Plugin

Experimental Windows bridge for Codex to supervise DSH Desktop 0.2.1-rc.3 tasks through local MCP, with blocking direction questions, plan checkpoints, review and revision, and named workspace routing. A Windows adaptation of Cryocell77/dsh-desktop-mcp-bridge (MIT); it uses the same package name and loader row, so it cannot be enabled alongside upstream.

Connects the hosted Knowra MCP server: search arXiv papers by topic, author or exact UTC dates without an account, and sign in with OAuth from the plugin page to read Scout reports on methods, evidence and limitations.

Interactive HTML or flowchart cards inside the conversation: the tool blocks until the user submits, then returns the clicked action, fields, comments and card state.

Composer infinity switch: while the highest reasoning effort is selected, one click appends a configurable notice to the system prompt.

Git tab in the DSH web UI right sidebar for the current session repository — switches branches from a dropdown of local and remote branches (with ahead/behind counts), stages files in staged/unstaged/untracked/conflicted groups, commits, fetches, pulls and pushes with automatic upstream setup, drafts a Conventional Commits message from the staged diff with the default model and falls back to a plain chore message when none is available, and renders git log --graph with per-branch colors plus a word-level highlighted diff.

Built on dsh-better-sidebar: the repository graph is a tab inside that sidebar, and this plugin still loads without it - install dsh-better-sidebar to see the tab. Forked from @linxin666/dsh-client-ui-git-graph 0.4.8, which linxin666 develops in zhu1090093659/dsh-web. Both halves own the /git route prefix, so this plugin and the upstream one cannot be installed at the same time - pick one. It discovers the Git repositories up to three levels below the workspace directory (at most 200, skipping node_modules, dist, build, out, coverage, target, vendor and dot-directories), lists them as a tree with branch, head, dirty-file and conflict counts, and draws the selected repository commit graph as an SVG lane canvas with curved cross-lane edges, merge diamonds, ref badges, row zoom and a pinned commit card. The tab writes to the repository: the Initialize Git repository button runs git init, and the conventions picker writes a GIT_CONVENTIONS.md document into the repository (or a workspace-wide .dsh/GIT_CONVENTIONS.md that nested repositories inherit) from six frameworks - Conventional Commits, GitHub Flow, Git Flow, Trunk-based, Conventional Branch, GitLab Flow. The host injects that document into the system prompt before every model step and flags the last 20 commit messages that break the rules the document states - subject length, type prefix, Chinese, English or mixed language, lowercase, trailing period and imperative mood, plus an optional AI attribution policy (ai: required / optional / forbidden) that reads the message trailers, so Assisted-by and Co-authored-by count as attribution while ai: forbidden only rejects AI-specific trailers. When the workspace document, the document of the checked repository and the documents of its parent directories disagree, the panel and the injected prompt both name every disagreement and the agent is told to ask before changing anything; one click then merges them into the workspace .dsh/GIT_CONVENTIONS.md, keeps the union of the machine-checkable rules, leaves a timestamped .bak copy of every rewritten file and replaces the absorbed repository documents with pointers to the workspace one. With the agentTool option enabled the host also registers a git_conventions tool, so a model step can run the same check and stage the same merge. The same graph can also open as a page in the main area of the conversation, opt-in and closable from the tab, beside the views the host ships. The composer branch chip, interactive graph dialog and worktree manager of the upstream plugin are carried over unchanged.

Forensic analysis of past model-routing decisions. Replays each session, classifies every turn's task complexity from its request shape (deterministic, no LLM), and reports two kinds of waste: an expensive model on cheap work (the difference is measurable), and a cheap model on expensive work that failed on its own merits — where every retry re-sends the whole context, so the downgrade cost more than the upgrade. Distinguishes provider-capacity retries (RATE_LIMIT/TIMEOUT/TRANSPORT) from model failures so it does not cry wolf. Derives routing rules from the evidence, priced against a built-in per-million table.

Four-tier prompt enhancer for the composer — the system tier always emits a LangGPT-structured system prompt (# Role / Profile / Skills / Rules / Workflows / Initialization), upgrading task drafts into the system prompt of the agent that runs them, while the other tiers rewrite as a light tidy-up, a single paragraph, or a five-section brief; the iterate workbench refines an editable base under an optional requirement, with fill-back and undo. Never changes intent, never invents requirements.

