Plugins

Evidence-based research workspaces with archived sources, knowledge graphs, exact citations and exportable reports. Experimental alpha. Needs Node 24.9 or later, uv and Python 3.13, plus a one-time setup step (setup-engine.mjs) that copies the engine and runs a locked uv sync; pnpm must also be allowed to run the build scripts of onnxruntime-node and protobufjs.

Multi-workspace agent federation with persistent ticket trees, parallel worker sessions, structured evidence reports, independent acceptance checks, retry/stop and restart recovery. Workers get shell and file tools in the ticket's directory, company_verify re-runs the ticket's acceptance commands in a shell on the host, and scope lists are task authorization rather than a sandbox. State is stored in ~/.dsh/ai-company; the optional company_repo_* tools run separately installed aoci/codegraph commands on any path they are given.

Runs macOS caffeinate with -w bound to the DSH host process id, so the sleep-prevention process exits when the host exits, and exposes /keepawake to report status and switch sleep-prevention modes. macOS only and on by default: from startup it keeps the Mac from idle sleep (and from system sleep on AC power) until DSH exits, while the display may still sleep; /keepawake running or off narrows it.

Read-only disk usage scan and interactive HTML cleanup report for Windows and macOS. Needs Python 3 on the machine; the bundled scan scripts run offline as child processes.

Detects repetitive loops in an agent's reasoning output and records, nudges, or cuts them, with a history report. Defaults to a nudge (a reminder injected into the next request); cutting and auto-resuming is opt-in, and should not be combined with another plugin's cutting such as dsh-loop-rescue. Needs DSH 0.2.1+ and Node.js 22.15+.

Session-archive layer for DSH. It injects one per-workspace Markdown file into the system prompt on every turn, searches past session transcripts with AND keywords, an optional time window and a workspace filter, and recovers the images and documents those past sessions carried by returning their original local paths or copying them to a chosen directory. It also reports index health: version, entry count, cache usage, frequent sessions and sessions stored in an archive format it does not recognise. Keyword search only, no embedding service.

Folds older oversized tool results into spill files and replaces them with short, deterministic stubs at step boundaries, with no LLM call: the stub keeps the first 200 and last 80 characters plus the sequence number, tool name, line/char/size counts and the original file path, and each fold is a real surface replacement accompanied by compaction/prune shadow pricing so token counts stay honest. Triggers on context pressure at 25% of the model window (falling back to 32768 tokens when the window cannot be resolved) or on five accumulated candidates; the last three turns, error results and rich content are never folded. The original is recovered from the session log rather than the spill file, so /folds, /unfold <seq> and /unfold all keep working after spill cleanup; session_fold_recall lets the model fetch a folded result by sequence number or by keyword, and session_fold_diagnose reports the effective settings. /spill-prune removes orphaned spill directories, preview only unless --apply. A Settings page exposes 12 of the 13 options as draft-then-save fields; they are volatile, so saved changes apply without restarting the plugin.

Puts a domain bar under the session header — a domain groups skills and tools and toggles as one block, so one click loads or unloads every member. Hovering a chip lists its members (skills, then tools) in a height-capped panel that scrolls. The right end of the bar is a compaction-threshold counter (− 5 +, 100k tokens per step) that rewrites the live compaction engine policies, so the compaction trigger line changes on the next turn without restarting dsh; a degraded fallback turns the number amber and is reported in the snapshot. Settings maintains domains and members with search, skill/tool tabs, all/selected/unselected filters and a copyable AI grouping prompt. Tool removal uses tools.restrict plus a guard; hiding a skill re-registers it with modelInvocable false, so it leaves the model catalog while /name still works. Sessions of another preset are left untouched.

DeepSeek Harness plugin: AI research toolkit — model release radar, model benchmark & community report, and AI news digest (4 tools + 3 skills).

Evolver for DeepSeek Harness, by EvoMap: recalls a proven strategy behind each prompt, reports its reuse at turn end, and distills verified work outside the conversation.

