feat(agent): a save start loads the save with Longshot's HOS_OOT_BOOT_SAVE, with no presses #91
No reviewers
Labels
No labels
kind
bug
kind
chore
kind
docs
kind
feature
kind
performance
kind
question
priority
critical
priority
high
priority
low
risk
high
risk
low
risk
medium
status
blocked
status
confirmed
status
duplicate
status
invalid
status
needs-info
status
wontfix
ci: approved
No milestone
No project
No assignees
2 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
MagicBeanDev/Beanstalk!91
Loading…
Reference in a new issue
No description provided.
Delete branch "dyst/Beanstalk:feat/boot-save"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
A farm run that starts from a repository save gets into play by pressing A through the title screen and the save list
about once a second, which races the game's start. Longshot's new
HOS_OOT_BOOT_SAVE(MagicBeanDev/Longshot#1577) starts the game straight in the save with no title screen and no presses; this makes it
the agent's start for
save:<name>, and keeps the presses for builds that don't have it.What changed
ScenarioRunner.StartEnv: asave:<name>start setsHOS_OOT_BOOT_SAVE=slot_00for the game (the repository save isthe run's only slot). It is the agent's own: a plan may not set it (
Plan.cs's andplan.py's deny lists, theplanner's text,
docs/test-plan.md), so atitlestart stays a title start.ScenarioRunner.PressIntoPlay:IntoPlayAsyncpresses no A while the game's own start carries it, which its bridge'sstatus.boot_savesays (waiting,picked,loaded). A build without the switch has noboot_save(and reads nosuch switch), and one whose start let go says
off,refusedorfailed: A is pressed through the title asbefore. So both sides of an A/B get into play whichever has the switch, and
IntoPlayLimitstill bounds the wait.docs/test-plan.md: thesave:<name>start says how it gets into play.Nothing changes for
titleandnewstarts, the steps, the probes or the reports.How it was tested
python -I agent/tests/test_agent_core.py --strict(Windows): 1957 checks, with new ones in the scenario isolationchecks:
StartEnvfor each start,PressIntoPlayover eachboot_savevalue and with none.coordinator/tests/test_plan.py: 186 passed, with the new refusal ofHOS_OOT_BOOT_SAVEin a plan. The planner'sdatabase tests need Postgres on Linux and were not run here (only its prompt text changed).
In the game (Windows 11, RTX 3080, 621.14), the agent's rule driven through the control bridge with the
kokiri-startsave (HOS_OOT_BOOT_SAVE=slot_00, A only whenPressIntoPlaywould press), twice against each sideof an A/B packed over the same base:
status.boot_saveloadedAll four ended at the same position and facing. Time from the first frame to play: 12.5 and 18.8 s with the switch,
25.5 and 16.1 s pressed: no difference resolved, so this is about a start with no presses, not speed (the first run
with the switch also spent 71 s starting a freshly copied runtime before its first frame).
Two cold reviews by a second agent; their fixes are in
9537048and8f924b8:RunScenarioAsync;[scenario.env],[matrix]and[matrix.pr_only], and both readers must refuse each for that switch (removing the agent's deny entry now fails the checks);TryGetValue;boot_save"failed" ends the start with its own reason;boot_saveit ended in.With these the agent's checks pass: 1981.
Not run: a whole
beanstalk-agent run --pr(it builds both sides and wants 150 GB free; this PC had 59 GB).Depends on
MagicBeanDev/Longshot#1577 for the switch itself. Merged before it, this changes nothing a tester sees: no build has
boot_saveyet, so every start presses as now.🤖 Generated with Claude Code
Possibly related, for context:
dyst referenced this pull request from MagicBeanDev/Longshot2026-11-18 11:42:46 +00:00
Two cold reviews by a second agent, fixed in
9537048and8f924b8: the start's environment checked as the sandbox gets it for each start, a plan's value left out, three bad-plan fixtures (scenario env, matrix, pr_only) refused by both readers for HOS_OOT_BOOT_SAVE, JSON read with TryGetValue (no exception on another type), boot_save failed ends the start with its own reason (Longshot#1577's failed now means no title and an unknown save), a timeout says how the start stood. Agent checks: 1981 passed. Longshot#1577's head fe0c661b with this rule in the game: 3 of 3 starts in play with 0 presses.WIP: feat(agent): a save start loads the save with Longshot's HOS_OOT_BOOT_SAVE, with no pressesto feat(agent): a save start loads the save with Longshot's HOS_OOT_BOOT_SAVE, with no pressesOn the overlap with #89: they complement each other. #89 makes the default plan's gameplay scenario start from save:kokiri-start; with this PR every such start sets HOS_OOT_BOOT_SAVE=slot_00, so the default plan reaches play with no presses on builds that have MagicBeanDev/Longshot#1577 (and is pressed through the title as now on builds that don't). The two touch the same lines as text only: both add entries to RunSessionAsync's own environment (#89 RX_VK_TEST_AGENT/RX_VK_TEST_LABEL, this PR the StartEnv line after it) and both add checks in ScenarioIsolationChecks' per-start loop. Either order works; whichever merges second keeps both sides. If #89 lands first I'll rebase this one onto it.
8f924b81c65abfdf15dfView command line instructions
Checkout
From your project repository, check out a new branch and test the changes.