AgentConfigScore is a deterministic linter and regression gate for persistent coding-agent instructions. It does not claim that A 100 means semantically perfect instructions.
I am looking for real repositories and sanitized before/after changes that answer a practical question: did the tool catch something useful, miss something important, or get in the way?
Fastest way to contribute
Open the structured Real-world case form. Choose one outcome:
- useful finding / prevented regression
- false positive
- false negative / missed regression
- setup or workflow feedback
Please include the AgentConfigScore version, the smallest sanitized instruction change, the observed rule IDs and score change, and what you expected instead. Never include credentials or private repository data.
Especially useful cases
- a refactor left stale paths in
AGENTS.md, CLAUDE.md, GEMINI.md, Cursor, or Copilot instructions
- a PR deleted a supported instruction file or reversed an exact directive
- different coding-agent instruction files drifted or contradicted each other
- a PR introduced unsafe shell guidance or secret-shaped text
- persistent instructions became significantly larger or duplicated
- vague, paraphrased, or tool-specific harmful guidance escaped the deterministic catalog
False positives are equally useful. The scanner intentionally prefers explainable, conservative rules over pretending to understand arbitrary prose.
Current evidence and limitations
Aggregated claims such as precision or recall will only be added after enough labeled cases exist. Until then, every accepted example should remain inspectable and reproducible.
AgentConfigScore is a deterministic linter and regression gate for persistent coding-agent instructions. It does not claim that A 100 means semantically perfect instructions.
I am looking for real repositories and sanitized before/after changes that answer a practical question: did the tool catch something useful, miss something important, or get in the way?
Fastest way to contribute
Open the structured Real-world case form. Choose one outcome:
Please include the AgentConfigScore version, the smallest sanitized instruction change, the observed rule IDs and score change, and what you expected instead. Never include credentials or private repository data.
Especially useful cases
AGENTS.md,CLAUDE.md,GEMINI.md, Cursor, or Copilot instructionsFalse positives are equally useful. The scanner intentionally prefers explainable, conservative rules over pretending to understand arbitrary prose.
Current evidence and limitations
Aggregated claims such as precision or recall will only be added after enough labeled cases exist. Until then, every accepted example should remain inspectable and reproducible.