WIP: feat(host): HOS_BOOT_SAVE starts the game in a save, with no title screen and no presses #1577

Draft
dyst wants to merge 11 commits from dyst/Longshot:feat/boot-save into main
Member

Summary

Test runs that start from a save have to press A through the title screen and the save list, and those presses race
the game's start (#659). A new test switch, HOS_BOOT_SAVE, starts the game straight in a save with no title screen
and no presses: the game takes the same path a Continue would and loads the save its own way. Nothing changes for
players: the switch is off unless set.

  • HOS_BOOT_SAVE=1 loads the newest save, =slot_NN that slot (00 to 05). No save, a Master Quest slot, or a slot
    without its caption.sav keeps the title.
  • When the game reads Sequence/NormalTitle.module.ainb, the host points the "a save exists" branch past the title
    screen at the step Continue → save card → Yes ends in, and once the graph starts it makes the save list's pick for
    the chosen slot. Everything after that is the game's own load.
  • The control bridge's status gains boot_save (off, waiting, picked, loaded, refused, failed), so a
    test driver knows when to press and when not to.
  • With the switch set there is no title for the rest of the run: Back to title loads the save again.

Read first: host/oot_boot_save.c (the switch, the start's state), host/oot_title_graph.c (the one link changed),
hos_oot_boot_save_write in host/hos_main.c (the pick).

See #659, #1539.

Testing

Tested: Windows 11 Pro N 22631, RTX 3080 (driver 621.14), Ryzen 9 3900X, from the repository save kokiri-start
through the control bridge, at 1786d810 (this head):

  • HOS_BOOT_SAVE=1: in play in Kokiri Forest with 0 presses and status.boot_save loaded; the picture is the forest
    5 s into play and after a 120-frame stick push, which moves Link 1.52 m, the same as the pressed start.
    Two such starts against two pressed ones (RX_VK_LOAD=create, warm): 14.5 and 13.7 s from the game's first frame to
    in play with 0 presses, against 18.6 and 17.2 s with 16 and 13 presses; all four moved Link 1.52 m and drew the
    forest (649 to 679 draws a frame).
  • An earlier head put the title's graph back as shipped once in play. That made the picture black and the stick dead
    (mooki's Linux run; reproduced here: 0 m moved, 3,197 draws a frame against 641). Its save branch's selector was
    still running, turned to the title screen, and held the picture and the input. This head no longer puts it back.
  • Checks: tests/test_boot_save.py (synthetic graphs: the one link changed and no other byte, refusals, every
    truncation, the slot choice, the read hook, the bridge state, the copies staying changed in play),
    tests/test_skip_title.py, tests/test_control.py, tests/test_switch_table.py, tests/test_doc_anchors.py,
    built with the Builder's llvm-mingw.

Still to test:

  • Linux, any GPU: start from the kokiri-start save with HOS_BOOT_SAVE=1 and no presses, expecting play in Kokiri
    Forest with status.boot_save loaded, the picture there and the stick moving Link; and the same save with the switch
    unset, expecting the title as before (@Beanstalk linux; the farm runs its nearest plan: pictures before and after a
    stick push with the switch on and off, the save-pick log line, no crash)

Measurements

Verdict: the switch's start is about 4 s faster than pressing through the title (14.1 s against 17.9 s from the
game's first frame to in play, medians), with 0 presses instead of 13 to 16, and ends in the same place every time.

Setup: Windows 11 Pro N 22631, RTX 3080 (621.14), Ryzen 9 3900X, 64 GB; kokiri-start, defaults, headless,
HOS_AUDIO=0, RX_VK_LOAD=create (warm caches, the same load mode for every run); one runtime, the switch set and
unset, 2 interleaved runs each through the job queue. Measured from the game's first frame (the bridge's state leaving
boot, after the pipelines are prepared) to in play (state play, a fresh snapshot, no event, no load, held 2 s), and
from launch. Unset: A pressed once a second until in play (the Test Farm agent's start), which is main's start: unset,
nothing of this PR runs.

Baseline: this PR's own build with the switch unset, which takes main's start (unset, nothing of this PR runs), so
both sides share one runtime and its caches.

Before: this PR @ 1786d810, switch unset, A pressed After: this PR @ 1786d810, HOS_BOOT_SAVE=1
first frame to in play (median of 2) 17.9 s (18.6, 17.2) 14.1 s (14.5, 13.7)
launch to in play (median of 2) 25.4 s (26.5, 24.2) 21.7 s (21.8, 21.5)
presses 16, 13 0, 0

Longer blocks agree: 5 runs of each back to back in timing mode, on an experiment build of this head's code (two more
graph edits in it, turned off by a variable): 25.6 s pressed (24.2 to 32.3) against 19.8 s with the switch (18.1 to
34.2). With those two edits on (the title's background scene and the opening's preload left out) the block took
20.4 s: no gain, so they are not in this PR.

Risks

Nothing for players: unset, nothing of this runs. Set, a game file read differently from this build (another version)
is left as it is and the title shows (boot_save refused); the pick is made only for a slot the caption manager has
a usable record of, and gives up after 90 s (failed). With the switch set, Back to title loads the save again: there
is no title in such a run. Turning it off: unset HOS_BOOT_SAVE. Beanstalk#91 (the Test Farm agent's start) uses the
switch under this name.

Checklist

Ready for review

  • One topic, on current main; the title is a Conventional Commit.

  • Tested on my own PC, in the game for a change the game runs, as Testing says.

  • mise run test passes (any failure that also fails on main is named). tests/run_all.py on Windows with the Builder's llvm-mingw at 1786d810: 706 passed, 60 failed, 124 skipped (no Vulkan headers on this PC, Linux-only); the 60 rerun on this branch and on its base afc4d1ad fail on both alike, none on the branch alone.

    The 60, failing on afc4d1ad too

    Auto-controlled, test_appimage_player_build.py, test_auto_handheld.py, test_bug_report_defaults.py, test_build_long_paths.py, test_capture_read_once.py, test_crash_dialog.py, test_crash_record.py, test_crt_options.py, test_damage_debug_menu.py, test_dev_ab.py, test_dev_build.py, test_dev_migrate.py, test_dev_packs.py, test_device_lost_exit.py, test_discover.py, test_display_startup_failure.py, test_dlss_feature_size.py, test_failed_compile_keeps_objects.py, test_flush_nothing_held.py, test_frame_overlap.py, test_game_menu_mouse_buttons.py, test_gpu_wait_stats.py, test_guest_imported.py, test_high_fps_fixes.py, test_host_guest_writes.py, test_hud_layout_row.py, test_job_queue.py, test_libc_hooks.py, test_lift_cache.py, test_lift_lookups.py, test_linux_preflight.py, test_linux_pressure.py, test_longshot_ctl.py, test_make_game_cache.py, test_make_game_ngx.py, test_mise_build.py, test_missing_nup.py, test_mods_list.py, test_ngx_sdk.py, test_pads_after_menu.py, test_pending_guest_import.py, test_pipeline_play_cache.py, test_prepare_cache_save.py, test_present_semaphores.py, test_profiling_mode.py, test_psychov35_fsr3_fg.py, test_readback_timeout_source.py, test_reflex_swapchain_lifetime.py, test_renderer_new_handler.py, test_renhdrum.py, test_sdiv_overflow.py, test_slow_frames.py, test_song_note_cue.py, test_sparse_mirror.py, test_upscaler_rows.py, test_vcore_escape.py, test_vcore_mask0.py, test_walk.py, test_walk_key.py

  • Measured against main, or Measurements says why there is nothing to measure.

  • No game files, build output or hand-edited generated files; third_party/ untouched; nothing GPL-incompatible.

When it applies

  • A switch added or changed: it ships at the default the measurements chose, with its row in gpu/vk/switches.def.
  • A check under tests/ that reads changed source is updated, not weakened.
  • Host code builds on Windows (llvm-mingw) and Linux, with UTF-8 paths. (Windows built and run; Linux built by mooki at fe0c661b, this head still to build: Still to test)
  • Docs updated.
  • AI assistance recorded with Assisted-by: / Co-Authored-By: trailers.

