agent_harness

Runs scoped research as a composite skill through the shared skill registry and router.

Create and maintain paper research packets in this repo with the required research/<paper>/ layout, architecture docs, and experiment implementations aligned to the agent-browser TypeScript stack. Use this whenever the user asks to add a paper, summarize research, create experiment plans, or implement a paper capability as a reference architecture.

ARIS-inspired deep research workflow with adversarial executor/reviewer loops, persistent memory, and claim-evidence assurance checks.

Toolkit for interacting with and testing local web applications using Playwright. Supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs.

Operating guide for the agent-browser workspace shell. Use this whenever the user asks how to inspect, navigate, or modify the active agent-browser workspace, Files surface, browser pages, sessions, clipboard history, render panes, or WebMCP tool flows. Prefer it before improvising tool chains because the active workspace, workspace files, and mounted session drives have specific semantics in this project.

Create an AgentEvals-style eval suite for a named agent under .agents/<agent-name>/.evals/. Use this whenever the user asks for an eval, regression suite, benchmark, or repeatable acceptance test for a workspace agent. Prefer it even when the user asks for a smoke test or acceptance check without naming AgentEvals directly.

Create a scoped agent folder with an AGENTS.md file under .agents/<agent-name>/. Use this whenever the user asks for a new agent, reusable agent instructions, a workspace-scoped AGENTS.md, or a named automation persona inside the current workspace. Prefer it even when the user only describes the role and not the file layout.

Create a reusable agent skill bundle under .agents/skills/<skill-name>/ that follows agentskills.io conventions. Use this whenever the user asks for a SKILL.md, reusable workflow skill, skill scaffold, or packaged agent capability inside the current workspace. Prefer it even when the user only describes the capability and not the folder structure.

Use when recalling, storing, or updating durable workspace memory through .memory markdown files.

Project context and canonical terminology for the agent_harness repository, especially the agent-browser product model. Use this whenever the user asks you to work in this repo, mentions agent-browser, workspaces, Research, Build, workspace switching, worktrees, page overlays, chat panel, terminal mode, workspace files, virtual filesystems, local model installation or browser-runnable ONNX models, active document surfaces, or Codespaces browser debugging. This skill exists to prevent terminology drift and incorrect architectural assumptions in this repository.

"Optimize an existing agent skill by running /sensei --gepa, tightening agentskills.io compliance, moving bulky guidance out of SKILL.md into references/, and creating agentevals.io-style evals. Use this whenever the user asks to improve a skill's triggering, shrink an overgrown SKILL.md, refactor a skill into progressive disclosure, or add repeatable evals for a skill under skills/ or .agents/skills/."

Simplify and harden existing codebases using Clean Code, Clean Architecture, SOLID, KISS, DRY, and package-first refactoring. Use this whenever the user asks to simplify code, refactor a module, reduce complexity, remove duplication, untangle architecture, replace custom infrastructure with a library, or clean up design patterns, even if they only say "make this cleaner" or "reduce technical debt".