Skip to content

BDSP AutoStory: D-pad taps where no wall holds every possible tile (00.05c no plan) - #427

Draft
dayfine wants to merge 16 commits into
bdsp/start-certaintyfrom
bdsp/0005c-plaza
Draft

dayfine wants to merge 16 commits into
bdsp/start-certaintyfrom
bdsp/0005c-plaza

Conversation

@dayfine

@dayfine dayfine commented Nov 12, 2026 •

Copy link
Copy Markdown
Owner

Depends on #400 (base bdsp/start-certainty; the diff shows only this change).

00.05c failed on lab-B (2026-11-12 02:52) with "no plan from Jubilife City (35, 24), (36, 24)" after the walk west on row 24 talked to a woman. The marks didn't cause it: from (36, 24) no direction ends at a wall. North is the TV station's clown, south a man, west the woman's tile, and east the plaza's avoid box, so the recovery planner, which uses holds to a wall only, had nothing to try. When walls alone find nothing, plan_recovered_route() now allows one-tile D-pad taps for holds that have no wall from every tile, up to four holds. Here that gives two taps east and then north to the TV station's wall. Plans that walls alone already find are unchanged.

This doesn't fix the root cause of the third failure. The position belief drifted by about 10 tiles on "west 17 (to wall)" along row 26: the time count read "about 13 tiles", but the screen-change profile matches the 4-tile south run just before it, and the failure frame shows the same spot beside the Wonder Station as the 01:22 and 01:55 failures. That fix and the hand measurement are tracked separately.

Testing: a red-first StoryTable case (00.05c: a plan from (35, 24), (36, 24) past the woman on row 24), plus the PokemonBDSP AutoStory suites, run with the headless runner.

Update (run 4, 03:57): the woman by the Wonder Station is most likely a scripted greeter, not an obstacle. She walked up and talked three times on one walk round the plaza, twice right after a turn with no step taken, and the third talk failed the step as "dialogue the map does not explain". Route steps now take "greeter" rows (00.05c's Jubilife routes name (30, 22) to (48, 34)); a dialogue no trainer explains there is heard out and walked on from, and the walk allows six of those instead of two. Tested with a red-first grid sim case where the greeter talks again after each talk (old walker fails at the third, new one arrives), plus the PokemonBDSP AutoStory suites. Also merges #400's tip b258693.

Update (run 5, 04:58): her talk was taken for "a trainer on the way" at the wall check, so the greeter path never ran. No battle followed, 2 s of B left her long explanation on screen, and the walk on into her ran out of time. Now, after the fight with whoever stops a walk, a dialogue still showing (or back within 5 s) is a talk with no battle. It's pressed through until the overworld has been quiet for 5 s (at most six times) and counted with the greeter cap. The open tile ahead of a single stop tile comes off our map, and the walk replans instead of walking on into her. Tested with a red-first grid sim case (a two-row corridor, a talk that outlasts two press-throughs, with and without her standing on the leg), plus the PokemonBDSP suites with Monte Carlo 00.05b/00.05c, using the headless runner.

🤖 Generated with Claude Code

https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss

dayfine and others added 3 commits November 12, 2026 03:06
…on row 24 (red)

On lab-B (2026-11-12 02:52) the walk west on Jubilife's row 24 talked to the
woman it took for (34, 24) and, from (35, 24) and (36, 24), found no plan
with the route costs, the default ones or without the plaza's avoid box.
Neither mark matters: the two tiles alone have no plan.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
plan_recovered_route() brought several possible tiles to one only by holds
to a wall from every tile. From (36, 24) in Jubilife no way ends at a wall:
the TV station's clown north, a man south, the woman's tile west, the
plaza's avoid box east, so 00.05c failed with "no plan" (lab-B 2026-11-12
02:52). When walls alone find nothing, a hold with no wall for every tile
may be one D-pad tap (whole tiles), up to four holds: two taps east, then
north to the TV station's wall. Plans that walls alone find are unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 12, 2026
dayfine and others added 2 commits November 12, 2026 03:31
…it; taps back to the wall find where