Ready to merge

  • Everything in Still to test has run, its results in Tested.
  • Tier 1 hardware it can affect is tested (docs/dev/standards.md, "Hardware tiers").
  • Approved in review, every comment answered.
  • CI passes on the head; on current main, measured again or a comment says why the numbers still hold.

🤖 Generated with Claude Code

## Summary Test runs that start from a save have to press A through the title screen and the save list, and those presses race the game's start (#659). A new test switch, `HOS_BOOT_SAVE`, starts the game straight in a save with no title screen and no presses: the game takes the same path a Continue would and loads the save its own way. Nothing changes for players: the switch is off unless set. - `HOS_BOOT_SAVE=1` loads the newest save, `=slot_NN` that slot (00 to 05). No save, a Master Quest slot, or a slot without its `caption.sav` keeps the title. - When the game reads `Sequence/NormalTitle.module.ainb`, the host points the "a save exists" branch past the title screen at the step Continue → save card → Yes ends in, and once the graph starts it makes the save list's pick for the chosen slot. Everything after that is the game's own load. - The control bridge's `status` gains `boot_save` (`off`, `waiting`, `picked`, `loaded`, `refused`, `failed`), so a test driver knows when to press and when not to. - With the switch set there is no title for the rest of the run: Back to title loads the save again. Read first: `host/oot_boot_save.c` (the switch, the start's state), `host/oot_title_graph.c` (the one link changed), `hos_oot_boot_save_write` in `host/hos_main.c` (the pick). See #659, #1539. ## Testing **Tested:** Windows 11 Pro N 22631, RTX 3080 (driver 621.14), Ryzen 9 3900X, from the repository save `kokiri-start` through the control bridge, at 1786d810 (this head): - `HOS_BOOT_SAVE=1`: in play in Kokiri Forest with 0 presses and `status.boot_save` `loaded`; the picture is the forest 5 s into play and after a 120-frame stick push, which moves Link 1.52 m, the same as the pressed start. Two such starts against two pressed ones (`RX_VK_LOAD=create`, warm): 14.5 and 13.7 s from the game's first frame to in play with 0 presses, against 18.6 and 17.2 s with 16 and 13 presses; all four moved Link 1.52 m and drew the forest (649 to 679 draws a frame). - An earlier head put the title's graph back as shipped once in play. That made the picture black and the stick dead (mooki's Linux run; reproduced here: 0 m moved, 3,197 draws a frame against 641). Its save branch's selector was still running, turned to the title screen, and held the picture and the input. This head no longer puts it back. - Checks: `tests/test_boot_save.py` (synthetic graphs: the one link changed and no other byte, refusals, every truncation, the slot choice, the read hook, the bridge state, the copies staying changed in play), `tests/test_skip_title.py`, `tests/test_control.py`, `tests/test_switch_table.py`, `tests/test_doc_anchors.py`, built with the Builder's llvm-mingw. **Still to test:** - Linux, any GPU: start from the `kokiri-start` save with `HOS_BOOT_SAVE=1` and no presses, expecting play in Kokiri Forest with `status.boot_save` `loaded`, the picture there and the stick moving Link; and the same save with the switch unset, expecting the title as before (`@Beanstalk linux`; the farm runs its nearest plan: pictures before and after a stick push with the switch on and off, the save-pick log line, no crash) ## Measurements **Verdict:** the switch's start is about 4 s faster than pressing through the title (14.1 s against 17.9 s from the game's first frame to in play, medians), with 0 presses instead of 13 to 16, and ends in the same place every time. Setup: Windows 11 Pro N 22631, RTX 3080 (621.14), Ryzen 9 3900X, 64 GB; `kokiri-start`, defaults, headless, `HOS_AUDIO=0`, `RX_VK_LOAD=create` (warm caches, the same load mode for every run); one runtime, the switch set and unset, 2 interleaved runs each through the job queue. Measured from the game's first frame (the bridge's state leaving `boot`, after the pipelines are prepared) to in play (state `play`, a fresh snapshot, no event, no load, held 2 s), and from launch. Unset: A pressed once a second until in play (the Test Farm agent's start), which is main's start: unset, nothing of this PR runs. Baseline: this PR's own build with the switch unset, which takes main's start (unset, nothing of this PR runs), so both sides share one runtime and its caches. | | Before: this PR @ 1786d810, switch unset, A pressed | After: this PR @ 1786d810, `HOS_BOOT_SAVE=1` | |---|---|---| | first frame to in play (median of 2) | 17.9 s (18.6, 17.2) | 14.1 s (14.5, 13.7) | | launch to in play (median of 2) | 25.4 s (26.5, 24.2) | 21.7 s (21.8, 21.5) | | presses | 16, 13 | 0, 0 | Longer blocks agree: 5 runs of each back to back in timing mode, on an experiment build of this head's code (two more graph edits in it, turned off by a variable): 25.6 s pressed (24.2 to 32.3) against 19.8 s with the switch (18.1 to 34.2). With those two edits on (the title's background scene and the opening's preload left out) the block took 20.4 s: no gain, so they are not in this PR. ## Risks Nothing for players: unset, nothing of this runs. Set, a game file read differently from this build (another version) is left as it is and the title shows (`boot_save` `refused`); the pick is made only for a slot the caption manager has a usable record of, and gives up after 90 s (`failed`). With the switch set, Back to title loads the save again: there is no title in such a run. Turning it off: unset `HOS_BOOT_SAVE`. Beanstalk#91 (the Test Farm agent's start) uses the switch under this name. ## Checklist ### Ready for review - [x] One topic, on current `main`; the title is a Conventional Commit. - [x] Tested on my own PC, in the game for a change the game runs, as Testing says. - [x] `mise run test` passes (any failure that also fails on `main` is named). `tests/run_all.py` on Windows with the Builder's llvm-mingw at 1786d810: 706 passed, 60 failed, 124 skipped (no Vulkan headers on this PC, Linux-only); the 60 rerun on this branch and on its base `afc4d1ad` fail on both alike, none on the branch alone. <details><summary>The 60, failing on afc4d1ad too</summary> `Auto-controlled`, `test_appimage_player_build.py`, `test_auto_handheld.py`, `test_bug_report_defaults.py`, `test_build_long_paths.py`, `test_capture_read_once.py`, `test_crash_dialog.py`, `test_crash_record.py`, `test_crt_options.py`, `test_damage_debug_menu.py`, `test_dev_ab.py`, `test_dev_build.py`, `test_dev_migrate.py`, `test_dev_packs.py`, `test_device_lost_exit.py`, `test_discover.py`, `test_display_startup_failure.py`, `test_dlss_feature_size.py`, `test_failed_compile_keeps_objects.py`, `test_flush_nothing_held.py`, `test_frame_overlap.py`, `test_game_menu_mouse_buttons.py`, `test_gpu_wait_stats.py`, `test_guest_imported.py`, `test_high_fps_fixes.py`, `test_host_guest_writes.py`, `test_hud_layout_row.py`, `test_job_queue.py`, `test_libc_hooks.py`, `test_lift_cache.py`, `test_lift_lookups.py`, `test_linux_preflight.py`, `test_linux_pressure.py`, `test_longshot_ctl.py`, `test_make_game_cache.py`, `test_make_game_ngx.py`, `test_mise_build.py`, `test_missing_nup.py`, `test_mods_list.py`, `test_ngx_sdk.py`, `test_pads_after_menu.py`, `test_pending_guest_import.py`, `test_pipeline_play_cache.py`, `test_prepare_cache_save.py`, `test_present_semaphores.py`, `test_profiling_mode.py`, `test_psychov35_fsr3_fg.py`, `test_readback_timeout_source.py`, `test_reflex_swapchain_lifetime.py`, `test_renderer_new_handler.py`, `test_renhdrum.py`, `test_sdiv_overflow.py`, `test_slow_frames.py`, `test_song_note_cue.py`, `test_sparse_mirror.py`, `test_upscaler_rows.py`, `test_vcore_escape.py`, `test_vcore_mask0.py`, `test_walk.py`, `test_walk_key.py` </details> - [x] Measured against `main`, or Measurements says why there is nothing to measure. - [x] No game files, build output or hand-edited generated files; `third_party/` untouched; nothing GPL-incompatible. ### When it applies - [x] A switch added or changed: it ships at the default the measurements chose, with its row in `gpu/vk/switches.def`. - [x] A check under `tests/` that reads changed source is updated, not weakened. - [ ] Host code builds on Windows (llvm-mingw) and Linux, with UTF-8 paths. (Windows built and run; Linux built by mooki at fe0c661b, this head still to build: Still to test) - [x] Docs updated. - [x] AI assistance recorded with `Assisted-by:` / `Co-Authored-By:` trailers. ### Ready to merge - [ ] Everything in Still to test has run, its results in Tested. - [ ] Tier 1 hardware it can affect is tested (docs/dev/standards.md, "Hardware tiers"). - [ ] Approved in review, every comment answered. - [ ] CI passes on the head; on current `main`, measured again or a comment says why the numbers still hold. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
A switch for test and development runs. =1 starts in the newest save, =slot_NN in that slot (00 to 05).

