History

2026-10-12 to 2026-10-13: from no firmware to a running custom image

Date	Step
2026-10-12	No X91 firmware is published anywhere; USB ID 10d6:1101 measured on the unit
2026-10-13	66 AGPTEK packages downloaded and 52 decrypted for reference; none fits the X91
2026-10-13	Upgrade mode entered by software (adfu_reboot); chip confirmed as ATJ2157
2026-10-13	Firmware dumped twice through the boot area and verified: x91_lfi_verified.bin
2026-10-13	Read-only rehearsal through the vendor driver: 0 of 38,811 sectors differ
2026-10-13	First write (the unchanged image), then a custom start-up picture: the write path works
2026-10-13	Stages 1 to 12 built and flashed one after the other; x91_tool.py written
2026-10-13	Stage 13 built (never flashed alone), stage 14 built

2026-10-14: status bar, clock, reader

Stage	On the player	Result
14	Mint	works: scrolling text, tag title, status bar at once in Music
15	Berry	works: main menu clock, screensaver, the Berry theme
16	Berry	reader in landscape; lines broke inside words
17	Berry f1c849f5…	works: lines end between words, page number, bar, keys
18	Berry 4737779c…	playlist list without names; a stutter while writing

2026-10-21 to 2026-10-22: playlists and lists

Stage	Image (Berry)	Result
18 rebuilt	d40fd160…	works: names, adding, new playlists, “Update list?” reads the .m3u files
19	13bacdaa…	glyphs and hidden rows fine; built-in playlists still shown
19 rebuilt	3189295f…	works
20	c2ce9554…	works: library at once, Remove from Playlist, Previous restarts; one message broke
21	861c23e5…	add, delete, volume fine; “Move song” failed (two files open)
21 rebuilt	a14cef26…	works
22	af61b64e…	fine apart from the picker’s sixth row and Previous with the screen off
23	59e444c8…	works
24	c793fc34…	menu names wrong or cut off, empty top lines, no counter in album folders
24 rebuilt	decde654…	works, except the eBook list’s name and the status bar’s delay in sub-menus
24 rebuilt again	efa340ee…	eBook name fine; status bar much faster, but a short blink when a menu or list opens
24 rebuilt a third time	82eed591…	never flashed on its own; part of every stage 25 and 26 image

2026-10-23: search, adding songs, naming playlists

Stage	Image (Berry)	Result
25	e64bdfb6…	search works; around its texts stale pixels: the underline broken under the typed letters, specks above the keys, uneven key edges
25 rebuilt	ee7d8549…	works: clean texts and keys, and fast
26	00112ce2…	“Add songs” works; after each add the screen first showed the digits typed and an empty line
26 rebuilt	2a25f3eb…	no digits, but with fewer than four matches the screen waited about 5 s (keys dead) after each add
26 rebuilt again	ec287d08…	the screen at once, keys at once; the new songs show in the playlist only after it is entered again
26, fourth build	89137f1b…	works: the new songs show in the playlist at once
27	770fd5fe…	typing works; the name’s line misplaced: the mark of the letter being typed too wide and about 20 pixels too far right, the caret behind the third letter of a longer name, stray letters to the right of the mark (the width function was asked with type 0x7d and read the UTF-16 as single bytes)
27 rebuilt	b1f169d1…	works: the width asked with type 0x7f, the mark and the caret in place, no stray letters

Stage 27 rebuilt in every theme: Amber 5fb39a7f…, Mint 2b084b26…, Ice 7f935828…, Berry b1f169d1…. Up to then the images of stages 24 to 26 stayed reproducible from frozen copies of their sources (design/playlist/stage24/, stage25/ and stage26/; stage 25 rebuilt: Amber 8945629a…, Mint 2c36fdb9…, Ice e91c8f53…, Berry ee7d8549…; stage 26, fourth build: Amber 76f68ef2…, Mint ded7d030…, Ice 26612f9f…, Berry 89137f1b…).

2026-10-23: release 1.0

Step	Result
The 27 stage builders merged into one build, design/build.py with the parts in design/graphite/	every change applied in its final form to the stock image; code a later stage took out again is never placed; a build of all four themes takes about a minute (the chain took several)
Compared with the stage 27 images	byte for byte the same in all four themes: Amber 5fb39a7f…, Mint 2b084b26…, Ice 7f935828…, Berry b1f169d1… (RELEASE in build.py, --check)
Removed	the stage builders, the frozen sources per stage, stage 9’s music_progress.S and its checker (all in the git history); the stage previews moved to design/history/

The merged build keeps one oddity of the stage builders: in CONFIG.TXT, the readable copy of the settings that the firmware does not read, stage 5 wrote the six-tile menu into the commented-out line above the active one, which still lists the five tiles of stage 3. Correcting it would change only that file, and so the images’ checksums.

Early assumptions that were wrong

Comments in the stage builders (in the git history) may still reflect these:

• The screensaver clock is drawn by code, not by its layout (stage 6 → 7).

• Pictures 345/346 are the brightness sun, not battery marks (stage 4 → 10).

• The track counter’s digit count is a byte of the number-box record (stage 22; stage 14 thought it fixed).

• The engine’s message numbers are the SDK’s plus 2 from 0x303 on (stage 18 used the SDK’s; fixed in stage 20).

• String fill mode 3 does not draw “only the dots” over what is on the screen: it leaves the driver’s strip as the previous string left it (stage 25; rebuilt with mode 1).

• The engine does have a seek: 0x317 (stage 20).

• Bytes 0 and 1 of 0x202fbc mean “the music engine runs”, playing or paused (stage 24).

• The volume curve is in the music engine, not in the audio driver (stage 21).

• The display driver’s text width (function 23) takes UTF-16 with the type the string drawing takes, 0x7d: it takes it only as 0x7f (stage 27).