Foreword¶

When integrating skills into DeepSeek Harness (DSH), the common practice is to select one based on README descriptions or experience. As the number of candidates grows, the problem becomes: which path should the current task actually take? Descriptions often differ from actual outputs, especially for tasks requiring finished products like PPTX, research figures, research reports, webpages, or videos, making it difficult to judge solely from descriptions.

Below, we introduce ForkProbe (jayden-x-l/forkprobe). Developed by maintainer Jayden-X-L, it is categorized as a workflow plugin. Its core idea is to run the same input through a baseline and multiple candidate skills in parallel, generate a local HTML report, and after you select a winner, allow the agent to continue along the winning path. The GitHub repository currently has approximately 70 stars, 4 forks, is at version v1.1, and is licensed under MIT.

What This Is¶

ForkProbe is an AI skill selection and trial-run tool, offering both a native DSH Cordis plugin (forkprobe-dsh) and a Python runner. In one sentence: Compare multiple skills on the same task and pick the winner.

Compared to “directly picking a skill to start,” it transforms selection into an observable process: recommend candidates → run in parallel with the same input → local report displays outputs and reviews → human selects winner → generate continuation handoff to proceed with the task. The report is generated locally first (local-first). After selecting a winner, you can save the handoff via the “Continue” button in the report and optionally share your selection anonymously (without uploading the original task or outputs).

Core Features¶

Parallel Trial Runs and Local Report¶

ForkProbe’s workflow can be summarized in five steps:

1. Recommend a few candidates from a curated directory, locally installed skills, the EverMind Skill Hub, GitHub, or a custom path.
2. Run the baseline (a reference group without additional skills) and multiple candidates using the same input.
3. Display each path’s complete output, time taken, token estimates, file previews, and AI review suggestions in the report.
4. You select the winner.
5. Generate a continuation handoff to let the agent proceed.

Seven Workflow Modes¶

The README breaks down the capabilities into seven modes, covering text to finished artifacts:

The image prompt mode generates files like prompt.md and style-card.md within the runner but does not call image APIs by default. When the host has generation capabilities, optional rendering can be performed according to render-queue.json. Other environments can also render externally and backfill rendered.png.

Native DSH Plugin Capabilities¶

As a DSH plugin, ForkProbe provides two tools:

• forkprobe_compare: After user confirmation of candidates, runs them in parallel and generates a report.
• forkprobe_resume: Resumes the selection in the report after the waiting window ends.

forkprobe_compare requires confirmed=true; candidate subagents do not receive tool permissions to avoid recursive calls or workspace modifications. For text candidate comparisons, it uses native subagent fan-out. For file-based tasks like scientific figures, research reports, webpages, and videos, it can still run the existing Python runner (--platform deepseek_harness) via a headless profile.

Installation and Enabling¶

The SkillHub directory page lists version v1.1.0 and categorizes it as a workflow. Before installation, it is recommended to review the GitHub repository source code and the MIT license. The plugin runs with the current dsh process permissions, so installing it trusts its code behavior.

Installing the DSH Native Plugin¶

Install ForkProbe into the DSH web profile:

dsh plugin --profile web add "github:Jayden-X-L/forkprobe"

If you need to use it in a headless profile, also run:

dsh plugin --profile headless add "github:Jayden-X-L/forkprobe"

After installation, restart the corresponding profile. The package.json requires Node >= 22.19, and the plugin depends on DSH ecosystem packages like @deepseek-ai/cordis.

Artifact Runner Dependencies¶

When comparing file-based artifacts via the Python runner, install the core dependencies:

pip3 install jinja2

Video mode additionally requires local FFmpeg (for media probing and cover generation via ffprobe / ffmpeg). DeepSeek Harness is currently in developer preview, and the README suggests pinning a verified version for stable tasks.

Typical Usage¶

Natural Language Trigger (Recommended)¶

No need to remember commands; directly explain your intent to the agent:

Help me compare a few skills to see which is more suitable for the current task.

