Verifiable research-report engine for DeepSeek Harness with a content-addressed evidence ledger, versioned sealed reports where every claim carries a verification verdict and the manifest hash seals the directory, and retrieval orchestration that reuses the ctx.web and ctx.jobs seams.
Install
# from npm (prebuilt)
dsh plugin --profile web add dsh-research-report
# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)
dsh plugin --profile web add github:PerryLink/dsh-research-report
Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).
README
dsh-research-report
Ship a report whose every claim anyone can re-check offline on their own machine — change one byte and the seal hash stops matching.
A verifiable research-report engine for DeepSeek Harness: claims bind to immutable evidence snapshots, numbers and quotes are checked byte-for-byte, and the result is sealed into a versioned report directory whose seal is the SHA-256 of dsh-v0.2.2-rc.4(2026-11-18);manifest.json.
Install · Configuration · FAQ · 简体中文 · English · Español · Português · हिन्दी
dsh plugin --profile <name> add dsh-research-report
Community plugin, not affiliated with DeepSeek. Canonical repository: https://github.com/PerryLink/dsh-research-report (Gitee is a mirror).
What it is
Give the model a topic and a set of sources, and dsh-research-report binds every claim to immutable evidence snapshots, checks the numbers and quotes byte-for-byte, then seals the result into a versioned report directory anyone can re-check independently. The directory holds report.md, manifest.json, and the audit journals; the seal is the SHA-256 of manifest.json, and dsh-research-verify recomputes it offline without mounting the plugin. It deliberately ships no deep-research loop: retrieval runs over ctx.web, long runs over ctx.jobs, and planning and synthesis stay with the model or an upstream plugin.
Three-step setup
Install (pick one channel):
# npm channel (published releases) dsh plugin --profile web add dsh-research-report # git channel (latest main; equivalent to git+https://github.com/PerryLink/dsh-research-report.git) dsh plugin --profile web add github:PerryLink/dsh-research-report # tarball channel: pnpm pack in this repo, then dsh plugin --profile web add ./dsh-research-report-<version>.tgzBoth online channels install the bundle row (see
cordis.patch.yml) into the profile'sdsh.profile.bundlesstack; restart to take effect.Minimal config (defaults are usable; relative roots resolve against the harness working directory):
- insert: - id: dsh-research-report name: dsh-research-report config: ledgerRoot: .research-ledger reportRoot: research-reportsVerify it mounts:
dsh --profile web --dump-config | grep dsh-research-reportThen run the minimal loop in a session and re-check the seal with the CLI:
evidence_add({ origin: "docs/market.md", title: "Market snapshot" }) # → ev-1a2b3c4d5e6f research_report({ topic: "...", sections: [...], claims: [...], evidenceRefs: ["ev-1a2b…"] }) ledger_query({ claimId: "c1" }) # bindings + verdictdsh-research-verify --report research-reports/<topic>/<YYYYMMDD-HHmmss>
Uninstall: dsh plugin --profile web remove dsh-research-report (or remove the row from the profile patch).
Every PerryLink plugin is listed in the built-in DSH Desktop market: Market → Sources → Add source → paste https://perrylink-dsh-catalog.perrylink.workers.dev/catalog-source.json → select. Installation still requires the market's npm identity check and your confirmation.
Core capabilities
Three things carry the trust: evidence is immutable, checks are recomputable, and gaps are never hidden.
- Evidence ledger — a content-addressed snapshot store (
<ledgerRoot>/objects/<sha256>+ JSONL journals). The same content is stored exactly once; snapshots are immutable; every read recomputes the hash, so tampering or deletion is detected instead of trusted. - Claim ↔ evidence binding — claims register with the evidence ids they rely on; the ledger keeps the binding and every verification verdict (latest wins).
