docs: add intro video and make the README agent-agnostic

[uipreliga]

Summary

Two things, both docs/copy only — no source or behavior changes.

1. Add the intro video. New ## Watch the intro section in the README with a clickable YouTube thumbnail and the linked title: Coder Eval: UiPath open-source framework to test AI Coding Agents. Placed after the feature list so it does not orphan those bullets under the heading.

2. Make the framing agent-agnostic. The README and the Pages stub both read as a Claude Code tool, even though the framework has driven four harnesses for a while. OpenCode was missing from several surfaces entirely. New tagline throughout: "Playwright for coding agents".

README

• Title: "…AI coding agents and their skills" (was "Claude Code skills").

• Tagline under the badges: Playwright for coding agents — one declarative test file, any agent runtime, a real sandbox, and a pass/fail gate in CI.

• Intro states the switch cost explicitly: changing harness is one field (agent.type); tasks, criteria, scoring, telemetry, and reports stay the same.

• Quick Start prerequisites were "the Claude CLI" only. Now a table of all four agents with agent.type, install command, and a link to each guide:

  Agent	agent.type	Runtime
  Claude Code (default)	claude-code	brew install claude
  OpenAI Codex	codex	pip install 'coder-eval[codex]'
  Antigravity (Gemini)	antigravity	pip install 'coder-eval[antigravity]'
  OpenCode	opencode	npm install -g opencode-ai

• Task Definition notes swapping agent.type or -D agent.type=opencode, and links Run-Limit Parity.

• Known limits: "bring your own agent runtime and model credentials", including an OpenRouter key for open-weight models via OpenCode.

• Acknowledgments credit all four harnesses, not just the Claude Agent SDK.

Left Claude Code specific on purpose: the "Use inside Claude Code" plugin marketplace section — it genuinely is — but it now says so, and notes the suites authored there run on every harness.

Pages stub + package metadata

• .github/pages-stub/index.html: <title>, <meta name="description">, and the visible lead paragraph. OpenCode was absent from all three. The old lead also read "agents and their Claude Code skills", which parsed wrong even under the Claude-only framing.
• pyproject.toml: the stub's own comment says its description mirrors the package metadata, so description and keywords gain OpenCode — otherwise that comment stops being true the moment the stub changes. This ships to PyPI on the next release; happy to drop it if you would rather keep metadata out of a docs PR.

Verification

make lint (388 passed) and make check green. No doc-surface rule tripped — CE026's Action snippet, its prerequisite steps, and the CE028-generated docs index tables are all untouched. Both YouTube thumbnail URLs verified 200.

Not covered

The coder-eval.com website source is not in this repo, so its copy still carries the old Claude-Code-first framing and needs a separate change.

🤖 Generated with Claude Code

https://claude.ai/code/session_016v9L8gXbq4MzH6V9JZPqcb

[github-actions]

Claude finished @uipreliga's task in 1m 43s —— View job

---

Code Review in Progress

Todo List:

• Read .github/code_review.md for review standards
• Read CLAUDE.md for project conventions
• Review the full diff (git diff origin/main...HEAD)
• Read full context of changed files
• Check cross-file consistency
• Analyze "what's missing"
• Provide structured review feedback

Starting review now...

[bai-uipath]

Fix what you agree with, otherwise lgtm. The agent-agnostic rework is the right call and the video placement is clean. The notes below are all about the framing that's left.

• The skills-only read survives, and that's the one costing us. A prospect benchmarking agents on their own domain task bounced off this page believing the framework was skills-only. The title still leads with skills, the audience line still says "built for CLI and skill builders", and the "not an agentic coding benchmark" disclaimer reads as "we don't do domain benchmarks". Fix: keep the SWE-bench contrast but recast it as "not a fixed leaderboard: you bring the tasks and the scoring", and widen the audience line to include benchmark authors.

• The docs site still says "Claude Code skills". docs/index.md (title, description, H1), docs/llms.txt, docs/comparison.md, and mkdocs.yml's site_description live in this repo, sync to coder-eval.com/docs, and outrank both surfaces this PR touches. Cheap to fold in here.

• The Quick Start asks for the contributor toolchain. --extra dev pulls pytest, ruff, pyright, and pre-commit onto someone who just wants one run, and the pip install path is parked under a heading that reads CI-only. uv itself is the right call given the 3.13 floor, so this is about which extras and which ordering, not about uv.

• Minor: the prerequisites table hands users pip install three lines above uv sync (should be --extra codex / --extra antigravity); "installs the harness adapters, never the agents themselves" isn't true for codex or antigravity, whose extras ship the runtime; and the agent roster is now restated in five places with nothing holding it to AgentType, which is how OpenCode went missing in the first place.