Sci-Research: Agent Multi-Pipeline para Pesquisa Científica
haiou312/sci-research
Seis pipelines multi-agente para pesquisa, notícias e relatórios com credibilidade verificada
Ver no GitHubAgent Python que orquestra seis pipelines independentes para gerar artigos de pesquisa citados, briefings de notícias, monitoramento de reputação e relatórios macro/mercado. Implementa 18 agentes especializados com grading de credibilidade, verificação de datas por URL e filtros editoriais para entregar conteúdo de alta qualidade em até 5000 palavras.
Instalação
npx convez-ai haiou312/sci-research1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar artigos de pesquisa comparativa com múltiplas entidades e referências acadêmicas
- Produzir briefings de inteligência de notícias diárias por país com grading T1-T4
- Monitorar risco reputacional de empresas com alertas automatizados
- Criar relatórios semanais macro/mercado com dados de FRED, BOE, BOJ e yfinance
Veredito

Agent robusto e production-ready com arquitetura bem pensada (pipelines independentes, grading multi-camadas, verificação de datas). Ideal para equipes que precisam de pesquisa estruturada com garantia de credibilidade; exige investimento em configuração e ajuste de prompts para cada caso de uso.
README
A tradução para o português deste README ainda está na fila.
sci-research
A Claude Code plugin with six independent multi-agent pipelines for research, news intelligence, branded briefings, reputation monitoring, and weekly macro & market reports.
Given a topic, a country, a company, or a date, this plugin orchestrates specialised agents to produce a polished, sourced deliverable — academic article, news briefing, branded Word document, reputational risk email, or weekly research report.
Six Pipelines
/sci-research | /news-scan | /daily-news-intelligence | /daily-briefing | /reputation-track | /weekly-report | |
|---|---|---|---|---|---|---|
| Purpose | Multi-entity comparison research article | Real-time news briefing | Single-country daily news briefing | Multi-country branded briefing (SPD Bank) | Company reputation risk monitor | Weekly macro & market report |
| Time focus | Historical + current | Last 7/30/90 days | Single date | Single date (reads existing reports) | Single date | Trailing 7 days |
| Sources | Academic, official, T1 media | Wires, financial, industry press | T1-T4 graded (per-URL date verification) | Country reports from Pipeline C | News (T1-T4) + Reddit + X | Pipeline C reports + FRED/BOE/BOJ/yfinance |
| Output | ≤5000-word article with APA refs | 1000-3000-word briefing + images | Five-section Markdown + docx (+email) | 13-15-story branded Word document (+email) | Inline HTML email (only when negative) | Multi-section Markdown + docx (+email) |
| Default lang | zh | zh | zh | zh | zh | zh |
| Languages | zh / en / ja | zh / en / ja | zh / en / ja | zh / en | zh / en | zh / en / ja |
All six pipelines are completely independent — they don't share agents and changes to one don't affect the others.
Why This Plugin
- 18 specialised agents across six pipelines, each agent narrowly scoped
- Parallel retrieval — one researcher / scanner per entity, running simultaneously
- Credibility-first — three independent grading systems (academic five-star, news five-star, T1-T4 date-verified)
- Per-URL date verification in
/daily-news-intelligenceand/weekly-report— neighbouring days are discarded - Editorial second-pass filter (
news-verifier) for daily news — originality / authority / impact / dedup - External-view China matrix —
/daily-news-intelligence --country "China"uses only Western media + international organisations + external governments by structural design (no Chinese-domestic outlets, no Chinese government domains) - Free-prose Writer (1.9.0+) — daily news Writer composes explanatory prose in the target language, not a mechanical translation
- Branded Word output via SPD Bank template (
/daily-briefing) - Gmail SMTP email delivery built into three pipelines (daily-news-intelligence, daily-briefing, reputation-track, weekly-report)
- 8 quality hooks enforce word limits, entity coverage, reference integrity, news freshness, daily-news Markdown format, weekly-report format, and email-send safety
- Multilingual — Chinese / English / Japanese output
Installation
Option 1: From GitHub Marketplace
# In Claude Code, add this repo as a marketplace
/plugin marketplace add haiou312/sci-research
# Install the plugin
/plugin install sci-research@haiou312-sci-research
Option 2: Local Development
git clone https://github.com/haiou312/sci-research.git
claude --plugin-dir /path/to/sci-research
After modifying plugin files, reload without restarting:
/reload-plugins
Verify Installation
/plugin
You should see sci-research listed with its version.