- Byte-level verification — every number and quoted span in a claim must be locatable verbatim in the bound snapshots. No bound evidence, or no checkable literal, marks the claim
unverified; bound evidence that cannot confirm or deny the claimed literals marks itinsufficient; a label whose snapshot value differs (the claimed value absent) marks itdisproven; tampered/missing snapshots mark itcontradicted. No semantics, no embeddings — auditable byte checks. - Optional numeric bridge — when a claim cites a structured workspace dataset (CSV/JSON) and
dsh-data-qualityis mounted, citations are cross-checked with tolerances through its frozenverifyCitationscontract; a dataset mismatch disproves the claim. - DOI evidence (zero network) — DOI origins are validated deterministically (
10.xxxx/xxxxstructure, a prefix whitelist, and a DOI character set); invalid DOIs fail loud. Optional journal/year metadata is accepted, andrequireJournalMetadatagates academic DOI evidence only when enabled. - Versioned sealed reports —
<reportRoot>/<slug(topic)>/<YYYYMMDD-HHmmss>/report.md+manifest.json+verification.jsonl+disconfirmation.jsonl; the seal hash is the SHA-256 of the manifest, which itself carries the report hash, every evidence hash, and the hash of each audit journal. - Pre-delivery re-audit & seal interception — before sealing, every bound claim is re-verified offline and journaled to
verification.jsonl; verdict drift, tampered/missing bound evidence, or a journal serialization failure blocks the seal (fail loud, no tunable). - Falsification ledger — every contradicted or disproven claim is recorded in
disconfirmation.jsonl(claim + evidence references + reason) and listed in the report's falsification-log appendix. - Negative knowledge — a disproven claim is remembered by its content hash (
disproofs.jsonl); the same text re-reported against unchanged evidence is forced back todisprovenand only re-verifies once the evidence changes. - Read-only verifier loop — after sealing, a deterministic
verifySealedReportfallback (zero network, zero model) recomputes the seal and audit hashes and re-checks every claim, writing the machine check toverifier-note.md; whenctx.jobsis mounted a read-only verifier job is also spawned (the model review is an enhancement, never a replacement). - Session-anchored evidence —
evidence_addaccepts an optionalsessionRef(sessionId+eventRange, validated loud); the anchor is stored, rendered in Appendix B, and registered in the manifest andverification.jsonl. Session-anchored evidence verifies honestly asunverified(session-anchored evidence requires a manual check against the session log). - Honest gaps — unverified, insufficient, contradicted, and disproven claims keep a visible marker in the report body and are listed in Appendix A. Nothing is silently passed.
- No deep-research loop — retrieval orchestration is deliberately reused:
ctx.webfor search/fetch,ctx.jobsfor long runs. Planning and synthesis stay with the model (or an upstream plugin).
Tool surface
evidence_add({ origin, content?, title? })— register one evidence snapshot. Withcontentthe text is stored verbatim; without it a URL origin is fetched throughctx.weband a workspace-relative path is read from disk (reads never escape the workspace). Returns the evidence id and SHA-256 hash.research_report({ topic, title?, sections, claims, evidenceRefs, gather?, depth?, background? })— assemble and seal a report: validate (unregistered claim references are rejected loudly), verify every claim, renderreport.mdwith visible markers, writemanifest.json, and return the seal hash.gather: trueruns ONE search round overctx.weband returns captured candidate evidence plus an explicit gap list — it never auto-assembles.background: truereturns{ kind: 'background', jobId }overctx.jobs.research_trail({ limit?, format? })— read-only claim → evidence → verdict trail: joins registered claims with their evidence and stored verdicts, marks claims with no resolvable evidence as unbound, lists dangling evidence ids, and returns Markdown for a reviewer plus structured rows. No new storage, no verdict changes.ledger_query({ claimId? | evidenceId? })— read-only binding/verdict queries; evidence is re-hashed on read so a tampered or missing snapshot is reported explicitly. With no id, returns a ledger summary.ctx.researchReport.assemble(request)— the frozen service surface for sibling plugins (seesrc/service.ts; gated byte-for-byte byscripts/verify-frozen-contract.mjs).
