writes are ed25519-signed.

Sweeps AI byproducts (scratch files, backups, logs, caches) into a reversible trash per project, and on project close classifies deliverables, archives reference results and writes a cleanup report. Requires Python 3 on PATH (the cleanup engine is a Python script; python3 / python / py are auto-detected, and without one cleanup cannot run and says so).

Cross-session and cross-project long-term memory: a non-blocking silent sedimentation dispatch on turn/end writes into a local SQLite store (FTS5 trigram full-text index with insert/update/delete sync triggers) that later turns recall from; entries are scoped per project alongside a global tree, and the plugin ships its own embedded web dashboard (it starts the HTTP server itself, no separate process) for browsing, graphing and controlled deletion.

Move a session to another folder, permanently delete it, or AI-rename it from the Web UI — by dragging a session row onto a folder title or from the session menu. Moving rewrites the session header cwd, relocates the persisted log and updates workspace accounting so history travels with it; deleting stops running agents and removes the log, projection-cache row and workspace accounting. The same three operations ship as agent tools: workbench_session_move, workbench_session_delete and workbench_session_rename_ai.

Adds an enhancement chip to the composer input action row, left of the model selector: it rewrites the draft in place using the model already selected in the session, in five styles, with one-click undo and an /enhance slash command.

A right-sidebar session panel that lists every Skill, MCP tool call and file read or write by turn and step, with per-failure error detail.

Category tabs (All, Appearance & UI, Usage & account and more) for the Installed list on the Plugins page, remembering the chosen tab; patch plugins named in an editable parent table are indented under the plugin they patch. Display-only, writes no files.

Recurring Excel template fill and reconciliation review: declare the file, sheet, header row, match key and amount columns once, preview every cell write before it lands, compare amounts as exact two-decimal values, list the unmatched rows as exceptions, and export a filled template copy with an HTML report and a JSON trace back to the source rows.

Read, write, convert and retime a subtitle track across SRT, WebVTT and ASS, audit it for over-long or unreadable cues, and transcribe audio through a configured speech recogniser.

Turn a raw recording into a published podcast with a local ffmpeg — detect chapter boundaries from silence, cut at those boundaries, normalise loudness to a broadcast target in two passes, and write an iTunes-compatible RSS feed whose enclosure byte lengths are measured from disk.

Read, build and take apart EPUB 3 books with no external dependency — the ZIP reader/writer and the XML scanner are both built in. Inspect metadata, reading order and TOC for structural problems, assemble a book from chapters, or split one back into per-chapter HTML.

Pack a directory into a deterministic tar, read an archive's contents without extracting it, unpack while refusing every member that could reach outside the destination, and verify what came out against what went in by SHA-256. No external dependency — the tar format is written and read in this plugin. Extraction is deliberately more permissive on reading than on writing, because a reader's job is to report what an archive says, while a writer must never emit what it cannot read back. A tar has no central index, so the walk tracks whether it reached the end-of-archive marker: without that, a half-file checksums fine and reads as complete.

Bounded self-managed long-term memory (Hermes-style MEMORY.md / USER.md with hard character limits and a safety scan), a review that runs after every turn and writes plain-language notes or refines an existing skill instead of adding a near-duplicate, a curator that finds and surgically repairs skills the loader would silently discard, cross-session search over past messages, and a learning card above the composer that appears only when there is news and asks before it deletes or rewords a memory.

Per-model input type and thinking level declarations on the 模型 settings page — 输入类型 checkboxes on every provider card, and a 思考等级 field in the provider 编辑 dialog. Writes the official pi-ai input and reasoningEfforts fields.

Swaps the composer's model control for a PI-Desktop style picker: a two-pane menu with search and provider grouping, a reasoning-effort slider, ArrowUp/ArrowDown prompt-history recall, and a prompt-rewrite button with undo.