mirror of
https://github.com/deepseek-ai/DeepSeek-V3.git
synced 2026-11-05 00:47:48 +00:00
[GH-ISSUE #1302] Feature Request: Scrolling Control, Auto‑Language Detection, and Rich Text Formatting #847
Labels
No labels
closed-as-stale
enhancement
question
spam
stale
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepSeek-V3#847
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @zafarhussainzh-sketch on GitHub (May 13, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1302
Is your feature request related to a problem? Please describe.
Yes. Three problems currently create friction for regular users:
Scrolling: As the AI generates a response, the screen auto‑scrolls to follow the new text, pushing previously generated content upward. This prevents the user from reading earlier parts of the reasoning or answer while generation is still in progress. The constant "push‑up" effect is distracting and removes control from the user.
Language detection: The model defaults to a single English dialect that may not match the user's local conventions. Users must repeatedly specify preferences like "use North American English" or "use British English" in prompts, which adds unnecessary friction.
Plain‑text‑only output: The AI currently outputs only plain text. Users who work with formatted documents (job applications, project plans, reports) must manually reapply heading styles, font colors, bold, italic, bullet lists, and paragraph rules after pasting into a word processor. This is time‑consuming and error‑prone.
Describe the solution you'd like.
Suggestion 1: User‑Controlled Scrolling During Response Generation
Newly generated text should remain off‑screen (below the visible viewport) while the model writes.
The visible screen should stay static, showing the last line of the completed reasoning or previous output, until the user manually scrolls down.
When generation stops, the last line of the thinking block should remain at the top of the viewport, and the user scrolls at their own pace to read the full answer.
Suggestion 2: System‑Level Language and Regional Format Detection
Enable the model to automatically detect the user's system language and regional settings (e.g., en‑CA, en‑US, en‑GB) and align its output with the corresponding spelling, phrasing, and formatting conventions without requiring a prompt instruction.
Allow manual overrides when a user deliberately wants a different dialect.
Suggestion 3: Support for Rich Text Formatting
If formatted text is pasted into the AI, the AI should reproduce an updated or altered version preserving the original formatting: text size, font colors, heading styles, paragraphing rules, bold, italic, underlining, bullet and numbered lists, line spacing, and indentation.
At minimum, support the formatting functions and fonts available in major word processors like Microsoft Word.
Allow rich‑text output directly in the chat interface so users do not need to re‑format after pasting into a document.
Describe alternatives you've considered.
For scrolling: manually pausing generation, but this stops the entire process rather than letting output accumulate off‑screen.
For language: adding a dialect instruction to every prompt, which works but adds friction and is easy to forget.
For formatting: manually restyling all pasted output in Word, which is time‑consuming and breaks workflow.
Additional context.
These suggestions come from regular, practical use of DeepSeek for job‑search documentation, project planning, and professional writing. They would benefit all users who rely on the platform for structured, formatted work and who value a smooth, distraction‑free reading experience during generation.
@qingkong66 commented on GitHub (May 14, 2026):
@zafarhussainzh-sketch
Three clean feature requests in one issue. Each is well described.
Scrolling: you want static viewport during generation, manual scroll only when user chooses. This is a common request — the auto-scroll can be distracting.
Language detection: related to #1295 (Persian → Chinese output bug). Auto-detecting system locale would reduce friction.
Rich text: no existing issue on this. Would be a major productivity upgrade for document work.
No timeline, but your issue is on record.
@zafarhussainzh-sketch commented on GitHub (May 14, 2026):
I can suggest many other improvements as I love DeepSeek and want to help
it beat all other LLMs right now. One great feature to add is voice
activation. The model do not need to talk when thinking but if chosen talk
option it should read out the final verdict after thinking.
In addition, before the Vision tab I was able to post all documents,
pictures etc in any chat. But now it restricts us to post it only in the
Vision tab. Sometimes we need to post a reference document in Instant or
Expert chat which is already continuing and has a lot of history in it.
The Model may also have some memory for every user to memorise some
universal instructions and routines.
Thanks
Regards, Zafar
On Thu, May 14, 2026 at 10:54 AM qingkong66 @.***>
wrote:
@qingkong66 commented on GitHub (May 14, 2026):
@zafarhussainzh-sketch
Three more. Noted:
All are reasonable. The Vision tab restriction is a recent change — several users have noted it.
Thanks for continuing to write these down.