Examples
These are the plugin's real entry points; install once, then call one.
Detailed example
Command
research_report({
topic: "示例行业概览",
sections: [{ heading: "市场规模", paragraphs: [{ text: "2025 年全球示例行业市场规模为 1,280 亿元。", claimIds: ["c1"] }] }],
claims: [{ id: "c1", text: "市场规模为 1,280 亿元", evidenceIds: ["ev-5dc0a84d0e48"] }],
evidenceRefs: ["ev-5dc0a84d0e48", "ev-c6cf024e0916", "ev-61f8f4533cb7"]
})
Output
report sealed: /tmp/rr-test-4U2mlF/reports/示例行业概览/20261019-124904
seal (sha256 of manifest.json): 5c85c97597096ced281ca191d7abaaf7434f18b397d35ac81c31aa364a3b12f9
claims: 1 verified / 0 unverified / 0 contradicted / 0 insufficient / 0 disproven (of 1); evidence bound: 3
Observed from README.md:112-118,150; test/index.spec.ts:166-177; src/tools/research-report.ts:170-175
Failure branch
Input
{"topic":"probe","sections":[{"heading":"Growth","paragraphs":[{"text":"Growth was 42%.","claimIds":["c1"]}]}],"claims":[{"id":"c1","text":"Growth was 42%.","evidenceIds":[]}]}
Output
edge-case (no throw): isError=false; kind=sealed; verdicts[0]={"claimId":"c1","status":"unverified","note":"claim binds no evidence snapshot"}; counts={"verified":0,"unverified":1,"contradicted":0,"insufficient":0,"disproven":0}; evidenceCount=0; report.md marker=Growth was 42%. [未核实]
Observed from plugins/dsh-research-report/src/verify.ts:173-175 (unverified verdict); plugins/dsh-research-report/src/assemble.ts:209-220 (visible mark); plugins/dsh-research-report/README.md:87 (honest gaps); exercised through plugins/dsh-research-report/test/harness.ts
Brief examples: research_report · ledger_query
dsh-research-verify CLI
The standalone binary (bundled as lib/cli.js, zero @deepseek-ai imports) audits any sealed report directory without mounting the plugin:
dsh-research-verify --report <dir> [--seal <sha256>] [--ledger <dir>] [--format json|sarif]
--report <dir>— the sealed report directory (manifest.json+report.md+ the audit journals).--seal <sha256>— the expected seal hash to compare the recomputed manifest hash against. Omitted = the recomputed hash is reported without comparison.--ledger <dir>— the evidence ledger root (objects/<sha256>+index.jsonl) enabling per-claim byte-level re-checks. Omitted = claim re-checks are skipped honestly.--format—json(default) orsarif(SARIF 2.1.0).
It recomputes the seal hash (SHA-256 of manifest.json), the report.md hash, and the audit-journal hashes, re-runs the byte-level + integrity check for every claim, and exits non-zero when any performed check fails. The same verifySealedReport / buildVerificationReport / renderSarif / renderVerificationJson functions are exported from the package for library use.
Configuration
All tunables are Schemastery Config fields; invalid values fail the profile load loudly. Relative roots resolve against the harness working directory (the workspace). cordis.patch.yml documents each key inline.
