Turns explicit DSH session experience into reviewable native Skill drafts that are saved only after user approval.
Install
# from npm (prebuilt)
dsh plugin --profile web add dsh-run2skill
# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)
dsh plugin --profile web add github:qkycir-123/dsh-run2skill
Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).
README
中文 | English
Have you run into any of these situations?
- You teach an agent the same workflow today, then teach it again in a new session tomorrow.
- You keep correcting the agent step by step: “Do not edit that file,” “write the test first,” or “this check is required.”
- The agent finally learns the right approach, but the lesson disappears when the conversation ends.
dsh-run2skill turns the corrections, constraints, and workflows you explicitly teach a DeepSeek Harness (DSH) agent into reviewable, reusable native Skills.
You teach once while doing real work. Run2Skill prepares a Skill draft from the reusable parts. Nothing is saved until you approve it.
Run2Skill does not let an agent silently create permanent rules for itself. Every draft shows its source, intended scope, and complete content before you decide whether to save or discard it. Only an approved draft is written to DSH's native Skill directory.
The current stable release is
0.3.1. It supports DSH Web0.1.1-rc.2only.
0.3.1 adds low-noise learning status, Synthesize this experience now, revision requests that generate a new draft, and bounded preservation of important long-workflow evidence. An immediate request still waits for the agent to stop and for complete facts, then follows the existing lookup, review, and publication safety gates. Internal batch counters are not shown in the settings page.
See the complete flow
- Find the draft — Run2Skill only notifies you when action is required and places the draft under Settings → Plugins → Run2Skill.
- Review its evidence and scope — inspect the rationale, filtered conversation evidence, target scope, and the complete
SKILL.mdthat would be written. - Request a revision if needed — leave one short instruction and Run2Skill generates a new complete draft; the old version is not published and the new one still needs review.
- Save only after approval — successful results appear under Recent activity and remain ordinary native DSH Skills.
Install
First confirm that you use DSH Web 0.1.1-rc.2, install Node.js ^22.19.0 || >=24.0.0, and make sure both dsh and pnpm are available in your terminal. Then run:
dsh plugin --profile web add dsh-run2skill@0.3.1
Restart DSH Web. Open Settings → Plugins; the plugin is loaded when the Run2Skill tab appears.
Run2Skill does not need a separate model key. When it analyzes a Skill draft, it uses the model already selected for the current DSH session. If the session has no available model, learning stops and reports the reason instead of silently switching providers.
Use it while you work
Continue talking to DSH normally. For example:
Save this workflow as a Skill so it can be reused later.
You can also give an explicit correction, describe a durable constraint, or teach a reusable ordered workflow during ordinary work. You do not need to stop and manually reconstruct a SKILL.md file.
After each conversation turn, Run2Skill records only low-cost observations; it does not call a model to review every turn. Batch detection runs at bounded batch boundaries, when the session becomes idle, or when you explicitly request saving or immediate synthesis. Only reusable experience proceeds to Skill lookup and draft generation.
When something needs your attention, DSH shows one native notification. To review a draft:
- Open Settings → Plugins → Run2Skill.
- Review the draft, its intended scope, evidence, and complete content.
- Approve and save it, request a revision, discard it, or retry a failed save.
- After a successful save, it is an ordinary native DSH Skill and remains usable even if Run2Skill is later uninstalled.
If a draft is not right, submit one revision request of at most 2,048 UTF-8 bytes. The Host uses the current complete draft to generate a new immutable version, records the parent/child relationship, and invalidates the old approval. You must inspect and approve the new version. This is not a free-form editor and does not bypass lookup, human review, or safe publication.
Drafts can target the current project (PROJECT) or the current user (USER). The current release writes only to DSH's default Skill storage. If you replace or disable that storage, Run2Skill stops safely instead of guessing a path.
What it learns
Run2Skill focuses on reusable experience that you state explicitly:
- Correction: “Do not fix the implementation first; write a failing test.”
