[codex] Add Optimize Agent Prompt skill

[shubh24]

What changed

• add a reusable optimize-agent-prompt skill for improving Browserbase Agent API system prompts
• add a zero-dependency Node harness for Agent creation/update, run polling, message pagination, session-log capture, scoring, stuck-run stopping, trace inspection, and comparison reports
• add API and evaluation references, UI metadata, and three realistic skill evals
• list the skill in the repository README

Why

Browserbase Agents can now be improved with the same outer-teacher pattern as Autobrowse: hold the task and schema fixed, inspect observable run traces, change one prompt heuristic, and require an unchanged confirmation run before convergence.

The public name is deliberately optimize-agent-prompt rather than the broader optimize-agent, so the trigger clearly targets prompt optimization rather than model, tool, or architecture changes.

Impact

Demo and POC authors can create an isolated experiment workspace and converge on an evidence-backed Agent system prompt without rebuilding the API orchestration for every project.

Validation

• node scripts/validate-skills.mjs — 17 skills passed, zero warnings
• node --check skills/optimize-agent-prompt/scripts/optimize_agent_prompt.mjs
• full mocked lifecycle: init → Agent create → run → messages → session logs → score → inspect → report
• renamed CLI workspace initialization and report generation
• git diff --check