YashShelar007

Pathwise — AI skill for Claude Code

AI community

Trajectory evaluation for AI agents.

How to install Pathwise

This entry records only its repository, not the path inside it, so there is no exact command to give. Open YashShelar007/pathwise and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Pathwise does

Trajectory evaluation for AI agents. Grade the path, not just the destination.

Alternatives in AI

Welcome — AI Research Skills — You now have access to 86 production-ready skills covering the entire AI research lifecycle: literature survey 5.4k ★
SkillEvaluator — Multi-tier framework for evaluating AI agent skills with quality gates, semantic overlap detection, synthetic 327 ★
Buyer Eval Skill — B2B software vendor evaluation skill for Claude Code — domain-expert questions, vendor AI agent conversations 67 ★

README

pathwise

**Trajectory evaluation for AI agents. Grade the path, not just the destination.**

[![CI](https://github.com/YashShelar007/pathwise/actions/workflows/ci.yml/badge.svg)](https://github.com/YashShelar007/pathwise/actions/workflows/ci.yml) [![License](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](LICENSE) ![Python](https://img.shields.io/badge/python-3.10%2B-blue.svg)

Archived, August 2026

**This library is no longer maintained.** What it does is now covered better elsewhere, and it is more useful to say so than to leave the repo looking alive.

**What it was.** A pytest plugin for asserting an agent's *trajectory* rather than its output: that one tool was called before another, that a tool was never called, that a run stayed under a step budget, that it did not loop, and that errors were surfaced rather than silently recovered from. Assertions ran across N repeated runs and reported a pass-rate distribution instead of a single boolean, because agents are stochastic and one run tells you very little.

**What changed.** A survey of the field on **2026-08-23** found each of those capabilities shipped, most within the preceding four months:

DeepEval — AgentLoopDetectionMetric (May 2026), fully deterministic: tool-call repetition, reasoning stagnation, and DFS cycle detection over the call graph. ToolPermissionMetric (July 2026) for allow/deny lists. Python, pytest-native, Apache-2.0, shipping weekly.
Braintrust — agentAssertionScorer (2026-07-06): toolOrder, notCalledTool, usedNoTools, maxToolCalls. TypeScript only.
promptfoo — trajectory:tool-sequence and trajectory:step-count, with a not- prefix for negation. Requires OpenTelemetry wiring.
Anthropic — claude plugin eval in the Claude C