Email Delivery (optional)
For pipelines C / D / E / F that support --email, configure Gmail SMTP via .env:
cp .env.example .env
# Edit .env with your Gmail address + app password
See .env.example for the required variables.
Usage
Pipeline A — /sci-research (Deep Research Article)
/sci-research <topic> --entities "Entity1,Entity2,..." --lang zh|en|ja
# Compare nuclear fusion progress across 4 countries
/sci-research 核聚变能源最新进展 --entities "中国,美国,EU,日本"
# mRNA vaccine landscape in English
/sci-research mRNA vaccine technology landscape --entities "US,EU,China" --lang en
Pipeline B — /news-scan (Real-Time News Briefing)
/news-scan <topic> --entities "Entity1,Entity2,..." --period 7d|30d|90d --lang zh|en|ja
# Recent Open Banking news, UK and China focus, 7 days
/news-scan 开放银行最新进展 --entities "中国,英国" --period 7d --lang zh
# Broad scan without entity filter
/news-scan AI regulation --period 90d --lang en
Pipeline C — /daily-news-intelligence (Single-Country Daily Briefing)
/daily-news-intelligence --country "<name>" [--date YYYY-MM-DD] \
[--lang zh|en|ja|zh+en|en+zh|zh+ja|ja+zh|en+ja|ja+en] \
[--out-dir <path>] [--min-per-category <N>] \
[--email <a@x.com,b@y.com>] [--email-attach both|docx|md|none] [--email-dry-run]
# Today's UK briefing in Chinese, dry-run the email
/daily-news-intelligence --country "United Kingdom" --email you@gmail.com --email-dry-run
# China briefing — external-view by design (Western media + international orgs + external govs only)
/daily-news-intelligence --country "China" --date 2026-05-11 --lang zh
# Japanese briefing for Japan, English output
/daily-news-intelligence --country "Japan" --lang en
# Bilingual (1.18.0+) — Chinese primary, ships 4 attachments (zh+en md+docx) + stacked zh+en email body
/daily-news-intelligence --country "China" --lang zh+en --email boss@company.com
Note on --country "China": Pipeline C scans China from an outside-observer perspective. Chinese-domestic outlets (Xinhua, People's Daily, Caixin, China Daily, SCMP, TechNode, etc.) and Chinese government domains (gov.cn, pbc.gov.cn, stats.gov.cn, …) are not queried. Source pool is Reuters / AP / AFP / Bloomberg / DJ Newswires (T1-wire); FT / WSJ / NYT / WaPo / Guardian / BBC / Telegraph / Times / Economist / Le Monde / Spiegel / FAZ / El País / Nikkei Asia (T1-flagship); NHK World / ABC Australia / Straits Times / Korea Herald / The Hindu (T2 regional); IMF / World Bank / WTO / OECD / BIS / IEA / US Treasury / USTR / State Dept / Commerce-BIS / White House / EU Commission / UK Gov / METI / MOFA Japan (T4 external institutions).
Pipeline D — /daily-briefing (Multi-Country Branded Word Document)
/daily-briefing [--date YYYY-MM-DD] [--countries "中国,英国,美国,欧洲,日本,韩国"] [--total 14] \
[--source-dir <path>] [--email <a@x.com>] [--email-subject <text>] [--email-dry-run] [--no-wait]
Reads existing Pipeline C reports from a directory, curates the most impactful 13-15 stories across countries, and emits a branded Word document via the SPD Bank template.
