TrendR

Trend Research — Automated literature review + platform trend monitoring + Obsidian knowledge management.

4 Agents · 8 Skills · 9-source search · 9-platform trends · Basic / Full install

Tell your agent one sentence. TrendR handles the rest.

You: "Survey the latest advances in agentic RAG 2025"

TrendR:
  → 9-source parallel search, 81 candidate papers found
  → Deep-read 11 papers, structured notes + comparison matrix
  → 14KB literature review (taxonomy, gap analysis, BibTeX)
  → Auto-archived to Obsidian, paper pool persisted
  → Notifies you: Done ✅

Inspired by karpathy/autoresearch — redesigned from “LLM training optimization” to “paper search + literature review.”

TrendR is a research-agent harness system, evolving toward a domain-specific agent OS.


What Problem It Solves

Step Manual TrendR
Cross-platform paper search 3–4 hrs 5 min (9 sources parallel)
Filter relevant papers 2–3 hrs Auto score 1–5 + dedup
Deep read + notes 8–12 hrs Structured extraction (problem / method / result / limitation)
Write literature review 6–8 hrs Auto-generated (taxonomy + gap analysis + trends)
BibTeX references 1–2 hrs Automatic
Archive to knowledge base 1 hr Auto-sync to Obsidian
Total ~20–30 hrs ~30 min wait

Architecture

System Overview

┌──────────────────────────────────────────────────────┐
│              User  (Telegram / 飞书 / Web / CLI)       │
└─────────────────────┬────────────────────────────────┘
                      ▼
┌──────────────────────────────────────────────────────┐
│           OpenClaw Gateway (runs locally)             │
│                                                      │
│  ┌─ main agent ─────────────────────────────────┐   │
│  │    receive → decompose → dispatch → synthesize│   │
│  └──────┬──────────────┬──────────────┬──────────┘   │
│         ▼              ▼              ▼               │
│  ┌────────────┐ ┌────────────┐ ┌──────────┐ ┌──────┐ │
│  │paper-scout │ │paper-      │ │review-   │ │verif-│ │
│  │search·score│ │analyzer    │ │lead      │ │ier   │ │
│  │dedup       │ │read·extract│ │orchestr. │ │verify│ │
│  └────────────┘ └────────────┘ └──────────┘ └──────┘ │
│                                                      │
│  ┌── Skills (executable Markdown knowledge files) ──┐ │
│  │  paper-scout · paper-analyzer · review-writer   │ │
│  │  verifier · trendr-watchdog · platform-hotspots │ │
│  │  chrome-cdp-setup · research-vault              │ │
│  └──────────────────────┬───────────────────────────┘ │
│                         ▼                            │
│  ┌── v2 engine (state machine / validators / watchdog)┐│
│  │  INIT→DISCOVERY→ANALYSIS→GAP_CHECK→WRITING→VERIFY│ │
│  │  Basic:  9×academic APIs (free, no extra MCP)    │ │
│  │  Full:   +Scrapling +Nano-pdf +Context7 +Zotero  │ │
│  │  Fallback: Playwright (JS gaps / login only)     │ │
│  └──────────────────────────────────────────────────┘ │
└──────────────┬───────────────────────┬───────────────┘
               ▼                       ▼
  ┌─────────────────────┐   ┌───────────────────────┐
  │  9 Academic APIs    │   │  Obsidian Vault        │
  │  arXiv·S2·OA·PubMed │   │  paper pool / reviews  │
  │  CrossRef·DBLP···   │   │  cards / daily logs    │
  └─────────────────────┘   └───────────────────────┘

v2 State Machine

INIT → DISCOVERY → ANALYSIS → GAP_CHECK → WRITING → VERIFY → DONE
                          ↑                          ↓
                          └────── coverage gaps ─────┘

VERIFY fail:
WRITING ← verify.json.pass=false (max 2 repair rounds)

Pipeline

User prompt
    │
    ▼
Phase 1 · Search ────── paper-scout: 3–5 APIs parallel
    │                   → candidates.csv (40–100, scored 1–5)
    ▼
Phase 2 · Deep Read ─── paper-analyzer: reads score ≥ 4
    │                   → notes/*.md + matrix.csv
    ▼
Phase 3 · Gap Check ─── enough coverage? → Ph.4 : loop back
    ▼
Phase 4 · Write ──────── review-lead: full literature review
    │                   → review.md (15–25KB) + references.bib
    ▼
Phase 5 · Verify ──────── verifier: citation/claim/taxonomy check
    │                   fail → Ph.4 (max 2 rounds) | pass → Ph.6
    ▼
Phase 6 · Persist ──────── Basic: ~/research/  Full: Obsidian+Zotero
    ▼
Notify user (Telegram / 飞书)

Contents

Core (Basic + Full)

Type Name Role
Agent paper-scout 9-source search + score + dedup
Agent paper-analyzer Deep read + structured notes + matrix
Agent review-lead Pipeline orchestration + survey writing
Agent verifier Citation validity / taxonomy consistency
Skill paper-scout 9 academic API playbooks (10KB)
Skill paper-analyzer Structured extraction templates
Skill review-writer Survey template + quality checklist
Skill verifier VERIFY rules + verify.json protocol
Skill research-vault Obsidian persistence + paper pool index
Skill trendr-watchdog Runtime supervision + auto-resume
Skill platform-hotspots 9-platform trend scraping
Skill chrome-cdp-setup Chrome 146+ CDP dual-instance + cookie sync
Runtime engine/ v2: state machine + validators + watchdog
Runtime cli.py Standalone CLI: run / resume / status

Full Mode Extras

Component Function Without it
Scrapling JS-rendered page crawling Static API only, lower coverage
Zotero Auto-import DOI to library BibTeX still generated locally
Obsidian + obsidian-cli Paper cards + review archive + daily logs Results saved to ~/research/
Nano-pdf Full-text PDF reading Abstract/metadata only
Context7 Precise library docs for codex-coder Falls back to web search

Fallback Layer

Component Trigger
Playwright JS rendering gaps, login-gated pages, or explicit user request only

9 Search Sources

All APIs are publicly free — called via web_fetch, no extra MCP needed.

# Source Coverage Key Required
1 arXiv CS / math / physics preprints No
2 Semantic Scholar 200M+ papers, citation graph Recommended (free)
3 OpenAlex 250M+ works, fully open No
4 PubMed 36M+ biomedical No
5 CrossRef 140M+ DOI registry No
6 DBLP Computer science bibliography No
7 Europe PMC 40M+ life sciences No
8 bioRxiv Biology preprints No
9 Papers with Code ML papers + code repos No

Agent auto-selects 3–5 most relevant sources per topic.


Platform Trend Monitoring

Beyond academic papers, TrendR monitors 9 platforms in real time via Chrome CDP:

You: "What's trending in AI today?"

TrendR:
  → Chrome CDP automation (dedicated instance with login state)
  → Zhihu · Xiaohongshu · X/Twitter · Reddit
  → YouTube · GitHub Trending · Hacker News · Product Hunt
  → Cross-platform tech trend summary

Compatible Runtimes

Platform Support Notes
OpenClaw Full Native multi-agent + browser automation
Standalone CLI v2 engine python cli.py run --topic "..." --depth B
Claude Code Skills readable via CLAUDE.md, WebFetch / Agent tool
Codex Skills readable via AGENTS.md, curl/fetch, sequential
Other agents Skills readable Standard Markdown, API URLs copyable

Anti-Forgetting Mechanism

When using non-frontier models (e.g. MiniMax M2.5), agents may skip reading Skill files. TrendR uses a 3-layer defense:

Layer Mechanism
AGENTS.md Hard rule: “task description must include ‘read skills/xxx/SKILL.md first’”
SOUL.md Top warning: “⚠️ Step 1: read skills/xxx/SKILL.md”
SKILL.md Complete copy-paste commands, not abstract instructions

Obsidian Vault Structure

[Vault]/Research/
├── _index/
│   └── paper-pool.csv       ← paper pool (cross-project, cumulative)
├── papers/
│   └── 2301.12345.md        ← paper card (YAML frontmatter + wiki-links)
├── reviews/
│   └── project-name/
│       ├── review.md
│       ├── references.bib
│       └── matrix.csv
├── daily/
│   └── 2026-03-10.md        ← daily research log
└── templates/

Paper pool CSV tracks state: candidateanalyzedcited_in_review


Installation

git clone https://github.com/gy-hou/trendr.git
cd trendr
chmod +x install.sh
./install.sh

Choose Basic for zero extra dependencies, or Full for Scrapling + Obsidian + Zotero + Nano-pdf.


Known Limitations

  • Not real-time: academic APIs have rate limits (arXiv: 3s/request); full search takes a few minutes
  • Network policy variance: some DNS/proxy routes academic domains to 198.18.x.x (fake-ip); TrendR has fallback search but coverage may drop
  • Non-frontier model forgetting: MiniMax M2.5 may occasionally skip Skill files despite 3-layer defense
  • Full-text reading (Basic): abstract only; Full mode with Nano-pdf enables PDF deep reading
  • No dual-AI review: extensible (see paper-distill-mcp dual-review mode)

Credits