Plugins

Complementary web-search provider: every web_search merges the DeepSeek Search API (primary) with serper.dev Google results (secondary) at a fixed 2:1 ratio. Both sides are direct HTTP retrievals with no auxiliary model turn: URL-deduped sources, long DeepSeek summaries clipped into snippets, per-backend timeouts and graceful degradation to whichever backend answers.

Read-only attachment library: lists every image and file attachment stored on this machine, its size, and which conversations (including those in the session trash) still use it, with Show in Finder and Open in Preview.

Pick the model your subagents run on — with a reasoning-effort ceiling and an output-token cap — enforced in code on every child DSH starts, whichever tool starts it.

Deep-space theme for the Harness Web GUI: two fully designed light and dark palettes over a seamless cross-dissolve nebula wallpaper, plus a model probe that reports which configured models actually answer and their real context window.

Read-only clinic for DSH session logs: 22 named diagnostics (token waste, retry and failure loops, context bloat, cross-session amnesia, idle MCP servers and skills), each reported with a replayable evidence locator such as session.v4.jsonl.zstd#L101-L110, and each finding turned into an executable prescription — a profile-patch YAML snippet or a command with current to suggested values, risk, verification and source — plus a feedback command that retires a prescription you reject, after which the report states it was removed by your feedback. Zero dependencies, default zero egress, session files are never written.

Scales the whole DeepSeek Harness interface from 50% to 200%, menus and pop-ups included, set from Settings - General with a slider, a percent field or presets. Works in dsh web and the desktop app, which has no zoom of its own.

Compiles a rough need into a versioned task book (goal / context / constraints / acceptance / decisions / open gaps, all numbered) and hands it to any AI window losslessly: the receiver must pass a read-back handshake — restate, list gaps, declare READY — before work starts; /answer closes gaps with an automatic version bump, /forge-list tracks which window holds which version.

Live usage gauge injected into the system prompt every turn: session tokens and estimated spend, a next-turn estimate, and two predictions the model can act on — which step fills the context window and which step burns the budget. /qm panels, zero-config by default, optional price table, and it publishes a summary.json budget contract sibling plugins read.

Tick which models appear in the DSH model selector, and fold provider groups in the settings panel.

Reports which PowerShell the DSH pwsh executor rows actually run — enabled rows only, with version-numbered Store paths refused as pins — and installs the self-updating Microsoft Store package as a background job after reading the connection cost, because Windows defers Store downloads on a connection marked as metered. Two tools and a prompt section on the host side, plus a configuration card, a diagnostics section and a re-detect action in the Web UI; Windows only, no runtime dependencies.

Read-only Skill inspector for DeepSeek Harness. It shows which Skills a conversation actually loaded — from observed loading evidence only — and which Skills the current environment can discover, and gives each Skill a detail page whose left column switches between the framework, the evidence, validation, evaluation, step evidence, run logic, the before/after of a modification when there is one, the definition facts and the raw SKILL.md. Declaration and run stay apart: the body is split into structure, declared workflow and progressive disclosure, the run logic is read from local receipts, and every declared step carries its own evidence — no observed evidence is reported as exactly that, not as executed or not executed, and the run's instruction hash sits beside the current definition as match, mismatch or unavailable. The evidence view gives each subject one of three states (declared, observed, unavailable), walks definition → load → use → outcome, names what each stage cannot reach, declines to attribute an outcome to the Skill, and can copy the snapshot out as byte-stable JSON. Validation checks the Agent Skills open standard and the platform profiles separately, each finding on a rule id and labelled standard compliance or platform compatibility. Evaluation keeps repeatable cases with the conditions each run was recorded under and the four stages of runtime evidence, comparing two runs of one case item by item without scoring them. The trigger boundary keeps should-trigger and should-not-trigger prompts, has you run them in a real new session, then records what was observed — whether a load record appeared and whether it was model-invoked or user-explicit; a typed /name is not a natural trigger, and an absent load record is reported as not observed, not as not triggered. The same projections are exposed to the agent as a read-only skill_review tool, which returns a structured review packet and says what it omitted when the packet will not fit whole. The inspector, the evidence chain and these reviews are computed on this machine from the definition and the session receipts; the Chinese reading version sends the body through the host's model when you ask for it, keeping code fences, paths and frontmatter verbatim and storing the result on this machine. A Skill can also be cloned into a local copy. It does not write back to the source Skill and does not run its scripts.

