MIT-licensed Python tool that integrates natively with DeepSeek Harness (dsh plugin) to compare candidate skills on one task, pick a winner from a local HTML report, and continue the task along the winning skill — plus artifact comparison for PPTX, figures, reports, web and video.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-08-28
License
MIT
01What can it help you accomplish?
Run several candidate skills side-by-side on the same task and pick the winner from real output
Local HTML report showing each candidate's full output, elapsed time, token estimates, file previews and AI judge recommendations
Developers and researchers who use DeepSeek Harness and want to choose between skills based on observed results rather than skill descriptions
Compare finished artifacts — PPTX decks, scientific figures, research packages, web pages and video — across candidate pipelines
Per-candidate artifact packages with previews, source files, QA and AI review in the report (e.g. `video.mp4`, figure packages, research packages, runnable pages)
Teams whose deliverables are PPTX, figures, research reports, runnable pages or video and need to judge finished files instead of outlines
After picking a winner in the report, continue the real task along the winning skill
A local verdict and continuation handoff saved by the Report's continue button, letting the current agent keep the winner's style, structure or file artifacts
Anyone who wants the DeepSeek Harness agent to resume work using the skill that won the comparison
02How to install into DeepSeek Harness
Prerequisites
- Python 3 with the core dependency `jinja2` (installed via `pip3 install jinja2`)
- The DeepSeek Harness CLI (`dsh`) and a DSH profile to install into — the README uses `web` (and `headless` when you also run file-based artifact runners)
- `DEEPSEEK_API_KEY` when running the headless DeepSeek Harness path with `--platform deepseek_harness`
- Local `FFmpeg` for video artifact comparison (installed via `brew install ffmpeg`)
Installation steps
- 01
Copy the repo into your agent skill directory: `cp -r forkprobe ~/.claude/skills/` (Claude Code) or `cp -r forkprobe ~/.agents/skills/` (Codex / local agent skill directory)
- 02
Install the core dependency: `pip3 install jinja2`
- 03
Install the native DSH plugin into the `web` profile: `dsh plugin --profile web add "github:Jayden-X-L/forkprobe"`
$ dsh plugin --profile web add "github:Jayden-X-L/forkprobe"
- 04
If you also use the headless profile for file-based artifact runners, install there too: `dsh plugin --profile headless add "github:Jayden-X-L/forkprobe"`
$ dsh plugin --profile headless add "github:Jayden-X-L/forkprobe"
- 05
Restart the corresponding profile, then ask DSH to use ForkProbe (e.g. have it recommend candidate skills, run a parallel comparison and open the report)
Verify the integration
Not specified by the author
03DSH integration and capability boundaries
First-class native DSH plugin — install with `dsh plugin --profile web add "github:Jayden-X-L/forkprobe"`, restart the profile, then the `forkprobe_compare` / `forkprobe_resume` tools run native subagent fan-out, AI judging and continuation handoff. A headless-compatible path also runs the Python artifact runners with `--platform deepseek_harness`.
Side-by-side comparison with AI judge
the same task input run against `baseline` and multiple candidate skills→local HTML report with each candidate's full output, elapsed time, token estimate, file previews and AI judge recommendation
`forkprobe_compare` requires `confirmed=true`; candidate subagents get no tool permissions, so they cannot recurse into ForkProbe or modify the workspaceMulti-source candidate discovery
curated directory, locally installed skills, EverMind Skill Hub, GitHub and BYO paths→a deduplicated candidate list sorted by scenario match, presented for confirmation before any run
external discovery uses cleaned task signals only — it does not use raw documents as search terms and never auto-installs or executes unconfirmed candidatesImage prompt / style direction comparison (v1.1)
image prompt / style pipelines→per-candidate prompt packages: `prompt.md`, `style-card.md`, `composition.md`, `negative-prompt.md`, `render-notes.md`
no image API is called inside the runner; optional Codex host rendering via a local `render-queue.json`, or user-side external render backfilled as `rendered.png`Artifact comparison (PPTX / figure / research / web / video)
candidate artifact pipelines (presentation generators, figure pipelines, research report pipelines, web builders, video editors)→finished artifacts with previews, QA and AI review in the report — PPTX, figure packages, research packages, runnable pages, `video.mp4`
04Who is it for? When not to use it?
