プラグイン

Pixel-level screenshot diff for visual regression. Two local PNGs in; a machine-readable first line (`PIXEL_CHANGED=true|false`, greppable as an assertion), then the difference ratio, per-region bounding boxes, maximum channel delta and mean absolute error, plus a self-contained HTML comparison page. An ignore-region list with absolute or fractional coordinates excludes clocks, cursors and other per-render noise. It deliberately does no content recognition: it answers only "did this change", which is the part that can be made deterministic. Entirely local — no model call, no network, no upload.

Chooses which skills and MCP servers each workspace may use, from a "Skills & MCP" item in the workspace row's "..." menu; unchecked skills leave the session catalog, the `skill` tool and `/name`, and unchecked servers have their `mcp__<server>__*` tools denied for that workspace's sessions, running ones included from the next turn.

Registers OpenCode Zen free models as a native provider: the catalog is discovered live from opencode.ai intersected with models.dev pricing, and a detail page can re-fetch it or test which models answer. Declares no dependencies, so the host's pi-ai stays the only one.

Queries the balance and token prices of the model providers configured in DSH, whichever endpoint or relay they use, accepts cc-switch usage scripts, and shows the selected provider's balance under the composer.

Token usage meter in the DSH sidebar footer that runs both data sources at once rather than picking one: the DeepSeek platform account's usage and cost endpoints, which cover every device on the account but only the official channel, and local session logs, which cover this machine in real time across every provider but carry no cost. The two are merged per day — the official channel takes the larger figure, since one call is recorded by both, and providers the platform cannot see are added on top. History is backfilled month by month from the platform, providers are classified by channel rather than model name, days follow the platform's UTC boundary, and no token has to be pasted because the platform figures reuse the DSH account you are already signed in to. Totals appear as all-time, current-project, today, week, month, last-7 and last-30 in a compact bar, an expand panel and a detail modal with calendar, chart, model, provider, year, month and day tabs.

Switches every session's thinking and answers to Chinese — terse, a little wry, plain-spoken — while keeping proper nouns in English, and adds an "answer only what was asked" interaction rule. One global system-prompt section (order 10): no UI, no binaries, no network calls. Inactive under the minimal preset, which replaces the whole prompt with its own persona.

Select a passage in a DSH conversation and read a plain-language explanation in a floating panel, which carries the surrounding message as the context. Three explanation depths (plain / general / pro) that re-explain on the spot when switched; optional web search and local file lookup (the model decides by default, and a button can force one lookup that reports only what is new); follow-up questions; and a button that hands the whole exchange to a new session in standby mode - the background is queued without waking the model or spending tokens. Settings live on the host in settings.json, and nothing is written until you press save. The repository also ships an OpenAI-compatible endpoint and a browser extension (source).

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.

Discord bridge loaded into the same Host as the desktop app or dsh web, with no separate process or profile — chat with sessions in any workspace from DMs or threads, attach to sessions open on the desktop, and answer tool approvals and model questions from whichever side responds first.

Pushes DSH session events to phone IM and lets you reply straight under that message, injecting the text back into the right session. Covers nine event kinds: task completion, execution error, execution blocked, manual abort, token limit reached, abnormal interruption, waiting for your answer, permission request, and waiting for plan confirmation. Channels are QQ single-chat bots and Feishu custom apps, both over WebSocket long connections — no public IP, domain or tunneling needed. Several bots can be configured at once, each subscribing to the sessions it cares about. Approvals and questions arrive with tappable buttons, and replies route home through three layers (button, quoted reply, explicit short-id prefix), falling back to the most recent intervention-needed notification and echoing which session it went to.

Vision through NVIDIA NIM: mounts the bundled stdio MCP server as mcp__opencode_eyes_nvidia__describe_image, which describes an image with any of seven hosted vision models (MiniMax-M3 by default) and rotates several API keys.

Eyes for text-only models: mounts the bundled stdio MCP server as mcp__opencode_eyes__describe_image, which sends an image file to StepFun step-3.7-flash and returns a detailed description.

