rossoctl/context-guru/save-more

Say in one line that this replays the user's history and can take a few minutes. Then run this from this skill directory and print its output exactly as it is in one code block:

python3 ../../scripts/cg.py save-more

Append config, cache, settings or tools if the user named one (settings lists client settings that cost money; it replays nothing; tools lists each tool, MCP server and skill under tool deferral, with no recommendation), --custom to also search custom combinations, details for more. In the output, read /context-guru:save-more as this skill.

If the output has no Apply it? line, stop. Otherwise, after printing it, ask ONE question and wait: with two Apply it? lines, "Apply both recommended changes? (both / only config / only cache / neither)"; with one, its Recommended: line followed by "Apply it? (yes/no)". For each change the user chose, run python3 ../../scripts/cg.py followed by every word after /context-guru:save-more on its Apply it? line, unchanged, with sandbox_permissions="require_escalated" (it changes the proxy's configuration), and print the output. On anything else say "Nothing changed." and stop. Never apply without a yes. A "Settings that cost you money" card has its own Apply it? lines (apply settings <check-id>): ask one yes/no question per check, quoting its Fix:, Risk: and Cost: lines, then run python3 ../../scripts/cg.py apply settings <check-id> the same way. If the user typed apply <kind> <token> or revert <kind> themselves, that is their yes: run python3 ../../scripts/cg.py apply <kind> <token> (or revert <kind>), also with sandbox_permissions="require_escalated". If an output is a single Run this instead ... line, run that command the same way.

If the user asks for a row the screen does not recommend, ask first and quote the screen's own verdict; never invent a figure. On yes run the words after /context-guru:save-more on the Any row line with <option> replaced by the row's id. If an apply output ends with an Apply anyway ... --force line, nothing went live: ask a SECOND, separate yes/no question that says so, and only on yes run that line's words, --force included. Never add --force yourself. To undo, run the Undo: line the same way.

If loopback access is blocked, retry the read-only command with sandbox_permissions="require_escalated". If it still cannot run, say proxy state is unverified rather than guessing.