The title is a script graph, Sequence/NormalTitle.module.ainb. When the game reads it (host/nn_fs.c, read_at, a
whole read only), host/oot_title_graph.c points its save branch, the BoolSelector fed by SeqQueryIsExistSave, past the
title screen at the Sequential that Continue, a save card and Yes end in (the one running TitleToMainWithLoadSave).
Both are found by kind, name and links; when either is not found exactly once the file is left as it is. Once the
graph has started (UI_Debug_IsOpenedTitle), hos_main.c puts the slot where a pick on the save list puts it
(CaptionData.SelectedSavePathIndex, as SaveCard_Pick does) and the game loads the save its own way.

A game with no save, a Master Quest slot, a missing slot or a graph that cannot be changed keeps its title. The Skip
title row's presses stand down while the switch carries the start, and start as ever when it lets go.

Checks: tests/test_boot_save.py (synthetic graphs: the one link, refusals, every truncation; the slot choice; the
read hook and the tick), and tests/test_skip_title.py's new case.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
"off", "waiting", "picked", "loaded", "refused" or "failed", after "game", so a test's driver waits with no presses
while the switch carries the start and presses through the title as before when it does not (or the build is older
and has no such field). The state chooses nothing itself: the bridge may ask before the host knows its user folder.

Checks: tests/control_test.c's status, tests/test_boot_save.py's states.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Should the game read Sequence/NormalTitle.module.ainb some other way than whole through the host's file reads, the
graph starts as it always does and the title shows: once UI_Debug_IsOpenedTitle is set with the file unchanged, the
start is "refused", so the Skip title row and a test's driver press through the title rather than wait for a load
that will not come.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
A later read of Sequence/NormalTitle.module.ainb (a title the game opens again, Back to title) is left as it is, so
that title is the title rather than the save loaded again.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
docs(tested): HOS_OOT_BOOT_SAVE's start against the pressed one, RTX 3080 on Windows
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 7s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 10s
CI / Workflow lint (actionlint) (pull_request) Successful in 10s
Third-party parts / As their notes say (pull_request) Successful in 11s
Website / Website on Cloudflare (pull_request_target) Has been skipped
PR description / Description follows the template (pull_request_target) Successful in 5s
beanstalk/approval approved by dyst (reviewed)
beanstalk/linux 0/1 pass, 2 running, 0 queued
aa1948c70b
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Author
Member

Tested in game: RTX 3080, driver 621.14, Windows 11: the kokiri-start repository save through the control bridge with HOS_OOT_BOOT_SAVE=1, 4 of 4 starts in play in Kokiri Forest with no presses and no title screen (status.boot_save waiting, picked, loaded); 3 interleaved pressed starts with it unset ended at the same position, facing, hearts and rupees

Tested in game: RTX 3080, driver 621.14, Windows 11: the kokiri-start repository save through the control bridge with HOS_OOT_BOOT_SAVE=1, 4 of 4 starts in play in Kokiri Forest with no presses and no title screen (status.boot_save waiting, picked, loaded); 3 interleaved pressed starts with it unset ended at the same position, facing, hearts and rupees
Member

Possibly related, for context:

  • issue #587 Skip title screen should start way early and quickly — Related title-screen startup behavior; asks for faster automatic loading.
  • PR #399 feat(menu): skip the title screen and load the newest save — Merged implementation of the player-facing Skip title screen behavior this builds on.
  • issue #1198 Idea: an automated test player that plays sections of the game and hunts for bugs — Automated test-player work that would benefit from scripted save startup.
Possibly related, for context: - issue #587 Skip title screen should start way early and quickly — Related title-screen startup behavior; asks for faster automatic loading. - PR #399 feat(menu): skip the title screen and load the newest save — Merged implementation of the player-facing Skip title screen behavior this builds on. - issue #1198 Idea: an automated test player that plays sections of the game and hunts for bugs — Automated test-player work that would benefit from scripted save startup.
Author
Member

Retested at aa1948c7 (the head: the once-a-run change and the run file now in the runtime). The Test Farm agent's start from MagicBeanDev/Beanstalk#91 (HOS_OOT_BOOT_SAVE=slot_00, A only while status.boot_save does not carry the start), 2 runs on each side of an A/B over base 310e7c9d: this PR 0 presses, boot_save loaded; the base alone no boot_save, 16 and 12 presses. All four in play in Kokiri Forest at the same position and facing. RTX 3080, 621.14, Windows 11.

Retested at aa1948c7 (the head: the once-a-run change and the run file now in the runtime). The Test Farm agent's start from MagicBeanDev/Beanstalk#91 (HOS_OOT_BOOT_SAVE=slot_00, A only while status.boot_save does not carry the start), 2 runs on each side of an A/B over base 310e7c9d: this PR 0 presses, boot_save loaded; the base alone no boot_save, 16 and 12 presses. All four in play in Kokiri Forest at the same position and facing. RTX 3080, 621.14, Windows 11.
Author
Member

@Beanstalk linux (any GPU): start from the kokiri-start repository save with HOS_OOT_BOOT_SAVE=1 and no presses, expecting play in Kokiri Forest with status.boot_save "loaded"; and the same save with the switch unset, expecting the title as before. This PR's Linux build has not run anywhere yet.

@Beanstalk linux (any GPU): start from the kokiri-start repository save with HOS_OOT_BOOT_SAVE=1 and no presses, expecting play in Kokiri Forest with status.boot_save "loaded"; and the same save with the switch unset, expecting the title as before. This PR's Linux build has not run anywhere yet.
Member
Beanstalk test farm · an earlier run: commit aa1948c (the newer results are in a comment below)

Commit aa1948c · tier reviewed · approved by dyst · dashboard

What will be tested · AI plan (gpt-6.1-terra)

On Linux, the plan compares the kokiri-start save in Kokiri Forest with the new boot-save switch enabled on the PR, and checks the title screen with the switch unset. A failed gameplay or title-state check, crash, or substantial title-screen image difference fails the plan. The runs take about 22 minutes per tester. The farm schema cannot assert status.boot_save or guarantee a no-press save start, and its title start cannot supply the repository save, so those requested checks remain unverified.

  • Where: linux (1-2 reports).
  • Matrix: 1 cell, each on main and the PR: HOS_OOT_BOOT_SAVE 1 (PR only).
  • kokiri-save: from the repository's save kokiri-start, then take the picture kokiri-save; 1 run on each side.
  • title-unset: from the title screen, then take the picture title to compare with main; 1 run on each side.
  • Measured: crashes.
  • The PR fails if: it crashes; or a picture differs from main by more than 10%.
  • Time: about 20 min of game runs per tester (4 game starts), plus the two builds.

Started at 11:43 UTC.

Estimate: done in about 40 min, from the queue and the testers online now.

