Skip to content

FRLG AutoStory 00.28-00.32: Rock Tunnel to Giovanni in the Rocket Hideout - #293

Draft
dayfine wants to merge 252 commits into
frlg/checkpoint-replayfrom
frlg/stage8-rocktunnel
Draft

dayfine wants to merge 252 commits into
frlg/checkpoint-replayfrom
frlg/stage8-rocktunnel

Conversation

@dayfine

@dayfine dayfine commented Nov 9, 2026 •

Copy link
Copy Markdown
Owner

Depends on #258

The table stopped at Route 10's Pokemon Center (00.27). 00.28 cuts the tree (10, 20), goes through Rock Tunnel without Flash, fights the trainers the way can't go round and saves by the south mouth; 00.29 goes down Route 10's east side to Lavender Town's nurse, and 00.30-00.32 go on by Bicycle to Celadon and through the Game Corner's Rocket Hideout to Giovanni. Every battle step names its trainer and party in the plan; the walker now models the Bicycle, dark caves, spinners, arrow floors and the hideout's lift, with the move timings in MovementMechanics, and the map resource lock moves to PokemonSimulation bdca47c (Rock Tunnel, Lavender, Routes 7/8, Celadon and the hideout). Tested with the new AutoStoryLavender, Bike, fault-sweep and walker sim suites plus the FRLG, Pokemon:: and StoryReplay suites, all non-image tests passing locally at a30397d; this merge was not run on the console.

🤖 Generated with Claude Code

https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss

dayfine and others added 8 commits November 9, 2026 10:20
Unverified authored-checkpoint drafts, outside the embedded StoryData set:
the Lavender Mart, Route 8, the Underground Path, a Pidgey for Fly, and the
Rocket Hideout to Giovanni. The README lists the whole 00.29-00.51 draft,
its conventions, the user's open decisions and the leg check.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
A frame source serves a recording's frames ("<ms>.png") by virtual time,
and replay_until / replay_events feed them to VisualInferenceCallbacks
or a ScreenEventObserver at the inference period on one thread, so a
console failure can become a deterministic test of the real detectors.

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

The walk's warp watcher and FRLG's screen events, replayed over each
run's end frame (CommandLineTests/PokemonFRLG/FrameReplay/), agree that
no fade showed: the walk ended off the door, not past a missed fade.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
…stration function at 50 lines

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

PokemonFRLG_Tests.cpp is at its 1000-line limit.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine and others added 9 commits November 9, 2026 10:56
…ock Tunnel

Console B, 2026-11-09: B1F row 37 met a wild battle 10 to 14 tiles into
every try and the walker gave up after 9 stops. Red: both fail.

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

Rock Tunnel B1F row 37 (console B, 2026-11-09): a wild battle 10 to 14
tiles into every try of "west 16". The only single hold that found the
player went east to the wall (45, 37), behind the leg's start, and the
way back walked all 32 tiles again, until the route gave up. The
reanchor search now also tries each first hold with the fewest holds
after it, and plan_reanchor() takes the way whose holds and way back
walk the fewest tiles together: there, south to row 38 and west to the
wall (1, 38), then 13 tiles back.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
The late-watcher case showed a fade missed by a watch that starts after
the hold; it now replays the walk's watch (WatchAlong over the hold, then
the same watcher on) and checks it sees a door's black that is over before
a 2000 ms hold ends, while a watcher started at the end of the hold still
misses it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
The walk waited for a warp's fade only once the hold onto the warp had
ended, so a fade that was over by then went unseen and the walk stopped on
"no fade in and out after walking onto the warp" though the player had
warped (00.27, console B, 2026-11-09 01:41, the warp to CeruleanCity: the
watch started about 1.3 s into the hold). The walk and the recovery walk
now show the hold's frames to the same BlackScreenOverWatcher
(Pokemon::WatchAlong) and wait on with it only if it has not fired. The
watcher and its thresholds are unchanged, so a dim textured floor is still
no fade.