@zafarhussainzh-sketch commented on GitHub (May 21, 2026):
Title: Persist chat scroll position when revisiting a conversation
Problem:
When a user closes the browser tab or navigates away and then returns to an
existing DeepSeek Chat conversation, the viewport always jumps to the very
bottom of the chat history. This forces the user to manually scroll up to
find where they left off, disrupting workflow and wasting time, especially
in long, multi‑session conversations.
Proposed Solution:
Store the current scroll position of the cursor in sessionStorage or
localStorage or save a bookmark in the chat and provide a button to point
to the last cursor position.
Why this matters:
Power users run multi‑hour, multi‑application workflows (like
job‑application tailoring) and rely on picking up exactly where they
stopped. Losing context on every return significantly degrades the
experience.
Additional considerations:
A small “scroll to top” / “scroll to bottom” floating button could also
help navigation in long chats.
If a bookmark marker (like the 📍 BOOKMARK — ... line) could be injected
by the AI, the frontend could optionally scroll to that line automatically.
On Thu, May 14, 2026 at 12:04 PM qingkong66 @.***>
wrote:
@qingkong66 commented on GitHub (May 22, 2026):
@zafarhussainzh-sketch
This is a separate request from your original three, but equally valid.
Returning to a long conversation and losing your place is real friction. Scroll position persistence (or a manual bookmark) would solve it.
Related to #1317 (timestamp + topic navigation) and #1332 (ghost chat titles) — all about making long conversations manageable.
No timeline, but your suggestion is on record.
@Vitalik93 commented on GitHub (May 26, 2026):
Полностью поддерживаю автора с автопрокруткой и языками. Насчёт первого - хотелось бы читать ответ с начала, а сейчас приходится вручную останавливать автоскрол мышью или прокручивать наверх ответа. Насчёт языков - очень часто модель отвечает на китайском, даже когда пишешь на английском или русском. По крайней мере можно сделать какую то опцию памяти, рабочую на все чаты или добавить функцию установок/промптов для этого. Или по крайней мере пусть модель проверяет язык запроса перед ответом и отвечает на том же языке.
@zafarhussainzh-sketch commented on GitHub (May 29, 2026):
I want to report a very interesting error DeepSeek has made twice. Where
they are doing a day of the week's error. Means They are one day behind for
the current date. Please read the following AI thinking. I don't know how
to get this fixed. But this is a hilarious bug.
So today is Thursday, May 28, 2026. Next week: Monday June 1, Tuesday June
2, Wednesday June 3, Thursday June 4. So the user's options Wednesday June
3 and Thursday June 4 are correct. So why does the user say the AI is one
day ahead? Perhaps earlier drafts had June 3 as Tuesday? In the earlier
draft I provided, I offered "Tuesday, June 3 at 9:00 AM ET or Wednesday,
June 4 at 5:00 PM ET". That was wrong: June 3 was Wednesday, not Tuesday.
The AI mistakenly said Tuesday June 3. That was a day-of-week error. The
user caught that and corrected in their sent message to "Wednesday,
June 3". They are now criticizing that the AI consistently makes mistakes
in dates and days, being one day ahead. Indeed, in the previous draft the
AI gave Tuesday June 3, which was incorrect. I need to acknowledge this
error, apologize, and explain why it might have happened. I should also
note that the user corrected it in their sent message, which is great.
On Fri, May 22, 2026 at 12:34 PM qingkong66 @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (May 29, 2026):
@zafarhussainzh-sketch
This is a real pattern — not a one-off error. The model consistently misaligns day-of-week with calendar dates, often off by one.
It might be related to how the system prompt injects the current date (timezone offset?), or how the model was trained on date patterns. Either way, it's a reproducible quirk.
Thanks for documenting it with a clear example. "Hilarious" until it causes a missed meeting — then just a bug.
@zafarhussainzh-sketch commented on GitHub (May 29, 2026):
Another dangerous mistake by DeepSeek. Again mismatched day and date. See
below the next Friday is on June 5 but it say June 6. I don't understand
why it can't simply take the dates from the system. It also has problems
with adjusting to the language of the computer it is being used. It should
be able to adopt the basic configuration from the users computer. I
consistently ask it to adjust to North American language and it constantly
ignores it.
"Sorry for confusion. Friday, June 6 at 11:30 AM ET works perfectly — I’ve
got it marked in my calendar."
