wmw343/dsh-resume-expert

dsh-resume-expert is a resume-generation plugin built specifically for DeepSeek Harness (DSH), shipped as a Cordis plugin with a bundle-manifest and listed on dsh-plugin.org. The package is written in TypeScript and exposes a host adapter layer (src/adapter.ts) that implements the six host contracts — llm, auth, storage, telemetry, config, and logger (plus an optional render hook) — while the inner core under src/plugin/ stays framework-agnostic and dependency-free so it can be compiled and tested in isolation. Installation requires DSH ≥ 0.1.5 and Node.js ≥ 22; the recommended command is the version-pinned dsh plugin --profile web add github:wmw343/dsh-resume-expert#v1.0.1, after which the user writes DEEPSEEK_API_KEY into ~/.dsh/.env via PowerShell Add-Content. The plugin never stores or persists the API key itself.

A typical workflow runs through four stages driven by the agent calling tools named resume_intake, resume_refine, resume_compose, resume_export, resume_set_photo, and resume_tailor. First, the intake tool turns a vague one-line request into an initial draft plus a completeness report and follow-up questions; then resume_refine absorbs answers across multiple turns covering name, education, internships, projects, and skills; resume_compose merges everything into a polished, submission-ready draft that strips placeholders, and resume_export renders a single-page A4 PDF through Chrome CDP (falling back to printable HTML when Chrome is missing); resume_set_photo embeds a local ID photo and resume_tailor rewrites the resume against a target job description. Built-in quality rules enforce quantified phrasing, bolded metrics, grouped skills, top-pinned achievements, and a city-priority plus 3-4-4 phone format.

Dependencies and caveats are explicit: the adapter talks directly to the DeepSeek API via an OpenAI-compatible endpoint, so no third-party LLM proxy is required, and session state is persisted to ~/.dsh/resume-expert-store.json by default — overridable with RESUME_EXPERT_STORE — so drafts survive dsh web restarts. The plugin holds no secrets, writes no external database, and serializes every input/output as JSON to keep iframe and subprocess isolation feasible. A suite of 471 automated acceptance checks (204 core, 70 real-model, 37 end-to-end, 72 real-browser, 58 integration, 30 launcher) guards quality, and the bundled npm test command runs 21 offline invariants — including cross-process survival, atomic writes without stray .tmp files, TTL handling, and corruption self-heal — without needing the network or a running dsh daemon. PDF output is verified by parsing the produced file with pypdf. The project is MIT-licensed; first-time users should run npm install && npm run build, set DEEPSEEK_API_KEY in ~/.dsh/.env, then start dsh web and speak a request such as "帮我做一份简历，我是××专业应届生，会××和××" to begin. Uninstallation uses dsh plugin --profile web remove dsh-resume-expert, and any custom store file must be deleted manually because removal does not clear session data.