diff --git a/crates/op-host-desktop/src/chat_session_launch_design.rs b/crates/op-host-desktop/src/chat_session_launch_design.rs
index e0d0dbb2d..2b1792996 100644
--- a/crates/op-host-desktop/src/chat_session_launch_design.rs
+++ b/crates/op-host-desktop/src/chat_session_launch_design.rs
@@ -127,7 +127,7 @@ fn resolve_design_thinking(model: Option<&str>, chat_default: ThinkingMode) -> T
 /// falls through to the orchestrator path).
 ///
 /// Mirrors `launch_if_pending`'s builtin chat branch but uses the
-/// design toolset and a 8192-token budget.
+/// design toolset and a 16384-token per-turn budget.
 pub(super) fn launch_design_loop_turn(
     host: &mut WidgetHostNative,
     user_text: String,
@@ -161,7 +161,14 @@ pub(super) fn launch_design_loop_turn(
         system_prompt: op_ai_skills::design_agent_system_prompt().to_string(),
         user_message: user_text,
         history,
-        max_output_tokens: 8192,
+        // Per-TURN output cap. The real budget-burner was hidden reasoning, not
+        // the DSL: a glm-5.2 loop run streamed ~94k thinking chars vs ~4k of
+        // actual batch_design, blowing the cap so a turn truncated mid-JSON, its
+        // tool call failed to parse, and the loop stopped early with an
+        // unfinished design. The root fix is at the wire (the loop now sends
+        // thinking:{type:disabled} for glm/minimax); 16384 (up from 8192) is
+        // headroom so a genuinely large section-batch still completes per turn.
+        max_output_tokens: 16384,
         thinking,
         effort,
         attachments,
diff --git a/crates/op-host-services/src/chat_agent_loop.rs b/crates/op-host-services/src/chat_agent_loop.rs
index 8e6128a0f..4c0672c8c 100644
--- a/crates/op-host-services/src/chat_agent_loop.rs
+++ b/crates/op-host-services/src/chat_agent_loop.rs
@@ -53,6 +53,14 @@ pub struct AgentLoopConfig {
     /// shared execution path for BOTH regular builtin chat and the design
     /// loop, so the finalize side-effect MUST be opt-in per provider.
     pub finalize_on_exit: bool,
+    /// When true AND the model is a MiniMax / GLM reasoning model, send
+    /// `thinking:{type:"disabled"}` in the OpenAI-compat body. Without it a
+    /// reasoning model (glm-5.2) streams tens of thousands of hidden reasoning
+    /// tokens per turn — measured 94k thinking chars vs 4k of actual
+    /// `batch_design` DSL — which burns the per-turn `max_output_tokens` and
+    /// truncates the design mid-build. The single-shot builtin chat path
+    /// already gates this on the same flag; the loop body was missing it.
+    pub disable_thinking: bool,
 }
 
 impl AgentLoopConfig {
@@ -658,13 +666,26 @@ pub async fn run_openai_agent_loop(
     messages.push(json!({ "role": "user", "content": cfg.user_prompt }));
 
