justinzyc271828-sys/dsh-ai-jobs

AI hiring radar for DeepSeek Harness. Index the public job boards of AI labs and AI-heavy companies, rank every posting against your own criteria, mine what they actually ask for (years, skills, degree, work mode, comp, visa), and push a daily digest of the openings worth reading.

No API keys, no accounts, no scraping of logged-in sites — every source is a public ATS board endpoint that the company itself publishes for its careers page.

Why

Job hunting for AI roles is a search problem over ~70 companies that each use a different ATS. The data is public but scattered, re-read by hand every week, and impossible to compare across companies. This plugin keeps one local index, so you can ask cross-company questions:

• Which labs are hiring research engineers right now, and what do they require?
• What share of core AI roles ask for a PhD? What is the median "N+ years"?
• Which openings appeared in the last 7 days, and which of them match my stack?
• What do Anthropic/OpenAI/Mistral ask for that I do not have yet?

What it does

Step	Detail
Collect	71 pre-verified boards (Greenhouse, Ashby, Lever) + Workable/SmartRecruiters for your own additions
Normalize	One posting shape across ATSes: title, locations, remote flag, department, dates, URL, JD text
Classify	core / adjacent / other AI relevance, role family, seniority — rule-based and auditable
Mine	Years of experience, degree, work mode, compensation range, visa sponsorship, 38-skill vocabulary
Rank	0–100 score from relevance, freshness, seniority, your skills of interest and preferred locations
Store	Local JSON index with JD snapshots, change history, and a digest cursor
Push	Markdown digest + optional daily reminder through the harness scheduler

Install

# from the marketplace / npm (once published)
dsh plugin --profile web add dsh-ai-jobs

# from a local checkout, for development
dsh plugin --profile desktop add ./dsh-ai-jobs

# from a packed tarball
npm pack && dsh plugin --profile desktop add ./dsh-ai-jobs-0.1.0.tgz

The bundle patch mounts one plugin row (ai-jobs); restart the app or toggle it on the Plugins page.

Quick start

ai_jobs_refresh                     # index every watched board (first run: ~10s for 3 big boards)
ai_jobs_search { "tier": "core", "postedWithinDays": 7, "minScore": 70 }
ai_jobs_requirements { "tier": "core", "limit": 200 }
ai_jobs_digest
ai_jobs_watchlist { "action": "add", "company": "Mistral", "ats": "ashby", "token": "mistral.ai" }

Tools

Tool	What it returns
ai_jobs_refresh	Re-index all/A-selected boards. Reports per-board ok/error, added/changed/closed counts.
ai_jobs_search	Filtered, scored postings: query, company, tier, seniority, remote, location, minScore, postedWithinDays, limit, sort.
ai_jobs_get	One posting in full: metadata, score, extracted requirements, JD head, stored JD path.
ai_jobs_digest	Markdown digest of what changed since the last digest (or the last N days).
ai_jobs_requirements	Cross-company aggregate: skill frequency, median years, degree mix, work-mode mix, comp ranges.
ai_jobs_watchlist	list / add / remove boards (any of greenhouse, ashby, lever, workable, smartrecruiters).
ai_jobs_status	Index health: boards, open/closed counts, tier mix, last refresh, last digest, storage path.
ai_jobs_schedule	create / list / delete the daily digest reminder (needs the schedule service).

Daily digest

Two ways:

1. ai_jobs_schedule { "action": "create", "time": "09:00:00", "timeZone": "Asia/Shanghai" } — one reminder per session, matching title ≤ 120 chars.
2. dailyDigest.enabled: true in config — the plugin arms the reminder itself the first time a session runs ai_jobs_refresh or ai_jobs_digest.

The digest only reports postings that are new since the previous digest cursor, so a daily push does not repeat yesterday's openings. ai_jobs_digest { "days": 7 } ignores the cursor.

