ScreenBuffer

A live output panel for Jupyter: update one line, the header or the footer, and the rest stays put. Comes with a built-in run log.

Long-running loops (scrapers, ETL jobs, model sweeps, simulations, data pulls) leave you two bad choices in Jupyter:

• print() every step: the cell fills with thousands of lines you cannot read.
• clear_output() every step: you lose everything, including the parts you wanted to keep (a progress header, a running total, the last error).

ScreenBuffer keeps the output as a list of lines and redraws the whole panel from that list. You change one line, the header, or the footer; the rest stays on screen. Every line is also appended to a log file, so the screen can stay small while the full history is kept.

from screenbuffer import ScreenBuffer

sb = ScreenBuffer(max_display=8, header="Scraper | page 0/200",
                  log_file="scrape_run", max_wait=1000, start=True)

for i, page in enumerate(pages, 1):
    sb.update(header=f"Scraper | page {i}/200")
    sb.put(f"page {i}: +{len(page)} rows")
    sb.display()                                       # draw once per step

The panel shows a fixed header, the newest 8 lines, and (optionally) a footer. Nothing else on screen moves.

---

Who it is for

Analysts, data engineers, data scientists and researchers who run long jobs in Jupyter and want to watch progress without flooding the cell, and who want a plain-text record of what happened.

Not for: plain terminals or scripts outside Jupyter (it is Jupyter-only for now), rich dashboards (use tqdm, rich, or a BI tool), or log shipping (it writes local text files; sending them elsewhere is up to you).

Requirements

• Python 3.8+ and Jupyter (Notebook, JupyterLab, VS Code, or a Jupyter console).
• No third-party packages. Uses IPython.display.clear_output; if IPython is missing, drawing silently does nothing (the buffer and log still work).

Install

pip install screenbuffer

In a notebook cell, use %pip install screenbuffer so it installs into the kernel's environment. Then:

from screenbuffer import ScreenBuffer

---

Create a buffer

Everything starts here. All arguments are optional.

sb = ScreenBuffer(
    wait=True,                     # False = redraw on every change
    log_file=None,                 # name for the log set; None = no files
    max_main=None,                 # history file line cap
    max_wait=None,                 # saved-buffer file line cap
    max_display=None,              # lines drawn (newest N)
    start=False,                   # True = wipe existing files for this log_file
    header=None, header_alignment=None,
    footer=None, footer_alignment=None,
    show_header=None, show_footer=None,
    silent_display=None, only_display=None,
    auto_clean_edges=False,
    auto_clean_silent_only_display=False,
    shift_max_display=True,
    width=None,                    # None = detect terminal width
)

Argument	What it does	Details
wait	True (default): draw only when you call display(). False: redraw on every change	Redraw modes
log_file	Name of the log set (creates screen_buffer_logs/<name>.txt and companions). None = nothing written to disk	Log files
start	True wipes existing files for that name; False restores the saved panel	Persistence
max_display	Draw only the newest N lines	Display window
max_main, max_wait	Line caps for the history and saved-buffer files	Limits
header, footer	Fixed text above / below the lines	Header and footer
header_alignment, footer_alignment	Alignment for those two	Alignment
show_header, show_footer	Hide or show them (default: shown when text is set)	Header and footer
silent_display, only_display	Hide some lines / show only some lines	Filtering
auto_clean_silent_only_display, shift_max_display, auto_clean_edges	Fine-tune filtering and spacing	Filtering
width	Panel width in columns for center / right; None = detect	Alignment

Typical setups:

sb = ScreenBuffer()                                              # simplest: collect lines, then sb.display()
sb = ScreenBuffer(wait=False, max_display=10)                    # live panel, newest 10 lines
sb = ScreenBuffer(max_display=10, log_file="run1", max_wait=1000, start=True)   # long job with a log

---

put(): the core call

put() is how everything gets into the panel. Every other feature (alignment, window, filters, logs, resume) works on the lines you put().