# Today's multi-country briefing with default countries and 14 stories
/daily-briefing --email you@gmail.com
# Specific date with custom country selection
/daily-briefing --date 2026-05-11 --countries "中国,日本,韩国" --total 12
Pipeline E — /reputation-track (Company Reputation Risk Monitor)
/reputation-track --company "<name|ticker>" [--date YYYY-MM-DD] [--lang zh|en] \
[--sources news,reddit,x] [--severity-min low|medium|high] \
[--email <a@x.com>] [--email-dry-run]
Resolves the company + executives, scans News + Reddit + X for adverse content, classifies category and severity. Silent when clean — only emails a report if negative findings exist.
# Scan Tesla for today's negative coverage
/reputation-track --company "TSLA" --email you@gmail.com
# Scan Alibaba for a specific date, low-severity threshold
/reputation-track --company "BABA" --date 2026-05-10 --severity-min low --lang en
Pipeline F — /weekly-report (Weekly Macro & Market Report)
/weekly-report [--end-date YYYY-MM-DD] [--start-date YYYY-MM-DD] \
[--countries "CN,US,UK,EU,JP,KR"] [--lang zh|en|ja] [--out-dir <path>] \
[--news-dir <path>] [--commodity-symbols "GC=F,SI=F,CL=F"] \
[--kr-bond-symbol 148070.KS] \
[--email <a@x.com,b@y.com>] [--email-attach both|docx|md|none] [--email-dry-run]
Aggregates the previous 7 days of /daily-news-intelligence reports for the Market Event section, then pulls live market data (FRED for money market & FX, BoE + BoJ + FRED for fixed income, yfinance for KR ETF proxy and commodities) for the data sections.
# Weekly report ending today
/weekly-report --email you@gmail.com
# Specific date range and country set
/weekly-report --end-date 2026-05-11 --start-date 2026-05-05 --countries "CN,US,JP" --lang en
Utility Commands
# Add more entities to an active research session (Pipeline A & B)
/add-entity "韩国,印度"
# Switch output language mid-session
/set-lang en
How It Works
Pipeline A — /sci-research
User Input (topic, entities, lang)
│
├─→ Researcher(A) ─┐
├─→ Researcher(B) ─┼─→ Comparator → Fact-Checker → Writer → Hooks → Article
└─→ Researcher(C) ─┘ (opus) (sonnet) (opus)
(sonnet, parallel)
| Agent | Model | Role |
|---|---|---|
researcher | sonnet | One instance per entity, multi-source retrieval |
comparator | opus | Cross-entity dimension selection + root-cause analysis |
fact-checker | sonnet | Claim verification with confidence labels |
writer | opus | Multilingual article synthesis with APA references |
Pipeline B — /news-scan
News-Scanner(×N) → news-imager → news-analyst → Hooks → Briefing
(parallel) (sonnet) (opus)
| Agent | Model | Role |
|---|---|---|
news-scanner | sonnet | One per entity, real-time news retrieval (no image extraction) |
news-imager | sonnet | Image extraction and relevance verification for top events |
news-analyst | opus | Dedup, timeline, impact matrix, trend signals |
Pipeline C — /daily-news-intelligence
daily-news-scanner → news-verifier → daily-news-writer → pandoc → email (optional) → publish (optional)
(sonnet) (sonnet) (opus)
| Agent | Model | Role |
|---|---|---|
daily-news-scanner | sonnet | English WebSearch + per-URL date verification (T4 → T1 → T2 → T3). Paywall fallback (Step 3.5) reroutes Hard-paywall hits as Corroborated, finds free Lead. |
news-verifier | sonnet | Editorial second-pass filter: originality / authority / impact / dedup |
daily-news-writer | opus | Consumes Verifier KEEP set, composes explanatory prose in target language, emits Markdown + APA refs |
Pipeline D — /daily-briefing
daily-news-reports/YYYY-MM-DD/*.md (existing country reports)
│
└─→ briefing-curator → generate-branded-docx.py → send-briefing-email.py
(opus) (python-docx) (Gmail SMTP)
Pipeline E — /reputation-track
reputation-resolver → reputation-scanner × 3 (parallel: news / reddit / x) → reputation-classifier → reputation-writer → email (only if findings)
(opus) (sonnet) (sonnet) (opus)
Silent exit when total_items_kept == 0. Reddit and X go through the apidirect MCP (single call per source) to avoid public-endpoint scraping blocks.