Most-used keys:
| Key | Default | Meaning |
|---|---|---|
enabled |
true |
Master switch; false mounts nothing. |
ledgerRoot |
.research-ledger |
Evidence-ledger directory (objects + JSONL journals). |
reportRoot |
research-reports |
Sealed-report root (versioned per topic and timestamp). |
workspaceRoot |
(host session working directory) | Root for workspace-path evidence; defaults to the session working directory, then the harness cwd. |
maxEvidenceBytes |
2097152 |
Hard cap on one evidence snapshot's UTF-8 bytes. |
maxEvidencePerReport |
200 |
Hard cap on evidence items bound into one report. |
| Key | Default | Meaning |
|---|---|---|
enabled |
true |
Master switch; false mounts nothing. |
ledgerRoot |
.research-ledger |
Evidence-ledger directory (objects + JSONL journals). |
reportRoot |
research-reports |
Sealed-report root (versioned per topic and timestamp). |
workspaceRoot |
(host session working directory) | Root for workspace-path evidence; defaults to the session working directory, then the harness cwd. |
maxEvidenceBytes |
2097152 |
Hard cap on one evidence snapshot's UTF-8 bytes. |
maxEvidencePerReport |
200 |
Hard cap on evidence items bound into one report. |
fetchTimeoutMs |
20000 |
Deadline (ms) for one ctx.web fetch during capture. |
requireJournalMetadata |
false |
When true, DOI-typed evidence must carry a journal name and publication year at registration (fails loud otherwise). |
Example override in your profile patch:
- insert:
- id: dsh-research-report
name: dsh-research-report
config:
reportRoot: research-reports/acme
maxEvidencePerReport: 100
requireJournalMetadata: true
FAQ
Both sit under the configured roots: ledgerRoot (default .research-ledger) and reportRoot (default research-reports), resolved against the harness working directory when relative. Snapshots are content-addressed and immutable, and the same content is stored only once; delete or rewrite an object and the next read's hash check fails, the claims bound to it verify as contradicted, and ledger_query reports integrity: missing/tampered — nothing is quietly treated as fine.
Never silently. The shipped dsh-base mounts search only, so URL capture fails loud (WEB_UNAVAILABLE/WEB_PROVIDER_UNAVAILABLE) until a fetch provider is configured; search-based gather lists the uncaptured sources in the gap list and never fabricates evidence or auto-assembles a report.
No. ctx.dataQuality is optional and is looked up with ctx.get at call time — never injected. When present it cross-checks dataset citations through its frozen verifyCitations contract; when absent that path is skipped honestly. This plugin registers no waterfall listeners at all.
Verification is pure byte work: no network, no model calls, linear in the number and quote literals of a claim and the size of the bound snapshots. The pre-seal re-audit is the same deterministic work repeated. The only external wait is evidence capture, bounded by fetchTimeoutMs (default 20000 ms); long reports can run with background: true over ctx.jobs.
That is the byte-level design at work: the built-in check locates number and quote literals, and cannot judge semantic equivalence across a rewrite. See the first entry under Known limitations — auditable beats clever.
Security boundaries
- Tamper-evidence by construction — every snapshot read recomputes SHA-256 against the index; a mismatch verifies the bound claims
contradictedandledger_queryreportsintegrity: tampered/missing. - Workspace confinement — local evidence reads resolve against the workspace root and refuse escapes (both sides are
path.resolved before comparison). - Fail-loud configuration — invalid bounds throw at mount; unregistered claim references, unknown evidence ids, and id/content conflicts throw at assemble.
- No credential handling, no hidden network — URL capture rides
ctx.web(provider selection, error taxonomy, and any SSRF policy stay with the deployment's web providers). - Reversible registrations — every contribution goes through
ctx.effect()/register(), so uninstall and hot reload are clean.
Permissions & data
dsh-research-report consumes only public seams: ctx.tools, ctx.systemPrompt, and optionally ctx.web / ctx.jobs / ctx.dataQuality (looked up with ctx.get at call time, never injected). It writes only inside the configured ledger and report roots (both default to workspace-local directories), reads workspace files only inside the workspace, and reaches the network exclusively through the harness web seam — never a direct fetch. Evidence snapshots are immutable and content-addressed; claim registrations are immutable; verdicts are append-only.
- Permissions declared in the manifest:
fs:workspace-write,network:ctx-web-only,session:read,session:write,native-code:none(see thedshWorkshopblock inpackage.json). - Session events:
research-report/evidence,research-report/verify, andresearch-report/sealare an adaptive mirror; they land in the session log only when the host knows the vocabulary, otherwise the ledger journals alone carry the durable truth (see Known limitations).