Scans a workspace for plaintext API keys, flags files inside git repositories, and reports findings masked so no full secret is ever printed.

Skill and plugin audit console for DeepSeek Harness. It scans local skill folders and installed DSH plugins, checks the packaging rules that commonly break installation, scores each asset, and exports a Markdown fix report.

Extra folders registered for a workspace become real workspace roots: the session runtime context lists them, a workspace_roots tool reports each root top-level entries, git branch and AGENTS.md/README/manifests, and file tools read, write, patch, move and delete inside them without approval prompts. Folders are edited per workspace from a sidebar hover action or from Settings. On Windows a confined shell stays inside the workspace, because ACL confinement cannot take extra roots.

Capability drift detection: audit a plugin once, store a baseline of what it could do, then after every upgrade report only what is NEW (new install scripts, shell/eval access, outbound domains, credential paths). Deletions and version bumps are not reported, and no score is computed.

Windows only: runs a build command through cmd.exe, collects the artifacts with sizes and SHA-256 prefixes, verifies expected outputs, and writes a report for each run.

Standalone rescue for a broken DeepSeek Harness profile: a CLI that runs without DSH diagnoses the crash logs (six classified causes, from real incidents), pre-flights the profile for command-name collisions and files plugins read at construction time, arms a safe mode that loads only official bundles until a healthy boot restores them, and keeps manifest snapshots for rollback. A resident watcher watches the crash-log directories via OS file events — idle costs nothing, and when the loader dies the report is written within seconds, one file overwritten per crash. Windows only: crash logs are read from %APPDATA% and the watcher autostarts from the Windows Startup folder.

DevLead mode for DeepSeek Harness: a human-only switch turns one ordinary session into a lead with eight devlead_* tools to register workspaces and create, read, continue and watch other ordinary sessions; it dispatches work to those sessions, reviews the delivery and reports, is woken through a durable notification digest when a watched session's turn settles, and can give a dispatched writing session its own Git worktree when the host provides the experimental worktree service.

Host-side progress checkpoints for long DSH tasks. After N minutes or M tool calls the model is nudged to post a one-line progress note and keep working, so you can correct it before it runs far off course.

Adds a DeepSeek Harness agent preset that hands implementation work to Codex subagents (auto-approved, workspace-write sandbox); the session agent only splits tasks, dispatches them and reports results. Requires DSH 0.2.2-rc.1+, the @deepseek-ai/dsh-subagent-codex bundle and a locally installed, logged-in Codex CLI.

A DeepSeek Harness agent preset where a Codex subagent analyses and decomposes the task, Claude Code subagents execute each work item, and the session only dispatches, checks and reports. Requires Codex installed and signed in locally, and Claude Code signed in locally.

Packages a working style into a versioned pack of rules, skills and presets injected into the system prompt at global, project or session scope, distills local agent sessions into aggregate reports and suggested rules you can adopt or undo, exports the pack as a tarball a teammate can install, and ships a reference FDE pack with 40 skills.

Experiment-orchestration SOP skill for long-running experiments. A top-level model drafts a decomposition plan — subtasks, the model for each, an acceptance figure with tolerance, and dependency order — and dispatches only after the user confirms that plan; each subtask then goes to a lightweight executor subagent as a self-contained prompt, and its report is confined to measured facts and the delta from the agreed figure, with judgement and decisions staying at the top level. Ships one plan template and five task templates (gather, execute, consistency check, read-only review, mechanical transform), acceptance gates split into a task layer and an experiment layer, and three rules that stop internal-consistency evidence from being presented as independent or external validation.

Render any Markdown report or DSH session excerpt into shareable PNG images with no npm dependencies, driving the Chromium-family browser already on the machine and falling back to PDF, then HTML. Reports come out as a summary card plus a full verbatim long image; a report written for dsh-fact-check also gets the tier badge, the two rails and the per-claim tiering, all taken from the report itself, and a report following one of the other report skeletons it defines gets the section that skeleton designates as its conclusion lifted onto the summary card. Conversations come out as a message flow, exportable by turn range, event range or the most recent N turns.