Pi Quiet Ask

An ordinary requested edit can score up to about 0.85 on destructive, so gate confirms at 0.9 destructive, 0.7 exfiltration, 0.85 beyond scope, or high impact with confidence. Defaults run as shadow warnings until you enforce.

Local patterns plus a Jev leak check catch known shapes and the remainder; a hit appends do-not-repeat advice to the tool result. Failure class with enough confidence appends one advice line and never blocks.

Built-ins first, then user packs, then trusted project packs, then config overrides. Later same-name packs replace earlier ones, single numbers override without copying, and bad packs are named and skipped.

Triage answers which option the context picks for ask_user calls and can suggest or auto-answer; graph tracks phase, progress, drift, and observed checks into a HUD plus steering. Both sit outside the generic pack actions.

pi 0.85 or newer and a TypeSafe API key. Install with `pi install git:github.com/HyunjunJeon/pi-quiet-ask`, then set TYPESAFE_API_KEY in the environment, in ~/.pi/agent/pi-quiet-ask.json, or in the project's .env. Without a key the extension registers nothing and costs nothing. With one, the footer shows quiet packs:5 triage:auto graph, and /quiet prints the full status.

Ten: block and confirm on tool_call, annotate on tool_result, steer at turn_end or agent_end, set_thinking, set_tools, warn, status, tag and allow. The split is between intervening actions and always-on ones. block, confirm, steer, set_thinking and set_tools are reported as would-block in shadow mode and recorded as shadow:block, while annotate, warn, status and tag run in both modes. An allow stops evaluation of further rules.

Sources are declared per pack: cwd, user_request, last_user_message, recent_turns for the last 6 turns at 600 characters each with the system prompt excluded, tool, arguments, output up to outputChars, is_error, prompt, assistant_text, turn_tools and run_tools briefs capped at a 300-character output head, recent_failures for the last 5 failed calls, turn_index, graph and evidence. cacheSeconds judges identical state plus questions once per interval, and the gate pack uses 120.