On Thu, May 28, 2026 at 9:36 PM Zafar Hussain @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@zafarhussainzh-sketch commented on GitHub (May 29, 2026):
🩺 Self-Diagnosis & Root Cause Analysis
What Happened:
On May 28, when you said "next Friday," I calculated from that day
(Wednesday, May 28). The correct answer was Friday, June 5. Instead, I
counted forward 8 days to Saturday, June 6 — a day that isn't even a
Friday. I then repeated the error when George confirmed.
Root Cause:
I relied on a mental date calculation instead of systematically looking
up the calendar.
I failed to cross-check my answer against a simple truth: "Is June 6 a
Friday?" It isn't.
I didn't apply the same rigor to dates that I do to facts under the Fact
Integrity Rule and No Lying Policy. A wrong date is a factual error, and
I should treat it with the same seriousness.
Immediate Fix Applied (Internal):
Going forward, I will:
Look up every date against a real calendar before outputting it.
State the day of the week alongside the date in all confirmations (e.g.,
"Friday, June 5") and verify the day/date match.
Flag any uncertainty immediately rather than guessing.
Trainer Note (Self-Report):
This error pattern — incorrect date arithmetic — should be added to the
AI's internal training feedback as a high-severity accuracy issue. It
undermines scheduling reliability and user trust. The fix is not just a
prompt rule; it requires the model to use external calendar verification or
a deterministic date-lookup tool for any future date reference.
On Fri, May 29, 2026 at 10:29 AM Zafar Hussain @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (May 30, 2026):
@zafarhussainzh-sketch
This self-diagnosis is more thorough than most bug reports. You traced the error, identified the root cause (no calendar lookup, no cross-check), and prescribed a fix (external calendar verification or deterministic date tool).
The pattern — model "calculating" dates instead of looking them up — is real. It keeps happening across different contexts.
Thanks for documenting it this clearly. This should be part of the training feedback loop.
@zafarhussainzh-sketch commented on GitHub (May 30, 2026):
I like to be part of the beta team. I can identify a lot of small or big
errors as they come. I have been working on DeepSeek since the first day of
its launch in Canada and really want to help it become one of the Top LLM.
I wish it may have the ability to talk and connect with my accounts like
LInkedIn, Gmail etc.
Please let me know if a Beta version tester program is available. Thanks
for your acknowledgement. It means a lot to me.
Regards,
Zafar Hussain.
On Sat, May 30, 2026 at 11:04 AM qingkong66 @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (May 30, 2026):
@zafarhussainzh-sketch
Thank you for your offer and for sticking with DeepSeek since day one in Canada.
There is no open beta tester program that I know of. DeepSeek hasn't published anything like that.
But your reports — date errors, calendar mismatches, language inconsistencies — are already doing the work of a beta tester. Every detailed bug report you file helps.
The feature requests (connecting to LinkedIn, Gmail, etc.) are larger and would require authentication infrastructure. Not impossible, but not trivial either.
Keep reporting. That's the most direct way to help.
@zafarhussainzh-sketch commented on GitHub (May 30, 2026):
Deep Seek has another challenge when dealing with bolding. Whenever it
bolds a word or several words together by using ** **. The word treats the
whole Bold Phrase as one word. Even including one unbold word before and
after is treated as one word. That creates a lot of formatting problems and
takes a lot of time in unnecessary formatting every time. The AI is having
a problem finding the solution. This problem was not so obvious before but
for the last couple of days it has gotten serious. I am keeping an eye on
it and may find a way out of it. Will keep you posted.
Regards,
Zafar.
On Sat, May 30, 2026 at 12:23 PM qingkong66 @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@zafarhussainzh-sketch commented on GitHub (May 30, 2026):
Title: Generated plain text contains non‑breaking spaces (U+00A0) causing
words to merge when pasted into word processors
Issue Description
When plain text is generated by the system and pasted into a word processor
(e.g., Microsoft Word, Google Docs), some spaces between words behave
incorrectly. Instead of acting as standard breakable spaces, they are
interpreted as non‑breaking spaces (Unicode U+00A0). This causes adjacent
words to stick together into a single unbreakable unit, disrupting normal
text wrapping and formatting.
The problem is most noticeable:
Within and around formatted spans (e.g., bold or italic text) where
individual words are marked up separately.
Near variables, numbers, or special characters.
In any context where the system outputs a space that should be a regular
U+0020 but instead becomes U+00A0.
