Skip to content

BDSP AutoStory intro: language/look pickers, resume, name limit, capture mode - #527

Draft
dayfine wants to merge 36 commits into
bdsp/starter-rng-adoptfrom
bdsp/intro-pickers
Draft

dayfine wants to merge 36 commits into
bdsp/starter-rng-adoptfrom
bdsp/intro-pickers

Conversation

@dayfine

@dayfine dayfine commented Nov 15, 2026

Copy link
Copy Markdown
Owner

Depends on #526 (base bdsp/starter-rng-adopt); merges in #305 (bdsp/autostory-intro-picks), so the diff shows #305's pickers plus this change.

The new-game intro always restarted the game and had no way to stop before a name the game would cut short, and nothing collected each language's name keyboard. This takes #305's language and look pickers (Game language picks the kana Japanese item for Japanese; PLAYER_LOOK picks the portrait) and adds three things: on an intro screen (user select, language or look select, keyboard, dialogue, rival list) the intro takes up from there without closing the game; before typing, the keyboard's "0/N" counter is read with OCR and a longer name fails at the name screen with nothing typed; and CAPTURE_INTRO_SCREENS restarts the game per language-select item, picks it and the look, and saves <tag>_keyboard.png (plus en_language-list.png and en_look-picker.png) to the Screenshots folder. With Run Starter RNG on, PLAYER_LOOK must match the Starter RNG character model.

Tested with new cases in the BDSP AutoStory Actions tests (resume progress, counter parsing and name fit, capture tags); compiled locally, CI runs the suite. Not run on a console yet: the counter's OCR box is measured on a Legends: Z-A keyboard frame, and an unread counter only logs and types.

dayfine and others added 30 commits November 9, 2026 10:35
…me intro

The intro pressed A on every screen, so it took the language the cursor
starts on and the first look (Lucas); run 3 (French, Dawn) needed both
picked by hand.

NintendoSwitch/Programs/MenuCursor: the D-pad presses from one item of a
list or grid to another, none across an edge, so they hold whether or
not the menu wraps; and a D-pad probe (press, opposite press, measure)
that tells a screen with a cursor from one without, for menus no
detector knows. Nothing in it is BDSP's, for PLZA's new game too.

BDSP: the language select's order and the appearance grid's shape are
data. Until a pick is made, screens that would take A get a probe
first: a list that moves up/down but not sideways takes the language
pick, a grid that moves sideways takes the look pick, then A. The
player's keyboard with a pick still to make fails. New options: Switch
language (where the list's cursor starts) and Player's look (models
1-8, PlayerModelOption's icons). Defaults keep the old behavior.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
The language select is a 2 x 5 grid of buttons, not a 9-item list
(lab-B frame, 2026-11-09): English, Espanol, Francais, Deutsch, Italiano
on the left; Japanese kana and kanji, Korean, Simplified and Traditional
Chinese on the right; the cursor starts on English. Right moves on it,
so the probe rule took it for the appearance grid.

A detector reads it from its ten buttons (grey, the cursor's blue) and
which one the cursor is on. The intro classifies it as its own screen,
moves from the read cursor to the game's language, reads it again (up
to three tries), then A. The look grid keeps the D-pad probe, only after
the language is picked. The Switch language option is gone: the cursor
is read, not assumed. The wrong-game guard is back to 60 s.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
A missing negative screenshot no longer skips the whole test, and a
black screen must not read as the language select.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
The Right/Left probe never saw the look select's cursor move (console B,
2026-11-09: 20 probes, then the name keyboard). The look select is a
2 x 4 grid of portraits, the cursor's background blue and the others
grey, like the language select's buttons (lab-B frame).

The language detector becomes IntroMenuDetector, one card-grid reader
with a layout per menu: language buttons, and two background strips
beside each portrait's face. The intro classifies the look select as
its own screen and picks on it as on the language select: read the
cursor, move to the chosen look, read again (3 tries), then A.

Model N is portrait N - 1, row by row, matching PlayerModelOption's
icons: 1 Lucas light, 2 Lucas dark, 3 Dawn dark-haired, 4 Dawn dark;
5 Lucas blond, 6 Lucas brown, 7 Dawn blond, 8 Dawn brown-haired. The
option text says so. The D-pad probe is gone from MenuCursor: nothing
uses it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Brings in the base stack's code-health splits, including the split of
NewGame.h into IntroSteps.h and IntroProbe.h.

Conflict resolutions:
- NewGame.h: keeps this branch's language/look picks (grids, IntroPicks,
  pick moves) and detail::intro_start; the intro state machine moved to
  the base's IntroSteps.h and IntroProbe.h, and NewGame.h no longer
  needs its header_symbols allow tag.
- IntroSteps.h: this branch's new steps (FAIL_PICK_MISSED, PICK_LANGUAGE,
  PICK_LOOK), the pending-pick fields, string_view screens and its detail
  namespace for the limits, record_intro_step and note_intro_screen; the
  allow tag is dropped, as detail names are not counted.
- IntroProbe.h: the X/B probe in detail, as this branch had it.
- NewGame.cpp: this branch's menu picking; includes unioned.
- Actions_Tests.cpp: this branch's screen list with the base's const.
- StoryExecutor.h, SourceFiles.cmake: union, cmake block regenerated.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
…ro-picks

Brings #128's seam work down through #167 and #226.
Conflict: PokemonBDSP_AutoStory.cpp includes, both NewGame.h (this
branch's IntroPicks) and Actions.h, as lab/dev-bdsp has them.
Semantic: ScreenState gained a field in the base; the language-select
and look-select states pass nullptr for it, as lab/dev-bdsp does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
…ro-picks