A push against someone the map lacks reads as stopped late, so its time
counts many tiles too many, and a walk on that moved no chunk counted its
pushed time as tiles too. Such a stop is now any tile of the leg up to a
tile past the count; D-pad taps back to the wall behind tell which, and the
blocker goes on the map. 00.05c's plaza avoid boxes follow lab-B's
2026-11-12 03:25 taps: row 26 is open to x 44, the woman stands west of it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
…tiles round its count

Route 204's walk below the Ravaged Path (the sim's wandering youngster)
found no plan from the whole leg across Jubilife's edge. Without a wall
to tap back to, keep the old ±1 window; taps that stop short of every
wall leave the stop tiles shifted back by the taps.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine added a commit that referenced this pull request Nov 12, 2026
…laza boxes from 03:25 taps) into lab/dev-bdsp
dayfine and others added 5 commits November 12, 2026 04:06
…she talks

00.05c failed a 4th time on lab-B (2026-11-12 03:57): a woman by the
Global Wonder Station walked up and talked three times on one walk, at
about (44, 28), (43, 30) and (39, 32), each right after a turn or a short
leg, and the third failed the step as "dialogue the map does not explain".
She is a greeter who comes to us near the plaza, not a trainer or a wall.
A route step's new "greeter" rows name her ground; a dialogue no trainer
explains while we may stand there is heard out as before (the trainer
press-through) and walked on from, and the walk allows six such talks
instead of two. 00.05c's four Jubilife routes take the box (30, 22) to
(48, 34). story_route_avoid() becomes story_route_tiles(world, step, key).

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

00.05c below_the_company_clown failed a 5th time on lab-B (2026-11-12
04:58): the Wonder Station's woman stopped "west 1 (to wall)" as "a trainer
on the way", no battle followed, 2 s of B left her long explanation on
screen, and the walk on to the wall ran into her until it ran out of time.
The greeter box never applied: that path only walked on.

After the fight with whoever stopped a walk (a trainer the map lacks, or
the one at the wall check), a dialogue still showing or back within 5 s is
a talk no battle followed: it is pressed through again until the overworld
is quiet 5 s (at most six times), counts as an unexplained talk (the
greeter cap applies), the open tile ahead of a single stop tile comes off
our map, and the walk plans again from where it stopped instead of walking
on into her. check_at_wall() and on_to_wall() return LegNext for that
replan. Grid sim: a two-row corridor with a talk that outlasts two
press-throughs, with and without her standing on the leg (red before).

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

00.05c below_the_company_clown failed a 6th time on lab-B (2026-11-12
06:12, lab c7a7ac1): the Wonder Station's woman stopped the check walk
of "west 1 (to wall)" as "a trainer on the way". fight_route_trainer()
saw no battle follow and pressed her talk to its end, so the hear-out
found no dialogue, and check_at_wall() walked on to the wall, which
started her talk again until it ran out of time.

RouteTrainerFight (and StoryStepIO::fight) now return whether a battle
followed. check_at_wall() and unknown_trainer() take a fight with no
battle as a talk no battle followed, as when a dialogue still shows:
counted, the tile ahead marked, and the walk plans again from where it
stopped instead of walking on.

Tests: a scripted wall leg (trainer stop at the wall check, no battle,
nothing shown after) and a grid-sim talker whose talk ends in the fight
while she stands on the leg; both red before. The greeter and talker
fakes report no battle.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
00.05c below_the_company_clown failed a 7th time on lab-B (2026-11-12
07:45, run 8, lab a6e3e8f): the plaza greeter's talk stopped "south 5
(open)" from (46, 27) with no tap counted, so the walk planned from the
whole column (46, 27) to (46, 33); her next talk, on a 1-tap recovery leg,
spread that over two columns, and no plan brought them to one wall: "no
plan from Jubilife City (44, 27), (45, 27), ...".

A talk no battle followed on a tapped leg now counts only the leg's own
tiles, not the trainer's extra one (a talk freezes us; nobody walks over).
And when no plan comes from the tiles we may stand on, the walk holds
toward a wall each of them stops at, choosing a way whose walls the time
walked can tell apart, keeps the walls whose distance fits that time
("found again"), and plans from there. Tiles are still marked only when
the stop is one tile.

Tests: a grid sim of the plaza from (46, 27) whose open-leg talks lose
the tap count (one talk; two talks, the second a tap on): the two-talk
case failed with "no plan from" before, as on lab-B.

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