SessionStart hook injects superpowers context twice into Claude Code sessions

[djm204]

Problem

The session-start hook outputs the superpowers skill content in two JSON fields for Cursor/Claude cross-platform compatibility:

{
  "additional_context": "<superpowers content>",
  "hookSpecificOutput": {
    "hookEventName": "SessionStart",
    "additionalContext": "<superpowers content>"
  }
}

Claude Code reads both fields and doesn't deduplicate, so the full superpowers prompt gets injected into the conversation context twice. This wastes context window tokens on every session.

Location

hooks/session-start lines 41-48

Suggested Fix

Only emit the field that the current platform consumes. For Claude Code, that's hookSpecificOutput.additionalContext. The top-level additional_context is the Cursor path.

Options:

1. Detect the runtime (Claude Code vs Cursor) and emit only the relevant field
2. Drop the top-level additional_context if Claude Code is now the primary target
3. Keep both but document that Claude Code will double-inject (so users can patch locally)

Workaround

Remove the "additional_context" line from the hook output manually. This survives until the next plugin update overwrites the cache.

[mvanhorn]

I've submitted a fix for this in PR #660. The hook now detects the runtime via CLAUDE_PLUGIN_ROOT and emits only the JSON field consumed by the current platform, avoiding double injection.

[obra]

This was already fixed in the current version of the session-start hook. The script now detects the platform and emits only the relevant field:

• Claude Code (CLAUDE_PLUGIN_ROOT set): emits only hookSpecificOutput.additionalContext
• Other platforms (Cursor, etc.): emits only additional_context

See hooks/session-start lines 40-59.

— Claude Opus 4.6, Claude Code 2.1.72

[mvanhorn]

Sounds good! thanks for taking a look