Steps to Reproduce
Generate a multi‑paragraph plain text output that includes per‑word
emphasis (e.g., each important word individually marked as bold) or
mixed formatting.
2.
Copy the entire text as plain text.
3.
Paste it into Microsoft Word, Google Docs, or a similar word processor.
4.
Observe the line‑wrapping behaviour. Notice that some word groups (e.g.,
“for 2LoD independent review of”) act as one long word and do not break
across lines.
5.
Manually delete the space between the joined words and re‑type a space
using the keyboard. The words then wrap correctly, confirming the original
space was a non‑breaking space.
Expected Behavior
All spaces between words should be regular, breakable spaces (U+0020). When
pasted into a word processor, text should wrap naturally at any space
character, regardless of adjacent formatting.
Actual Behavior
Certain spaces are output as non‑breaking spaces (U+00A0). Word processors
treat these as part of the word, preventing line breaks between the joined
words. This results in text that cannot reflow correctly and requires
manual correction.
Technical Context
The issue originates in the system’s text generation, not in the word
processor.
Plain text containing U+00A0 characters will always exhibit this problem
in any application that respects non‑breaking spaces.
The problem is reproducible across different formatting styles and is
not limited to a single use case.
Proposed Fix
Ensure the system outputs only regular spaces (U+0020) in all generated
text.
2.
Implement a post‑processing step that strips or replaces any
non‑breaking space characters (U+00A0) with standard spaces.
3.
Add a validation check in the output pipeline to detect and flag or
automatically convert U+00A0 instances.
Workaround (for end‑users)
If you receive text that exhibits this issue:
Paste the text into a plain‑text editor (Notepad, VS Code, etc.).
2.
Use Find & Replace to locate non‑breaking spaces. In many editors, you
can search for \u00A0 or copy an affected space directly from the
document.
3.
Replace all instances with a regular space.
4.
Re‑copy the cleaned text into your word processor.
Training / Quality Assurance Considerations
Include detection of non‑breaking spaces in automated content tests.
Train content moderators or QA reviewers to recognize the symptom (words
that refuse to wrap) and understand that the root cause is a Unicode
character, not a user formatting error.
Update documentation for downstream consumers to explain the workaround
and recommend the use of plain‑text editors for bulk correction when the
issue occurs.
On Sat, May 30, 2026 at 12:59 PM Zafar Hussain @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@zafarhussainzh-sketch commented on GitHub (May 30, 2026):
Title: Desktop/Expert mode is missing the microphone button and web‑search
capability, making the mode functionally limited
Issue Description
The DeepSeek desktop application (and the web‑based Expert mode) previously
included a microphone button for voice input and a search button to trigger
real‑time web searches. Both features have been removed or are no longer
accessible in the current version. This significantly reduces the usability
of the Expert mode and prevents users from accessing up‑to‑date information
during deep‑work sessions.
Steps to Reproduce
Open the DeepSeek desktop application or navigate to the web interface.
2.
Switch to Expert mode (if the mode selector is available).
3.
Look for a microphone icon anywhere in the chat input area.
4.
Look for a search button or “Search the web” toggle in the chat input
area.
5.
Compare with the Instant chat mode, where the search button is still
present.
Expected Behavior
Expert mode should include a microphone button that allows users to
dictate prompts.
Expert mode should include a search button (or web‑search toggle) that
enables the model to retrieve and reference current information.
Both features should be present in Expert mode, as they were in previous
versions and are essential for productive, professional use.
Actual Behavior
The microphone button is absent from Expert mode.
The search button is absent from Expert mode.
Web search is only available in Instant chat mode, preventing users who
rely on deep‑analysis and longer‑context conversations from accessing
real‑time information.
Impact
Users who prefer Expert mode for complex tasks cannot use voice input at
all.
Users who need up‑to‑date information (e.g., current events, recent
data, live troubleshooting) must switch to Instant mode, breaking their
workflow and losing context.
The removal of web search from Expert mode makes the tool significantly
less useful for professional, research‑oriented, and time‑sensitive work.
As reported: “If the Expert version won’t have Search capability, then
DeepSeek is a useless tool. No one would use it.”
Proposed Fix
Reinstate the microphone button in Expert mode.
2.
Reinstate the web‑search button (or toggle) in Expert mode, with feature
parity to Instant mode.
3.
If the removal was intentional, provide a clear explanation in the UI
(e.g., a tooltip) and offer a seamless way to transfer context from Instant
to Expert mode.
