persona-feedback hits v0.2.0 with per-persona browsers

The plugin's earlier approach relied on subagents inheriting the main conversation's MCP tools, which meant a single shared Playwright connection and one browser. Adding --isolated looked like a fix, but it only prevented the profile from being saved to disk, so the mixing persisted and was hard to notice unless one persona ran at a time. Documentation noted that plugin subagents ignore the mcpServers frontmatter field, so giving each subagent its own MCP server was not possible.

The move to Playwright CLI changed that by naming sessions with an execution timestamp and persona ID, so each persona gets its own browser. Closing and opening sessions is handled by the main agent, while personas only operate the session they are handed, and a two-persona test writing different values into the same form showed no crossover. The same update also consolidated duplicate observations about a screen and changed how "confused" is counted, switching from seconds between actions to the persona's own reports because most of that time was the AI thinking.

The security trade-off is that personas now receive Bash to run Playwright CLI commands, while Write and Edit are withheld. Only the allow-list line runs automatically, and other commands prompt the main session for confirmation. How well this holds up may depend on users refusing unexpected confirmations, since hooks are also unavailable to plugin agents and the restriction rests on instructions.