autoresearch-tasks

What problem does it solve?

It helps you discover autonomous-agent-friendly work for a repository while minimizing the risk of the agent “cheating” by deleting or shrinking the system to game the evaluation.

Core Features & Use Cases

• Oracle-first task design: Produces tasks around building or leveraging robust, behavior-preserving test oracles instead of vague metrics that are easy to Goodhart.
• Repo-aware workflow: Instructs the agent to inspect common repo signals (README/docs, build manifests, CI, devcontainers, routes/CLIs, schema/migrations, and existing tests).
• Tiered oracle strategy: Classifies repositories into Tier 1/2/3 based on oracle strength and then generates task packages appropriate to each tier (including oracle bootstrapping for Tier 2 and safe inventory for Tier 3).
• Anti-Goodhart constraints: Enforces guardrails like locking observable surface manifests, preventing deletion-based shortcuts, flagging large negative diffs, and requiring objective evaluator descriptions.

Quick Start

Ask an agent: “Use autoresearch-tasks to analyze this repository and produce ranked task packages with robust, behavior-preserving oracles and explicit anti-Goodhart safeguards.”