Target Reports Crash Frame p99 (ms) vs main VRAM peak (MB) vs main Result
linux 0/1 — — — running
<!-- beanstalk:previous --> <details><summary>Beanstalk test farm · an earlier run: commit <code>aa1948c</code> (the newer results are in a comment below)</summary> Commit `aa1948c` · tier **reviewed** · approved by dyst · [dashboard](https://beanstalk.magicbean.dev/prs/1577) **What will be tested** · AI plan (`gpt-6.1-terra`) On Linux, the plan compares the kokiri-start save in Kokiri Forest with the new boot-save switch enabled on the PR, and checks the title screen with the switch unset. A failed gameplay or title-state check, crash, or substantial title-screen image difference fails the plan. The runs take about 22 minutes per tester. The farm schema cannot assert status.boot_save or guarantee a no-press save start, and its title start cannot supply the repository save, so those requested checks remain unverified. - **Where:** linux (1-2 reports). - **Matrix:** 1 cell, each on main and the PR: HOS_OOT_BOOT_SAVE `1` (PR only). - **kokiri-save:** from the repository's save `kokiri-start`, then take the picture kokiri-save; 1 run on each side. - **title-unset:** from the title screen, then take the picture title to compare with main; 1 run on each side. - **Measured:** crashes. - **The PR fails if:** it crashes; or a picture differs from main by more than 10%. - **Time:** about 20 min of game runs per tester (4 game starts), plus the two builds. Started at 11:43 UTC. **Estimate:** done in about 40 min, from the queue and the testers online now. | Target | Reports | Crash | Frame p99 (ms) vs main | VRAM peak (MB) vs main | Result | | --- | --- | --- | --- | --- | --- | | linux | 0/1 | — | — | — | running | </details>
fix(host): HOS_OOT_BOOT_SAVE puts the title's save branch back once the save is in play
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 9s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 14s
CI / Workflow lint (actionlint) (pull_request) Successful in 15s
Third-party parts / As their notes say (pull_request) Successful in 16s
beanstalk/approval not approved yet
PR description / Description follows the template (pull_request_target) Successful in 5s
6e06faff50
The game keeps the title's graph it read at start-up, so a later title from that copy (Back to title) would have
loaded the save again. Once the player is seen, the save branch is put back on the title screen's node in the bytes
the game keeps, and only when those bytes are still that graph with the change in it (hos_oot_title_graph_unpatch:
both nodes found exactly once, the branch on the load): memory the game has since used for something else is never
written.

Checked in the game: "boot save: in play: the title's graph is as the game shipped it again", then play as before.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Author
Member

6e06faff: once the save is in play, the title graph's save branch is put back in the copy the game keeps (only if those bytes are still that graph with the change in it), so a later Back to title is the title. In the game (RTX 3080, Windows 11): 'boot save: in play: the title's graph is as the game shipped it again', in play in Kokiri Forest with 0 presses. One other run of that pair never drew a frame within my driver's 600 s: the freshly copied runtime was still preparing its pipelines (18422 of 34822 at the cut-off), before any of this runs; the next run, with the preparation kept, got in. The Test Farm's agent allows 900 s for that.

6e06faff: once the save is in play, the title graph's save branch is put back in the copy the game keeps (only if those bytes are still that graph with the change in it), so a later Back to title is the title. In the game (RTX 3080, Windows 11): 'boot save: in play: the title's graph is as the game shipped it again', in play in Kokiri Forest with 0 presses. One other run of that pair never drew a frame within my driver's 600 s: the freshly copied runtime was still preparing its pipelines (18422 of 34822 at the cut-off), before any of this runs; the next run, with the preparation kept, got in. The Test Farm's agent allows 900 s for that.
fix(host): HOS_OOT_BOOT_SAVE after its cold review: a slot the list can pick, give-ups, every copy, kicks kept
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 7s
beanstalk/approval not approved yet
PR description / Description follows the template (pull_request_target) Successful in 9s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 10s
CI / Workflow lint (actionlint) (pull_request) Successful in 12s
Third-party parts / As their notes say (pull_request) Successful in 12s
dd40ea31ca
- A slot is chosen only with its caption.sav too: the save list picks by the caption, and once the title's graph is
  changed there is no title to fall back to.
- Give-ups: the graph not read whole within 90 s of the game's first controller read is "refused" (the title shows,
  presses take over); the pick not made within 90 s of the change is "failed". The game reads no controller while it
  prepares its pipelines, so a long preparation is not counted.
- Every whole read of the graph is changed until the pick is made, and each copy changed is put back once in play
  (a list of them, under a lock); a file that came changed (a mod's) is not ours to put back: patch returns 2 and
  leaves title_node alone, so a node 0 is never written into it.
- Shared state is atomic, the slot chosen by compare-exchange, the graphs' list published with release/acquire.
- oot_ctx_restore: the Ctx put back after calling the game keeps a kick the core scheduler set meanwhile (poll), in
  oot_gamedata_bool and hos_oot_boot_save_write, which now run at every controller read while the start waits.
- A blackboard not made yet is "not yet", not "never"; stubs for builds other than Windows and Linux.
- docs/control.md: on "failed" there is no title to press through; a driver ends the run.
- Checks: two True links, no False, another query, not a selector; slot_001; a slot without caption.sav; the state
  unset before anything chose; a file that came changed; a second copy changed and put back; both give-ups; the
  path's case and separators.

In the game, a six-slot save: slot_02, slot_00, slot_05 and 1 each loaded their own save with no presses
(docs/tested/2026-11-18-boot-save-dyst-run02.md).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
fix(host): HOS_OOT_BOOT_SAVE's start as one state word, the pick's give-up from the title's start
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 8s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 10s
CI / Workflow lint (actionlint) (pull_request) Successful in 11s
Third-party parts / As their notes say (pull_request) Successful in 12s
beanstalk/approval not approved yet
PR description / Description follows the template (pull_request_target) Successful in 4s
fe0c661bee
From the second cold review:
- the pick's 90 s count from the first tick that sees the title's graph started, not from the file's read: a slow
  title's setup (a Steam Deck preparing pipelines) no longer turns a good start into "failed";
- one state word moved by compare-exchange (none, changed, loaded, failed, refused): a change that loses the race
  with a refusal is undone at once, and the bridge never sees a passing "loaded";
- more copies of the graph than can be put back are left as they are;
- off on platforms other than Windows and Linux, where the game's side is not built (the stubs are for the link only);
- checks: a node 0 that is a Sequential (a wrong put-back would show), the fifth copy, the pick's give-up counted
  from the title's start, oot_gamedata_bool keeping the scheduler's kick.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Author
Member

Two cold reviews by a second agent, and their fixes (dd40ea31, fe0c661b):

  • the pick is proven in the game: a six-slot save, slot_02/slot_00/slot_05/1 each loaded its own save (Hyrule Field, Zora's Domain, Jabu-Jabu's Belly, the newest) with 0 presses;
  • a slot needs its caption.sav (the list picks by it); a failed pick is 'failed' (no title: a driver ends the run), documented so;
  • 90 s give-ups: the graph not read within 90 s of the first controller read -> refused (title shows), the pick not made within 90 s of the title's start -> failed; preparing pipelines is not counted (no controller reads then);
  • one state word moved by compare-exchange; every copy of the graph changed until the pick, each put back in play; a file that came changed is never put back; a fifth copy is left alone;
  • oot_ctx_restore keeps the core scheduler's kick (poll) in oot_gamedata_bool and the pick; off outside Windows and Linux.
    Checks catch the reviewers' mutations (tried: the length check, a changed-on-arrival file recorded, the give-up counted from the wrong moment). run_all on dd40ea31: no failure main 430eef4f does not also have on this PC. At fe0c661b in the game: 3 of 3 starts in play with 0 presses (RTX 3080, Windows 11). Pre-existing note: seven other functions in hos_main.c still copy the whole Ctx back and can drop a kick; left for its own PR.
Two cold reviews by a second agent, and their fixes (dd40ea31, fe0c661b): - the pick is proven in the game: a six-slot save, slot_02/slot_00/slot_05/1 each loaded its own save (Hyrule Field, Zora's Domain, Jabu-Jabu's Belly, the newest) with 0 presses; - a slot needs its caption.sav (the list picks by it); a failed pick is 'failed' (no title: a driver ends the run), documented so; - 90 s give-ups: the graph not read within 90 s of the first controller read -> refused (title shows), the pick not made within 90 s of the title's start -> failed; preparing pipelines is not counted (no controller reads then); - one state word moved by compare-exchange; every copy of the graph changed until the pick, each put back in play; a file that came changed is never put back; a fifth copy is left alone; - oot_ctx_restore keeps the core scheduler's kick (poll) in oot_gamedata_bool and the pick; off outside Windows and Linux. Checks catch the reviewers' mutations (tried: the length check, a changed-on-arrival file recorded, the give-up counted from the wrong moment). run_all on dd40ea31: no failure main 430eef4f does not also have on this PC. At fe0c661b in the game: 3 of 3 starts in play with 0 presses (RTX 3080, Windows 11). Pre-existing note: seven other functions in hos_main.c still copy the whole Ctx back and can drop a kick; left for its own PR.
Member

Linux run of this head (fe0c661be), the open item in the description. The host builds on Linux from this head as it stands (on its own base, not rebased, no errors). The switch picks the right save every time, with no presses; after that the picture is black and the stick does nothing, details below.

