PaperInsight AgentFlow is a clean, agent-first PDF paper analysis toolkit. It keeps the refactored workflow only:

1. Use MinerU API to turn each PDF into Markdown.
2. Use IDE agents or web-enabled tools to match the paper and fill journal + latest impact factor.
3. Use Longcat to extract the remaining metrics from the Markdown, one paper per thread.
4. Merge identity data and metrics data into Excel/JSON reports.
5. Return the final report path after the run finishes.

• A focused CLI for the new agent-first workflow.
• MinerU parser integration with retry and SSL EOF download fallback.
• Longcat-based metric extraction.
• Incremental finalization that merges 03_paper_data.json and 04_metrics_result.json.
• Minimal tests covering the new flow only.

• paperinsight/cli.py: CLI entrypoint.
• paperinsight/agentflow/: prepare, identity import, metrics extraction, finalize.
• paperinsight/parser/mineru.py: MinerU API adapter.
• paperinsight/core/extractor.py: Longcat-driven metric extraction.
• paperinsight/core/reporter.py: Excel/JSON export.
• paperinsight/models/schemas.py: shared paper schema.
• docs/AGENTFLOW.md: stage-by-stage artifact contract.
• docs/PROJECT_LAYOUT.md: compact module map.

git clone <your-repo-url>
cd paper-analysis-toolkit-agentflow
pip install -r requirements.txt
pip install -e .

Runtime config is loaded from ~/.paperinsight/config.yaml. The repo only keeps config/config.example.yaml as a template.

Required keys for the refactored flow:

• mineru.token
• llm.api_key
• llm.provider=longcat

Copy the example if you need a fresh local config:

New-Item -ItemType Directory -Force "$HOME/.paperinsight" | Out-Null
Copy-Item config/config.example.yaml "$HOME/.paperinsight/config.yaml"

Sensitive fields are encrypted before saving by paperinsight.utils.config_crypto.

paperinsight agent prepare ./pdfs

Output per paper:

• 01_parse.md
• 01_parse_meta.json
• 02_identity_job.json

Run-level artifacts:

• manifest.json
• jobs/identity_jobs.jsonl
• jobs/identity_results.jsonl
• jobs/identity_prompt.md

Fill jobs/identity_results.jsonl with one JSON line per paper, then run:

paperinsight agent import-identity <run_dir>

This generates:

• 03_identity_result.json
• 03_paper_data.json

paperinsight agent extract-metrics <run_dir>

This generates:

• 04_metrics_result.json
• 04_metrics_meta.json

Recommended usage: one paper per clean thread so the model stays inside context limits.

paperinsight agent finalize <run_dir> --json

Finalize behavior:

• prefers 04_metrics_result.json
• overlays identity fields from 03_paper_data.json
• falls back to 03_paper_data.json when metrics are missing
• exports incomplete rows instead of silently dropping papers

Generated outputs:

• reports/paperinsight_report_<timestamp>.xlsx
• reports/paperinsight_report_<timestamp>.json

The CLI prints the final report path directly.

A typical run looks like this:

agent_runs/
  run_20260319_120000/
    manifest.json
    jobs/
      identity_jobs.jsonl
      identity_results.jsonl
      metrics_summary.json
      finalize_summary.json
    papers/
      0001_sample_abcd1234/
        01_parse.md
        01_parse_meta.json
        02_identity_job.json
        03_identity_result.json
        03_paper_data.json
        04_metrics_result.json
        04_metrics_meta.json
        05_final_paper_data.json
    reports/
      paperinsight_report_20260319_122253.xlsx
      paperinsight_report_20260319_122253.json

python -m pytest tests/test_agentflow_prepare.py tests/test_api_integrations.py tests/test_project_layout.py -q

This cleaned project intentionally does not keep the old desktop shell, legacy web crawlers, packaging scripts, PRD archives, or unrelated regression suites. The repo now starts from the refactored agent-first workflow only.