Raw code, UUIDs, and 'DONE' tokens leak into answers

Category: Bugs Priority: P1 Severity: major Area: ai-answers Reports: 4 Status: Open

Summary

Four users report internal pipeline artifacts leaking into answer text instead of rendered output: raw UUIDs appear intermixed with the answer where source references should render, "programming language" shows up in the output of any fiqh/shar'i search, follow-up prompts repeatedly fail with the model replying just "DONE", and a first-time French user's very first exchange returned a broken answer that was literally "Done." (his email subject). All four point to answer post-processing/formatting failing to strip or render internal tokens and citation markup — the citation-id/UUID leak and the DONE end-of-stream token are two faces of the same rendering pipeline. Three of the four users attached screenshots and one repro is "any Islamic legal question", so this should be easy to reproduce. Abdellatif replied to all four threads that fixes are in progress.

What users say

Mohammed Azzaoui mail.azzaoui@gmail.com

• Date: 2026-08-22 | Language: Arabic | Thread: 1a028810405a4c59 | Replied: yes

Formal bug report (subject: "بلاغ عن خلل تقني: ظهور أكواد غريبة في نصوص الإجابات" — technical fault report: strange codes appearing in answer texts): strange programmatic codes/identifiers (UUIDs) appear intermixed with the texts and answers instead of properly rendered references and sources. Phone screenshots attached. Abdellatif replied the team is fixing it.

تظهر أكواد ومعرفات برمجية غريبة (UUIDs) متداخلة مع النصوص والإجابات بدلاً من عرض المراجع والمصادر بالشكل الصحيح. لقد أرفقت لقطات شاشة من هاتفي توضح الخلل أثناء استخدام التطبيق.

Strange codes and programmatic identifiers (UUIDs) appear intermixed with the texts and answers instead of displaying the references and sources correctly. I have attached screenshots from my phone showing the fault while using the app.

zsblaihed zsblaihed@yahoo.com

• Date: 2026-08-16 | Language: Arabic | Thread: 1a00a217dcc1d4bc | Replied: yes

When searching any Islamic legal (shar'i) question, programming language/code appears in the output; screenshot attached (image0.png). In a follow-up he says the app is more than wonderful.

يوجد خطأ في التطبيق عند البحث عن أيـ مسأله شرعيه ، تظهر لغة البرمجه ( مرفق لكم صوره )

There is an error in the app: when searching any Islamic legal question, programming language appears (a picture is attached for you).

Omar Shah omarshah29@hotmail.com

• Date: 2026-08-07 | Language: English | Thread: 19fdcacc9a4a0e9a | Replied: yes

After Qaf answers one prompt, asking more on it continuously produces an error where the model just says "DONE" (screenshot attached). The same email asks about usage limits and database freshness; in follow-ups he praises Qaf's accuracy and shares his own Claude hadith extension. Abdellatif apologized (the report initially went to spam) and said he will work on fixing it.

1. Many times when it answers one prompt/question and you ask more on it, it continuously keeps giving the error by saying "DONE" (screenshot attached)

Ilyass Saidi isaidi2004@gmail.com

• Date: 2026-08-07 | Language: French | Thread: 19fdcbbf7e678847 | Replied: yes

First-time user; his very first exchange produced a broken answer — the email subject is literally the AI's reply: "Done.". Screenshot attached (image.png).

Bonjour, que la paix soit sur vous, c'est la première fois que je teste qaf.ai et je me retrouve déjà avec un problème. Ci-joint la réponse que l'IA me donne après 1 réponse.

Hello, peace be upon you. It's the first time I'm testing qaf.ai and I already have a problem. Attached is the answer the AI gives me after 1 response.

Details / repro

• Symptoms, all consistent with the answer post-processing/rendering layer: 1. Raw citation UUIDs printed inline in the answer body instead of being resolved into reference/source chips (1a028810405a4c59, screenshots). 2. "Programming language" (raw code/markup) visible in answers; repro claimed for any shar'i question search (1a00a217dcc1d4bc, screenshot image0.png). 3. Follow-up prompts in an existing chat repeatedly return only the internal end token "DONE" (19fdcacc9a4a0e9a, screenshot; Android — "Sent from my Galaxy"). 4. Answer consisting solely of "Done." on the first-ever exchange (19fdcbbf7e678847, screenshot image.png, French UI).
• The "DONE"-only replies (Aug 7) and the UUID/code leaks (Aug 16, Aug 22) span the whole post-launch period; the Aug 22 report shows the citation rendering is still broken.
• Overlap: the DONE-on-follow-up failure resembles the silent follow-up failure in the answers-fail-to-complete cluster (19fdcdf7827d8a99) — likely the same conversation-continuation bug surfacing with or without the leaked token.

Threads

• 1a028810405a4c59 — Raw UUIDs intermixed with answers instead of rendered references (Mohammed Azzaoui)
• 1a00a217dcc1d4bc — Programming code appears in answers to any fiqh search (zsblaihed)
• 19fdcacc9a4a0e9a — Follow-up prompts repeatedly fail with "DONE" (Omar Shah)
• 19fdcbbf7e678847 — First exchange returns broken answer "Done." (Ilyass Saidi)