The overworld sim's walker watches for the fade from where the console
does (SimWalkerRules::watch_fade_along_hold), and fails on a warp whose fade
it missed, as the console does. OverworldWarpFade walks out of Viridian
Mart with a 200 ms fade that is over before the hold ends: seen when
watched from the start of the hold, missed from its end.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
replay_watcher() shows the hold's frames to the walk's BlackScreenOverWatcher
through a WatchAlong, then waits on with it, as walk_overworld_route() does
since its fade watch covers the hold; a fade over before the hold ends
reads "fade over during the hold".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine and others added 9 commits November 9, 2026 11:07
… player

Rock Tunnel 00.28 with a wild battle every 7 to 25 steps (sim) still lost
the open legs' progress after the first fix:
- Walk on: after an open leg stops, if the route's next holds bring every
  tile the player may be on to one tile (and none of those ways crosses a
  trainer's sight off the planned way), the walker walks on the rest of
  the leg from the middle candidate instead of finding the player at
  walls first. The player still faces the leg's way, so no turn tap.
  A stop on the walked-on leg counts from every tile it may start on.
- A leg onto a warp stopped by a wild battle, where no trainer may see the
  player, holds on to the warp as off the grass: below Rock Tunnel 1F's
  ladder (20, 13) the holds that found the player went down a column
  Hiker Lucas looks over.
- plan_reanchor() weighs open-leg tiles 4x over held ones, and puts ways
  whose holds cross a trainer's sight last.
- The sim looks for trainers before rolling for a wild battle, as the
  game does (pokefirered ProcessPlayerFieldInput()).

OverworldReanchor rules and two OverworldSim rule tests follow the new
choices.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
… walk on through open legs

The 00.28 sim with a wild battle every 7 to 25 steps now passes at all 12
seeds (before these changes: 0 of 12).
- A step into a trainer's line that sets off a wild battle took its
  battle screen for the trainer's, and the battle step after it found no
  message. The step now watches 5 s for a wild battle, runs from one seen
  with no message before it, and steps on (program and sim). The game
  looks for trainers before it rolls for a wild battle, so the trainer
  didn't see the player on that step.
- Walking on after a stop may go through open legs of the plan before the
  hold that finds the player (Rock Tunnel 1F row 31: west, then south 1,
  then the hold west to (12, 32)); a stop on any of them counts from every
  tile it may start on.
- A hold to a wall where wild battles start, stopped right away after
  another interrupt (Cooper, then a wild battle), is walked again.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
SimScript::encounter_rates rolls each step onto a wild tile as pokefirered
TryStandardWildEncounter() does (8 - rate / 10 cooldown steps at 5%, then
rate * 16 / 1600 plus a buff that grows with each roll lost; a battle
clears both), seeded per run. Rock Tunnel 1F/B1F use rate 7, Route 10 21
(wild_encounters.json). SimState counts wild battles and their time.

New test: 00.28 at those odds from 30 seeds must all pass, and logs the
runs' numbers. At the same seeds, before this PR's walker fixes 13/30
passed; now 30/30. Over passing runs, before -> after: time 1037 -> 912 s
mean (p90 1190 -> 1004), tiles 535 -> 406 (p90 679 -> 437), tiles walked
again 158 -> 3 (p90 302 -> 34), wild battles 22.6 -> 17.7, time lost per
battle 4.5 -> 0.3 s (p90 6.6 -> 2.0).

With trainers checked before wild battles in the sim, Cooper is always
fought: the 7-to-25-step test no longer allows him unfought.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Console B run 2 of 00.28 lost the one-tile step north onto B1F row 34
after a wild battle on row 38; the walk east "to the wall (17, 34)" then
ran along row 35 to the rock (20, 35) and martha_line found no trainer.
below_martha_line now ends at (16, 35) and martha_line walks up column 16
(or 17, a tile short) into her line; below_allen_line starts with the
walk east to the wall (17, 33). Sim test replays the console's battle and
the short step.

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

Console B lost the one-tile step north onto Rock Tunnel B1F row 34 in
00.28 run 2: the game missed the 40 ms turn tap, the leg's short hold then
only turned the player (pokefirered PlayerTurnInPlace(), 8 frames), and
the leg was taken as walked. OverworldWalkSteps::turn_to() taps, reads the
facing (read_player_facing()) and taps again while it shows another way,
up to 3 taps; an unreadable facing is taken as turned. Both walkers,
step_once() and face_wall() turn with it.

Sim: SimScript drops turn taps by number or at a rate; sim_turn_to()
mirrors turn_to(), switched by SimWalkerRules::confirm_turn. New tests:
a dropped tap before a one-tile leg (red without the rule), and 00.28 at
the game's odds with 1 in 10 turn taps dropped, 30 seeds.

