From 0a0702f15db4d3f739049b5f0a9882ad1bcc631d Mon Sep 17 00:00:00 2001
From: avi <avi@localhost>
Date: Wed, 16 Sep 2026 18:28:59 -0500
Subject: [PATCH] =?UTF-8?q?Summarize=20progress=20counts=20reasoning=20tok?=
 =?UTF-8?q?ens:=20thinking=20models=20stream=20most=20of=20the=20run=20in?=
 =?UTF-8?q?=20reasoning=5Fcontent,=20which=20the=20ticker=20ignored=20?=
 =?UTF-8?q?=E2=80=94=20UI=20sat=20at=20'waiting=20for=20model'=20then=20ju?=
 =?UTF-8?q?mped=20to=20done=20with=20no=20percentage?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 backend/shonar/services/ai/ollama.py        |  8 ++++--
 backend/shonar/services/ai/openai_compat.py | 11 ++++++--
 backend/tests/test_ai_adapters.py           | 28 +++++++++++++++++++++
 3 files changed, 43 insertions(+), 4 deletions(-)

diff --git a/backend/shonar/services/ai/ollama.py b/backend/shonar/services/ai/ollama.py
index 424f96a..7edf1ea 100644
--- a/backend/shonar/services/ai/ollama.py
+++ b/backend/shonar/services/ai/ollama.py
@@ -117,9 +117,13 @@ async def _collect_reply(resp: httpx.Response, ticker) -> str:
         except ValueError:
             continue
         piece = (obj.get("message") or {}).get("content") or ""
-        if piece:
+        # Like the openai_compat adapter: a thinking channel is real
+        # work and must move the progress ticker even when the server
+        # ignored think:false.
+        thinking = (obj.get("message") or {}).get("thinking") or ""
+        if piece or thinking:
             parts.append(piece)
-            total += len(piece)
+            total += len(piece) + len(thinking)
             ticker(total)
         if obj.get("done"):
             break
diff --git a/backend/shonar/services/ai/openai_compat.py b/backend/shonar/services/ai/openai_compat.py
index 0fe4d9a..49cb1bb 100644
--- a/backend/shonar/services/ai/openai_compat.py
+++ b/backend/shonar/services/ai/openai_compat.py
@@ -48,11 +48,18 @@ async def _collect_reply(resp: httpx.Response, ticker) -> str:
             chunk = _json.loads(data)
             delta = chunk["choices"][0].get("delta") or {}
             piece = delta.get("content") or ""
+            # Thinking models (qwen3 on llama.cpp) stream a long
+            # reasoning_content channel BEFORE any content: counting only
+            # content left the progress bar frozen at "waiting for model"
+            # for the entire run, then jumping straight to done. Reasoning
+            # is real work — count it toward progress (never into the
+            # reply itself).
+            thinking = delta.get("reasoning_content") or ""
         except (ValueError, KeyError, IndexError, TypeError):
             continue  # keep-alives / usage chunks / odd frames: not content
-        if piece:
+        if piece or thinking:
             parts.append(piece)
-            total += len(piece)
+            total += len(piece) + len(thinking)
             ticker(total)
     content = "".join(parts)
     if not content:
diff --git a/backend/tests/test_ai_adapters.py b/backend/tests/test_ai_adapters.py
index 5c7d260..ff295c7 100644
--- a/backend/tests/test_ai_adapters.py
+++ b/backend/tests/test_ai_adapters.py
@@ -155,6 +155,34 @@ async def test_openai_compat_streams_with_progress():
     assert all(0 <= p <= 99 for p in seen_pcts)  # never claims 100 early
 
 
+async def test_openai_compat_thinking_moves_progress_before_content():
+    """qwen3-on-llama.cpp streams reasoning_content for most of the run;
+    progress must climb during that phase, not sit at "waiting for model"
+    and then jump straight to done."""
+    import json as _json
+
+    def sse(request: httpx.Request) -> httpx.Response:
+        think = [f"reasoning chunk {i} " for i in range(40)]  # ~800 chars
+        lines = [f"data: {_json.dumps({'choices': [{'delta': {'reasoning_content': p}}]})}"
+                 for p in think]
+        answer = _json.dumps({"short": "Done.", "key_points": []})
+        lines.append(f"data: {_json.dumps({'choices': [{'delta': {'content': answer}}]})}")
+        lines.append("data: [DONE]")
+        return httpx.Response(
+            200, content=("\n\n".join(lines) + "\n\n").encode(),
+            headers={"content-type": "text/event-stream"})
+
+    p = OpenAICompatProvider("http://llm:8000", model="qwen",
+                             http_client=mock_client(sse))
+    seen_pcts = []
+    res = await p.summarize("t", on_progress=seen_pcts.append)
+    assert res.short == "Done."  # reasoning never enters the reply
+    # The ticker must report real mid-run values while only reasoning
+    # streamed — otherwise the UI shows no percentage for the whole job.
+    assert len(seen_pcts) > 10
+    assert seen_pcts[-1] > 50
+
+
 async def test_openai_compat_partial_json_gets_defaults():
     import json as _json