[next-dev-loop] Fix some papercuts

[gaearon]

Alternative to #95135.

Summary

This fixes a bunch of papercuts I discovered by running this skill repeatedly and analyzing logs:

• Set the min version to Support glob in --wait vercel-labs/agent-browser#1483 which fixes --wait.
• Push the agent to read the docs so it doesn't guess browser commands incorrectly.
• Ask agent to derive session name from CWD so that it doesn't conflict for git worktrees or same-named directories.
• Explain the daemon warning better so the agent doesn't restart sessions for no reason.
• The mcp prefix was confusing the agent and it was invoking the commands wrong.
• Add a few more gotchas that the agent often runs into.

Test Plan

Harness

I've been running this in an eval that runs a small next project and navigates around it. The eval penalizes the score if the log shows the agent stumbling, waiting unnecessarily, getting errors, etc. This version was most successful so far:

Real Project

I've also tried it with a real project as a part of CC migration skill. It started the browser and later was able to talk via MCP and do edits seemingly without problems. More real testing would be nice — I don't have much experience with skills.

[github-actions]

Tests Passed

Commit: d24c7df

[github-actions]

Stats cancelled

Commit: d24c7df
View workflow run