Superpowers 7.0.2 fixes skill loading on Pi, OpenCode and Copilot for Windows

Superpowers 7.0.2, released November 3, fixes its bootstrap on six harnesses, from Copilot on Windows to Pi and OpenCode. The bootstrap loads the using-superpowers skill when a session starts. Where it failed, your agent began without its introduction to using skills.

Copilot CLI and VS Code Copilot run hook commands through PowerShell, which rejected the session start command with a parser error. The hook file now carries a PowerShell form of that command.

On Pi, the extension turned the bootstrap off after a session's first prompt, so every later prompt went without it. That change also rewrote the whole prompt cache on the second prompt, in more than four of five sessions measured.

Pi's skills filter on the Superpowers package entry works now. An empty list still loaded all 15 skills, because the extension added the whole skills folder back after Pi applied the filter.

OpenCode registers skills by bare names, like brainstorming. A model that copied the namespaced form from the bootstrap got skill not found and carried on without the skill. The bootstrap now tells it to use bare names.

In OpenCode 2, compaction and generate requests went out without the bootstrap. So compaction, the largest request of a session, missed the prompt cache. Both requests now carry it.

In Hermes Agent, each delegate task subagent looked like a new session and got the full bootstrap, so some loaded using-superpowers instead of doing their task. Children now skip it.

Muse ran the session start hook, a bash script, with sh. On Debian and Ubuntu, sh is dash, so the hook failed with exit 2 and the bootstrap never loaded. It now runs with bash.

The subagent-driven development helpers now act on the repository of the current directory, as git does. Outside one, they stop instead of writing into the plugin cache. A brand new project needs a first commit first.

task-brief now finds tasks numbered like A1, 6b and 2.1. And it refuses a brief that an unclosed code fence ran to the end of the plan, where one short task had come back holding five.

The systematic-debugging skill installed four pressure-test prompts beside its SKILL.md, opening with: this is a real scenario, you must choose and act. An agent could find them mid-task. They now live under tests.

If you run Superpowers on Pi or OpenCode, I would update to 7.0.2 first. Those fixes put the bootstrap back on every prompt and stop the cache misses. The full notes are on the release page.