Picks up #226's code-health fixes after the #128 merge. No conflicts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Retargets onto the consolidated BDSP console stage (#375's branch). No conflicts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Picks up the consolidated stage's StorySim split and code-health fixes. No conflicts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Picks up the consolidated stage's test-file split. No conflicts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
…not a crash

gray_box() sorted and indexed an empty vector when the box was outside the
frame or the image empty (SIGSEGV at 0x0, 2026-11-12). read_move_name now
returns "" for it and the menu falls back to the PP guess.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
The intro's language and look pickers. StoryEnv lives in StoryCheckpoint.h
here; it carries the picks, and AutoStory gets the PLAYER_LOOK option.
…pture mode

- On an intro screen (user select, language or look select, keyboard,
  dialogue, rival list) the intro takes up from there without closing the
  game; elsewhere it restarts as before. intro_resume() sets which picks are
  still ahead.
- Before typing the player's name, the keyboard's counter ("0/N") is read
  with OCR; a longer name fails at the name screen with nothing typed.
- CAPTURE_INTRO_SCREENS: per language-select item, restart, pick it and the
  look, save <tag>_keyboard.png (and en_language-list.png,
  en_look-picker.png on the first round) to the Screenshots folder.
- PLAYER_LOOK must match the Starter RNG character model when both run.
Capture mode no longer assumes the name keyboard follows the look pick.
From the look's A it records frames as <tag>_after-look_NN.png: one on
each screen change (pixel RMSD over CAPTURE_CHANGE vs the last saved) and
at least one per second, until the screen holds still for 3 s (not before
8 s) or 30 s pass. It presses nothing: on the keyboard every button edits,
confirms or closes it. The look picker is saved every round as
<tag>_before-look-pick.png, the language list on the first as
en_before-language-pick.png; a missing frame is logged, not a failure.
The story-run intro is unchanged.
…and the language load's visits

jp1 (lab-B 2026-11-15) took the intro up on the look select of a zh-Hant
game while Japanese was asked for; jp2 failed on the language select's
visit cap while Japanese loaded. The cap gets a name, still 10.
…tay up through the language's load

jp1: taken up on a later intro screen, the intro kept the game's language
(zh-Hant) while Japanese was asked for. It now takes the intro up only on
the user select and the language select, and restarts on any other screen;
intro_resume() and picks_may_be_past go with it.

jp2: the pick did confirm, but Japanese's load keeps the language select up
for the pick and about 10 more looks (~6 s; 9 in the capture run, at the cap
of 10). The cap is now 30.
…d look picks

On Japanese the look select showed for a frame before it read as one, and
the A taken there picked the default look.
The settings step matched every Options row against English label and
value images, so a Japanese game failed on Text Speed. Japanese also has
a Kana/Kanji row under Text Speed. options_layout() now gives the rows per
language: English keeps its label and value images; other languages
confirm the row by the selected row with its value highlighted and the
value by the column of its red ink (sliders still by their number);
Japanese leaves the Kana/Kanji row as it is and finds Autosave a row
further down.
PlayerHouse1F mid-room frames show only 10.6-14.4% black on an edge strip,
so indoors() read them as outside and the walker lost 3 tiles in 00.01
above_the_door. On 2,669 lab screenshots warm outdoor frames never exceed
3% and rooms never fall below 10.6%.
…r a step

00.01 below_barrys_warning north 18 counted 3 while the player walked about 5,
and no frames were kept to see why. TAPS legs now log per tap the change, the
shift, unshifted/shifted difference, their ratio and the verdict, and save the
before/after frames (at most 10 pairs a walk) under Screenshots/AutoStory/route/
when a tap reads still with a change of at least 60% of the walk's largest.
No change to what counts as moved.
The camera looks down at an angle, so a step north up Twinleaf's path
scrolls the near rows further than the far ones (7 pixels of the copy at
the top, 10 at the bottom). One shift for the whole view lined up those
steps only 1.69 to 2.73 times better than none, so 00.01
below_barrys_warning's north 18 read two taps as still and stopped after 3.

walk_view_shift() now finds the best shift for each of four horizontal
bands of a smoothed copy, and a tap scrolled when at least 3 bands line up
twice as well as none, shifted the same way. The 3x3 smoothing keeps the
point-sampled copy's grass from aliasing on part-pixel scrolls. On the two
saved console pairs every band lines up 2.49 to 5.18 times better; cloud
shadows, a TV and a turn still read as standing. The tap log prints each
band's shift.
…le move menu test needs no OCR

The intro waits on unknown screens between the language and look picks
(5d64b5f), so the pick-steps table expects WAIT for that cutscene. The
unreadable menu test put one button in the frame, which reached Tesseract
and failed on CI without its data; both buttons now sit off the frame.
…inleaf's exit with taps 13 and 14 read still

Red: below_barrys_warning's north 18 taps stop BLOCKED at tap 14 and the
row fails short_leg, as on lab-B 2026-11-15.
The north 18 leg counted taps by the view shift, which read moving taps
still by the "201" banner (lab-B 2026-11-15) and failed short_leg. The
leg is now a to_dialogue run of 19 tiles onto the warning's trigger row
(Route 201 row 18, coord_event_0 from pret/pokediamond narc_0327), as
00.03's runs into Rowan's and Dawn's scenes; the step onto the trigger goes.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant