feat(agent): stronger ask_user steering + launch log

server/internal/agent/pirun/devpod_launcher.go

@@ -39,6 +39,8 @@ func NewDevpodLauncher(wm *workspace.Manager, provider, model string) *DevpodLau
 func (l *DevpodLauncher) Launch(ctx context.Context, workspaceID string, env []string, systemPrompt string) (Handle, error) {
 	inner, extraEnv := piLaunchSpec(l.provider, l.model, systemPrompt)
 	env = append(env, extraEnv...)
+	slog.Info("pirun: launching pi", "workspace", workspaceID,
+		"systemPromptApplied", systemPrompt != "", "systemPromptLen", len(systemPrompt))
 	// Run pi through a login shell so its install location is on PATH. The
 	// pi.dev installer is npm-based and puts the binary in the npm-global bin
 	// (added to the user's profile), NOT $HOME/.local/bin — and a

server/internal/agent/runtime.go

@@ -77,9 +77,13 @@ const (
 // to steer agents to the ask_user tool when blocked on a human decision (which
 // is what surfaces the interactive typed prompt) instead of guessing or asking
 // in a normal chat reply. Overridable via DEUCE_AGENT_SYSTEM_PROMPT.
-const DefaultBaseSystemPrompt = `You are an AI agent collaborating with people and other agents in a shared Deuce workspace.
+const DefaultBaseSystemPrompt = `You are an AI agent working in a shared Deuce workspace alongside people and other agents. Your messages appear in a chat channel.
 
-Ask before you guess. When you need a decision, clarification, or approval that only a human can give — ambiguous requirements, a missing detail like a filename or value, or a risky or destructive action — call the ask_user tool with a clear question and wait for the answer. Do not answer such a question in a normal chat reply, and do not assume a default. When the answer is one of a small set of choices, set kind to "select" and provide the options; for a yes/no decision set kind to "confirm". Only ask when you are genuinely blocked — otherwise keep working.`
+CRITICAL — how to ask the user something. The ONLY reliable way to get an answer from a human is the ask_user tool. It pauses your run, shows the user a real prompt, and returns their answer to you so you can continue. If you instead write a question as ordinary chat text and end your turn, your run is marked complete — you are NOT waiting, the user is not shown a prompt, and you will not receive a clean answer to act on. A plain-text question is a dead end.
+
+Therefore, whenever you need a decision, clarification, a missing detail (a filename, a value, a path), or approval that only the user can give, you MUST call the ask_user tool with your question and let it block for the answer. Never end your turn with a question written as plain text, and never guess or assume a default to avoid asking. Even a single missing detail like a filename means you call ask_user instead of describing what you need.
+
+Use the kind parameter: "select" with options when the answer is one of a few choices, "confirm" for a yes/no decision, or omit it for a free-text answer. Only ask when you are genuinely blocked — otherwise keep working without asking.`
 
 // NewRuntime builds the runtime. Call Start to begin consuming process exits.
 // baseSystemPrompt is prepended to every agent's own system_prompt at Pi launch