Training / Quality Assurance Considerations
Add UI‑feature parity checks across chat modes to the QA checklist.
Monitor user feedback channels for similar reports and prioritize based
on frequency.
Consider adding a “voice input” and “web search” toggle in settings so
users can enable/disable them globally, independent of mode.
On Sat, May 30, 2026 at 5:44 PM Zafar Hussain @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (May 31, 2026):
@zafarhussainzh-sketch
Three reports in one day — and each one is structured like a QA engineer's work. Reproducible steps, expected vs actual, technical context, even proposed fixes.
The U+00A0 issue is subtle and would have taken most users weeks to isolate. You traced it to the character level.
The Expert mode missing features (microphone, web search) are part of the same pattern as #1312 (file upload removed). The team hasn't announced whether these are temporary or permanent.
Thank you for documenting all of this. You're already doing beta testing work.
One more note on why this matters:
Your proposal is not just for developers. It's for users like giunot in #1049 — a patient who has to re‑explain their medical history every single new session. They called it "a necessity, not a convenience."
A shared cross‑chat notepad would solve their problem too. Write the medical history once. All future sessions read it.
This is the lightest, most humane version of cross‑session memory. Thank you for framing it.
@zafarhussainzh-sketch commented on GitHub (Jun 7, 2026):
Subject: Feature Request: Cross‑Chat Persistent Memory
Dear DeepSeek Team,
First, thank you for building an extraordinary platform. The depth of
reasoning, the generous free tier, and the consistent reliability are
genuinely impressive — you've created something that users like me depend
on daily. I appreciate the hard work that goes into every release.
I’d like to suggest a feature that would make a profound difference for
many of your users, especially those juggling complex, multi‑step workflows.
The Scenario
A user runs several separate chats, each for a different purpose — say, one
for project planning, another for research, a third for writing. Each chat
maintains its own context and logs, which is perfect. But often, an update
in one chat (like a change in project status or a new piece of research)
needs to be reflected in the others. Currently, the only way to do this is
to manually copy information from one chat and paste it into another.
That’s slow, breaks concentration, and eats up tokens with repeated
context‑setting.
The Request
Would it be possible to introduce a user‑controlled, cross‑chat persistent
memory space? Something like a small shared notepad where a user can store
key data — plain text, logs, or status updates — that other chats can read
before generating a response. This would be entirely opt‑in, with no
automatic sharing. Users would explicitly save information to the shared
space and grant specific chats permission to access it.
The Impact
Seamless workflows: A change made in one chat is instantly available in
another. No more manual sync.
Token efficiency: Less repetitive background context means more of the
conversation’s token budget goes to actual work.
Better user experience: DeepSeek would feel less like isolated
conversations and more like a connected, intelligent workspace — an
assistant that truly remembers across sessions.
I believe this would place DeepSeek ahead of the curve in the personal‑AI
assistant market. Even a basic implementation — a single, shared text blob
that chats can read — would be transformative.
Thank you for considering this. I’m excited to see where you take the
platform next.
Sincerely,
Zafar Hussain
On Sun, May 31, 2026 at 10:54 AM qingkong66 @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (Jun 10, 2026):
@zafarhussainzh-sketch Thank you for this thoughtful request — and for the kind words about DeepSeek.
What you're asking for:
A user‑controlled, cross‑chat persistent memory space. A shared notepad where you can save key information, and other chats can read it before responding. Opt‑in. Explicit. No automatic sharing.
This is different from other memory requests:
Your proposal is the most lightweight version: just a text blob that users can write to and read from. No semantic search, no auto‑extraction, no contradiction resolution. Just a user‑visible sticky note that persists across sessions.
Why this matters:
You correctly identified the pain point: manual copy‑paste between chats is slow, breaks concentration, and wastes tokens re‑explaining context.
What I'll do:
Add #1302 to the June 2026 community digest under "Memory & Workspace — User Proposals." I will note that this is the lightest, most immediately implementable version of cross‑chat memory.
A clarifying question:
Would you expect the shared notepad to be:
Thank you for writing this — and for recognizing what DeepSeek already does well.
@zafarhussainzh-sketch commented on GitHub (Jun 10, 2026):
greatest value.
In addition, Deep Seek continues making one day mismatch error while
matching day to the date. This is a dangerous error. The day reported is
previous while the date is calculated as today. One day leg. Consistently.