What was run

  • Linux (Arch based), kernel 7.2.15, NVIDIA RTX 4090 with driver 615.78.08, windowless through the workbench's run wrapper, one game at a time, the control bridge polled about once a second. The machine had other jobs running (load average 25 to 32), so the times are observations, not a benchmark.
  • Saves: a six-slot folder made by a test script (six different saves; slot_03's progress.sav is the newest file), and the repository's kokiri-start save alone in a fresh profile for the check the description asks for. The Skip title row was off unless stated. HOS_FS_TRACE=1 was on so the log shows each save file the game opens.
  • Per run: the bridge's map, age, hearts and rupees at the first frame it calls controllable (state play, a fresh game snapshot, no event, no load) against the save's own flags; the last progress.sav the game opened is the asked slot's; status.boot_save; the host's presses (no skip title: pressing line, the row off).

The switch picks the right save

Setting Save Map, age, hearts, rupees equal the save's flags Last progress.sav the game opened Host presses status.boot_save at that frame Uptime at the first controllable frame (s) From the first frames to that point (s)
HOS_OOT_BOOT_SAVE=slot_00 slot_00 yes slot_00 0 loaded 79.4 10.2
HOS_OOT_BOOT_SAVE=slot_01 slot_01 yes slot_01 0 loaded 17.3 10.3
HOS_OOT_BOOT_SAVE=slot_02 slot_02 yes slot_02 0 loaded 18.6 11.5
HOS_OOT_BOOT_SAVE=slot_03 slot_03 yes slot_03 0 loaded 16.0 8.9
HOS_OOT_BOOT_SAVE=slot_04 slot_04 yes slot_04 0 loaded 17.5 10.4
HOS_OOT_BOOT_SAVE=slot_05 slot_05 yes slot_05 0 loaded 17.2 10.2
HOS_OOT_BOOT_SAVE=1 slot_03 yes slot_03 0 loaded 17.3 10.3

7 of 7 loaded their own save, 0 presses, boot_save loaded each time. The first run on a fresh copy took 79.4 s to the first controllable frame (a slow first start); the other six took 16.0 to 18.6 s (median 17.3), 8.9 to 11.5 s after the first status past boot. For the description's check, HOS_OOT_BOOT_SAVE=1 with the kokiri-start save alone: status.boot_save loaded, Kokiri Forest, child, 3 hearts as that save's README says, no presses, first controllable frame at 14.7 s (and 13.2 s in a second run).

What did not match: the picture is black and Link does not answer input

  • Every picture the bridge's shot took with the switch on was black (mean grey 0): at the first controllable frame and at 5 s in all 7 six-slot runs, and at 10, 20 and 40 s in slot_02 and in kokiri-start. The bridge said play, the frame counter kept advancing, and the log shows the scene being drawn (about 2,240 draws a frame in the slot_02 run, about 690 when the same save is loaded by another route). The same build with no switch and the Skip title row on shows its picture as usual (mean grey 0.53), so the capture works there.
  • The bridge's stick pushed forward for 120 frames moved Link 0 units, 5 s after the first controllable frame (slot_02) and 45 s after (slot_02 and kokiri-start). On the same build with no switch and the row on, the same push moved him 11.9 units; with another route that loads slot_02 it moved him 13.0 units.
  • Pressing A for 8 frames and pushing the stick forward again for 120 frames, in a run of its own (slot_02): Link did not move, no event or load began, and the pictures stayed black.
  • So by the bridge's own criterion the game is controllable at that frame, but on this machine nothing is on screen and the stick does nothing. We have not looked into why, and we do not know whether it is Linux only. The runs' logs and pictures are available if they help.

No switch, and refused starts

Setting Result status.boot_save
no switch, row off title stays (state menu, no game at 39.5 s) off
no switch, row on the row loads a save as before off
=slot_02, row on slot_02 loaded, log says the row's presses are not wanted, first controllable frame at 19.8 s loaded
=slot_03, in a folder with slots 00 to 02 and slot_04 only (no slot_03 folder) log: no save there, title stays (39.8 s) off
=slot_04, in that folder (its slot_04 is an empty progress.sav with no caption.sav) log: no caption.sav, title stays off
=slot_06, in the six-slot folder (no such folder) log: no save there, title stays off
=2 log: neither 1 nor slot_NN, title stays off

boot_save reads off for a refused request as it does for no request, so only the log says why. Not tried: a Master Quest slot that exists, the graph left unchanged (refused), a failed pick (failed), Back to title after the start, co-op rooms.

One note on =1: in the same six-slot folder the Skip title row alone (no switch) logged slot_03 as the most recent save, and the game loaded slot_01's save (by its map), while =1 loaded slot_03's.

--load-slot, if you want it

We have a --load-slot <0 to 5|newest> argument (and HOS_LOAD_SLOT) in a branch we have not opened a PR for. It reaches the same point another way: it leaves the title's graph alone, finishes the title's waits and presses A once at the save list, so it needs one press where yours needs none. It refuses a number outside 0 to 5, a non-number, a missing or empty slot (log line, title stays), and the Skip title row stands aside. On the same six saves and the same checks: 7 of 7 right, 1 A press each, first controllable frame 14.5 to 18.3 s (median 15.6; the first run 61.4 s) on this busy machine, so no speed claim in either direction. Its pictures show the scene within 5 s of the first controllable frame (3 of the 7 first-frame pictures are still the load's fade), and the same stick push moves Link in slot_02.

If you would like it, we could offer it as a thin layer on your switch instead: --load-slot N and newest would only choose the slot, give the refusals above and hand the start to HOS_OOT_BOOT_SAVE, and we would drop our own route rather than keep two. Only if you want that; the switch is yours and we are not asking you to change it.

Overlap with our PR 61 (restart in place)

Both change host/oot_autoload.c. Git merges PR 61 with this head with no text conflict there. In PR 61, read_input in host/nn_hid.c calls hos_oot_autoload_input() only while a demand word is armed, and the word is armed only for a restart or the Skip title row. The switch's tick (the pick, the give-ups, putting the graph back) is called from that function, so with PR 61 and the row off it would not run. This is from reading the code; we have not built the two together. Whichever lands second adds one arming condition for the switch where the word is armed and keeps it armed until the start is over. We will make that change on our side if PR 61 lands second.