     for _turn in 0..cfg.max_turns.max(1) {
-        let body = json!({
+        let mut body = json!({
             "model": cfg.model,
             "stream": true,
             "max_tokens": cfg.max_output_tokens,
             "messages": messages,
             "tools": tools_json,
         });
+        // Turn OFF hidden reasoning for MiniMax / GLM. Without this a glm-5.2
+        // design turn spends its whole `max_tokens` on `reasoning_content` and
+        // truncates the `batch_design` mid-JSON — the single-shot builtin body
+        // gates the same field on the same flag (`chat_builtin_http`), but the
+        // loop body was missing it, so every loop turn leaked thinking.
+        if cfg.disable_thinking
+            && (crate::chat_builtin_http::is_minimax_model(&cfg.model)
+                || crate::chat_builtin_http::is_glm_model(&cfg.model))
+        {
+            if let Some(obj) = body.as_object_mut() {
+                obj.insert("thinking".into(), json!({ "type": "disabled" }));
+            }
+        }
         let resp = reqwest::Client::new()
             .post(&cfg.url)
             .bearer_auth(&cfg.api_key)
diff --git a/crates/op-host-services/src/chat_agent_loop_tests.rs b/crates/op-host-services/src/chat_agent_loop_tests.rs
index b70e75abd..063bd511f 100644
--- a/crates/op-host-services/src/chat_agent_loop_tests.rs
+++ b/crates/op-host-services/src/chat_agent_loop_tests.rs
@@ -242,6 +242,7 @@ fn anthropic_loop_executes_tool_and_continues_with_tool_result() {
         executor: executor.clone(),
         max_turns: 5,
         finalize_on_exit: true,
+        disable_thinking: false,
     };
     let (outcome, deltas) = run_loop_collect(cfg, true);
     assert_eq!(outcome, Ok(true));
@@ -318,6 +319,7 @@ fn loop_skips_finalize_when_disabled_for_plain_chat() {
         executor: executor.clone(),
         max_turns: 5,
         finalize_on_exit: false,
+        disable_thinking: false,
     };
     let (outcome, _deltas) = run_loop_collect(cfg, true);
     assert_eq!(outcome, Ok(true));
@@ -347,6 +349,7 @@ fn anthropic_loop_stops_at_turn_cap_with_max_tokens_reason() {
         executor: executor.clone(),
         max_turns: 2,
         finalize_on_exit: true,
+        disable_thinking: false,
     };
     let (outcome, deltas) = run_loop_collect(cfg, true);
     assert_eq!(outcome, Ok(true));
@@ -406,6 +409,7 @@ fn openai_loop_executes_tool_and_continues_with_role_tool_message() {
         executor: executor.clone(),
         max_turns: 5,
         finalize_on_exit: true,
+        disable_thinking: false,
     };
     let (outcome, deltas) = run_loop_collect(cfg, false);
     assert_eq!(outcome, Ok(true));
@@ -518,6 +522,7 @@ fn anthropic_loop_replays_screenshot_result_as_image_content_block() {
         executor: executor.clone(),
         max_turns: 5,
         finalize_on_exit: true,
+        disable_thinking: false,
     };
     let (outcome, _deltas) = run_loop_collect(cfg, true);
     assert_eq!(outcome, Ok(true));
@@ -615,6 +620,7 @@ fn openai_loop_replays_screenshot_result_as_image_url_part() {
         executor: executor.clone(),
         max_turns: 5,
         finalize_on_exit: true,
+        disable_thinking: false,
     };
     let (outcome, _deltas) = run_loop_collect(cfg, false);
     assert_eq!(outcome, Ok(true));
diff --git a/crates/op-host-services/src/chat_builtin_http.rs b/crates/op-host-services/src/chat_builtin_http.rs
index bfc912bd4..001a4844f 100644
--- a/crates/op-host-services/src/chat_builtin_http.rs
+++ b/crates/op-host-services/src/chat_builtin_http.rs
@@ -180,6 +180,7 @@ impl ChatProvider for ConfiguredBuiltinProvider {
                     executor,
                     max_turns: MAX_TOOL_TURNS,
                     finalize_on_exit: provider.finalize_on_exit,
+                    disable_thinking,
                 };
                 match provider.kind {
                     BuiltinAgentKind::Anthropic => run_anthropic_agent_loop(cfg, &tx).await,
@@ -237,7 +238,7 @@ impl ChatProvider for ConfiguredBuiltinProvider {
 
 /// MiniMax M 系("MiniMax-M*"、旧 "abab*")是推理模型,其思考由 MiniMax 专属的
 /// `thinking` body 字段控制。据模型名判定,以便只对它发关思考字段。
-fn is_minimax_model(model: &str) -> bool {
+pub(crate) fn is_minimax_model(model: &str) -> bool {
     let m = model.to_ascii_lowercase();
     m.starts_with("minimax") || m.starts_with("abab")
 }
@@ -247,7 +248,7 @@ fn is_minimax_model(model: &str) -> bool {
 /// 一个设计子任务 thinking_len≈3 万、text_len=0，整段 parse 失败）。它接受和
 /// MiniMax 同样的 `thinking:{type:"disabled"}`（curl 对 ark glm-5.2 验证：关思考后
 /// reasoning_tokens=0、content 为干净 JSON）。按名判定，只对 GLM 下发。
-fn is_glm_model(model: &str) -> bool {
+pub(crate) fn is_glm_model(model: &str) -> bool {
     model.to_ascii_lowercase().contains("glm")
 }