dsh-plugin-potential
Skill0 starsChanged 3 months agoMITView on GitHub
Evaluate whether any project could be packaged as a DeepSeek Harness plugin: inventory its capabilities, map them to plugin forms (skill provider, tools, policies, services, adapters), score fit, and produce a verdict report. Assessment only — never builds the plugin or modifies the target.
In this repository4 other files it gives its agents.
- dsh-enhancement-analysisSkill
- harness-runtime-optimizerSkill
- publish-screenSkill
- repo-reconSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 67
Every claim carries file:line evidence (from the recon report or your own reading).
- Distinctive: line 68
Never oversell: ⛔ stays ⛔.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
DSH Plugin Potential
Read-only assessment of one question: could this project's functionality live as a DSH plugin, and is it worth building? The deliverable is a verdict report the user decides on. This skill never writes plugin code, never scaffolds a bundle, never changes the target's code or config — evaluation is the entire job (its single write is its own report under the target's dsh-scan/). A credible ⛔ "keep it standalone" is a first-class result, same as a ✅.
Direction check (read first)
Two directions get confused; this skill is only the first:
- project → plugin (this skill): the project's capability becomes an extension other DSH users install. Worked example: dsh-lens itself — repo-analysis workflows packaged as a skill-provider plugin.
- project → DSH host (
dsh-enhancement-analysis): the project migrates its pipeline onto DSH to gain its primitives.
The same repo can score differently in each direction. Run both if relevant; keep the reports separate; never blend verdicts across them.
Plugin forms (what "becoming a plugin" can mean)
Names verified against the DSH v0.1.0-rc source; re-verify on drift (DSH is pre-1.0, no compatibility promise):
- Skill provider — package expertise as SKILL.md guidance + resources behind
ctx.skills.registerProvider(); zero behavior code. Best fit for methodologies, workflows, analysis checklists. - Tool / capability plugin — expose the project's functions as agent-callable tools on
ctx.tools(asdsh-web,dsh-shell,dsh-subprocessdo). Best fit for libraries with clean input→output boundaries. - LLM adapter —
ctx.llm.registerAdapter(): a model-provider integration becomes per-route swappable (dsh-llm-deepseek,dsh-llm-pi-ai). - Policy plugin — intercept waterfall events (
agent/request,agent/request-error,llm/stream): retry, routing, metering, guard rules (dsh-llm-retry,dsh-token-meter,dsh-guard). - Service plugin on a stable key — caches, storage, notifications; consumers resolve by key, never import implementations.
- Execution-world provider — swap
ctx.fs/ctx.subprocessto move execution into a VM/container with zero tool changes.
Deployment shape (bundle anatomy): package.json + cordis.patch.yml (patch rows over a base profile) + src/index.ts with apply(ctx) + README; installable via dsh plugin --profile <name> add.
Locating the DSH source
Resolve the local checkout before fact-checking plugin forms, in order: the $DSH_ROOT environment variable; conventional checkout locations (~/git/deepseek-harness, ~/src/deepseek-harness); the GitHub repository deepseek-ai/deepseek-harness as last resort. Record the resolved path and commit in the report header; if only GitHub was available, say so.
Rubric (score every candidate capability, 1–5 per dimension)
| Dimension | Question | Low-score signals |
| 封装性 | Is it a separable unit with clear inputs/outputs? | welded into a monolith or UI, global mutable state |
| 接口适配 | Does it map cleanly onto one plugin form? | needs a human judgment loop, no function boundary |
| 需求面 | Would other DSH users/projects install it? | single-team internal value, no recurring cross-project need |
| 状态适配 | Does its state fit session-log / ctx.fs primitives? | heavy external mutable state, db schema lock-in |
| 移植成本 (reverse-scored) | Language gap, rewrite size, dependency weight? | non-Node core without a bridge, huge dep tree |
| 维护风险 (reverse-scored) | Exposure to pre-1.0 API drift? | touches unstable internals, pins exact versions |
Reverse-scored = lower cost/risk earns the higher score. State the reverse convention in the scorecard header so readers don't misread it.
Workflow
- Gather inputs — optional accelerator: recon report at
<project>/dsh-scan/recon-deep-<project>.md(fromrepo-recon, inside the target project). No report? Read the project directly: README + docs first, then entry points and the modules behind them. Focus: what the project does for its users, not how it's built. - Capability inventory — list the distinct functional units an outside agent could want (a fetcher, a linter, a workflow, a provider integration…), each with file:line evidence and a one-line description. This inventory, not the architecture, drives everything downstream.
- Map to plugin forms — for each capability pick a best-fit form and an alternate from the six above. Capabilities that fit no form go to the "stays standalone" list — listed, not forced.
- Score the rubric — per capability, per dimension, with a one-line justification for any score ≤2 or ≥4. Verdict per capability: ✅ 强候选 / 🟡 部分(state exactly which slice extracts and which stays)/ ⛔ 不适合.
- Overall verdict + pilot pick — one sentence the user can act on, plus (if any ✅ exists) the ONE capability to pilot first and why.
- Descriptive sketch only — for the pilot pick: plugin name, form, interface surface (what it registers, on which ctx key or event), install story, dependencies. Prose and tables. Never code, never scaffolding.
- Honest costs + boundaries — language gap, rewrite size estimate, adapter count, rc-stage drift risk; plus what deliberately stays out and why.
- Report from
templates/report-template.md→<project>/dsh-scan/plugin-potential-<project>.md— inside the analyzed project (createdsh-scan/if missing; a user-specified path overrides). Report carries the 未深入 list and the non-DSH findings section. Deliver Top findings + the verdict table to the user; wait for their decision. - Handoff — user approves a pilot? Building it is a separate later effort, outside this skill. The question was actually "should this project run on DSH"? Hand to
dsh-enhancement-analysis. Either way, this skill stops at the report.
Rules
- Assessment only: no code, no scaffolds, no PRs, no changes to the target's code or config — ever, not even after user approval; the only write is the report under
<project>/dsh-scan/; approval triggers handoff, not implementation - Every claim carries file:line evidence (from the recon report or your own reading); DSH plugin-form facts re-verified against the source, never guessed
- Never oversell: ⛔ stays ⛔; costs stay explicit (language gap, rewrite size, dependency weight, rc-stage drift)
- Findings outside the plugin question (security, storage races, rendering drift, …) go to the non-DSH findings section, not the trash
- Reports in 中文; tables for mappings and scores; ≤300 lines
- Report header records: target repo path + branch + HEAD, recon report path if used, DSH baseline (resolved path + commit + date)
- DSH facts may drift between releases — re-check the source if the analysis is older than ~2 weeks or predates the latest tag
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