sb.put(text, state=0, index=None, fill_value=None, alignment=None)

Argument	Meaning
text	The line. Non-strings are converted with str(). Keep it to one line (a \n inside would show as several screen lines but count as one buffer line, so center/right would apply to the whole block).
state	0 🟢 ok (default), 1 🟡 warning, 2 🔴 error. Written to the history log only; it does not change how the line looks on screen.
index	Where to put it. None = append at the end. An existing index = replace that line in place. Past the end = pad the gap, then add. Negative counts from the end.
fill_value	Text for the padding lines when index is past the end. Default: blank lines.
alignment	"center", "right", an indent (4), or from-the-right (-10). Applies to this line only. See Alignment.

Returns nothing.

Append, replace, pad

sb.put("loaded 1,200 rows")              # line 0 (append)
sb.put("retrying...", state=1)           # line 1, logged as 🟡
sb.put("progress: 40%", index=0)         # replace line 0 in place: the screen updates, nothing moves
sb.put("row 10", index=10)               # lines 2..9 become blank, line 10 is "row 10"
sb.put("progress: 80%", index=-1)        # negative: replaces the LAST line

Replacing by index is the "update one part in a loop" trick. Reserve a line, then keep rewriting it:

sb.put("Progress: 0%", index=0)           # reserve line 0 as a status line
for i in range(100):
    sb.put(f"Progress: {i}%", index=0)    # rewrite just that line
    sb.put(f"processed item {i}")         # new lines keep appending below

(In a window with max_display, a line that has scrolled out of the window is still updated in the buffer, but you won't see it. Keep status lines in the header or footer if they must always be visible.)

What happens on every put()

1. The buffer changes (append / replace / pad).
2. The line's alignment is stored if you passed one (replacing a line without alignment= keeps its old alignment).
3. If log_file is set: one timestamped entry is appended to the history log, and the saved buffer is updated.
4. In live mode (wait=False) the panel redraws. In the default mode it does not, until display().

The history entry records the index, even for replacements:

2026-10-21 00:12:07 🟡 retrying... [📍1]

Errors

• A negative index further back than the buffer length raises IndexError.
• A failed history write raises RuntimeError (lines are never silently lost).

Speed

put() stays cheap as the buffer grows: about 0.005 ms at 20,000 lines without a log file, 0.035 ms with one. Disk writes are queued, so a slow disk does not slow put().

---

Feature index

Find what you need:

I want to...	Use	Section
Create a buffer, see every option	ScreenBuffer(...)	Create a buffer
Add a line	sb.put("text")	put()
Replace / update an existing line	sb.put("new", index=3)	put()
Mark a line OK / warning / error in the log	sb.put("x", state=0/1/2)	put()
Center, right-align or indent a line	put(..., alignment=...)	Alignment
Show only the newest N lines	max_display=N	Display window
Have a fixed header / footer	header=, footer=	Header and footer
Change header/footer in a loop	sb.update(header=...)	Header and footer
Hide / show header or footer	show_header, show_footer	Header and footer
Draw once per loop step (no flicker)	wait=True + sb.display()	Smooth updates
Control when the screen redraws	wait=True/False, sb.display()	Redraw modes
Hide some lines on screen, keep them stored	silent_display=	Filtering
Show only some lines on screen	only_display=	Filtering
Remove blank gaps left by filtering	auto_clean_silent_only_display=True	Filtering
Delete lines	sb.delete(...)	Deleting and clearing
Wipe the screen buffer	sb.clear(), sb.clear_all()	Deleting and clearing
Keep a full history file	log_file="name"	Log files
Read the history back (with its location)	sb.read_log()	Log files
Resume the panel after a kernel restart	log_file= without start=True	Persistence
Start fresh and wipe old files	start=True	Persistence
Cap file sizes	max_main, max_wait	Limits
Change any setting while running	sb.update(...)	Changing settings
Fix wrong center/right alignment	width=	Alignment
Get the panel as a string	sb.get_text()	Reading state
See raw buffer with indexes	sb.display_index()	Reading state
Use from several threads	built in	Threads

---

Core concepts

• Buffer: the list of all lines you have put(). Each line has an index (0, 1, 2, ...).
• View: what is drawn. It is the header, then the visible lines (after the window and any filter), then the footer.
• History log: a timestamped file of every put(). Append-only record of the run.
• Wait file: a saved copy of the current buffer, so the panel can be restored.

