rossoctl/context-guru/preset-picker

Choose a context-reduction preset

Run python3 ../../scripts/codex_plugin.py configure --show. Explain the choices without selecting one for the user: off changes no content; conservative performs deterministic cleanup; medium adds a paid relevance-model pass; high adds summarization; xhigh adds the deepest paid sweep; ultrahigh is xhigh plus tool deferral, and it is experimental. On OpenAI, GLM and Kimi models it uses Context Guru's own search tool (#825). Ask which one they want, then run python3 ../../scripts/codex_plugin.py configure --preset NAME. Run the change with sandbox_permissions="require_escalated", explaining that it writes the user's context-guru configuration and restarts the local proxy. Report the emitted settings and whether the proxy restarted successfully. For ultrahigh the output also has a probe= line: the proxy was asked to start deferral's provider check. On models that are not Claude the proxy ignores it: no check is needed there.