Pipeline F — /weekly-report
weekly-news-aggregator (Stage A) ─┐
market-data-collector (Stage B) ──┼─→ weekly-report-writer → pandoc → email (optional)
┘
| Agent | Model | Role |
|---|---|---|
weekly-news-aggregator | sonnet | Reads previous 7 days of Pipeline C reports, deduplicates, groups by country |
market-data-collector | sonnet | Runs FRED / BoE / BoJ / yfinance scripts in parallel, aggregates JSON |
weekly-report-writer | opus | Localises section headings per language-spec, composes Markdown |
Source Credibility Systems
Pipeline A (Academic + Official)
| Grade | Source Type | Examples |
|---|---|---|
| ★★★★★ | Peer-reviewed journals | Nature, Science, The Lancet |
| ★★★★★ | International org reports | WHO, OECD, IPCC, World Bank |
| ★★★★☆ | Government reports | DOE, EU Commission, 国务院白皮书 |
| ★★★★☆ | Wire services | Reuters, AP, AFP, Xinhua |
| ★★★☆☆ | Quality journalism | Scientific American, BBC Science |
| ★★☆☆☆ | Preprints | arXiv, medRxiv (flagged as non-peer-reviewed) |
Pipeline B (News & Media)
| Grade | Source Type | Examples |
|---|---|---|
| ★★★★★ | Wire services | Reuters, AP, AFP, Xinhua |
| ★★★★☆ | Financial / business media | FT, Bloomberg, CNBC, Caixin, BBC |
| ★★★☆☆ | Industry vertical media | Finextra, TechCrunch, 36氪 |
| ★★☆☆☆ | Think-tank commentary | Brookings, PIIE, VoxEU |
| ★☆☆☆☆ | Social media / blogs | Avoided unless verified expert |
Pipeline C (T1-T4 Date-Verified)
| Tier | Source Type | Examples |
|---|---|---|
| T1-wire | International wires (Universal) | Reuters, AP, AFP, Bloomberg, DJ Newswires |
| T1-flagship | Global prestige newspapers | FT, WSJ, Economist, NYT, WaPo, Guardian, BBC, Telegraph, Times, Le Monde, Spiegel, FAZ, El País, Nikkei Asia |
| T2 | Regional flagships | NHK World, ABC Australia, Straits Times, Korea Herald, The Hindu, CNBC, Politico Europe, DW, Al Jazeera |
| T3 | Sector verticals | TechCrunch, MIT Tech Review, Finextra, S&P Global, STAT, MLex |
| T4-official | Primary institutional releases | Fed, ECB, BoE, BoJ, Treasury, USTR, State Dept, EU Commission, IMF, World Bank, WTO, OECD, BIS, IEA |
For country = China: T1-wire is Universal only (no Xinhua / China News Service); T1-flagship Country-of-coverage is empty (no Caixin / People's Daily / SCMP); T3 has no Country: China rows; T4 uses an external-institution table (IMF, World Bank, WTO, OECD, BIS, IEA, US Treasury, USTR, State Dept, US Commerce/BIS, White House, EU Commission, UK Gov, METI, MOFA Japan). Chinese government domains are never queried.