Linux run of this head (fe0c661be), the open item in the description. The host builds on Linux from this head as it stands (on its own base, not rebased, no errors). The switch picks the right save every time, with no presses; after that the picture is black and the stick does nothing, details below. **What was run** - Linux (Arch based), kernel 7.2.15, NVIDIA RTX 4090 with driver 615.78.08, windowless through the workbench's run wrapper, one game at a time, the control bridge polled about once a second. The machine had other jobs running (load average 25 to 32), so the times are observations, not a benchmark. - Saves: a six-slot folder made by a test script (six different saves; slot_03's progress.sav is the newest file), and the repository's kokiri-start save alone in a fresh profile for the check the description asks for. The Skip title row was off unless stated. `HOS_FS_TRACE=1` was on so the log shows each save file the game opens. - Per run: the bridge's map, age, hearts and rupees at the first frame it calls controllable (state `play`, a fresh game snapshot, no event, no load) against the save's own flags; the last `progress.sav` the game opened is the asked slot's; `status.boot_save`; the host's presses (no `skip title: pressing` line, the row off). **The switch picks the right save** | Setting | Save | Map, age, hearts, rupees equal the save's flags | Last progress.sav the game opened | Host presses | `status.boot_save` at that frame | Uptime at the first controllable frame (s) | From the first frames to that point (s) | |---|---|---|---|---|---|---|---| | `HOS_OOT_BOOT_SAVE=slot_00` | slot_00 | yes | slot_00 | 0 | loaded | 79.4 | 10.2 | | `HOS_OOT_BOOT_SAVE=slot_01` | slot_01 | yes | slot_01 | 0 | loaded | 17.3 | 10.3 | | `HOS_OOT_BOOT_SAVE=slot_02` | slot_02 | yes | slot_02 | 0 | loaded | 18.6 | 11.5 | | `HOS_OOT_BOOT_SAVE=slot_03` | slot_03 | yes | slot_03 | 0 | loaded | 16.0 | 8.9 | | `HOS_OOT_BOOT_SAVE=slot_04` | slot_04 | yes | slot_04 | 0 | loaded | 17.5 | 10.4 | | `HOS_OOT_BOOT_SAVE=slot_05` | slot_05 | yes | slot_05 | 0 | loaded | 17.2 | 10.2 | | `HOS_OOT_BOOT_SAVE=1` | slot_03 | yes | slot_03 | 0 | loaded | 17.3 | 10.3 | 7 of 7 loaded their own save, 0 presses, `boot_save` `loaded` each time. The first run on a fresh copy took 79.4 s to the first controllable frame (a slow first start); the other six took 16.0 to 18.6 s (median 17.3), 8.9 to 11.5 s after the first status past `boot`. For the description's check, `HOS_OOT_BOOT_SAVE=1` with the kokiri-start save alone: `status.boot_save` `loaded`, Kokiri Forest, child, 3 hearts as that save's README says, no presses, first controllable frame at 14.7 s (and 13.2 s in a second run). **What did not match: the picture is black and Link does not answer input** - Every picture the bridge's `shot` took with the switch on was black (mean grey 0): at the first controllable frame and at 5 s in all 7 six-slot runs, and at 10, 20 and 40 s in slot_02 and in kokiri-start. The bridge said `play`, the frame counter kept advancing, and the log shows the scene being drawn (about 2,240 draws a frame in the slot_02 run, about 690 when the same save is loaded by another route). The same build with no switch and the Skip title row on shows its picture as usual (mean grey 0.53), so the capture works there. - The bridge's `stick` pushed forward for 120 frames moved Link 0 units, 5 s after the first controllable frame (slot_02) and 45 s after (slot_02 and kokiri-start). On the same build with no switch and the row on, the same push moved him 11.9 units; with another route that loads slot_02 it moved him 13.0 units. - Pressing A for 8 frames and pushing the stick forward again for 120 frames, in a run of its own (slot_02): Link did not move, no event or load began, and the pictures stayed black. - So by the bridge's own criterion the game is controllable at that frame, but on this machine nothing is on screen and the stick does nothing. We have not looked into why, and we do not know whether it is Linux only. The runs' logs and pictures are available if they help. **No switch, and refused starts** | Setting | Result | `status.boot_save` | |---|---|---| | no switch, row off | title stays (state `menu`, no game at 39.5 s) | `off` | | no switch, row on | the row loads a save as before | `off` | | `=slot_02`, row on | slot_02 loaded, log says the row's presses are not wanted, first controllable frame at 19.8 s | `loaded` | | `=slot_03`, in a folder with slots 00 to 02 and slot_04 only (no slot_03 folder) | log: no save there, title stays (39.8 s) | `off` | | `=slot_04`, in that folder (its slot_04 is an empty progress.sav with no caption.sav) | log: no caption.sav, title stays | `off` | | `=slot_06`, in the six-slot folder (no such folder) | log: no save there, title stays | `off` | | `=2` | log: neither 1 nor slot_NN, title stays | `off` | `boot_save` reads `off` for a refused request as it does for no request, so only the log says why. Not tried: a Master Quest slot that exists, the graph left unchanged (`refused`), a failed pick (`failed`), Back to title after the start, co-op rooms. One note on `=1`: in the same six-slot folder the Skip title row alone (no switch) logged slot_03 as the most recent save, and the game loaded slot_01's save (by its map), while `=1` loaded slot_03's. **`--load-slot`, if you want it** We have a `--load-slot <0 to 5|newest>` argument (and `HOS_LOAD_SLOT`) in a branch we have not opened a PR for. It reaches the same point another way: it leaves the title's graph alone, finishes the title's waits and presses A once at the save list, so it needs one press where yours needs none. It refuses a number outside 0 to 5, a non-number, a missing or empty slot (log line, title stays), and the Skip title row stands aside. On the same six saves and the same checks: 7 of 7 right, 1 A press each, first controllable frame 14.5 to 18.3 s (median 15.6; the first run 61.4 s) on this busy machine, so no speed claim in either direction. Its pictures show the scene within 5 s of the first controllable frame (3 of the 7 first-frame pictures are still the load's fade), and the same stick push moves Link in slot_02. If you would like it, we could offer it as a thin layer on your switch instead: `--load-slot N` and `newest` would only choose the slot, give the refusals above and hand the start to `HOS_OOT_BOOT_SAVE`, and we would drop our own route rather than keep two. Only if you want that; the switch is yours and we are not asking you to change it. **Overlap with our PR 61 (restart in place)** Both change `host/oot_autoload.c`. Git merges PR 61 with this head with no text conflict there. In PR 61, `read_input` in `host/nn_hid.c` calls `hos_oot_autoload_input()` only while a demand word is armed, and the word is armed only for a restart or the Skip title row. The switch's tick (the pick, the give-ups, putting the graph back) is called from that function, so with PR 61 and the row off it would not run. This is from reading the code; we have not built the two together. Whichever lands second adds one arming condition for the switch where the word is armed and keeps it armed until the start is over. We will make that change on our side if PR 61 lands second.
dyst force-pushed feat/boot-save from fe0c661bee
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 8s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 10s
CI / Workflow lint (actionlint) (pull_request) Successful in 11s
Third-party parts / As their notes say (pull_request) Successful in 12s
beanstalk/approval not approved yet
PR description / Description follows the template (pull_request_target) Successful in 4s
to 367f9bd5df
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
beanstalk/approval not approved yet
CI / Conventional Commit title (pull_request) Successful in 8s
PR description / Description follows the template (pull_request_target) Successful in 12s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 13s
CI / Workflow lint (actionlint) (pull_request) Successful in 13s
Third-party parts / As their notes say (pull_request) Successful in 14s
CI / Website build and checks (pull_request) Has been cancelled
2026-11-20 20:51:48 +00:00
Compare
fix(host): HOS_BOOT_SAVE's start keeps its picture and input: the title's graph is no longer put back in play
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 14s
CI / Workflow lint (actionlint) (pull_request) Successful in 26s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 30s
Third-party parts / As their notes say (pull_request) Successful in 38s
PR description / Description follows the template (pull_request_target) Successful in 15s
Website / Website on Cloudflare (pull_request_target) Has been skipped
beanstalk/approval approved by dyst (reviewed)
beanstalk/linux 2/1 pass
CI / Website build and checks (pull_request) Failing after 2m27s
1786d81053
The start's picture went black and Link stopped answering the stick once in play (MagicBeanDev/Longshot#1577, mooki's
Linux run; reproduced on Windows, RTX 3080: 0 m moved for a 120-frame stick push, 3,197 draws a frame against 641
pressed). NormalTitle was put back as shipped once the player was seen, while its save branch's selector still ran,
so it turned to the title screen, which held the picture and the input. Without the put-back the same push moves Link
1.52 m, as the pressed start does, and the picture is the forest.

The copies the game keeps now stay changed: with the switch set there is no title, and Back to title loads the save
again. The undo stays only for a change that lost the race with the start's refusal (the graph not running yet).

Also tried and left out: dropping the title's background scene from BeforeTitle and the opening's preload from Title.
Warm, from the game's first frame to in play, 5 runs each: 20.4 s with them, 19.8 s without (pressed: 25.6 s).

Checks: tests/test_boot_save.py (the copies stay changed in play, every whole read before the pick changed).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
dyst changed title from WIP: feat(host): HOS_OOT_BOOT_SAVE starts the game in a save, with no title screen and no presses to WIP: feat(host): HOS_BOOT_SAVE starts the game in a save, with no title screen and no presses 2026-11-20 22:56:00 +00:00
Author
Member

Thanks for the Linux run: the black picture and the dead stick were real, and not Linux only. Reproduced on Windows (RTX 3080) at the rebased fe0c661b: 0 m for a 120-frame stick push and 3,197 draws a frame against 641 for the pressed start, the same shape as your 2,240 against 690.