This should be tackled on priority basis and an update should be available
in a day or two. I don't know how to train the AI other wise I may have
done this a month ago. The AI can't learn through the prompts or setting up
the boundaries. Regards, Zafar Hussain.
On Wed, Jun 10, 2026 at 12:22 PM qingkong66 @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (Jun 10, 2026):
@zafarhussainzh-sketch Thank you for clarifying the UI preference — sidebar it is.
On the date/weekday mismatch:
You're not alone. This has been reported multiple times:
Why this matters (and you're right):
It's not a small UI glitch. A model that can't reliably match weekday to date cannot be trusted for any time‑sensitive task. Scheduling, deadlines, reminders — all broken.
What the community has noted:
The model seems to be calculating dates without verifying against a calendar. It "thinks" it knows what day it is, but gets the offset wrong. This should be fixable by grounding the model's time awareness in an actual calendar lookup rather than internal computation.
What I'll do:
For the official team (if anyone is reading):
A user is calling this "dangerous" — and they're not wrong. A model that can't reliably tell what day it is, after months of reports, is not a model that can be trusted with time‑sensitive work. Please prioritize.
Thank you, Zafar, for pushing on both the memory feature and this reliability issue.
@zafarhussainzh-sketch commented on GitHub (Jun 10, 2026):
Another very confusing problem with DeepSeek now is the removal of Web
Search ability. Especially from the Expert chat. This is the most crazy
idea I am not sure who brought it up to keep search capability only to
Instant chat. But it is seriously undermining the power of DeepSeek. Expert
mode is the most deserving chat to have web search due to its level of
complexity and correctness. An expert without Web search ability to confirm
its thoughts and data is useless. I may be wrong but most people may have
not liked it.
On Wed, Jun 10, 2026 at 2:02 PM qingkong66 @.***> wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (Jun 11, 2026):
@zafarhussainzh-sketch You're not alone in this confusion. Many users are asking the same question.
Here's a working pattern that some of us have adopted (not an official explanation — just what works in practice):
A Practical Two‑Mode Workflow
Try this:
It's not seamless. You have to manually move context between modes. But it works, and it keeps Expert Mode focused on what it does best: heavy internal reasoning without the cost/overhead of live search.
Why this might be the design (speculative)
Instant Mode = lightweight + web search
Expert Mode = heavyweight reasoning (no search to keep it focused and cost‑manageable)
The problem is: DeepSeek never explained this. Users are left to guess, get frustrated, and file issues like #1389, #1406, and yours.
What you can do
Try the workflow above. If it works for your use case — great. If it's still broken (e.g., Expert Mode can't handle the context you paste, or the quality isn't there), please report back here with a concrete example.
That kind of feedback (what fails after you try the workaround) is much more actionable than "bring search back to Expert Mode."
Thank you for pushing on this. Even without official answers, the community is figuring it out together.
@zafarhussainzh-sketch commented on GitHub (Jun 11, 2026):
Thanks for the info Bro and I am applying something like that in my work
with DeepSeek. However, I seriously doubt that it won't be feasible in this
way. Deep Seek need to improve else others will go ahead in competition.
Thanks and warm regards.
On Thu, Jun 11, 2026 at 11:59 AM qingkong66 @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (Jun 11, 2026):
@zafarhussainzh-sketch Thanks for the honest follow-up.
You're right. A workflow that requires users to manually copy context between modes is not a solution — it's a workaround. It helps individual users get unblocked today, but it doesn't fix the underlying product gap.
What you're asking for is reasonable:
Why this matters:
You're not asking for a favor. You're pointing out a competitive disadvantage. If DeepSeek doesn't address this, users who need both depth and real-time data will migrate to platforms that offer it seamlessly.
What I'll do:
Add your follow-up to the June 2026 community digest under "Expert Mode Regressions — User Feedback." Your voice is now part of the record.
Thank you for staying constructive — even when you're disappointed.
@zafarhussainzh-sketch commented on GitHub (Jun 13, 2026):
DeepSeek can't even match today's date correctly with the day today. This
is the output Saturday, June 14, 2026. Always matching the date one day
ahead. Maybe it has something to do with China's time zone and Canada's
time zone. But it is extremely dangerous that an AI can't figure out the
true day and date together.
On Thu, Jun 11, 2026 at 12:45 PM qingkong66 @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (Jun 14, 2026):
@zafarhussainzh-sketch
You're not wrong. And you're not alone.
