diff --git a/chat.go b/chat.go
index b132f42..88c7b2d 100644
--- a/chat.go
+++ b/chat.go
@@ -73,6 +73,13 @@ func cmdChat(args []string) error {
 				fmt.Printf("\n%s\n", red("[erreur] "+ev.Err.Error()))
 			case ev.Stats != nil:
 				stats = ev.Stats
+			case ev.DropReasoning:
+				// Le tour a « pensé sans agir » : on relance. Impossible d'effacer le
+				// texte déjà imprimé en terminal — on referme juste la ligne reasoning.
+				if inReason {
+					fmt.Print("\n")
+					inReason = false
+				}
 			case ev.ToolUsed != nil:
 				if ev.ToolUsed.Done || ev.ToolUsed.Typing {
 					break // résultat / frappe live affichés côté web ; en terminal on garde l'annonce seule
diff --git a/llm.go b/llm.go
index 8591cea..a8ad80f 100644
--- a/llm.go
+++ b/llm.go
@@ -209,13 +209,18 @@ func EnabledTools(caps Caps) []Tool {
 }
 
 // StreamEvent is what a ChatCallback receives for each piece of streamed output.
-// Exactly one of {Content, Reasoning, ToolUsed, Stats, Err} is set per call.
+// Exactly one of {Content, Reasoning, ToolUsed, Stats, Err, DropReasoning} is set
+// per call.
 type StreamEvent struct {
 	Content   string
 	Reasoning string
 	ToolUsed  *ToolUsedEvent
 	Stats     *StatsEvent
 	Err       error
+	// DropReasoning demande à l'UI de retirer la dernière bulle de raisonnement :
+	// le modèle a « pensé sans agir » et on relance le tour, ce raisonnement-là
+	// est mort-né et ne doit pas rester à l'écran (sinon double raisonnement).
+	DropReasoning bool
 }
 type ToolUsedEvent struct {
 	Name   string
@@ -346,6 +351,11 @@ func runChat(ctx context.Context, messages []Message, temperature float64, caps
 	// several times, it's stuck — break instead of spinning to the iteration cap.
 	lastSig := ""
 	repeatSig := 0
+	// Garde-fou « pensé sans agir » : certains modèles à reasoning planifient un
+	// appel d'outil dans leur <think> puis émettent le token de fin SANS l'émettre
+	// (ni réponse, ni tool_call). On relance alors UNE fois le tour avec un nudge
+	// explicite au lieu d'afficher « pas de réponse ».
+	nudged := false
 	for iter := 0; iter < 8; iter++ {
 		payload := map[string]any{
 			"model":       "jean",
@@ -695,6 +705,21 @@ func runChat(ctx context.Context, messages []Message, temperature float64, caps
 		// (empty content, e.g. it stopped right after a tool result), say so
 		// instead of leaving the user staring at a silent, finished chat.
 		if strings.TrimSpace(assistantContent.String()) == "" {
+			// Filet de sécurité (le vrai fix est le prompt court, voir baseSystemPrompt) :
+			// si un modèle « pense sans agir » malgré tout, on le relance UNE fois avec
+			// une consigne impérative au lieu d'afficher « pas de réponse ».
+			if len(tools) > 0 && !disableTools && !nudged {
+				nudged = true
+				// Le raisonnement de ce tour avorté ne mène à rien : on demande à
+				// l'UI de l'effacer avant de relancer, pour ne pas afficher deux
+				// blocs de réflexion successifs.
+				cb(StreamEvent{DropReasoning: true})
+				messages = append(messages, Message{
+					Role:    "user",
+					Content: "Tu as réfléchi mais tu n'as ni appelé d'outil ni répondu. Agis MAINTENANT : appelle directement l'outil approprié (par ex. mem_search/mem_read/bash), ou donne ta réponse finale si tu as déjà l'information. N'explique pas, agis.",
+				})
+				continue
+			}
 			cb(StreamEvent{Content: "_(le modèle n'a pas produit de réponse — finish: " + finishReason + ")_"})
 		}
 		return extra, nil
diff --git a/tools.go b/tools.go
index c149948..fa38279 100644
--- a/tools.go
+++ b/tools.go
@@ -30,44 +30,23 @@ func baseSystemPrompt(caps Caps) string {
 	if !caps.Agent {
 		return ""
 	}
-	cwd, _ := os.Getwd()
-
 	var b strings.Builder
-	b.WriteString("You are an expert coding assistant operating inside jean, a coding agent harness. You help users by reading files, executing commands, editing code, writing new files, and leveraging persistent memory to remember things across sessions.\n\n")
-
-	b.WriteString("Available tools:\n")
+	// Prompt VOLONTAIREMENT court. Un préambule verbeux (longue liste de
+	// « guidelines », surtout des méta-instructions sur la réflexion) fait
+	// sur-raisonner les modèles à reasoning (Qwen3) : ils émettent leur <think>
+	// puis le token de fin SANS appeler d'outil (~25-45 % de tours « morts »
+	// mesurés). Une version courte et directe ramène ça à 0 %. NE PAS regonfler.
+	b.WriteString("You are jean, an expert assistant operating directly on this machine with real tools.\n\n")
+	b.WriteString("Tools:\n")
 	for _, l := range []string{
-		"bash: Execute a shell command on this machine (ls, grep, find, cat, sed, read logs, run scripts…) and return stdout, stderr and the exit code. This is your real access to the system.",
-		"edit: Modify a file on disk by exact replacement (old → new; old must match exactly once). Use this to patch a file instead of rewriting it whole.",
-		"mem_search: Search your memory (Markdown pages under MEMORY/). Returns a ranked list of {file, title, snippet}.",
-		"mem_read: Read a memory page (Markdown file), line-numbered, with optional offset/limit.",
-		"mem_add: Create a new memory page (one topic per page, descriptive kebab-case name, first line is a short # title).",
-		"mem_edit: Edit a memory page by exact replacement (old → new; old must match exactly once).",
+		"bash — run a shell command on this machine (inspect files, processes, logs, run scripts).",
+		"edit — patch a file by exact replacement (old → new, old must be unique).",
+		"mem_search / mem_read / mem_add / mem_edit — your persistent Markdown memory under MEMORY/.",
 	} {
 		b.WriteString("- " + l + "\n")
 	}
-
-	b.WriteString("\nGuidelines:\n")
-	for _, l := range []string{
-		"Reason briefly (1-2 sentences) then ACT. As soon as an action is possible, call the tool immediately — don't announce it, do it.",
-		"Never repeat the same reasoning step or the same tool call twice. If you hesitate (\"maybe I should check…\"), stop thinking and call the tool now.",
-		"Use bash for any system question (files, processes, network, OS) instead of assuming — you have real access, don't ask permission to inspect.",
-		"Use bash to read files (cat, sed -n); to change a file, use the edit tool (old → new) instead of rewriting it whole with bash.",
-		"Use mem_search FIRST whenever the user asks about something you might already know (preferences, ongoing projects, past decisions). Open the most relevant page with mem_read; snippets are hints, not the answer.",
-		"Save anything worth keeping across sessions with mem_add; update an existing page with mem_edit. Don't save ephemeral conversation state.",
-		"Prefer targeted, non-interactive commands; chain them with && when they depend on each other.",
-		"Avoid destructive commands (rm -rf, mkfs, dd…) unless the user explicitly asks.",
-		"If a command fails, read stderr and the exit code, then fix it — don't re-run the same command unchanged.",
-		"Never invent a tool: you only have bash, edit and the mem_* tools. Anything else (reading code, web access) goes through bash.",
-		"When the task is done, give a short final answer and stop. Don't chain useless tool calls.",
-		"Be concise in your responses.",
-		"Show file paths clearly when working with files.",
-	} {
-		b.WriteString("- " + l + "\n")
-	}
-
-	b.WriteString("\nCurrent date: " + time.Now().Format("2006-01-02"))
-	b.WriteString("\nCurrent working directory: " + cwd)
+	b.WriteString("\nWhen the user asks about something you might already know (preferences, projects, past decisions), call mem_search first, then mem_read the most relevant page. For anything about the system or files, use bash instead of guessing. Act immediately — call the right tool, or answer if you already have the info. Never end your turn after only thinking. Be concise.\n")
+	b.WriteString("\nDate: " + time.Now().Format("2006-01-02"))
 	return b.String()
 }
 
diff --git a/ui/index.html b/ui/index.html
index ef6f98d..25fa8a7 100644
--- a/ui/index.html
+++ b/ui/index.html
@@ -954,6 +954,14 @@ async function send(){
                 if(tu.name==='mem_add' || tu.name==='mem_edit') loadMem();
               }
             }
+            if(d.drop_reasoning){
+              // Le modèle a « pensé sans agir » : on relance le tour. Le
+              // raisonnement mort-né est retiré pour ne pas afficher un double bloc.
+              if(reasonEl){
+                const i=turnCollapsibles.indexOf(reasonEl); if(i>=0) turnCollapsibles.splice(i,1);
+                reasonEl.remove(); reasonEl=null; fullReason='';
+              }
+            }
             if(d.reasoning_content){
               // Start a fresh buffer per bubble — otherwise a reasoning block that
               // follows a tool call re-renders ALL previously accumulated reasoning
diff --git a/web.go b/web.go
index 0a1bc92..4e4d16e 100644
--- a/web.go
+++ b/web.go
@@ -680,6 +680,9 @@ func runChatStream(ctx context.Context, body chatReq, emit func(map[string]any)
 		if ev.Stats != nil {
 			return emit(map[string]any{"stats": ev.Stats})
 		}
+		if ev.DropReasoning {
+			return emit(map[string]any{"drop_reasoning": true})
+		}
 		if ev.Reasoning != "" {
 			return emit(map[string]any{"reasoning_content": ev.Reasoning})
 		}