- Durable constraint: “All GitHub copy in this project must be written in Chinese.”
- Workflow: “Check the upstream version, run compatibility probes, then update the evidence.”
- Explicit save request: “Save this process as a Skill.”
It does not turn every successful agent action into a permanent rule. It learns what you clearly taught, rather than trying to guess your intent.
You remain in control
You can turn Automatic learning off under Settings → Plugins → Run2Skill:
- On: explicit corrections, durable constraints, and workflows may produce Skill drafts.
- Off: ordinary automatic learning pauses, while an explicit “save this as a Skill” request still works.
Run2Skill is local-first. It does not store model keys or copy the whole session. Only the necessary, filtered, truncated, and redacted context is sent to the session's selected model for analysis. Every Skill draft requires your approval before it is saved.
The settings page also provides Clear all cache. It removes Run2Skill-owned intermediate cache, pending Skill drafts, failures, and non-sensitive diagnostics, but never removes:
- original DSH session records;
- already published native Skills;
- provider, agent, or other DSH settings.
See Storage and upgrades for the detailed retention and migration rules.
Update and uninstall
Install another explicit version:
dsh plugin --profile web add dsh-run2skill@<version>
Uninstall:
dsh plugin --profile web remove dsh-run2skill
Restart DSH Web after either command. Uninstalling does not delete published Skills and keeps Run2Skill data by default. Use Clear all cache on the settings page before uninstalling if you want to remove that data.
Troubleshooting
- The Run2Skill tab is missing: confirm that you use the DSH
webprofile and restarted DSH Web after installation. - “Run2Skill is currently limited” (
DEGRADED): do not manually delete storage files. Retry, then report the issue if the state persists. - “This Run2Skill version is incompatible” (
INCOMPATIBLE): verify your DSH version and restore a compatible Run2Skill version. - No Skill draft appears: confirm that the current session has an available model, or say “Save this workflow as a Skill.”
- A
PROJECTdraft cannot be saved: the current session must belong to a project workspace recognized by DSH. - You changed DSH Skill storage: the current release supports only the default storage and stops safely instead of guessing where to write.
Report problems through GitHub Issues. Do not attach keys, complete sessions, private paths, or logs containing sensitive information.
Learn more
- Changelog
- DSH compatibility
- Storage and upgrades
- Product requirements
- Architecture baseline
- Design document index
- Single-owner design for one Skill-save intent
- Contributing
- Maintainer compatibility probes
This project is licensed under the MIT License. Licenses for dependencies embedded in the client bundle are listed in THIRD_PARTY_NOTICES.md.
Links
More in this category
zhu1090093659/dsh-web#packages/dsh-skill-explorer★ 6553
Skill center for the dsh web GUI: browse all loaded skills grouped by source, enable or disable model invocation, create new skills, and delete into a recoverable trash.
GanyuanRan/Aegis★ 1156
Software-engineering method pack for coding agents, with skills for baseline-first planning, systematic debugging, prompt hygiene, verification before completion, and repair/retirement tracking.
superdesigndev/superdesign-skill★ 488
Design skill for UI and marketing graphics on the Superdesign canvas: reads the repo for context, extracts its design system, then generates and iterates branchable design drafts, flow pages, and reusable components through the Superdesign CLI.
linhay/harmony-next.skills★ 341
HarmonyOS NEXT skill bundle for DeepSeek Harness with offline API references and DevEco, HDC, and emulator automation guidance.
Vladimir-Human/humanizer-ru#dsh★ 119
Cleans AI traces from Russian text: finds chatbot copy-paste artefacts (ChatGPT, Gemini, Grok, Perplexity, DeepSeek) and rewrites the text into natural prose on request; 39 regex markers with an evidence registry, offline, text-only bundle.
LayneChai/superpowers-dsh★ 110
Superpowers for DeepSeek Harness: TDD, debugging, planning, and collaboration skills adapted from obra/superpowers, installed as a Cordis plugin registering 14 skills with zero runtime dependencies.
Community comments
Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.