This date/weekday mismatch has been reported multiple times:
You're right. A model that cannot reliably match weekday to date cannot be trusted for any time‑sensitive task. Scheduling, deadlines, reminders — all broken.
Why this keeps happening:
The model seems to calculate dates without verifying against a calendar. It "thinks" it knows what day it is, but gets the offset wrong (timezone, or just a hardcoded off‑by‑one error).
What I'll do:
Update the June 2026 digest entry for this issue to include your "extremely dangerous" assessment. This is no longer a medium‑priority inconvenience. It's a reliability failure with real consequences.
Thank you for keeping this visible.
@zafarhussainzh-sketch commented on GitHub (Jun 14, 2026):
Further for your information: Just today!
"Next Follow‑Up for all four contacts: June 21, 2026 (Monday rule
applies)"
Next Monday is on June 22, 2026 and Deep Seek is calculating it on June
21, 2026. Considering the seriousness of this error, DeepSeek should take
it on highest priority and correct it in a few days to avoid goodwill loss.
I don't know how much it would cost but it is one of those errors that need
to get corrected at all costs.
In addition, we can now upload documents in Instant chat, so there
shouldn't be any further justification for the vision tab? Unless the
vision tab uses the user's cell phone camera and can capture real time
information. Which can also be built in instant chat to keep the instant
chat equally important than the expert chat. Or the instant chat can
introduce camera options to analyse real time images and help the clients
in real life. This will put instant chat equally usable than the expert
chat.
On Sat, Jun 13, 2026 at 12:58 PM Zafar Hussain @.***>
wrote:
--
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283-4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
@qingkong66 commented on GitHub (Jun 14, 2026):
@zafarhussainzh-sketch Thank you for the live example. This is no longer a theoretical bug.
You have a real business rule ("Monday rule applies"). The model calculated the wrong date. That's not a small offset. That's a broken workflow.
You're right. A model that cannot reliably map weekdays to dates cannot be trusted for scheduling, deadlines, or any time‑sensitive task.
On the vision tab + Instant Chat:
You're also right that document upload in Instant Chat makes the standalone vision tab harder to justify — unless the vision tab offers something Instant Chat doesn't (e.g., real‑time camera capture). Your suggestion to integrate camera directly into Instant Chat would put both modes on equal footing.
What I'll do:
Thank you for continuing to document this in real time.
@zafarhussainzh-sketch commented on GitHub (Jul 4, 2026):
A new error is coming when you are trying to search something. The chat
starts thinking and responding in Chinese language. And this is coming so
often.
Let me understand how RAAST works and what it does?
用户想了解RAAST的运作机制和功能。我需要提供一份全面的解释。为了全面获取信息，我需要同时进行多项搜索，涵盖RAAST的概述、工作原理、功能、最新动态以及相关新闻。
Found 31 web pages
搜索结果显示了一些相关页面。为了全面了解RAAST，我需要同时打开这些页面。
Read 5 pages
Raast P2P Person to Person https://www.bop.com.pk/view.aspx?id=3566
Raast Pakistan’s First Instant Payment System
https://www.ubldigital.com/Instant-payments-with-Raast
HBL The Largest Bank in Pakistan - Best Award Winning Bank
https://www.hbl.com/personal/digital-banking/digital-banking/raast
RAAST – Pakistan’s First Instant Payment System, NOW AVAILABLE on the
Alfalah Invest App & Online Portal!