Good for
- Developers and researchers who use DeepSeek Harness and want to choose between skills based on observed results rather than skill descriptions
- Teams whose deliverables are PPTX, figures, research reports, runnable pages or video and need to judge finished files instead of outlines
- Anyone who wants the DeepSeek Harness agent to resume work using the skill that won the comparison
05Compatibility, maintenance and safety notes
- External discovery (GitHub, EverMind Skill Hub) only receives cleaned scenario terms — raw tasks, documents and local paths are never sent. Local skill scanning only reads `SKILL.md` metadata and never auto-installs or executes skills.
- Winner selection may be shared anonymously; only `task_type`, `candidate_skill_names` and `final_choice` are uploaded, and sharing can be disabled with `FORKPROBE_TELEMETRY=0` or by unchecking it in the report.
- Anonymous sharing events are sent to a `workers.dev` endpoint that some networks cannot reach; events stay in a local outbox and retry automatically — you can override the receiving endpoint with a self-hosted domain.
- DeepSeek Harness is currently a developer preview; the README recommends pinning a verified version for stable production tasks.
- Video artifact mode requires local `FFmpeg`, and the headless DeepSeek Harness path needs `DEEPSEEK_API_KEY`.
MIT · actively maintained (README version v1.1, last push 2026-08-19)
06Frequently asked questions
How do I install ForkProbe as a DeepSeek Harness plugin?
Copy the repo into your agent skill directory, install the core dependency (`pip3 install jinja2`), then run `dsh plugin --profile web add "github:Jayden-X-L/forkprobe"`. If you also use the headless profile for file-based artifact runners, install there too with `dsh plugin --profile headless add "github:Jayden-X-L/forkprobe"`, then restart the corresponding profile.
Is the DSH integration MCP-based?
No. The README describes a native DSH plugin (`forkprobe-dsh`, stable since v1.0): it provides two tools — `forkprobe_compare` for parallel trial runs after your confirmation and `forkprobe_resume` to resume the report's selection after the waiting window. It supports native subagent fan-out, an AI judge and continuation handoff.
What's the difference between the native plugin and the headless compatible path?
The native plugin handles text candidates, AI judging, report selection and continuing with the same agent. The headless compatible path runs the existing Python runners for file-based tasks — scientific figures, research reports, web pages and video — via the official headless profile with `--platform deepseek_harness` (it needs `DEEPSEEK_API_KEY`).
What happens after I pick a winner?
ForkProbe records a local verdict and generates a continuation handoff. The report's continue button saves the handoff locally and the current agent keeps working along the winning skill's style, structure or file artifacts. You can also choose to anonymously share the selection — only `task_type`, `candidate_skill_names` and `final_choice` are uploaded.
Is my task content sent anywhere?
Task content stays in local reports and logs. GitHub and EverMind Skill Hub discovery only receives cleaned scenario terms, and local skill scanning only reads `SKILL.md` metadata. Anonymous winner sharing is opt-in and can be disabled with `FORKPROBE_TELEMETRY=0`.
07Related DSH workflows
distilly
by titanwings
Meta-skill turning source material into a versioned Person Profile skill with incremental merge, correction layer, rollback, per-host install.
dsh-market
by dsh-market
插件市场 (plugin marketplace): 打开 Settings → Plugin Market 浏览/搜索/一键安装社区插件；安装源限制为 curated awesome-dsh-plugin 注册表（其余拒绝），默认禁用构建脚本（pnpm>=10 需显式开启）。
ai_animation
by unclecheng-li
Generate standalone animated HTML diagrams, presentations, notes, protocol visualizations, and UI demos from prompts.
aegis
by ganyuanran
Make AI coding agents architecture-aware: baseline-first, evidence-verified, drift-checked, and safe across long tasks.
08Data and sources
将 ForkProbe 直接安装到 DSH `web` profile：
dsh plugin --profile web add "github:Jayden-X-L/forkprobe"
插件提供两个工具：`forkprobe_compare` 负责确认后的并行试跑，`forkprobe_resume` 负责在等待窗口结束后恢复 Report 中的选择。
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-08-28. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.