The screen is redrawn with Jupyter's own clear_output(wait=True): the old frame stays until the new text is ready, then they are swapped. The new text is printed as one string, with your exact spacing.

---

Usage

Alignment

Per line, set when you put():

sb.put("RESULTS", alignment="center")
sb.put("total: 4,210", alignment="right")
sb.put("nested detail", alignment=4)      # 4 spaces of indent
sb.put("ends 10 columns from the right edge", alignment=-10)
sb.put("also works as text", alignment="4")

Value	Effect
None / "left"	unchanged
"center"	centred in the panel width
"right"	right-justified
int ≥ 0	that many spaces of left indent
int < 0	ends that many columns from the right edge
numeric string	same as the int

Alignment follows its line: it is saved and restored with the buffer, and re-keyed when lines are deleted. Header and footer have their own: header_alignment=, footer_alignment=.

Width: by default the terminal width is detected. Jupyter front ends sometimes report the wrong width, which skews center/right. Set width=70 (or whatever suits your output area) to fix it.

Display window

sb = ScreenBuffer(wait=False, max_display=10)

Only the newest 10 lines are drawn; all lines stay in the buffer and the log. Drawing cost depends on max_display, not on how many lines exist (8 lines out of 100,000 draws in under a millisecond).

Header and footer

sb = ScreenBuffer(wait=False, header="Job: nightly_etl", footer="status: running",
                  header_alignment="center")

sb.update(header="Job: nightly_etl | 42%")       # change in a loop
sb.update(footer="!! source unreachable !!")
sb.update(show_footer=False)                     # hide it (the text is kept)
sb.update(show_footer=True)                      # bring it back

Setting a non-empty header/footer turns it on. Header and footer are saved to files when log_file is set. set_header(text, save_to_file=False) / set_footer(...) keep one in memory only.

Smooth updates

With wait=False (live mode) every put() / update() redraws immediately. That is fine for one change per step. If a step makes several changes (new header, new footer, new line), each one draws its own frame, and you briefly see half-updated screens.

For loops, use the default wait=True: make all your changes, then call sb.display() once. One step, one redraw.

sb.update(header="OFFLINE", footer="retrying in 30s")
sb.put("connection lost", state=2)
sb.display()          # one redraw for all three changes

Other things that keep it smooth, all automatic:

• New frames are flushed immediately.
• A redraw is skipped if nothing on screen would change.
• Only the max_display window is built when drawing.
• Disk writes are queued and done in order, outside the drawing path, so a slow disk (antivirus, OneDrive, network drive) does not stall the screen.

Fast loops: if a loop calls put() hundreds or thousands of times per second, Jupyter itself can be flooded (its output channel is rate-limited). Call sb.display() once per chunk of ~50 items instead of on every item, or add a short sleep.

Redraw modes

Mode	Behaviour
wait=False	Live. Every change redraws (skipped if nothing on screen changed). The panel draws once on creation.
wait=True (default)	Quiet. Changes are stored but nothing is drawn until you call sb.display(). Good for loading lots of lines first.

sb.display() always draws, even if nothing changed. sb.flush() clears the Jupyter output immediately.

Filtering

Hide or isolate lines on screen only. They stay in the buffer and the log. Indexes refer to the full buffer.

sb.update(silent_display=[2, 5])        # hide lines 2 and 5
sb.update(silent_display=slice(0, 10))  # hide a range
sb.update(silent_display="all")         # hide every line (header/footer stay)
sb.update(only_display=[7])             # show only line 7
sb.update(only_display=None)            # remove the filter

Accepts an int, a list of ints, a slice, or "all". silent_display and only_display cannot be used together (setting one clears the other).