Known limitations
- Byte-level, not semantic — the built-in check locates number/quote literals verbatim; paraphrased claims without a checkable literal verify as
unverified, and a true claim whose number is absent while its label appears with a different value readscontradicted. This is a deliberate v1 choice (auditable beats clever). - Session events are adaptive — the plugin declares typed
research-report/evidence,research-report/verify, andresearch-report/sealsession events, but the 0.1.5-alpha.1Session.appendstill exposes noignorableoption and no plugin event-registration surface, so appends activate only when the host build knows the types (otherwise the persistence layer would refuse the log on restore). The ledger journals are always the durable source of truth. - Default profiles mount no fetch provider — the shipped
dsh-basemounts search only, so URL capture fails loud (WEB_UNAVAILABLE/WEB_PROVIDER_UNAVAILABLE) until a fetch provider is configured; search-basedgatherlists uncaptured sources in the gap list. - Single-workspace scope — ledger and report roots resolve against the harness working directory at mount; multi-workspace deployments should configure absolute roots per profile.
Interoperability with other DSH plugins
This plugin does not contend structurally for another plugin's resources:
- One service key,
ctx.researchReport— provided by the Service Definition insrc/service.ts; no second service key is registered. - No waterfall listeners — the plugin registers no
ctx.on()listener, so it takes part in no ordering-sensitive event chain; if one is ever added, the allow/passthrough branch MUST callnext(). - No HTTP routes, no client surface — no
webServerprefix and no client slot; it is a pure host plugin (the repo has noclient/). - The patch layer only inserts its own row — the bundle patch inserts
dsh-research-reportand never overrides a built-in row'sconfig. - Optional siblings are looked up on demand —
ctx.web/ctx.jobs/ctx.dataQualityare resolved withctx.getat call time; when missing, the affected path fails loud with an explicit reason instead of silently taking another route. - Frozen contracts are gated —
ctx.researchReport.assemble(src/service.ts) andCitationCheck*(src/verify.ts) are byte-frozen blocks shared with sibling plugins, andscripts/verify-frozen-contract.mjsguards drift in CI.
Compatibility
| Surface | Status |
|---|---|
| Harness | DeepSeek Harness dsh-v0.2.2-rc.4 (verified 2026-11-18). npm dev/test line 0.1.7-rc.2; the compat matrix pins all four declared peer lines. On 0.1.5-alpha.1 (adapted 2026-11-18) the session envelope keeps its ignorable field for stored-log read compatibility only — Session.append still cannot stamp it, so audit-gate behavior is unchanged. |
| Node | ^22.19.0 || >=24.9.0, ESM only ("type": "module") |
| Platforms | Anywhere DSH runs (pure host; the network is reached through ctx.web only) |
| Model | Text models fully supported; verification is byte comparison, no extra model capability required |
Peer dependencies: @deepseek-ai/cordis ^4.0.5-alpha.1, @deepseek-ai/schemastery ^3.18.5-alpha.1, plus @deepseek-ai/dsh-session, @deepseek-ai/dsh-tools, @deepseek-ai/dsh-system-prompt, @deepseek-ai/dsh-web, @deepseek-ai/dsh-jobs (ranges in package.json). Optional siblings (never required): ctx.web for capture, ctx.jobs for background assembly, ctx.dataQuality for dataset citation cross-checks.
