Skip to content

Personal AI, On Personal Devices

Python

feat: display thinking tokens in X-Ray footer and stream usage data

Jon Saad-Falcon3cf4bbae30f4
diff --git a/frontend/src/components/Chat/XRayFooter.tsx b/frontend/src/components/Chat/XRayFooter.tsx
index 2b1dde95..564369b9 100644
--- a/frontend/src/components/Chat/XRayFooter.tsx
+++ b/frontend/src/components/Chat/XRayFooter.tsx
@@ -35,7 +35,19 @@ export function XRayFooter({ usage, telemetry }: Props) {
rows.push({ label: 'Engine', value: `${telemetry.engine}${modelDetail ? ` (${modelDetail})` : ''}` });
}
if (usage) {
- rows.push({ label: 'Tokens', value: `${usage.completion_tokens} generated \u00B7 ${usage.prompt_tokens} prompt` });
+ const tokenParts = [`${usage.completion_tokens} generated`, `${usage.prompt_tokens} prompt`];
+ // Estimate thinking tokens: if total generated >> visible output, the
+ // difference is internal reasoning (e.g. Qwen3.5 thinking mode).
+ if (telemetry?.tokens_per_sec && telemetry.total_ms && usage.completion_tokens > 50) {
+ const visibleEstimate = Math.ceil(
+ (telemetry.total_ms / 1000) * telemetry.tokens_per_sec,
+ );
+ if (usage.completion_tokens > visibleEstimate * 1.5) {
+ const thinking = usage.completion_tokens - visibleEstimate;
+ tokenParts.push(`~${thinking} thinking`);
+ }
+ }
+ rows.push({ label: 'Tokens', value: tokenParts.join(' \u00B7 ') });
}
if (telemetry?.tokens_per_sec) {
rows.push({ label: 'Speed', value: `${Math.round(telemetry.tokens_per_sec)} tok/s` });
diff --git a/src/openjarvis/engine/ollama.py b/src/openjarvis/engine/ollama.py
index 049c489d..5e95ef9a 100644
--- a/src/openjarvis/engine/ollama.py
+++ b/src/openjarvis/engine/ollama.py
@@ -41,6 +41,8 @@ class OllamaEngine(InferenceEngine):
host = env_host or self._DEFAULT_HOST
self._host = host.rstrip("/")
self._client = httpx.Client(base_url=self._host, timeout=timeout)
+ # Last stream usage — captured from Ollama's final chunk
+ self._last_stream_usage: Dict[str, int] = {}
def generate(
self,
@@ -174,6 +176,15 @@ class OllamaEngine(InferenceEngine):
if content:
yield content
if chunk.get("done", False):
+ # Capture usage from final chunk
+ self._last_stream_usage = {
+ "prompt_tokens": chunk.get("prompt_eval_count", 0),
+ "completion_tokens": chunk.get("eval_count", 0),
+ "total_tokens": (
+ chunk.get("prompt_eval_count", 0)
+ + chunk.get("eval_count", 0)
+ ),
+ }
break
except (httpx.ConnectError, httpx.TimeoutException) as exc:
raise EngineConnectionError(
diff --git a/src/openjarvis/server/routes.py b/src/openjarvis/server/routes.py
index 5fc4e62c..30dab503 100644
--- a/src/openjarvis/server/routes.py
+++ b/src/openjarvis/server/routes.py
@@ -222,8 +222,9 @@ async def _handle_stream(engine, model: str, req: ChatCompletionRequest):
yield "data: [DONE]\n\n"
return
- # Send finish chunk
- finish_chunk = ChatCompletionChunk(
+ # Send finish chunk with usage data if available
+ import json as _json
+ finish_data = ChatCompletionChunk(
id=chunk_id,
model=model,
choices=[StreamChoice(
@@ -231,7 +232,21 @@ async def _handle_stream(engine, model: str, req: ChatCompletionRequest):
finish_reason="stop",
)],
)
- yield f"data: {finish_chunk.model_dump_json()}\n\n"
+ finish_dict = _json.loads(finish_data.model_dump_json())
+
+ # Pull usage from the engine if it tracked it during streaming
+ raw_engine = engine
+ # Unwrap InstrumentedEngine if present
+ if hasattr(raw_engine, "_inner"):
+ raw_engine = raw_engine._inner
+ # Unwrap MultiEngine if present
+ if hasattr(raw_engine, "_engine_for"):
+ raw_engine = raw_engine._engine_for(model)
+ stream_usage = getattr(raw_engine, "_last_stream_usage", None)
+ if isinstance(stream_usage, dict) and stream_usage.get("total_tokens", 0) > 0:
+ finish_dict["usage"] = stream_usage
+
+ yield f"data: {_json.dumps(finish_dict)}\n\n"
yield "data: [DONE]\n\n"
return StreamingResponse(