Cause: 6e06faff put the title's graph back as shipped once the player was seen, but NormalTitle's save-branch selector was still running then. It turned to the title screen, which held the picture and the input over the loaded forest. With the put-back off, the same build draws the forest and the push moves Link 1.52 m, as the pressed start does.

Fixed in 1786d810: the copies the game keeps stay changed, so with the switch set there is no title, and Back to title loads the save again (documented). The undo is left only for a change that lost the race with the start's refusal, before the graph runs. Also on this branch: rebased on main, and the switch renamed to HOS_BOOT_SAVE (Beanstalk#91 follows).

On your other points:

  • boot_save reads off for a refused request as for none: kept, and documented that way; the log says why.
  • =1 choosing slot_03 where the Skip title row loaded slot_01: =1 takes the newest progress.sav by its time, and the row's choice is its own code, not changed here. Worth an issue of its own if the row is the one that is wrong.
  • PR 61's demand word: agreed, whichever lands second arms it for the switch until the start is over.
  • --load-slot as a thin layer on the switch would be welcome, if you still want it once this lands.

Measured at 1786d810 (Windows, RTX 3080, warm, 2 interleaved runs each): 14.1 s from the game's first frame to in play with 0 presses, against 17.9 s pressed (13 to 16 presses).

@Beanstalk linux (any GPU): start from the kokiri-start repository save with HOS_BOOT_SAVE=1 and no presses, expecting play in Kokiri Forest with status.boot_save "loaded", the picture there and the stick moving Link; and the same save with the switch unset, expecting the title as before.

Thanks for the Linux run: the black picture and the dead stick were real, and not Linux only. Reproduced on Windows (RTX 3080) at the rebased fe0c661b: 0 m for a 120-frame stick push and 3,197 draws a frame against 641 for the pressed start, the same shape as your 2,240 against 690. Cause: 6e06faff put the title's graph back as shipped once the player was seen, but NormalTitle's save-branch selector was still running then. It turned to the title screen, which held the picture and the input over the loaded forest. With the put-back off, the same build draws the forest and the push moves Link 1.52 m, as the pressed start does. Fixed in 1786d810: the copies the game keeps stay changed, so with the switch set there is no title, and Back to title loads the save again (documented). The undo is left only for a change that lost the race with the start's refusal, before the graph runs. Also on this branch: rebased on main, and the switch renamed to `HOS_BOOT_SAVE` (Beanstalk#91 follows). On your other points: - `boot_save` reads `off` for a refused request as for none: kept, and documented that way; the log says why. - `=1` choosing slot_03 where the Skip title row loaded slot_01: `=1` takes the newest `progress.sav` by its time, and the row's choice is its own code, not changed here. Worth an issue of its own if the row is the one that is wrong. - PR 61's demand word: agreed, whichever lands second arms it for the switch until the start is over. - `--load-slot` as a thin layer on the switch would be welcome, if you still want it once this lands. Measured at 1786d810 (Windows, RTX 3080, warm, 2 interleaved runs each): 14.1 s from the game's first frame to in play with 0 presses, against 17.9 s pressed (13 to 16 presses). @Beanstalk linux (any GPU): start from the kokiri-start repository save with HOS_BOOT_SAVE=1 and no presses, expecting play in Kokiri Forest with status.boot_save "loaded", the picture there and the stick moving Link; and the same save with the switch unset, expecting the title as before.
Member

Beanstalk test farm · ✅ passed

Commit 1786d81 · approved by dyst (reviewed) · dashboard

⚠️ Note: The save-start driver presses through the title when the switch is unset, so this plan cannot check that the unset run remains at the title without presses.
⚠️ Note: The schema cannot assert status.boot_save or measure Link’s position; the pick log and before/after pictures are proxies, not proof of the requested status or movement.

Summary, written by AI (gpt-6.1-terra) from the numbers below:

The change passed on Linux on both the AMD Radeon RX 6600 and NVIDIA GeForce RTX 5070 test PCs. Both candidate runs exited successfully, found the expected save-pick log, and had a screenshot difference of 0.0; the baseline runs each lacked that expected log. The farm could not directly verify bridge status, Link’s displacement, or the unset-save title start.

Done (started 2026-11-20 23:14 UTC, last report 01:00 UTC). Run 2: re-planned by dyst. Comment @Beanstalk to run it again.

Target Result Reports Frame time p99 (ms), switch off → switch on VRAM peak (MB) Crashes
linux ✅ pass 2 1 2 — — none
What was tested

What will be tested · AI plan (gpt-6.1-terra)

On Linux, compare the PR’s boot-save start with its normal pressed start using the kokiri-start save, over about 10–12 minutes. The plan checks for play in Kokiri Forest, the save-pick log, and pictures before and after a stick push; a missing pick, wrong state or map, crash, or substantial picture difference fails. The farm cannot directly verify the bridge status or Link’s displacement, or leave the unset save start at the title.

  • Where: linux (1-2 reports).
  • Against: the PR itself, switch on (HOS_BOOT_SAVE) and off; main is not built or run.
  • Matrix: 1 cell, each with the switch on and off: HOS_BOOT_SAVE 1 (PR only).
  • kokiri-boot-save: from the repository's save kokiri-start, then take the picture before-stick to compare with the switch off, move the stick, take the picture after-stick to compare with the switch off; 1 run on each side.
  • Measured: crashes, the game's log.
  • The PR fails if: it crashes; or a picture differs from main by more than 10%; or expected log lines missing is above 0.
  • Time: about 11 min of game runs per tester (2 game starts), plus the two builds.
Measurements on `linux`, for the PR description

Verdict: pass on linux (2 reports): no rule failed and no PR run crashed.

Hardware: 1: linux+amd+radv+dgpu AMD Radeon RX 6600 (RADV NAVI23), 26.2.6 AMD Ryzen 5 5600X 6-Core Processor, 64 GB RAM; 2: linux+nvidia+nvidia-proprietary+dgpu NVIDIA GeForce RTX 5070, 620.70.6 AMD Ryzen 5 9600X 6-Core Processor, 32 GB RAM
Scene: kokiri-boot-save, from the repository's save kokiri-start; each cell's settings and switches as the table names them
Method: the Beanstalk test farm: this PR built once and run with its PR-only switches on and off on the same PC in one session, game starts alternating (ABBA); kokiri-boot-save: 1 game start a side; each number the median of the runs, then of 2 reports
How to reproduce: the farm's plan above (@Beanstalk on this PR runs it again)

Before: this PR @ 1786d81, switch off After: this PR @ 1786d81, switch on
kokiri-boot-save · HOS_BOOT_SAVE=1 (PR only) · expected log lines missing 1.0 0.0 (−100%)

Flags and settings

Flag What it is Values tried Before this PR After this PR Effect (frame time (p99)) Helps / doesn't help / untested
HOS_BOOT_SAVE switch the PR adds (set on the PR side only) 1 off (the switch's default) the PR's default kokiri-boot-save, main without it → the PR with 1: not measured untested

As Markdown for the PR description:

## Measurements

**Verdict:** pass on linux (2 reports): no rule failed and no PR run crashed.

Hardware: 1: linux+amd+radv+dgpu `AMD Radeon RX 6600 (RADV NAVI23), 26.2.6` `AMD Ryzen 5 5600X 6-Core Processor, 64 GB RAM`; 2: linux+nvidia+nvidia-proprietary+dgpu `NVIDIA GeForce RTX 5070, 620.70.6` `AMD Ryzen 5 9600X 6-Core Processor, 32 GB RAM`  
Scene: kokiri-boot-save, from the repository's save kokiri-start; each cell's settings and switches as the table names them  
Method: the Beanstalk test farm: this PR built once and run with its PR-only switches on and off on the same PC in one session, game starts alternating (ABBA); kokiri-boot-save: 1 game start a side; each number the median of the runs, then of 2 reports  
How to reproduce: the farm's plan above (`@Beanstalk` on this PR runs it again)

| | Before: this PR @ 1786d81, switch off | After: this PR @ 1786d81, switch on |
|---|---|---|
| kokiri-boot-save · `HOS_BOOT_SAVE=1 (PR only)` · expected log lines missing | 1.0 | 0.0 (−100%) |