Configuration

{
  "lang": "zh",
  "timeZone": "Asia/Shanghai",
  "concurrency": 4,
  "skillsOfInterest": ["LLM", "inference", "agents", "RAG"],
  "preferredLocations": ["San Francisco", "Remote"],
  "remoteOk": true,
  "digestPerCompany": 4,
  "digestLimit": 30,
  "digestMinScore": 0,
  "dailyDigest": { "enabled": true, "time": "09:00:00" },
  "companies": [
    { "company": "Cohere", "ats": "ashby", "token": "cohere", "tags": ["frontier-lab"] }
  ]
}

Key	Default	Meaning
dataDir	$DSH_HOME/ai-jobs (falls back to ~/.dsh/ai-jobs)	Index, JD snapshots, digest state
lang	zh	Digest language (zh / en)
concurrency	4	Parallel board fetches
timeoutMs	25000	Per-request timeout
jdMaxChars	20000	JD text kept per posting
maxDetailPerBoard	60	Greenhouse JD fetches per board per refresh
tags	from seed	frontier-lab, ai-native, ai-user — used by scoring and filters
companies	[]	Extra boards appended to the shipped seed

Data sources

All endpoints are public and unauthenticated:

ATS	Endpoint	Seed boards
Greenhouse	boards-api.greenhouse.io/v1/boards/{token}/jobs	33
Ashby	api.ashbyhq.com/posting-api/job-board/{token}	37
Lever	api.lever.co/v0/postings/{token}?mode=json	1
Workable	apply.workable.com/api/v1/widget/accounts/{token}	—
SmartRecruiters	api.smartrecruiters.com/v1/companies/{token}/postings	—

The shipped seed covers Anthropic, OpenAI, xAI, Mistral, Cohere, Perplexity, Thinking Machines, Cursor, Cognition, Harvey, Sierra, Decagon, ElevenLabs, Suno, Midjourney, Runway, Baseten, Cerebras, Fireworks, Modal, Anyscale, Together, Scale, Pinecone, Weaviate, Chroma, LangChain, LlamaIndex, Databricks, Snowflake, Stripe, Cloudflare, Datadog, Elastic, MongoDB, Figma, Notion, Vercel, Plaid, Ramp, Airbnb, Instacart, Lyft, Pinterest, Reddit, Discord, Roblox, Duolingo, Khan Academy, Spotify and more. Companies on Greenhouse store their JD behind a second request; the plugin fetches it only for new or changed postings.

ai_jobs_watchlist adds any other company whose careers page exposes one of these boards. Token discovery is manual by design — the seed was verified one board at a time.

How ranking works

score = tier(55 core / 28 adjacent / 8 other)
      + freshness (30 → 0 over ~40 days)
      + seniority (-6 intern … +7 principal)
      + skills-of-interest overlap (max 12)
      + location preference (max 8) / remote (max 6)
      + comp disclosed (3)

Titles that mention AI but hire for a non-engineering function (counsel, sales, policy, marketing, recruiting, finance, procurement …) are classified other before the AI patterns run. People-management titles that carry AI vocabulary (program / product / project manager, TPM) end up adjacent: they are real openings, but they are not AI individual-contributor roles.

Data & privacy

• Everything is stored locally under $DSH_HOME/ai-jobs: store.json (index), jd/*.txt (JD snapshots), digest cursor.
• The plugin sends no telemetry and has no backend. It only calls the public board endpoints listed above, with a descriptive User-Agent.
• Nothing is scraped from logged-in sites; no credentials are read or stored.
• Deleting dataDir removes everything.

Legal note

Board endpoints are published by the companies for their own careers pages. Keep the default politeness settings (concurrency: 4, one refresh per run, JD fetched only for new postings) and do not republish bulk data. Requirements, compensation, and availability change; always confirm on the company's own careers page before acting.

Development

npm run smoke        # schema contract + refresh/search/digest against two small boards
npm run live-check   # multi-board end-to-end check against the real APIs
npm run probe        # re-verify candidate ATS tokens
npm run seed         # rebuild data/companies.json

Layout: src/text.js (HTML → text), src/sources.js (ATS adapters), src/classify.js (tier/seniority/requirements/score), src/store.js (index + JD snapshots), src/digest.js (markdown), src/tools.js (8 tools), src/index.js (bundle entry).

Zero runtime dependencies; Node built-ins only.

License

MIT