Imported from sarahllk/forkfund-mvp-Fintech-BAM (
AGENTS.md). Install upstream withnpx skills add sarahllk/forkfund-mvp-Fintech-BAM. Copyright stays with the author.
ForkFund — Agent Roles
This file describes the intended responsibilities of any AI coding agents working on this repository.
Scope
ForkFund is a prototype MVP with entirely synthetic data. Agents must not:
- add real API integrations (banking, government, POS providers)
- add authentication or user management
- add a database (SQLite, Postgres, etc.)
- add Docker or cloud deployment configuration
- commit secrets, API keys, or real personal/financial data
- add "Co-Authored-By: Claude" or similar AI attribution to commits
Module ownership
| Module | Responsibility |
|---|---|
src/connectors/ |
Simulate reading from PSD2, POS, accounting, KvK sources using CSV data |
src/standardization/ |
Transform raw connector DataFrames into the unified schema |
src/scoring/ |
Compute a 0–100 credit score with weighted sub-scores and explanations |
src/passport/ |
Assemble the final Restaurant Credit Passport from scored data |
src/lender/ |
Filter and rank lenders from data/synthetic/lenders.csv |
src/utils/data_generator.py |
Generate and write all synthetic CSV datasets |
pages/ |
Streamlit UI only — no business logic, only calls into src/ |
tests/ |
pytest unit tests for src/ modules |
Data rules
- All CSV data in
data/synthetic/is generated bysrc/utils/data_generator.py - Do not manually edit generated CSVs — re-run the generator instead
data/schemas/unified_schema.mdis the authoritative field reference
Style rules
- No comments explaining what the code does
- One-line docstrings only (no multi-paragraph blocks)
- All scores and sub-scores must be in range [0, 100]
- Connector functions must return a
pd.DataFrame - Do not add features beyond the current sprint's requirements
Sources of truth
Before making any claim about scoring logic, dataset parameters, or project scope, verify against:
| File | Authoritative for |
|---|---|
README.md |
Project overview and user-facing description |
docs/scoring_model.md |
Scoring dimensions, weights, and grade/risk bands |
docs/data_model.md |
Dataset sizes and generation parameters |
src/scoring/engine.py |
Canonical implementation — code takes precedence over documentation |
Workflow
- Human defines the task.
- Agent proposes and explains changes before modifying files.
- Human reviews and approves, rejects, or redirects.
- Human commits and pushes — agents do not commit or push autonomously.
Commit message conventions
When drafting commit messages for human review, use the imperative mood:
Fix scoring_model.md to match 7-dimension engine implementation
Update scenario labels to reflect actual KvK registration dates
Add screenshots and replace deployment URL placeholder in README
- One-line summary, 50 characters or fewer where possible
- No AI attribution in commit messages (see Scope above)
- Reference the affected module, file, or workflow step