Project rules for DSH: check commit messages, PR descriptions and versions before they ship, and ask before destructive or outward-facing commands.

Reminds the agent to speak up when it runs several model steps without emitting any text, so long tool runs stop going silent.

Evaluation guard for DSH: workspace-only filesystem, no network/subagents, wall-clock limits, and task installation for meow-ai-arena and benchmark tasks.

Telegram remote for DeepSeek Harness — attach a live web session from your phone and answer its questions there, with a Russian UI. Russian fork of hi-wenw/dsh-telegram-channel (MIT).

A global permission watchdog: a hard budget on auto-approved permission requests. At most 10 per rolling 60 seconds, at least 3 seconds apart; anything past either limit is deferred to the human answerer and waits. It never refuses and never changes what is asked -- it only decides who answers and how fast. Installs disabled; enable it with the plugin manager toggle.

Git workbench in the DSH session header and overlay: view changes, commit, fetch, pull, push, and browse branches and history in existing local repositories.

Git service plus right-sidebar UI — a Changes tab (stage/unstage, per-file diffs, commit/amend, push and rebase-then-push sync, confirmed discard), a paged canvas commit-graph tab, a Worktrees tab to list, enter and create checkouts, a composer branch selector, and opt-in LLM commit-message generation.

Recursive delegation for dsh agent trees: a preset plus Host limits bound what each subagent may do, subagents work in separate _agents/ workspaces, and the parent decides which results to integrate.

Agent preset for DeepSeek Harness: the standard tool stack plus four working rules written into the persona prompt — a project keeps at most four documents (intro, decisions, change log, current state and next steps), those documents describe only the current state with no AI edit history or superseded decisions, no git commit or push or pull request until the user explicitly asks, and a careful-coding workflow (restate the goal and done-criteria first, surgical diffs, evidence before claiming success) with plain-language reports.

Keeps a DeepSeek Harness run going on its own: continues a turn that ended abnormally, and continues a turn logged as completed while the model tool call was emitted as plain text instead, which leaves the agent silently idle. Revives a model request that stops streaming. A wait on an approval or on a question is logged and notified, never answered.

Local 3D studio orchestration: detects Blender, COLMAP and ComfyUI on every run and picks the route they support, and also ships a Three.js fallback that generates models from primitive specs, edits or scales GLB files, writes an offline viewer page and renders multi-view PNGs without Blender.

Model failover for DeepSeek Harness: pick one or more providers and the pool becomes every model inside them; when a request fails (rate limit, quota, timeout, server or transport error) the next untried model in the pool retries automatically, and the original error surfaces once the pool is exhausted. Configure from a seat beside the composer model picker or the model_failover tool, with no restart needed.

Finds wasted GitHub Actions runner time and proposes small, reviewable workflow patches, with explicit approval before any edit and a DSH settings page for analysis rules, history limits, storage, and cost estimation.

A continue button in the composer tool row, immediately left of the send key: one click writes an interruption-aware prompt into the draft and submits it, so an interrupted task resumes without retyping the context. The prompt names the interruption and appends the current task plus the tail of the last reply; Alt-click or right-click sends a bare 继续, and a non-empty draft is appended to rather than overwritten.

Live Markdown rendering inside the DSH composer plus Markdown-rendered user messages; the submitted text stays literal Markdown.

Publishes whichever model the profile has selected in chat or on the Models page as an OpenAI-compatible HTTP endpoint: /v1/models, /v1/chat/completions (streaming SSE and non-streaming), /health. A request may name any routable model; one that omits `model` gets the current chat selection, so switching model in chat switches the endpoint. Switched on, it can also lend the model the Harness tools switched on individually, run those calls through the Harness tool runtime and feed results back up to a round limit. The listener is the plugin's own, so 127.0.0.1, 0.0.0.0 and a NAT-forwarded port all behave the same with no change to the request path. Settings cover bind address, port, API key with a generator, fallback provider and model, reasoning effort, CORS origin and timeout.