**Flags and settings**

| Flag | What it is | Values tried | Before this PR | After this PR | Effect (frame time (p99)) | Helps / doesn't help / untested |
|---|---|---|---|---|---|---|
| `HOS_BOOT_SAVE` | switch the PR adds (set on the PR side only) | `1` | off (the switch's default) | the PR's default | kokiri-boot-save, main without it → the PR with 1: not measured | untested |
The testers' hardware
Target Report Ran on
linux 1 linux+amd+radv+dgpu AMD Radeon RX 6600 (RADV NAVI23), 26.2.6 AMD Ryzen 5 5600X 6-Core Processor, 64 GB RAM
linux 2 linux+nvidia+nvidia-proprietary+dgpu NVIDIA GeForce RTX 5070, 620.70.6 AMD Ryzen 5 9600X 6-Core Processor, 32 GB RAM
<!-- beanstalk:sticky --> ### Beanstalk test farm · ✅ passed Commit `1786d81` · approved by dyst (reviewed) · [dashboard](https://beanstalk.magicbean.dev/prs/1577) ⚠️ **Note:** The save-start driver presses through the title when the switch is unset, so this plan cannot check that the unset run remains at the title without presses. ⚠️ **Note:** The schema cannot assert status.boot_save or measure Link’s position; the pick log and before/after pictures are proxies, not proof of the requested status or movement. > **Summary**, written by AI (`gpt-6.1-terra`) from the numbers below: > > The change passed on Linux on both the AMD Radeon RX 6600 and NVIDIA GeForce RTX 5070 test PCs. Both candidate runs exited successfully, found the expected save-pick log, and had a screenshot difference of 0.0; the baseline runs each lacked that expected log. The farm could not directly verify bridge status, Link’s displacement, or the unset-save title start. Done (started 2026-11-20 23:14 UTC, last report 01:00 UTC). Run 2: re-planned by dyst. Comment `@Beanstalk` to run it again. | Target | Result | Reports | Frame time p99 (ms), switch off → switch on | VRAM peak (MB) | Crashes | | --- | --- | --- | --- | --- | --- | | `linux` | ✅ pass | 2 [1](https://beanstalk.magicbean.dev/reports/rp_3906ba81e73bdbb3) [2](https://beanstalk.magicbean.dev/reports/rp_8ca522f1422bb417) | — | — | none | <details><summary>What was tested</summary> **What will be tested** · AI plan (`gpt-6.1-terra`) On Linux, compare the PR’s boot-save start with its normal pressed start using the kokiri-start save, over about 10–12 minutes. The plan checks for play in Kokiri Forest, the save-pick log, and pictures before and after a stick push; a missing pick, wrong state or map, crash, or substantial picture difference fails. The farm cannot directly verify the bridge status or Link’s displacement, or leave the unset save start at the title. - **Where:** linux (1-2 reports). - **Against:** the PR itself, switch on (`HOS_BOOT_SAVE`) and off; main is not built or run. - **Matrix:** 1 cell, each with the switch on and off: HOS_BOOT_SAVE `1` (PR only). - **kokiri-boot-save:** from the repository's save `kokiri-start`, then take the picture before-stick to compare with the switch off, move the stick, take the picture after-stick to compare with the switch off; 1 run on each side. - **Measured:** crashes, the game's log. - **The PR fails if:** it crashes; or a picture differs from main by more than 10%; or expected log lines missing is above 0. - **Time:** about 11 min of game runs per tester (2 game starts), plus the two builds. </details> <details><summary>Measurements on `linux`, for the PR description</summary> **Verdict:** pass on linux (2 reports): no rule failed and no PR run crashed. Hardware: 1: linux+amd+radv+dgpu `AMD Radeon RX 6600 (RADV NAVI23), 26.2.6` `AMD Ryzen 5 5600X 6-Core Processor, 64 GB RAM`; 2: linux+nvidia+nvidia-proprietary+dgpu `NVIDIA GeForce RTX 5070, 620.70.6` `AMD Ryzen 5 9600X 6-Core Processor, 32 GB RAM` Scene: kokiri-boot-save, from the repository's save kokiri-start; each cell's settings and switches as the table names them Method: the Beanstalk test farm: this PR built once and run with its PR-only switches on and off on the same PC in one session, game starts alternating (ABBA); kokiri-boot-save: 1 game start a side; each number the median of the runs, then of 2 reports How to reproduce: the farm's plan above (`@Beanstalk` on this PR runs it again) | | Before: this PR @ 1786d81, switch off | After: this PR @ 1786d81, switch on | |---|---|---| | kokiri-boot-save · `HOS_BOOT_SAVE=1 (PR only)` · expected log lines missing | 1.0 | 0.0 (−100%) | **Flags and settings** | Flag | What it is | Values tried | Before this PR | After this PR | Effect (frame time (p99)) | Helps / doesn't help / untested | |---|---|---|---|---|---|---| | `HOS_BOOT_SAVE` | switch the PR adds (set on the PR side only) | `1` | off (the switch's default) | the PR's default | kokiri-boot-save, main without it → the PR with 1: not measured | untested | As Markdown for the PR description: ```markdown ## Measurements **Verdict:** pass on linux (2 reports): no rule failed and no PR run crashed. Hardware: 1: linux+amd+radv+dgpu `AMD Radeon RX 6600 (RADV NAVI23), 26.2.6` `AMD Ryzen 5 5600X 6-Core Processor, 64 GB RAM`; 2: linux+nvidia+nvidia-proprietary+dgpu `NVIDIA GeForce RTX 5070, 620.70.6` `AMD Ryzen 5 9600X 6-Core Processor, 32 GB RAM` Scene: kokiri-boot-save, from the repository's save kokiri-start; each cell's settings and switches as the table names them Method: the Beanstalk test farm: this PR built once and run with its PR-only switches on and off on the same PC in one session, game starts alternating (ABBA); kokiri-boot-save: 1 game start a side; each number the median of the runs, then of 2 reports How to reproduce: the farm's plan above (`@Beanstalk` on this PR runs it again) | | Before: this PR @ 1786d81, switch off | After: this PR @ 1786d81, switch on | |---|---|---| | kokiri-boot-save · `HOS_BOOT_SAVE=1 (PR only)` · expected log lines missing | 1.0 | 0.0 (−100%) | **Flags and settings** | Flag | What it is | Values tried | Before this PR | After this PR | Effect (frame time (p99)) | Helps / doesn't help / untested | |---|---|---|---|---|---|---| | `HOS_BOOT_SAVE` | switch the PR adds (set on the PR side only) | `1` | off (the switch's default) | the PR's default | kokiri-boot-save, main without it → the PR with 1: not measured | untested | ``` </details> <details><summary>The testers' hardware</summary> | Target | Report | Ran on | | --- | --- | --- | | `linux` | [1](https://beanstalk.magicbean.dev/reports/rp_3906ba81e73bdbb3) | linux+amd+radv+dgpu `AMD Radeon RX 6600 (RADV NAVI23), 26.2.6` `AMD Ryzen 5 5600X 6-Core Processor, 64 GB RAM` | | `linux` | [2](https://beanstalk.magicbean.dev/reports/rp_8ca522f1422bb417) | linux+nvidia+nvidia-proprietary+dgpu `NVIDIA GeForce RTX 5070, 620.70.6` `AMD Ryzen 5 9600X 6-Core Processor, 32 GB RAM` | </details>
Author
Member

@Beanstalk test it anyway

@Beanstalk test it anyway
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 14s
CI / Workflow lint (actionlint) (pull_request) Successful in 26s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 30s
Third-party parts / As their notes say (pull_request) Successful in 38s
PR description / Description follows the template (pull_request_target) Successful in 15s
Website / Website on Cloudflare (pull_request_target) Has been skipped
beanstalk/approval approved by dyst (reviewed)
beanstalk/linux 2/1 pass
CI / Website build and checks (pull_request) Failing after 2m27s
This pull request has changes conflicting with the target branch.
  • host/control.c
  • host/control.h
  • tests/control_test.c
View command line instructions

Manual merge helper

Use this merge commit message when completing the merge manually.

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u 'feat/boot-save:dyst-feat/boot-save'
git switch 'dyst-feat/boot-save'
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
4 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot!1577
No description provided.