Skip to content

Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.

feat(agent): show model reasoning as a "Thinking" disclosure

mrbizarro529c3ea76f71
diff --git a/agent/engine.py b/agent/engine.py
index 11479cd..3ac960e 100644
--- a/agent/engine.py
+++ b/agent/engine.py
@@ -81,6 +81,13 @@ class ChatResult:
finish_reason: str = "stop"
usage: dict = field(default_factory=dict)
model: str = ""
+ # Some chat servers (mlx-lm with reasoning models like Qwen 3.6,
+ # DeepSeek R1) return the model's chain-of-thought in a separate
+ # `reasoning` field on the response message. We surface it so the
+ # UI can show "what the model was thinking" — invisible reasoning
+ # is the #1 cause of "agent appears stuck" reports. Empty for
+ # non-reasoning models / non-supporting servers.
+ reasoning: str = ""
def _normalize_for_wire(messages: list[dict]) -> list[dict]:
@@ -199,13 +206,9 @@ def chat(messages: list[dict], config: EngineConfig,
# and `message.content` (final answer). When max_tokens is too small,
# reasoning consumes the budget and content comes back empty with
# finish_reason="length" — the agent gets nothing and the chat
- # appears "stuck". Two recovery paths:
- # 1. Empty content + non-empty reasoning + finish="length" → raise
- # a clear error so the user knows to bump max_tokens.
- # 2. Empty content + non-empty reasoning + finish="stop" → use
- # the reasoning as content. This is unusual (the model
- # decided to put its answer in the thinking block) but better
- # than silently dropping output.
+ # appears "stuck". We surface reasoning ALWAYS (so the UI can show
+ # "what the model was thinking") and fall back to using it as
+ # content only when content is empty.
reasoning = msg.get("reasoning") or ""
if not content and reasoning:
if finish == "length":
@@ -218,8 +221,10 @@ def chat(messages: list[dict], config: EngineConfig,
# No length issue but content is empty — fall back to reasoning
# so SOMETHING surfaces in the chat.
content = reasoning
+ reasoning = "" # already promoted to content; don't double-show
return ChatResult(content=content, finish_reason=finish, usage=usage,
+ reasoning=reasoning,
model=data.get("model", config.model))
diff --git a/agent/runtime.py b/agent/runtime.py
index b527867..27d0602 100644
--- a/agent/runtime.py
+++ b/agent/runtime.py
@@ -318,9 +318,16 @@ def run_turn(session: Session, user_message: str | None,
return
assistant_content = result.content or ""
- session.messages.append({"role": "assistant", "content": assistant_content})
+ # Persist reasoning alongside content so it survives panel restart
+ # and the UI can re-render it on session load. Most chat models
+ # don't return reasoning; the field is empty in that case.
+ msg_dict: dict = {"role": "assistant", "content": assistant_content}
+ if getattr(result, "reasoning", None):
+ msg_dict["reasoning"] = result.reasoning
+ session.messages.append(msg_dict)
yield emit(TurnEvent("assistant", {
"content": assistant_content,
+ "reasoning": getattr(result, "reasoning", "") or "",
"step": step_i,
"model": result.model,
"usage": result.usage,
diff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.py
index b4a44a2..25ea462 100644
--- a/mlx_ltx_panel.py
+++ b/mlx_ltx_panel.py
@@ -3824,6 +3824,13 @@ def _render_session_messages(sess: agent_runtime.Session) -> list[dict]:
"tool": action.get("tool"),
"args": action.get("args", {}),
}
+ # Surface model reasoning to the UI as a separate field so
+ # the renderer can show it as a collapsible "Thinking" block.
+ # Reasoning is what makes the agent feel "stuck" when it
+ # isn't — the user sees no content and assumes nothing's
+ # happening, but the model is reasoning.