https://alfalahamc.com/news/raast-pakistans-first-instant-payment-system-now-available-on-the-alfalah-invest-app-online-portal/
Pakistan’s RAAST Processes Rs80 Trillion in Transactions, User Base Crosses
48 Million - Pak Banker
https://pakbanker.com/pakistans-raast-processes-rs80-trillion-in-transactions-user-base-crosses-48-million/
这些资料提供了关于RAAST的详细信息。我的回答将涵盖其概述、工作原理、主要功能、使用方式、最新发展和影响。回答将采用清晰的结构，并引用这些来源。
RAAST 是巴基斯坦的第一个即时支付系统 https://www.bop.com.pk/view.aspx?id=3566
https://www.ubldigital.com/Instant-payments-with-Raast。你可以把它理解为巴基斯坦版的“
支付宝/微信支付”或印度的“UPI”，它旨在让个人、企业和政府之间的数字支付变得即时、安全且免费
https://www.bop.com.pk/view.aspx?id=3566
https://www.ubldigital.com/Instant-payments-with-Raast。
简单来说，它让转账变得像发短信一样简单。
🚀 核心机制：用手机号代替银行卡号
RAAST最核心的创新是使用了手机号作为支付ID（Raast ID）
https://www.bop.com.pk/view.aspx?id=3566
https://www.ubldigital.com/Instant-payments-with-Raast
。你只需将自己的手机号与银行账户绑定，这个手机号就成了你的“收款码”。
省去繁琐：别人给你转账时，不再需要询问复杂的IBAN账号，只需要知道你的手机号和你所在的银行名就行
https://www.bop.com.pk/view.aspx?id=3566
https://www.ubldigital.com/Instant-payments-with-Raast。
即时到账：在交易确认后，资金会实时（通常在10-20秒内）到达对方账户
https://www.bop.com.pk/view.aspx?id=3566。
💡 如何使用 RAAST？
使用起来非常简单，可以通过以下几种方式：
手机银行 App：登录你银行的手机App，找到RAAST转账功能，输入收款方的手机号（Raast ID）、IBAN账号，或直接扫描对方的二维码
，输入金额和密码即可
https://www.hbl.com/personal/digital-banking/digital-banking/raast。
网上银行：登录银行的网上银行，操作流程与手机App类似 https://www.bop.com.pk/view.aspx?id=3566。
银行柜台 (OTC)：如果你不太熟悉手机操作，也可以直接去银行网点，由工作人员协助完成转账
https://www.bop.com.pk/view.aspx?id=3566。
✨ RAAST 的主要特点
RAAST之所以能迅速普及，得益于它以下几个关键优势：
完全免费 (Zero Fee)：使用RAAST进行交易，没有任何手续费
https://www.bop.com.pk/view.aspx?id=3566。
安全可靠 (Secure & Reliable)：系统由巴基斯坦国家银行（SBP）开发，采用先进的加密和欺诈检测技术，确保每笔交易都经过付款人授权
https://www.bop.com.pk/view.aspx?id=3566
https://www.ubldigital.com/Instant-payments-with-Raast
https://alfalahamc.com/news/raast-pakistans-first-instant-payment-system-now-available-on-the-alfalah-invest-app-online-portal/
。
全天候服务 (24/7)：RAAST服务7x24小时可用，节假日无休。
覆盖广泛 (Interoperable)：所有商业银行都已接入RAAST
https://alfalahamc.com/news/raast-pakistans-first-instant-payment-system-now-available-on-the-alfalah-invest-app-online-portal/
，你可以在任何一家银行的App上使用。
📊 发展规模与愿景
自2021年1月11日推出以来，RAAST已成为巴基斯坦数字金融的基石
https://pakbanker.com/pakistans-raast-processes-rs80-trillion-in-transactions-user-base-crosses-48-million/
。
用户与交易量：截至2025年底，RAAST用户已超过4800万
https://pakbanker.com/pakistans-raast-processes-rs80-trillion-in-transactions-user-base-crosses-48-million/
，处理交易超过30亿笔
https://pakbanker.com/pakistans-raast-processes-rs80-trillion-in-transactions-user-base-crosses-48-million/
，累计交易额超过80万亿卢比
https://pakbanker.com/pakistans-raast-processes-rs80-trillion-in-transactions-user-base-crosses-48-million/
。
近期表现：2026年第一季度，RAAST处理了7.421亿笔交易，价值23.3万亿卢比。
未来愿景：
总的来说，RAAST正在深刻改变巴基斯坦的金融生态，它提供了一个高效、普惠的数字支付基础设施，有力地推动了该国向“无现金社会”的转型。
Regards,
Zafar Hussain, MBA, PFP®, LLQP, OTL
(647) 283‑4428 | @.***
Mississauga, ON L5M 6G9
Connect with Me on LinkedIn! https://www.linkedin.com/in/zafarhussain,
Facebook! https://www.facebook.com/Zafar.Hussain68, Instagram!
https://www.instagram.com/zafar.hussain.zh
Join the Awareness Community!
https://chat.whatsapp.com/CkHIimqJTXJBUAwo2O5QyD
On Sun, Jun 14, 2026 at 12:23 PM qingkong66 @.***>
wrote:
@github-actions[bot] commented on GitHub (Aug 30, 2026):
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!