Ecosystem & quality
- Gitee mirror (sync mirror; GitHub is canonical)
- dshfind plugin page (installs and ranking)
- OpenSSF Scorecard report (driven by .github/workflows/scorecard.yml)
- dsh-plugin-doctor verification report
- DSH Market (listed)
- 1024 store channel:
npm i -g dsh1024once, thendsh1024 plugin --profile web add dsh-research-report(counts toward the deepseek1024.com install ranking) - Ecosystem knowledge base — measured data, not marketing: plugin development guide · plugin-selection data · maintenance criteria
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:frozen-contract
node scripts/check-readme-sync.mjs
pnpm run check:lockfile
pnpm pack
typecheckresolves@deepseek-ai/*through the installed 0.1.7-rc.2 peers;typecheck:ciclearsskipLibCheckand enablesverbatimModuleSyntaxagainst the published types. Both must stay green.- Tests use the real
Context/Session/ToolRuntime/LocalJobRegistry/WebRuntimefrom the 0.1.7-rc.2 peers; only network backends are scripted providers registered through the realctx.webregistries. - Release:
node scripts/release.mjs <x.y.z>(bumps, stamps CHANGELOG, re-runs the gate, commits + tags; never pushes). - README images are generated programmatically by
scripts/gen-readme-assets.py(Python + Pillow): the raster banner isdocs/assets/dsh-research-report-hero-{zh,en}-{light,dark}.webp(1440×576) and the hand-written structure diagram isdocs/assets/dsh-research-report-flow-v5-{light,dark}.svg(1100×660), safe to re-run, shared by all five languages.
Directory layout:
| Path | Contents |
|---|---|
src/index.ts |
Function-plugin contract (name/inject/Config/apply, no default export) |
src/service.ts |
ctx.researchReport Service Definition + the byte-frozen assemble contract block |
src/ledger.ts |
The content-addressed evidence ledger (objects + JSONL journals, re-hash on read) |
src/verify.ts |
The byte-level checker + the frozen CitationCheck* bridge types |
src/assemble.ts |
Pure assembly: request validation, report.md rendering, deterministic manifest.json, seal |
src/gather.ts |
Capture (ctx.web / workspace files) and the gather candidate round |
src/tools/ |
The three consumers: evidence_add, research_report, ledger_query |
test/ |
vitest over the real host runtime; only network backends are scripted providers |
fixtures/ |
Three local documents covering the evidence → seal → tamper → contradicted flow |
Topics: dsh, dsh-plugin, deepseek-harness, cordis, research, evidence-ledger, verifiable-report, audit, citation-verification
Family & thanks
This project is part of the DeepSeek Harness plugin family from PerryLink (all Apache-2.0). If it is useful, a star helps the next person find it — nothing is gated behind it.
- @PerryLink — original author and maintainer: plugin architecture, evidence ledger, byte-level verification, sealed reports, five-language documentation, CI and release automation.
License
LICENSE (Apache License 2.0) © 2026 dsh-research-report contributors
Links
More in this category
loopx-project/loopx#dsh-loopx-plugin★ 6729
LoopX, a provider-neutral, local-first state kernel and control plane for long-horizon agents: keeps Goal, Todo, gate, evidence, quota, recovery, and handoff state above DeepSeek Harness, while the plugin bootstraps the CLI and skills, admits bounded same-session continuation, and adds a loopback GoalBar for the exact bound loop. Its HTTP routes have no DSH access check, so a web page open in your browser could call them.
Q00/ouroboros#integrations/dsh-plugin★ 6387
Config-only bundle that mounts Ouroboros through the DSH MCP client, exposing 36 interview, Seed, execution, evaluation, and evolution workflow tools in DSH.
chuspeeism/dashi-taskboard#deepseek-harness★ 3513
Embeds the active installed Codex Taskboard runtime in the DeepSeek Harness sidebar, using its launcher runtime descriptor instead of a fixed port.
NanmiCoder/dsh-agent-teams★ 3049
AgentTeams multi-agent teams.
EthanYoQ/AI-Novel-Writer#dsh-ai-novel-writer★ 2752
Installs a dedicated AI novel-writing preset and workbench: revisioned local project assets, a compact side drawer, and native approval-gated single-file changes.
shanliuling/dsh-image-gen★ 1655
Native conversational image generation for DeepSeek Harness: ask the agent to create an image, and it handles generation and keeps the result directly in the conversation.
Community comments
Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.