dsh-industry-research

---

仅供研究，不构成投资建议 — Research only, not investment advice. This pack does research support only: no trading, no price prediction, no paid/login-walled sources.

Compatibility

• DeepSeek Harness 0.1.0-rc.6 (peers pinned to 0.1.0-rc.6).
• Node ^22.19.0 || >=24.0.0, ESM only ("type": "module").
• Peer dependencies: @deepseek-ai/cordis ^4.0.1, @deepseek-ai/schemastery ^3.18.0, and @deepseek-ai/dsh-tools, @deepseek-ai/dsh-skill, @deepseek-ai/dsh-skill-filesystem at 0.1.0-rc.6 (all composed by the official dsh-base bundle).
• Optional, looked up at execution time (never injected): ctx.web (public-source retrieval — composed by dsh-base) and ctx.researchReport (the sibling dsh-research-report sealing engine).

What you get

• industry_map — build or update an industry chain map (chain.json): upstream/midstream/downstream nodes, edges, and metric slots. A metric with a value must carry a sourceRef; a value-less slot is an explicit gap. Seed notes/files and an optional ctx.web digest are registered as citable sources (S1, S2, …).
• industry_track — public-source policy/news tracking over the official ctx.web seam: per-topic search, host allow/block lists, since filtering, bounded snapshot fetches with SHA-256 provenance hashes, and append + dedupe into timeline.jsonl with a retention cap. Fails loud when the web capability is unreachable.
• company_scan — company cards (card.json + card.md) from your workspace data files (SHA-256-hashed, with Markdown outlines and figure-candidate lines so every number cites a file and a line), plus an optional ctx.web citation complement. Text formats only in v1 (no PDF).
• industry_report — one auditable report from the map + timeline + cards. With a mounted ctx.researchReport engine the evidence/sections/claims are sealed by its assemble and per-claim verdicts come back; without one, the builtin fallback writes versioned reports/<YYYYMMDD-HHmmss>/report.md + manifest.json with a SHA-256 source-traceability table and honestly labels engine: 'builtin-fallback'.
• Two methodology skills — industry-research-method (chain decomposition, supply/demand framework, sourcing discipline, gap-declaration discipline) and company-research-method (company framework, public-source list, compliance wording).
• Typed Cordis events — industry-research/map, industry-research/track, industry-research/report emitted after each committed artifact.

Quick start

git channel

# From a scratch profile (pins the commit; runs the self-contained `prepare` build)
dsh plugin --profile demo add "github:PerryLink/dsh-industry-research#<sha>"
# The profile's pnpm-workspace.yaml gains an allowBuilds entry for dsh-industry-research on first add.

npm channel

dsh plugin --profile demo add dsh-industry-research

Both channels install the bundle row (see cordis.patch.yml) into the profile's dsh.profile.bundles stack and take effect on restart.

Then, in a session:

加载 industry-research-method 技能，然后研究白酒行业：
先 industry_map 建产业链图，再 industry_track 跟踪政策，最后 industry_report 出报告。

Install & uninstall

dsh plugin --profile demo add dsh-industry-research       # install
dsh plugin --profile demo remove dsh-industry-research    # uninstall

Verify the row mounts: dsh --profile demo --dump-config | grep dsh-industry-research.

Configuration

All tunables are Schemastery Config fields; invalid values fail the profile load loudly.

Key	Default	Description
enabled	true	Master switch; false mounts nothing.
industryRoot	industry-research	Artifact root, relative to the session workspace (or absolute).
fetchTimeoutMs	20000	Per-request timeout (ms) for ctx.web search/fetch calls.
timelineMaxEntries	500	Entries retained per industry timeline.jsonl (oldest dropped).
sourceAllowlist	[]	Host allowlist for tracked sources (empty = all hosts).
sourceBlocklist	[]	Host blocklist for tracked sources (wins over the allowlist).
offline	false	Offline mode: never touch ctx.web; local workspace data only.
skillsDir	(unset)	Explicit skills root override; defaults to the packaged skills/.
track.maxResultsPerTopic	10	web_search maxResults per topic.
track.maxFetchesPerCall	10	Snapshot-fetch budget per industry_track call.
scan.maxFileBytes	1048576	Per-file read cap for company data files.
scan.maxFigureCandidates	100	Figure-candidate budget per company scan.