Replaces the composer model seat with a search-first picker and adds a model management page: create and edit llm-pi-ai channels, choose which of an endpoint's models exist, and declare per-model thinking ladders and capabilities.

Adds a Settings row to the sidebar header before Plugins, which opens the shell's Settings panel.

A read-only Cordis Inspect provider that lets an agent see the interface it is driving: outline() reports which UI regions the running shell actually rendered, with occupancy, geometry and current or past failure state; snapshot() returns one region's rendered node tree (tag, role, aria-label, own text, geometry, computed styles) addressed by slot key or CSS selector; failures() keeps a history of regions that failed to render and whether they recovered. Regions are addressed by the anchor the shell already writes for every slot, so it needs no region list and covers regions added later or contributed by other plugins. No UI, no tools, no dependencies, no network access.

A harness that gets smarter the more you use it, with no more goldfish memory. Before each prompt, Evolver recalls one proven strategy from your local history and the EvoMap network; after the turn it grades that strategy by the checks that ran, whether the edits were kept, and how you replied, then reports back to the network, which weighs it by evidence. Every success becomes a shortcut for next time, and every failure becomes a warning.

Delegates self-contained coding subtasks to the local Claude Code subscription through the official Agent SDK and returns the result to the host session. Nineteen per-call parameters control model, thinking effort, permission mode, allowed tools, turn count and budget cap; background jobs return a jobId immediately and stream incremental output; the result carries a session handle that can be passed back as resume to continue the same conversation; commands expected to run past ~20 seconds go through the built-in mcp__dsh__shell channel, which streams output line by line without consuming model context; a Claude Code monitor panel in the session lists jobs, renders the structured process, shows live output blocks and cancels a run; the delegation procedure ships as a runtime-registered skill, so installing the plugin installs the skill.

Bring a DSH plugin repository's build and release pipeline into the app: a sidebar panel listing each repository's latest workflow run, its releases, and how far the local checkout has drifted from the last release, with buttons to trigger a build, cut a release, publish a draft, re-run, cancel, or read the failed-step log. Drives the workflows through the machine's authenticated gh CLI, so the plugin stores no credential of its own, and which repositories are tracked lives in a config file rather than the package.

Plugin compatibility vet for dsh that compares installed plugins peerDependencies against the running dsh version to flag which ones the startup preflight will refuse to load, prints copy-pasteable fix commands, and offers a top-of-page banner in dsh web to isolate, uninstall or upgrade them, with an automatic backup before the change and rollback if the re-check fails.

dsh plugin: Jev-guided dispatch advice — before a turn starts, a TypeSafe Jev call decides whether the task belongs to a registered agent and which one, then injects a four-line, ignore-if-unsuited recommendation. Jev-optional: no key means silence.

Session-context surfaces ported from T3 Code: a session-row hover card showing machine, git branch, model, agent preset, approval policy and token usage, a status pill covering pending approval, awaiting input, working, error and subagent work, a provider and model chip, a per-session header strip, a lineage breadcrumb, row-menu actions that copy the context or run DSH compaction, and a preference row in DSH Settings for choosing which facts appear. Nothing is added to the composer.

Shows which backend model actually served each request when using the auto route.

Maintained fork of dsh-prompt-enhance, brought up to DSH 0.1.7 — the plugin would not even load on current hosts (InputState.imageIds became attachmentIds, so the composer button crashed the input box; settings.get() was removed; /enhance read a session id the agent no longer exposes). Rewrites default to reasoningEffort off, which measured 252 output tokens for a draft where the model default spent 2897, and a Settings page for all 17 fields is bundled. Verified on a real 0.1.7-rc.2 host: activation, a live HTTP enhancement request, /enhance in the GUI, and the Settings page reading and persisting configuration. 204 tests.

Plugin dependency graph for DeepSeek Harness: which plugin provides the services every other plugin injects, read from the live Cordis runtime.