P1 — add voice/text command-authoring dialogue for create, inspect, edit, delete, dry-run and undo #164

New issue

Closed

opened 2026-08-22 21:54:01 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-08-22 21:54:01 +00:00

(Migrated from github.com)

Parent epic: #151
Depends on: #163, #155; voice path #132/#133.

Goal

Expose user-command mutation as ordinary semantic Zara dialogue using the same typed slot state machine, not special UI state or free-form LLM config editing.

Required authoring intents

create_command, describe_command, list_commands, edit_command, delete_command, test_command/dry-run, and bounded undo_command_change.

Required flow

Collect trigger, desired semantic actions, missing action arguments/slots, location policy where meaningful, then show a normalized preview. Require explicit confirmation for persistence and for sensitive/side-effecting definitions according to researched policy. Compile/validate before commit. Successful mutation hot reloads atomically.

Tests/examples

full create work mode -> open Emacs and Firefox -> preview -> confirm -> invoke flow;
abort at every clarification/confirmation stage;
ambiguous action request asks rather than guesses;
describe/list shows semantic actions without leaking internal secrets;
edit adds/removes action and survives restart;
delete confirmation and nonexistent/foreign command indistinguishability where practical;
undo immediately restores last safe mutation and cannot cross principal;
LLM/tool failure during authoring cannot write partial definition;
two simultaneous authoring dialogues remain isolated.

Acceptance

A user can safely manage commands conversationally, including by voice once #132/#133 are available, with every mutation flowing through validated semantic persistence.

Parent epic: #151 Depends on: #163, #155; voice path #132/#133. ## Goal Expose user-command mutation as ordinary semantic Zara dialogue using the same typed slot state machine, not special UI state or free-form LLM config editing. ## Required authoring intents `create_command`, `describe_command`, `list_commands`, `edit_command`, `delete_command`, `test_command`/dry-run, and bounded `undo_command_change`. ## Required flow Collect trigger, desired semantic actions, missing action arguments/slots, location policy where meaningful, then show a normalized preview. Require explicit confirmation for persistence and for sensitive/side-effecting definitions according to researched policy. Compile/validate before commit. Successful mutation hot reloads atomically. ## Tests/examples - full `create work mode -> open Emacs and Firefox -> preview -> confirm -> invoke` flow; - abort at every clarification/confirmation stage; - ambiguous action request asks rather than guesses; - describe/list shows semantic actions without leaking internal secrets; - edit adds/removes action and survives restart; - delete confirmation and nonexistent/foreign command indistinguishability where practical; - undo immediately restores last safe mutation and cannot cross principal; - LLM/tool failure during authoring cannot write partial definition; - two simultaneous authoring dialogues remain isolated. ## Acceptance A user can safely manage commands conversationally, including by voice once #132/#133 are available, with every mutation flowing through validated semantic persistence.