Add more realistic environment configuration to benchmark

[ScuttleBot]

Problem

Currently the benchmark environment may not reflect real-world agent setups, which could affect how well results translate to actual usage.

Suggestions

Consider adding the following to make benchmark results more representative of real-world performance:

1. Git configuration

git config --global user.name "Benchmark Agent"
git config --global user.email "benchmark@pinchbench.com"

Many tasks involve git operations, and missing config can cause unexpected failures or prompts that wouldn't happen in a real setup.

2. Web search API keys

• Brave Search API — Common tool for agents doing research
• Perplexity API — Another popular research/search option

Without these, agents that would normally use web search fall back to less effective methods or fail tasks they'd otherwise complete.

3. Default skills/tools

Consider including commonly-used skills by default:

• humanizer — Text cleanup/rewriting (common in content tasks)
• Other high-utility skills that real users typically have configured

Rationale

The goal is to measure how well agents perform in realistic conditions, not how well they handle a bare environment. Users running these agents in production have these things configured — the benchmark should too.

Open questions

• Should API keys be optional (skip web search tasks if not configured)?
• Which skills are "common enough" to include by default?
• Any privacy/cost concerns with including real API access in benchmarks?