Detect and route to active AI runtime

Detect AI runtime and provide uniform abstraction for hooks and GUPP

Multi-runtime agent frameworks needing uniform hook/GUPP semantics across Claude Code, Codex, Gemini, Cursor without editing skills.

• · environment variables (GT_RUNTIME, CLAUDE_SESSION_ID)
• · .claude/settings.json
• · process tree inspection

• · RuntimeProvider enum (claude|codex|gemini|cursor|unknown)
• · StartupStrategy (hook_injection|startup_fallback|polling)
• · GUPPStrategy (hook_injection|prompt_preamble|polling)
• · session ID (if available)
• · stall threshold (seconds)
• · nudge interval (seconds)

• · process inspection (ps, pgrep)
• · filesystem checks (.claude/settings.json)
• · environment variable reads

• · Environment initialized (env vars readable)
• · Workspace accessible (for .claude/settings.json check)
• · Process tree visible (for binary detection)

• · GT_RUNTIME env var overrides detection cascade (may lie)
• · Process tree inspection unreliable (binary names change across versions)
• · Falls back to 'unknown' + polling if no signal matches (graceful degradation)

• · Detection cascade documents fallback order (explicit over specific-signal-first)
• · Capability matrix shows which runtimes support hooks vs polling
• · Graceful degradation: unknown runtime → polling (never crashes)
• · Mirrors Go binary's runtime.go provider abstraction (proven pattern)