Add raphaelchristi/harness-evolver

agents/raphaelchristi__harness-evolver/README.md

@@ -0,0 +1,52 @@
+# Harness Evolver
+
+**Harness Evolver** is a LangSmith-native autonomous optimizer for LLM agent codebases, distributed as a Claude Code plugin and npm package (`npx harness-evolver@latest`).
+
+Point it at any agent project and it will iteratively improve the agent — prompts, routing, tool calls, retrieval, and orchestration architecture — through a rigorous, data-driven loop grounded in LangSmith evaluation.
+
+## What It Does
+
+- **Sets up ground truth**: Creates a LangSmith Dataset with train + held-out splits and an LLM-as-judge rubric. Captures a baseline score.
+- **Evolves iteratively**: Spawns self-organizing proposer sub-agents, each investigating a specific failure lens derived from trace data. Proposers work in isolated git worktrees — nothing touches main until it wins.
+- **Gates rigorously**: Every candidate must beat the current best on held-out examples, pass constraint checks, and clear an efficiency gate. Regressions are blocked.
+- **Compounds learning**: Winning patterns are consolidated into evolution memory and promoted back into CLAUDE.md for future iterations.
+- **Archives everything**: Losers are archived with diffs and scores so future proposers can avoid dead ends or branch from promising failures.
+
+## Skills
+
+| Skill | What it does |
+|---|---|
+| `/harness:setup` | Explore project, configure LangSmith, create dataset, write `.evolver.json` |
+| `/harness:evolve` | Run the propose-evaluate-merge optimization loop |
+| `/harness:health` | Diagnose dataset quality and auto-correct issues |
+| `/harness:status` | Rich ASCII progress chart with iteration history |
+| `/harness:deploy` | Tag, push, and finalize the winning evolved version |
+| `/harness:certify` | Validate and certify the agent against its evaluation rubric |
+
+## Quick Start
+
+```bash
+cd my-llm-project
+export LANGSMITH_API_KEY="lsv2_pt_..."
+claude
+
+/harness:setup      # explores project, configures LangSmith
+/harness:health     # check dataset quality
+/harness:evolve     # run the optimization loop
+/harness:status     # check progress
+/harness:deploy     # tag and finalize
+```
+
+## Real Results
+
+Tested on a RAG agent (Agno + Gemini Flash Lite): **0.575 → 1.000 correctness (+74%)** in 7 iterations. 4 candidates merged, 3 rejected by gate checks.
+
+## Runtime
+
+Works with Claude Code, Cursor, Codex, and Windsurf. Requires `LANGSMITH_API_KEY`.
+
+## Links
+
+- Repository: https://github.com/raphaelchristi/harness-evolver
+- npm: https://www.npmjs.com/package/harness-evolver
+- Paper: https://arxiv.org/abs/2603.28052 (Meta-Harness, Lee et al.)

agents/raphaelchristi__harness-evolver/metadata.json

@@ -0,0 +1,14 @@
+{
+  "name": "harness-evolver",
+  "author": "raphaelchristi",
+  "description": "LangSmith-native autonomous optimizer for LLM agent codebases. Iteratively improves prompts, routing, tools, and architecture via multi-agent propose-evaluate-merge loops with git worktree isolation.",
+  "repository": "https://github.com/raphaelchristi/harness-evolver",
+  "version": "6.4.2",
+  "category": "developer-tools",
+  "tags": ["agent-optimization", "langsmith", "claude-code", "multi-agent", "evaluation", "prompts", "harness", "llm-ops", "git-worktrees", "autonomous"],
+  "license": "MIT",
+  "model": "claude-sonnet-4-6",
+  "adapters": ["claude-code", "system-prompt"],
+  "icon": false,
+  "banner": false
+}