dsh-enhancement-analysis
Skill0 starsChanged 3 months agoMITView on GitHub
Analyze any project directly (recon report optional), map pain points onto DeepSeek Harness capabilities; produce a prioritized enhancement report.
In this repository4 other files it gives its agents.
- dsh-plugin-potentialSkill
- harness-runtime-optimizerSkill
- publish-screenSkill
- repo-reconSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
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 39
No recon report? Read the project directly: README + docs first, then main pipeline files.
- Distinctive: line 53
Every claim carries file:line (from the recon report or your own reading).
- Distinctive: line 54
Never oversell: ⚠️ items stay ⚠️.
- Distinctive: line 58
No code, no PRs, no repo changes until the user approves the report
≈ “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 Enhancement Analysis
Standalone: works on any project with or without a prior recon report. Output: a DSH enhancement report the user can approve before any code moves. Never write code in this phase.
Step 0 — What is DSH and why analyze for it (read this first, even with zero prior context)
DSH (DeepSeek Harness) is DeepSeek's open-source agent harness (deepseek-ai/deepseek-harness). Its thesis: "Everything is a Plugin" — built on the Cordis framework, it has NO privileged core. The model adapter, tool registry, session log, even the agent loop itself are all plugins.
Five primitives define it (names verified against the v0.1.0-rc source; re-verify on drift):
- Event waterfall — agent-loop stages (
agent/request,agent/request-error,agent/turn-stopping, …) and thellm/streamwaterfall over raw model calls: any plugin can intercept, rewrite, or wrap them - Plugin = Service — plugins attach to stable service keys (
ctx.tools,ctx.llm,ctx.fs,ctx.subprocess); consumers resolve by key, never import implementations - Session = log — the transcript is the source of truth; replay resumes any run
- Subagent delegation —
ctx.subagents.start()runs one-shot children with structured-output schemas; the in-processforkprovider shares parent history,spawnproviders start fresh (native parallel fan-out and A/B) - Swappable execution world — swap
ctx.fs/ctx.subprocessproviders and the whole execution environment moves to a VM/container with zero tool changes
What DSH enables that mainstream harnesses cannot (this is the goal lens for the whole analysis):
- Claude Code / OpenClaw / Cursor: core loop is NOT replaceable — you can only add tools/prompts around it. DSH's loop is itself a plugin; any stage can be intercepted or swapped
- LangGraph / hand-rolled SDK pipelines: routing, model choice, and checkpointing are compiled into code. In DSH, model routing is a request-hook (hot-swap mid-run), checkpointing is native (session replay), and business rules hang off the waterfall without touching core code
- Native A/B: delegate one task to N child agents in parallel and diff outputs — no harness conventionally offers this as a primitive
Goal of this skill: find where these unlocks matter for the target project, honestly bucketed (✅ equivalent / ✅✅ stronger / ⚠️ hard port), with costs stated. Not everything should move to DSH — a credible ⚠️/no-go is a valid result.
Locating the DSH source
Resolve the local checkout before fact-checking, 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, or the checkout is older than the latest tag, say so.
Workflow
- Gather inputs
- Optional accelerator: recon report at
<project>/dsh-scan/recon-deep-<project>.md(fromrepo-recon, inside the target project) if one exists - No recon report? Read the project directly: README + docs first, then main pipeline files; build the workflow map and pain-point list yourself, with file:line evidence
- Focus areas to cover when reading directly: entry points, LLM/agent call sites, pipeline stage structure, retry/error handling, model configuration, scheduling
- DSH source for fact-checks: see "Locating the DSH source" above; primitive detail in
references/dsh-capabilities.md
- Optional accelerator: recon report at
- Stage the pipeline — redraw the project workflow as numbered stages (trigger → stages → storage → output). Mark every LLM call point and every pain point onto it
- Map to DSH primitives — for each stage/pain point pick from the five primitives above +
references/dsh-capabilities.md(waterfall interception,agent/requestmodel routing, subagent delegation, plugin services, session-as-log replay, execution-world swap, existing plugins such asdsh-llm-retry/dsh-token-meter/ guard policies) - Bucket every verdict — ✅ equivalent (drop-in) / ✅✅ stronger (unlocks what the current stack cannot) / ⚠️ hard port (honest cost) — one-line justification each
- Honest costs — language gap, rewrite or wrapper size, plugin/adapter count, release-maturity risk (DSH is pre-1.0: APIs may change without compatibility shims)
- Priority matrix — severity × effort × strategic value → ranked roadmap (table)
- Migration path — default two-step: (a) wrap the project entry behind DSH as a facade (logging/model routing/notifications), zero risk to existing surfaces; (b) port ONE representative stage as a pilot behind a decision gate; only then consider full migration
- Generate report from
templates/report-template.md→<project>/dsh-scan/dsh-enhance-<project>.md— inside the analyzed project (createdsh-scan/if missing; a user-specified path overrides). The report must carry the 未深入 list and the non-DSH findings section. Deliver Top findings + the bucket table to the user; wait for approval - Handoff to runtime optimization — if the project already runs on DSH or the user approves a pilot/migration, recommend
harness-runtime-optimizernext. The enhancement report defines where to intervene; the runtime optimizer verifies whether the intervention actually improves quality/cost/latency/reliability.
Rules
- Every claim carries file:line (from the recon report or your own reading); when unsure re-verify against the DSH source, don't guess
- Never oversell: ⚠️ items stay ⚠️; costs stay explicit (language, hops, rewrite size)
- Findings outside the DSH capability domain (security, storage races, rendering drift, …) go to the non-DSH findings section, not the trash — they are often the most urgent fixes
- Reports in 中文; tables for mappings; ≤300 lines
- Report header records: target repo path + branch + HEAD, recon report path if used, DSH baseline (resolved path + commit + date)
- No code, no PRs, no repo changes until the user approves the report
- DSH facts may drift between releases — re-check the DSH 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.