Nearly half the agent cost, with accuracy intact. Just enough tools is a DeepSeek Harness plugin that uses Jev to reveal tools and skills progressively. The main model starts with a clean planning step. Jev then selects which capabilities to add as the task unfolds.

Cross-session experience memory for DSH. Lessons are evidence-graded, injected when the turn is about the same thing, and shown just before an action only when the record declares which call it applies to. A rule that holds whatever the turn is about — "always answer in Chinese" — can be marked **standing** and is then carried every turn: measured on 24 isolated real-model turns, a standing rule is followed 6/6 against 0/18 for the controls (empty store, same record unmarked, unrelated standing record). A guess still cannot become always-on: a standing record needs a verified passage like any other. What nobody looks up is retired, reversibly. Five tools, seven slash commands, zero runtime dependencies.

Pre-execution safety valve for DSH that judges with TypeSafe Jev: it hooks tools/pre-execute and decides every bash/pwsh call by first applying offline static rules, then by asking the Jev model — the System One model from TypeSafe, which returns a structured decision instead of prose — one yes-no question ("will this command irreversibly delete or overwrite real data?"), turning the answer into allow, revise, block or escalate; revise hands the model a safer rewrite, escalate offers a one-shot human token; exhausted credit or a missing API key degrades loudly instead of failing silent, and a missing key is requested in the conversation itself (`guard key set`, read from stdin); every verdict is appended to a shared audit log and the text people read is bilingual zh-CN/en.

Per-command proxy for DeepSeek Harness: the agent decides which shell commands leave through a proxy, backed by a kernel the plugin installs and supervises itself.

File-defined subagent roles for DeepSeek Harness: roles live in project and global .dsh/roles files, a delegating agent sees one catalog line per role, and each role's tool policy decides which tools its child can see.

Live token cost for every chat in the DSH Web UI: the chat, its subagents and the whole session tree, priced from a bundled models.dev snapshot of seven providers and 139 priced models, with cache read and write rates wherever the provider publishes them and DeepSeek from its own peak and off-peak table. Every finished answer carries its own price in the transcript, and a chat that is not currently open is priced from the log instead of disappearing. Spend is costed per interval from an append-only JSONL ledger in the project folder, so logged work keeps the tariff it was billed at. On top of it: budget-aware work plans packed against a money limit, routing scenarios that show which models are worth connecting, a generated plan document in English, Chinese or Russian, and a warning when spend belongs to no plan unit.

Task kanban where agents claim and hand off work through taskboard_* tools, which cannot move a task to done or canceled — acceptance stays with a human.

Configurable statistics row below the composer via a ${variable} display template — which segments appear and how they are labelled is fully user-defined.

Security auditing and session forensics for dsh: full tool-invocation-chain recording (who/which tool/files/timestamps), redacted playback, and rule-based scanning.

Registers /wrapped, which reads the session's own stats and renders a shareable dark SVG summary card — tool vs. thinking time, decode speed, and a tool/thinking-ratio verdict.

See how you have been using AI: `/session-insights` reviews your DSH sessions over time, shows what worked, which problems kept coming back, and what results were actually produced, then creates a retrospective report on your machine.

Check which DeepSeek Harness plugins actually loaded, and look up public evaluation boards on trapstreet.run

The DeepSeek time-of-use meter for dsh: what this session cost, which tariff is running, when it flips, and the account balance behind it — one line under the composer.