Or specify the ForkProbe process more explicitly:

Please use forkprobe to recommend candidates, wait for my confirmation, then run them in parallel and generate a report for me to select the winner.

After installing the DSH plugin, you can also say:

Please use ForkProbe to recommend a few Skills suitable for this rewrite, wait for my confirmation, then run them in parallel using native DSH subagents, open the Report for me to select the Winner, and continue along the winning result.

Text Comparison (Command Line)¶

Suitable for text artifacts like academic polishing or style rewriting. First prepare the input file, then run compare.py:

echo "Please polish this text and preserve the original meaning." > /tmp/forkprobe-input.txt

python3 scripts/compare.py \
  --input /tmp/forkprobe-input.txt \
  --skill baseline \
  --skill writing-anti-ai \
  --skill humanizer-zh \
  --skill remove-ai-flavor-writing-skill \
  --judge \
  --output /tmp/forkprobe-report.html

Recommend First, Then Confirm and Execute¶

For scenarios like research reports or image prompts, a two-step approach is recommended. First, recommend candidates:

python3 scripts/recommend.py --input /tmp/forkprobe-research-task.txt

After confirming the candidates, run the corresponding artifact runner, for example, for a research report:

python3 scripts/research_artifact.py \
  --input /tmp/forkprobe-research-task.txt \
  --pipeline baseline-research-report \
  --pipeline source-first-research \
  --pipeline analyst-style-report \
  --confirmed \
  --run \
  --judge \
  --render-report \
  --report-output /tmp/forkprobe-research-report.html

DSH Headless Compatible Path¶

After preparing DEEPSEEK_API_KEY, you can specify the platform to run:

FORKPROBE_PLATFORM=deepseek_harness \
DEEPSEEK_API_KEY=your-key \
python3 scripts/compare.py --input /tmp/forkprobe-input.txt --skill baseline --judge --output /tmp/forkprobe-report.html

ForkProbe will first use the command specified by FORKPROBE_DSH_CLI, then the global dsh, and finally npx @deepseek-ai/dsh.

Suitable Scenarios and Notes¶

Suitable For¶

• Unsure which skill to use for the current task and wanting to see real outputs before deciding.
• Need to compare a baseline with multiple skills rather than relying solely on skill descriptions.
• Deliverables are PPTX, research figure packages, research reports, image prompt packages, runnable webpages, or video productions, requiring file previews and QA.
• Wanting to find candidates from local skills, the EverMind Skill Hub, GitHub, or custom paths, then conduct small-scale trial runs.

Not Suitable For¶

• Simple, deterministic tasks where the answer or tool path is already clear—direct execution is usually faster.

Usage Notes¶

1. Permissions and Security: The plugin and runner execute with the current DSH process permissions. For external GitHub candidates, it is recommended to check the license, dependencies, and artifact paths before execution. ForkProbe does not automatically install or execute unconfirmed candidates.
2. Candidate Discovery Scope: Local scanning reads SKILL.md metadata for matching, including paths like ~/.dsh/skills and project-internal .dsh/skills; it does not use raw documents as search terms.
3. Platform Status: DeepSeek Harness is in developer preview. For production environments, pinning versions is recommended. Artifact runners default to workspace-write, which can be overridden with FORKPROBE_DSH_PERMISSION_MODE.
4. Community Directory Positioning: SkillHub (skillhub.cn) is a community directory for Skills aimed at Chinese users and has no official affiliation with DeepSeek / High-Flyer. The DSH ecosystem philosophy is “everything is a plugin,” and ForkProbe is one type of workflow extension.

Summary¶

ForkProbe transforms skill selection from “guessing based on descriptions” to “parallel trial runs on the same task with local report decision-making.” For DSH users, the native plugin covers text candidates and winner continuation. File-based tasks can continue to use the headless runner. For further details on the capability matrix and report artifacts, refer to the release page and README.

• SkillHub Directory: https://www.skillhub.cn/plugins/Jayden-X-L/forkprobe
• GitHub Repository: https://github.com/Jayden-X-L/forkprobe
• Release Page: https://jayden-x-l.github.io/forkprobe/