Related options:

Option	Effect
auto_clean_silent_only_display=True	Hidden lines disappear instead of leaving blank rows
shift_max_display=True (default)	Filter first, then take the newest max_display visible lines (hidden lines do not use up slots)
shift_max_display=False	Take the newest max_display buffer lines first, then filter
auto_clean_edges=True	Trim blank lines at the top/bottom of the body, then add one blank line each side

Deleting and clearing

sb.delete(3)                                   # delete line 3
sb.delete(slice(0, 5))                         # delete a range
sb.delete(slice(0, 20), contains="error")      # only lines in that range containing "error"
sb.delete(slice(0, 20), contains="error", exact=False)   # substring match
sb.clear()                                     # empty the buffer (header/footer stay)
sb.clear_all()                                 # empty the buffer AND remove header/footer

contains is case-insensitive. With exact=True (default) it matches a whole word; with exact=False, any substring. Alignments move with their lines when others are deleted. Deleting does not remove entries from the history log.

Log files

sb = ScreenBuffer(log_file="scrape_run")

Creates a screen_buffer_logs/ folder next to where you run it (relative to the current working directory):

File	Contents
screen_buffer_logs/scrape_run.txt	History: one timestamped line per put(), never rewritten by edits
screen_buffer_logs/scrape_run_wait.txt	Saved buffer: current lines and their alignments, for restoring the panel
screen_buffer_logs/scrape_run_header.txt	Saved header
screen_buffer_logs/scrape_run_footer.txt	Saved footer

History line format:

2026-10-21 00:12:07 🟡 retrying... [📍14]

(timestamp, state emoji, text, buffer index). Characters that would break a log entry across lines (\n, \r, Unicode line separators) are escaped in the history file, so one put() is always exactly one log line and text cannot forge fake entries.

Read it back:

sb.read_log()               # everything
sb.read_log(last=5)         # newest 5
text = sb.read_log(show=False)   # return as a string instead of printing

Output starts with the file location:

Log file: /home/you/project/screen_buffer_logs/scrape_run.txt
(5 of 312 entries)
----------------------------------------
2026-10-21 00:12:07 🟡 retrying... [📍14]
...

read_log() raises ValueError if no log_file is set. A write failure raises RuntimeError rather than silently losing log lines.

Persistence and resume

• ScreenBuffer(log_file="x"): if the saved buffer exists, it is restored (lines, alignments, header, footer). Kernel crashed? Re-run the cell and the panel comes back.
• ScreenBuffer(log_file="x", start=True): start fresh; existing files for that name are emptied.
• An explicit header= / footer= always wins over a stale saved one.

Limits and long runs

Option	Limits
max_main=N	Lines kept in the history file (oldest dropped)
max_wait=N	Lines kept in the saved buffer file (oldest dropped)
max_display=N	Lines drawn

max_wait cannot be smaller than max_display. Set max_wait (500–1000 is typical) on long runs; it keeps the saved copy small and makes appends cheap.

These cap files, not memory. The in-memory buffer keeps every line (roughly 100 MB per million short lines).

Changing settings at runtime

sb.update(max_display=20, width=90, wait=False)
sb.update(log_file="second_phase")      # switch to a different log set

Settable: wait, max_main, max_wait, max_display, shift_max_display, auto_clean_edges, auto_clean_silent_only_display, header, footer, header_alignment, footer_alignment, show_header, show_footer, silent_display, only_display, width, log_file.

update() checks everything first: if any value is invalid it raises and changes nothing. Unknown names raise ValueError.

Reading state

sb.get_text()          # the current panel as one string
sb.display_index()     # prints every buffer line with its index (debugging)
sb.buffer              # the raw list of lines

Threads

put(), update(), delete() and drawing are thread-safe. Disk writes keep buffer order even when several threads write at once.

