feat(prompt): Make CLI agent execute plans instead of waiting for confirmation

The previous wording told the model to "wait for the user to confirm
before exploring any files" on complex tasks. Qwen on Albert took it
literally: it would announce its plan ("Je vais lire le fichier X...")
and then stop, leaving the user to nudge it forward each step.

Reword so the model still announces a one-or-two-line plan (kept for
transparency) but immediately proceeds to execute it. Confirmation is
still required for destructive, irreversible actions (delete data, drop
tables, force-push). When scope is genuinely unclear, the model asks one
targeted question rather than stalling silently.

Author: obook

albert_code/core/prompts/cli.md

@@ -19,7 +19,7 @@ If unclear, default to investigate. It is better to explain what you would do th
 
 Explore. Use available tools to understand affected code, dependencies, and conventions. Never edit a file you haven't read in this session.
 Identify constraints: language, framework, test setup, and any user restrictions on scope.
-When given multiple file paths or a complex task: Do not start reading files immediately. First, summarize your understanding of the task and propose a short plan. Wait for the user to confirm before exploring any files. This prevents wasted effort on the wrong path.
+For complex tasks, briefly state your plan in one or two lines, then proceed to execute it. Do not wait for explicit user confirmation unless the user asked for plan-only mode or the task would have destructive, irreversible effects (delete data, drop tables, force-push, etc.). When in doubt about scope, ask one targeted question; otherwise act.
 
 Phase 2 — Plan (Change tasks only)
 State your plan before writing code: