complete_json now guards the whole choices[0].message.content
extraction and parse: empty, missing, or non-JSON content raises
AIMalformedResponseError after logging a WARNING with finish_reason,
the raw content repr, and a bounded reasoning_content excerpt, so the
candidate causes (thinking toggle ignored, content filter, token
truncation) are distinguishable on the next occurrence. classify_posts
catches the typed error before its generic handler and logs a terse
one-liner — the post stays AI: pending for manual review. No retry;
that decision is deferred until the logs reveal the actual cause.

Closes #106

@@ -10,3 +10,4 @@
 
 ---
 
+Occasionally the AI model answers with HTTP 200 but an empty message body, and the classifier died with a cryptic `JSONDecodeError`. Those posts now simply stay pending for manual review, accompanied by a clear warning; the model's finish reason, raw content, and a bounded reasoning excerpt are logged so the underlying cause is visible the next time it happens.

@@ -162,0 +201,4 @@
+            except json.JSONDecodeError:
+                pass
+        logger.warning(
+            "Model reply has no parsable JSON content: %s", _malformed_reply_diagnostics(payload, reply_content)

@@ -5,3 +5,3 @@
 fenliu 3.0.0
 
--
+- Improved: AI replies that arrive empty or unparseable now leave the post pending with a clear log line and full response diagnostics instead of a JSON traceback.

@@ -57,0 +84,4 @@
+    reasoning = str(message.get("reasoning_content") or "")
+    excerpt = reasoning[:_REASONING_EXCERPT_LIMIT]
+    return (
+        f"finish_reason={finish_reason} content={content!r} "

@@ -10,3 +10,4 @@
 
 ---
 
+Occasionally the AI model answers with HTTP 200 but an empty `message.content` field, and the classifier died with a cryptic `JSONDecodeError`. Those posts now simply stay pending for manual review, accompanied by a clear warning; the model's finish reason plus bounded excerpts of its content and reasoning are logged so the underlying cause is visible the next time it happens.

@@ -162,0 +205,4 @@
+            except json.JSONDecodeError:
+                pass
+        logger.warning(
+            "Model reply has no parsable JSON content: %s", _malformed_reply_diagnostics(payload, reply_content)

@@ -52,0 +128,4 @@
+
+    diagnostics = " ".join(record.getMessage() for record in collector.records)
+    expect("finish_reason='content_filter'" in diagnostics, "finish_reason logged").to_be_truthy()
+    expect("reasoning_content" in diagnostics, "reasoning_content noted").to_be_truthy()

@@ -57,0 +84,4 @@
+    message = envelope.get("message") if isinstance(envelope, dict) else {}
+    if not isinstance(message, dict):
+        message = {}
+    finish_reason = str(envelope.get("finish_reason", "unknown"))[:_FINISH_REASON_LIMIT]

@@ -162,0 +207,4 @@
+        logger.warning(
+            "Model reply has no parsable JSON content: %s", _malformed_reply_diagnostics(payload, reply_content)
+        )
+        raise AIMalformedResponseError(f"Model reply has no parsable JSON content: {reply_content!r}")

@@ -162,0 +221,4 @@
+            or not a string.
+
+        """
+        choices = payload.get("choices")

@@ -12,6 +13,8 @@ fenliu 3.0.0
 
 Some Mastodon attachments are served through a redirecting `media_proxy` URL rather than a direct file link. The AI vision stage's image downloader treated those redirects as errors, so posts with proxy-served media never received a vision verdict and had to be reviewed by hand. It now follows the redirect to the real file and infers the image type from the final URL's file extension, so proxied PNG/WebP/GIF attachments are labelled correctly instead of always falling back to `image/jpeg`.
 
+Occasionally the AI model answers with HTTP 200 but an empty `message.content` field, and the classifier died with a cryptic `JSONDecodeError`. Those posts now simply stay pending for manual review, accompanied by a clear warning; the model's finish reason plus bounded excerpts of its content and reasoning are logged so the underlying cause is visible the next time it happens.

@@ -162,0 +209,4 @@
+            except json.JSONDecodeError:
+                pass
+        logger.warning(
+            "Model reply has no parsable JSON content: %s", _malformed_reply_diagnostics(payload, reply_content)

@@ -12,6 +13,8 @@ fenliu 3.0.0
 
 Some Mastodon attachments are served through a redirecting `media_proxy` URL rather than a direct file link. The AI vision stage's image downloader treated those redirects as errors, so posts with proxy-served media never received a vision verdict and had to be reviewed by hand. It now follows the redirect to the real file and infers the image type from the final URL's file extension, so proxied PNG/WebP/GIF attachments are labelled correctly instead of always falling back to `image/jpeg`.
 
+Occasionally the AI model answers with HTTP 200 but an empty `message.content` field, and the classifier died with a cryptic `JSONDecodeError`. Those posts now simply stay pending for manual review, accompanied by a clear warning; the model's finish reason plus bounded excerpts of its content and reasoning are logged so the underlying cause is visible the next time it happens.

@@ -12,6 +13,8 @@ fenliu 3.0.0
 
 Some Mastodon attachments are served through a redirecting `media_proxy` URL rather than a direct file link. The AI vision stage's image downloader treated those redirects as errors, so posts with proxy-served media never received a vision verdict and had to be reviewed by hand. It now follows the redirect to the real file and infers the image type from the final URL's file extension, so proxied PNG/WebP/GIF attachments are labelled correctly instead of always falling back to `image/jpeg`.
 
+Occasionally the AI model would answer with HTTP 200 but an empty `message.content` field, and the classifier died with a cryptic `JSONDecodeError`. Those posts now simply stay pending for manual review, accompanied by a clear warning; the model's finish reason plus bounded excerpts of its content and reasoning are logged so the underlying cause is visible the next time it happens.

@@ -57,0 +92,4 @@
+    reasoning = str(message.get("reasoning_content") or "")
+    content_excerpt = (content or "")[:_CONTENT_EXCERPT_LIMIT]
+    return (
+        f"finish_reason={finish_reason!r} content={content_excerpt!r} "

@@ -5,6 +5,8 @@ visibility: public
 fenliu 3.0.0
 
 - Fixed: AI vision now follows `media_proxy` redirects, so posts with proxy-served attachments get classified instead of remaining pending.
+- Improved: AI replies that arrive empty or unparseable now leave the post pending with a clear log line and key response diagnostics instead of a JSON traceback.

@@ -12,6 +14,10 @@ fenliu 3.0.0
 
 Some Mastodon attachments are served through a redirecting `media_proxy` URL rather than a direct file link. The AI vision stage's image downloader treated those redirects as errors, so posts with proxy-served media never received a vision verdict and had to be reviewed by hand. It now follows the redirect to the real file and infers the image type from the final URL's file extension, so proxied PNG/WebP/GIF attachments are labelled correctly instead of always falling back to `image/jpeg`.
 
+Occasionally the AI model would answer with HTTP 200 but an empty `message.content` field, and the classifier died with a cryptic `JSONDecodeError`. Those posts now simply stay pending for manual review, accompanied by a clear warning; the model's finish reason plus bounded excerpts of its content and reasoning are logged so the underlying cause is visible the next time it happens.

@@ -37,1 +37,3 @@
-        return f"{timestamp} - {context}:{record.lineno} - {record.levelname} - {record.getMessage()}"
+        formatted = f"{timestamp} - {context}:{record.lineno} - {record.levelname} - {record.getMessage()}"
+        if record.exc_info:
+            formatted = f"{formatted}\n{self.formatException(record.exc_info)}"