+ if m.get("reasoning"):
+ entry["reasoning"] = m["reasoning"]
out.append(entry)
continue
out.append({"kind": role, "content": content})
@@ -7974,6 +7981,54 @@ HTML = r"""<!doctype html>
margin: 14px 0;
}
+ /* ---- Reasoning disclosure (chain-of-thought) ----
+ Shown above the assistant content as a collapsed details block.
+ Quiet by default so the answer stays the focus; expanding reveals
+ the model's thought process. Reasoning often runs 1–4k chars on
+ a planning turn — scroll caps prevent runaway height. */
+ .agent-reasoning {
+ margin: 0 0 8px 0;
+ border: 1px solid var(--border);
+ border-radius: 8px;
+ background: rgba(255,255,255,0.02);
+ }
+ .agent-reasoning > summary {
+ list-style: none;
+ cursor: pointer;
+ padding: 7px 12px;
+ display: inline-flex;
+ align-items: center;
+ gap: 7px;
+ font-size: 11px;
+ color: var(--muted);
+ user-select: none;
+ }
+ .agent-reasoning > summary::-webkit-details-marker { display: none; }
+ .agent-reasoning > summary svg {
+ transition: transform 0.18s ease;
+ flex-shrink: 0;
+ }
+ .agent-reasoning[open] > summary svg { transform: rotate(90deg); }
+ .agent-reasoning > summary .label {
+ font-weight: 600;
+ color: var(--text);
+ }
+ .agent-reasoning > summary .meta {
+ color: var(--muted);
+ font-weight: 400;
+ }
+ .agent-reasoning .agent-reasoning-body {
+ padding: 0 14px 12px 26px;
+ font-size: 11.5px;
+ line-height: 1.6;
+ color: rgba(255,255,255,0.62);
+ font-family: ui-monospace, SFMono-Regular, Menlo, monospace;
+ white-space: pre-wrap;
+ word-break: break-word;
+ max-height: 320px;
+ overflow-y: auto;
+ }
+
/* ---- Tool cards (calls + results, expandable) ----
Visually quieter than they used to be: 2px left accent (was 3px),
subtle background, smaller font, less vertical padding. The point
@@ -14738,6 +14793,13 @@ function renderMessage(m) {
body.appendChild(renderAttachmentChips(m.attachments));
}
+ // Reasoning disclosure — surfaces the model's chain-of-thought as a
+ // collapsible "Thinking" block above the content. Closed by default
+ // so the chat reads as the answer; expand to see the model's work.
+ // The #1 fix for "agent appears stuck" — reasoning models spend most
+ // of their time here and the user couldn't see it.
+ if (m.reasoning) body.appendChild(renderReasoningBlock(m.reasoning));
+
const content = document.createElement('div');
content.className = 'agent-msg-content agent-md';
content.innerHTML = mdToHtml(m.content || '');
@@ -14750,6 +14812,29 @@ function renderMessage(m) {
return row;
}
+// Reasoning block — shows the model's chain-of-thought (Qwen 3.6 / R1 /
+// any mlx-lm reasoning model). Collapsed by default; click to expand.
+// Reads as a quiet bracket of italic text above the answer so users
+// see SOMETHING happened during the long inference even if the final
+// content was short. Crucial for reasoning models where the model
+// spends 80% of its tokens here.
+function renderReasoningBlock(text) {
+ const wrap = document.createElement('details');
+ wrap.className = 'agent-reasoning';
+ const sum = document.createElement('summary');
+ const len = (text || '').length;
+ sum.innerHTML =
+ '<svg viewBox="0 0 24 24" width="11" height="11" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round"><path d="M9 5l7 7-7 7"/></svg>' +
+ '<span class="label">Thinking</span>' +
+ `<span class="meta">${len.toLocaleString()} chars</span>`;
+ wrap.appendChild(sum);
+ const body = document.createElement('div');
+ body.className = 'agent-reasoning-body';
+ body.textContent = text;
+ wrap.appendChild(body);
+ return wrap;
+}
+
function renderAttachmentChips(attachments) {
const wrap = document.createElement('div');
wrap.className = 'agent-msg-attachments';