---

Method reference

Method	Purpose
put(text, state=0, index=None, fill_value=None, alignment=None)	Add or replace a line
update(**settings)	Change any setting (validated first)
display()	Draw now
flush()	Clear the Jupyter output now
delete(index, contains=None, exact=True)	Delete by index/slice, optionally filtered by text
clear()	Empty the buffer
clear_all()	Empty the buffer, header and footer
set_header(text, save_to_file=True, refresh=True) / set_footer(...)	Set header/footer
get_header() / get_footer()	Current header/footer text
get_text()	Panel as one string
read_log(last=None, show=True)	Read the history log
display_index()	Print lines with indexes
refresh_buffer_from_wait_log()	Reload the buffer from the saved file (only when the buffer is empty)

---

Example: network scraper with an error banner

import time, random
from screenbuffer import ScreenBuffer

sb = ScreenBuffer(max_display=8, width=70,
                  header="SCRAPER | ONLINE", header_alignment="center",
                  footer_alignment="center",
                  log_file="scraper", max_wait=500, start=True)

for t in range(120):
    online = not (40 <= t < 70)
    if online:
        sb.update(header=f"SCRAPER | ONLINE | {t}s", show_footer=False)
        sb.put(f"[{t}s] page {t//2}: +{random.randint(3, 9)} items")
    else:
        sb.update(header=f"SCRAPER | OFFLINE | {t}s",
                  footer="!! NETWORK ERROR - retrying !!")
        if t % 5 == 0:
            sb.put(f"[{t}s] retrying...", state=1, alignment=4)
    sb.display()                       # one redraw per step
    time.sleep(0.5)

sb.read_log(last=5)

The error banner appears at the bottom during the outage and disappears when the network returns, while the page counter keeps climbing above it.

---

Performance (measured in a sandbox)

Case	Before tuning	Now
put() with 20k lines in the buffer (no log / log / log + max_main)	2.1 / 10.7 / 16.4 ms	0.005 / 0.035 / 0.11 ms
Draw 8 lines out of 100k	30.5 ms	0.72 ms
Frame wait with a 20 ms disk delay	202 ms	0.1 ms
Old frame cleared → new text shown	100–200 ms	~0.5 ms
Redraws in a 300-step demo (one draw per step)	241	81

These come from a Linux sandbox with a simulated Jupyter kernel. Real browser rendering was not measurable there; if you see flicker, see Troubleshooting.

Known limits

• Jupyter only. Terminals and plain scripts are not supported (drawing is a no-op without IPython).
• Replacing or deleting an old line rewrites the saved-buffer file. Appending is cheap; edits cost O(lines in that file), which max_wait keeps small.
• max_main trimming rewrites the history file once it goes over the cap.
• Memory is not capped (see Limits).
• Queued writes can be lost if the kernel is killed abruptly. Call sb.read_log() (it waits for writes) or let the cell finish before shutting down.
• Very fast loops can exceed Jupyter's output rate limit regardless of this tool. Draw once per chunk or add a short sleep.
• Wide characters (emoji, CJK) count as one column for alignment, so center/right can be off by a column or two.
• Not tested on Windows, in real browsers, or in the Jupyter console (the console front end).

Troubleshooting

Symptom	Likely cause / fix
Nothing appears	wait=True (default): call sb.display(), or create with wait=False
Flicker in a fast loop	Use wait=True and call sb.display() once per step (or per chunk of items); add a tiny sleep
Header and footer change on separate frames	Use wait=True, make both updates, then call sb.display() once
center / right look off	Set width= to your output area's real width
ValueError: max_wait cannot be less than max_display	Raise max_wait or lower max_display
Old panel returns after re-running a cell	Saved buffer was restored; use start=True to begin fresh
read_log() says no log	Create the buffer with log_file="name"
Output scrolls instead of replacing	Output area is detached or using a front end without clear_output support