Detailed rules:
- Pipeline A:
rules/research/source-credibility.md - Pipeline B:
rules/research/news-source.md - Pipeline C:
skills/daily-news-intelligence/references/rubric.md+skills/daily-news-intelligence/agents/daily-news-scanner.md§ Source Matrix
Quality Hooks
| Hook | Pipeline | Trigger | What It Does |
|---|---|---|---|
word-count-check | A | PostToolUse:Write | Blocks if Pipeline A article >5000 words / 7500 CJK chars |
entity-coverage-check | A | PostToolUse:Write | Warns if any entity lacks dedicated section / ≥3 mentions |
reference-validator | A | PostToolUse:Write | Warns if inline [N] doesn't match references list |
news-freshness-check | B | PostToolUse:Write | Warns if no Pipeline B source from last 7 days |
daily-news-format-check | C | PostToolUse:Write | Blocks Pipeline C Markdown if format violates spec (count invariants, [N] continuity, URLs, no global refs section) |
weekly-report-format-check | F | PostToolUse:Write | Blocks Pipeline F Markdown if format violates spec |
email-send-guard | C / D / E / F | PreToolUse:Bash | Blocks inline smtplib / MIMEMultipart / sendmail Bash commands that bypass the sanctioned send-*-email.py scripts |
research-summary | A & B | Stop | Logs session metadata (async, non-blocking) |
Project Structure
sci-research/
├── .claude-plugin/
│ ├── plugin.json # Plugin metadata
│ └── marketplace.json # Marketplace manifest
├── commands/ # 6 main + 2 utility slash commands
│ ├── sci-research.md
│ ├── news-scan.md
│ ├── daily-news-intelligence.md
│ ├── daily-briefing.md
│ ├── reputation-track.md
│ ├── weekly-report.md
│ ├── add-entity.md
│ └── set-lang.md
├── skills/ # 6 independent skill workflows; agents/*.md live under each skill as prompt templates (NOT registered as `sci-research:*` subagents — see CLAUDE.md § 项目定位 point 6)
│ ├── sci-research/ # Pipeline A
│ │ ├── SKILL.md
│ │ └── agents/
│ │ ├── researcher.md # Per-entity retrieval
│ │ ├── comparator.md # Cross-entity dimensions
│ │ ├── fact-checker.md # Claim verification
│ │ └── writer.md # Article writer
│ ├── news-scan/ # Pipeline B
│ │ ├── SKILL.md
│ │ └── agents/
│ │ ├── news-scanner.md # Per-entity news (no images)
│ │ ├── news-imager.md # Image extraction + verify
│ │ └── news-analyst.md # Dedup + timeline + impact
│ ├── daily-news-intelligence/ # Pipeline C
│ │ ├── SKILL.md
│ │ ├── agents/
│ │ │ ├── daily-news-scanner.md # Single-date scan + date gate + cross-category dedup + Cat5↔Cat6 routing
│ │ │ ├── news-verifier.md # Editorial second-pass filter
│ │ │ ├── daily-fact-extractor.md # Verifier KEEP → YAML fact manifest
│ │ │ ├── daily-news-writer.md # Free-prose target-language writer
│ │ │ └── daily-editor.md # 5-pass post-Writer editor
│ │ └── references/ # Pipeline C specs
│ │ ├── email-spec.md
│ │ ├── language-spec.md
│ │ ├── output-spec.md
│ │ ├── rubric.md
│ │ ├── schemas.md
│ │ └── verification.md
│ ├── daily-briefing/ # Pipeline D
│ │ ├── SKILL.md
│ │ ├── agents/
│ │ │ └── briefing-curator.md # Multi-country curator
│ │ ├── template/briefing-template.docx # SPD Bank brand template
│ │ ├── references/email-spec.md
│ │ └── scripts/
│ │ ├── generate-branded-docx.py
│ │ └── send-briefing-email.py
│ ├── reputation-track/ # Pipeline E
│ │ ├── SKILL.md
│ │ ├── agents/
│ │ │ ├── reputation-resolver.md # Ticker/name → exec list
│ │ │ ├── reputation-scanner.md # Per-source (news/reddit/x)
│ │ │ ├── reputation-classifier.md # Per-item negativity grader
│ │ │ └── reputation-writer.md # HTML email body composer
│ │ └── references/ # Pipeline E specs
│ │ ├── entity-resolution.md
│ │ ├── source-matrix.md
│ │ ├── negativity-rubric.md
│ │ ├── html-template.md
│ │ ├── email-spec.md
│ │ └── schemas.md
│ └── weekly-report/ # Pipeline F
│ ├── SKILL.md
│ └── agents/
│ ├── weekly-news-aggregator.md # 7-day event aggregator