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

Code health: OverworldSim kept at 7 public methods (sim_turn_to() reads the
state) and OverworldSim_Tests under 1000 lines (the turn tap cases moved to
OverworldWalk_Tests, as a table).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
A lab deployment without PokemonFRLG/PlayerSprites/red_normal.png failed
00.28 with a FileException; no test checked the files the readers load.
autostory_resource_paths() lists the files AutoStory and its readers and
detectors load, and the test fails naming each one missing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine and others added 9 commits November 11, 2026 17:57
Giovanni's script fades to black and removes him after his battle; the
next step's first tap, sent 0.4 s after the battle's text, was lost
(console B, 09.02 run 7). battle() now waits for a quiet overworld.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
09.02 run 8 (console B): "west 2" on Rocket Hideout B2F turned on its
first tap and showed no step on its second. The taps took that as a wall
ahead and kept only (17, 5), a tile from the player, so the legs after it
missed the arrow and the MOON STONE's talk found nothing. On foot, an open
leg's unseen step now stops at either tile and the walker finds the player
again; a leg onto a spinner that took its tiles counts its stop as a
candidate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
…ckpoint (red test)

Console B run 9 (as run 7) failed walking round the SILPH SCOPE to (21, 5)
right after Giovanni, with the ball adjacent east of the player on (19, 5).

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

09.02 now ends with Giovanni's battle and scene, the player left on (19, 5)
facing south. 09.03 starts there with its own save, turns east to the SILPH
SCOPE on (20, 5) and picks it up, instead of walking round it to (21, 5).

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

Folds 09.03 back into 09.02 as it was before the split: after Giovanni the
player turns east on (19, 5), picks up the SILPH SCOPE and uses an Escape
Rope out to Celadon (34, 22). Tests that named 09.03 now name 09.02.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
beside_giovanni tapped all 14 tiles from the right Grunt. Split it in
three: west_of_right_guard taps west 2 to (17, 15), up_column_17 runs
north 10 to the wall at (17, 5) (both Grunts beaten, the wall stops any
overshoot), and beside_giovanni taps east 2 to (19, 5), where his sight
starts. New sim test: the run leg ends on (17, 5) facing north.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
With 09.03 folded back into 09.02, chain_story_routes() still expected 40
checkpoints, so the golden and chain tests failed before planning a route.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Each tapped leg that is straight, has no arrow and no unbeaten trainer's
line on or beside it, and ends against a wall is now its own run step:
09.01 b4f_east_to_wall (east 3 to (6, 3)) and b3f_row_20_to_wall (west 5
to (10, 20)); 09.02 b3f_to_top_corridor (north 7 to (12, 2)) and
row_17_to_wall (east 15 to (30, 17)). The legs either side stay taps.
test_run_steps_to_walls() walks each run step, up_column_17 included, on
the sim to its wall, facing its way, with no battle.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine and others added 6 commits November 11, 2026 21:43
Conflict in cpp-ci-serial-programs-base.yml: the configure line keeps main's -DPA_SDL=ON and this branch's flags (no PA_FRLG_STORY_EXPORT).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Conflicts in the test registration: the base now registers shared suites through add_tests_AutoStory() with one header per suite, so FrameReplay gets Pokemon_FrameReplay_Tests.h (as in #307) and registers there; SourceFiles.cmake keeps both lists.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Conflict in Pokemon_AutoStory_Tests.h's includes: kept both, minus Pokemon_ScreenStory_Tests.h and Pokemon_StoryStep_Tests.h, which the base no longer has.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Brings in the facing reader that reads only the chosen player's sprite
sheet. This branch's walker reads the facing in more places (turn_to(),
the recovery walker, the Bicycle check, room legs, pacing), so each
takes the player's gender from StoryEnv::player_gender too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
dayfine and others added 8 commits November 12, 2026 07:01
Folded leaf PR #307; tree equals the leaf tip.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VUubhHeYSMGSSDXZkKZss
Conflicts: shared headers take autostory/engine's side (LegPlanner's tap_stop_risk, CheckpointRetry tests, StoryRunCounts in place of AutoStoryStats); FRLG screen events keep this branch's dark-cave check and middle-square test case on the shared CommonTools screen events and the nested Input/Expectation case shape.

Follow autostory/engine's move of the screen events to CommonTools/ScreenEvents (and SwitchHomeObserver to NintendoSwitch/Inference) in this branch's files.

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