Tools & surfaces

industry_map({ industry, seed?, seedFiles?, web?, chain? })

Called with chain: validates (dangling edges, unsourced values, illegal tiers, duplicate ids — loud failure with the full issue list) and persists chain.json, then lists the explicit gap slots. Called without chain: returns the current map, the registered sources, and an optional ctx.web chain-structure digest so the model can iterate. Emits industry-research/map.

industry_track({ industry, topics?, since? })

Searches each topic through ctx.web, filters by the allow/block lists and since, fetches page snapshots (SHA-256) within the call budget, and merges into timeline.jsonl (dedupe by normalized URL, capped). Sources whose snapshot failed are kept as citation-only entries with the reason in note. Fails loud naming the missing capability when ctx.web is unmounted or offline: true. Emits industry-research/track.

company_scan({ name, dataFiles?, web? })

Reads the workspace data files (.md/.txt/.csv/.tsv/.json; v1 does not parse PDF), hashes them, extracts Markdown outlines and figure-candidate lines, optionally attaches ctx.web citations, and persists the card. Rejected files come back with reasons; everything the card cannot establish is an explicit gap.

industry_report({ industry, sections?, companies?, draft? })

Assembles evidence (E-chain, E-timeline, E-company-<slug>) and either validates your draft (sections + claims; every claim's evidenceIds must reference registered evidence) or builds the mechanical auto-draft (sourced metrics and recent timeline entries become claims). Engine path: sealed directory + sealHash + per-claim verdicts. Fallback path: versioned Markdown + manifest, claims honestly marked unverified. Emits industry-research/report.

Skills

• industry-research-method — industry research methodology: upstream/midstream/downstream decomposition, supply/demand framework, sourcing discipline (every number: value + unit + source + asOf), and gap-declaration discipline (declare, never fabricate).
• company-research-method — company research methodology: the business-structure / financials / risks card framework, the public-source priority list (company disclosures → regulators → authoritative media), and compliance wording.

Both load on demand through the standard skill tool (加载 industry-research-method 技能).

Data layout

<workspace>/<industryRoot>/<industry>/chain.json      industry_map
<workspace>/<industryRoot>/<industry>/timeline.jsonl  industry_track
<workspace>/<industryRoot>/<industry>/sources.json    citable source registry (S1, S2, …)
<workspace>/<industryRoot>/<industry>/notes/          seed notes
<workspace>/<industryRoot>/<industry>/reports/<ts>/   industry_report (report.md + manifest.json)
<workspace>/<industryRoot>/companies/<name>/card.*    company_scan

Permissions & data

dsh-industry-research consumes only public seams: ctx.tools, ctx.skills, and — looked up optionally — ctx.web and ctx.researchReport. It performs no network access of its own (all retrieval goes through ctx.web with the deployment's provider selection and your configured timeouts), stores no credentials, and writes only inside the session workspace under industryRoot. Only public sources are used; paid or login-walled sources are out of scope — pass your own exports in as dataFiles.

Security boundaries

• Workspace containment — industry/company names are validated path segments; data files are containment-checked against the session cwd (both sides resolved).
• Provenance by construction — sources carry SHA-256 hashes; unsourced numbers are validation errors; corrupt timeline lines are skipped with a surfaced count, never silently.
• Honest degradation — missing ctx.web / ctx.researchReport produce loud failures or honestly labeled fallback artifacts, never silent fabrication.
• Reversible registrations — everything goes through ctx.effect() / ctx.on() / register().
• Research-only compliance — tool descriptions, cards, and reports carry 「仅供研究，不构成投资建议」; data points carry asOf and sources.

Known limitations

• Cordis events, not session-log events — the rc.6 Session.append signature has no ignorable marker for out-of-repo event types and no external event-registration surface, so appending custom industry-research/* session events would make the persistence coordinator refuse the log on restore. The durable record is the workspace artifacts; observability rides the typed Cordis events; model-visible tool results ride the durable tool/result session event.
• No ctx.attachment writes — the rc.6 attachment seam accepts images only (PNG/JPEG/WebP/GIF); Markdown reports therefore stay versioned workspace files, referenced by absolute path in the tool result.
• v1 reads text formats only — no PDF parsing; ask users to convert PDFs to text/Markdown first.
• Chinese-edition skills — the packaged methodology skills ship in Chinese; an English edition is future work.
• Foreground tools — industry_track is bounded by track.maxFetchesPerCall and fetchTimeoutMs and runs in the foreground; a background-jobs mode is future work.
• The auto-draft is mechanical — it summarizes artifacts and turns sourced data points into claims; narrative quality comes from model-authored drafts.

Development

pnpm install
pnpm run typecheck && pnpm run typecheck:ci
pnpm test
pnpm run build
pnpm run verify:self-contained && pnpm run verify:artifacts
pnpm run verify:readme-sync
pnpm run pack:check

Topics

dsh, dsh-plugin, deepseek-harness, cordis, industry-research, company-research, research, report

Contributors

• Initial design and implementation by the dsh-industry-research development session (DeepSeek Harness).

External contributions are welcome — open an issue or a pull request.

License

Apache-2.0 — see LICENSE.