│ ├── market-data-collector.md # FRED/BoE/BoJ/yfinance
│ └── weekly-report-writer.md # Weekly report writer
├── contexts/sci-research.md # Research mode behavioral context
├── hooks/hooks.json # Hook configuration (8 hooks)
├── scripts/
│ ├── hooks/ # Hook implementations (Node.js)
│ │ ├── word-count-check.js
│ │ ├── entity-coverage-check.js
│ │ ├── reference-validator.js
│ │ ├── news-freshness-check.js
│ │ ├── daily-news-format-check.js
│ │ ├── weekly-report-format-check.js
│ │ ├── email-send-guard.js
│ │ └── research-summary.js
│ └── send-report-email.py # Gmail SMTP (Pipelines C / E / F)
├── rules/research/ # 3 quality rules
│ ├── source-credibility.md # Pipeline A grading
│ ├── output-format.md # Pipeline A format
│ └── news-source.md # Pipeline B grading & dedup
├── examples/ # Sample outputs
│ ├── nuclear-fusion-zh.md # Pipeline A (Chinese)
│ └── mrna-vaccine-en.md # Pipeline A (English)
├── .env.example # Gmail SMTP environment template
├── CLAUDE.md # Project guidance for Claude Code
├── README.md
└── LICENSE
Customization
| Goal | Edit |
|---|---|
| Pipeline A word limit | scripts/hooks/word-count-check.js (WORD_LIMIT, CHAR_LIMIT_ZH) |
| Pipeline B news-freshness window | scripts/hooks/news-freshness-check.js (7-day constant) |
| Pipeline A comparison dimensions | skills/sci-research/agents/comparator.md § Dimension Selection |
| Pipeline A source grading | rules/research/source-credibility.md |
| Pipeline B news source rules | rules/research/news-source.md |
| Pipeline C source matrix / date verification | skills/daily-news-intelligence/agents/daily-news-scanner.md + skills/daily-news-intelligence/references/rubric.md |
| Pipeline C external-view China rules | skills/daily-news-intelligence/agents/daily-news-scanner.md § Source Matrix § T4-official + Step 2.1 |
| Pipeline C output format / Markdown contract | skills/daily-news-intelligence/references/output-spec.md |
| Pipeline C language localisation | skills/daily-news-intelligence/references/language-spec.md |
| Pipeline C email delivery | skills/daily-news-intelligence/references/email-spec.md + scripts/send-report-email.py |
| Pipeline D brand template | skills/daily-briefing/template/briefing-template.docx |
| Pipeline D curator rules | skills/daily-briefing/agents/briefing-curator.md |
| Pipeline E negativity rubric | skills/reputation-track/references/negativity-rubric.md |
| Pipeline E HTML email template | skills/reputation-track/references/html-template.md |
| Pipeline E entity resolution | skills/reputation-track/references/entity-resolution.md + skills/reputation-track/agents/reputation-resolver.md |
| Pipeline F market-data sources | skills/weekly-report/agents/market-data-collector.md |
| New output language | skills/sci-research/agents/writer.md + skills/news-scan/agents/news-analyst.md + skills/daily-news-intelligence/agents/daily-news-writer.md + commands/set-lang.md + skills/daily-news-intelligence/references/language-spec.md |
| Adding hook / changing email-send guard | scripts/hooks/email-send-guard.js + the relevant SKILL.md email step |
Examples
See the examples/ directory for complete sample outputs:
- nuclear-fusion-zh.md — Nuclear fusion progress, China / US / EU / Japan comparison (Chinese)
- mrna-vaccine-en.md — mRNA vaccine technology landscape, US / EU / China comparison (English)
Requirements
- Claude Code CLI
- Node.js ≥ 18 (for hook scripts)
- Python 3 (for email delivery scripts; only required when
--emailis used) pandoc(for Markdown → docx conversion in Pipelines C / F)- Internet access (for WebSearch / WebFetch)
- Gmail SMTP credentials (only for pipelines that support
--email; see.env.example)
Acknowledgements
Plugin structure inspired by everything-claude-code by Affaan Mustafa.
License
MIT