feat(testing): show test context and Dev metrics in agent runs #1539

Open
futondestroyer wants to merge 7 commits from futondestroyer/Longshot:codex/agent-test-hud into main
Member

Summary

Contributors can start a gameplay test from a disposable copy of the selected checkpoint. The window and screenshots show the build version, public agent identifier, short test label, CPU and GPU at top center. Dev FPS and performance metrics stay at bottom left during the test.

The contributor launcher lives in scripts/longshot_ctl.py and scripts/longshot_bridge.py. The renderer draws the same test information on the live window and capture targets. Control-bridge captures preserve clean comparison pixels from the same present. The merged Beanstalk runner sends the labels, drives a copied checkpoint into play and reads the clean comparison image.

The rebase retains the current presenter and moves the launch guide to docs/dev/reference/control-bridge.md. A periodic test capture can move to the next eligible frame when the presenter has the UI lock. Frame generation can then continue while the display releases a busy frame set.

Testing

Tested: The initial rebase passed nine focused Windows fixtures covering the launcher, title auto-continue, screenshot allocation, overlay order, frame pacing and bug-report capture. test_bug_report_defaults.py failed the same auto-profile case on the PR and exact main 89c3d21e, where the fixture expects an Auto-controlled shadow_res entry.

The lock regression check failed as intended on test-only 0cd0a04c. On b324a3cb, the capture resources, overlay order, frame pacing and offscreen UI recording checks passed. Independent source review verified both exact Graphify maps and followed the replay and presenter calls through frame-set release and submission.

On exact b324a3cb, Windows 11 with an RTX 4070 and NVIDIA 621.01 loaded the disposable Kokiri checkpoint in a live window and offscreen. Both runs displayed the test HUD in their bridge screenshots and provided a clean same-present companion. The live client displayed the same banner and Dev metrics. Expanded HDR10 with built-in 2× frame generation stayed responsive during a camera pan. Both overlays appeared in a generated-frame capture and the game log confirmed the HDR10 swapchain and made-frame output. The stock and candidate ordinary-launch controls each produced a single clean PNG. All five sessions exited 0 through the bridge.

Still to test:

  • Linux RADV or Steam Deck headless build and disposable save:kokiri-start scenario with clean exit. Verify the annotated and clean same-present captures locally and that the comparison uses the clean PNG. Test Farm after approval of the exact rebased head.
  • Live Linux window readability and alternate GPU presentation. By hand on the windowed slot and bridge capture.
  • Developer HUD timing off and on, plus capture cost, on matched warmed runs. By hand on the same scene and hardware.
  • Full current-head game-free suite with named baseline failures, and Graphics guide source review. By hand or through the corresponding CI jobs.

Measurements

Verdict: The game-free lock check reproduced the blocking wait before the fix and passes afterwards. A matched timing run is still needed for release review.

HOS_TEST_AUTOLOAD defaults off and is enabled by the disposable contributor launch. RX_VK_TEST_AGENT is unset by default. It enables the test HUD when provided. RX_VK_TEST_LABEL defaults to Gameplay test.

Risks

The opt-in test capture adds an overlay draw and, for bridge captures, a second readback and PNG. Timing coverage is still open. Periodic test captures wait for the next opportunity if the display is busy. Regular launches keep the existing settings and screenshot path.

Screenshots

These bridge captures show the same disposable Kokiri checkpoint and camera, with SDR and frame generation off. The second is from the exact rebased windowed agent session.

Stock main at 89c3d21e

Windows test HUD at b324a3cb with test banner, hardware and Dev metrics

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).
  • 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 check under tests/ that reads changed source is updated, not weakened.
  • Screenshots of the same spot on main and here, for a change that shows on screen.
  • Docs updated.
  • Host code builds on Windows (llvm-mingw) and Linux, with UTF-8 paths.
  • 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.
## Summary Contributors can start a gameplay test from a disposable copy of the selected checkpoint. The window and screenshots show the build version, public agent identifier, short test label, CPU and GPU at top center. Dev FPS and performance metrics stay at bottom left during the test. The contributor launcher lives in `scripts/longshot_ctl.py` and `scripts/longshot_bridge.py`. The renderer draws the same test information on the live window and capture targets. Control-bridge captures preserve clean comparison pixels from the same present. The merged [Beanstalk runner](https://git.magicbean.dev/MagicBeanDev/Beanstalk/pulls/89) sends the labels, drives a copied checkpoint into play and reads the clean comparison image. The rebase retains the current presenter and moves the launch guide to `docs/dev/reference/control-bridge.md`. A periodic test capture can move to the next eligible frame when the presenter has the UI lock. Frame generation can then continue while the display releases a busy frame set. ## Testing **Tested:** The initial rebase passed nine focused Windows fixtures covering the launcher, title auto-continue, screenshot allocation, overlay order, frame pacing and bug-report capture. `test_bug_report_defaults.py` failed the same `auto-profile` case on the PR and exact main `89c3d21e`, where the fixture expects an Auto-controlled `shadow_res` entry. The lock regression check failed as intended on test-only `0cd0a04c`. On `b324a3cb`, the capture resources, overlay order, frame pacing and offscreen UI recording checks passed. Independent source review verified both exact Graphify maps and followed the replay and presenter calls through frame-set release and submission. On exact `b324a3cb`, Windows 11 with an RTX 4070 and NVIDIA 621.01 loaded the disposable Kokiri checkpoint in a live window and offscreen. Both runs displayed the test HUD in their bridge screenshots and provided a clean same-present companion. The live client displayed the same banner and Dev metrics. Expanded HDR10 with built-in 2× frame generation stayed responsive during a camera pan. Both overlays appeared in a generated-frame capture and the game log confirmed the HDR10 swapchain and made-frame output. The stock and candidate ordinary-launch controls each produced a single clean PNG. All five sessions exited 0 through the bridge. **Still to test:** - Linux RADV or Steam Deck headless build and disposable `save:kokiri-start` scenario with clean exit. Verify the annotated and clean same-present captures locally and that the comparison uses the clean PNG. Test Farm after approval of the exact rebased head. - Live Linux window readability and alternate GPU presentation. By hand on the windowed slot and bridge capture. - Developer HUD timing off and on, plus capture cost, on matched warmed runs. By hand on the same scene and hardware. - Full current-head game-free suite with named baseline failures, and Graphics guide source review. By hand or through the corresponding CI jobs. ## Measurements **Verdict:** The game-free lock check reproduced the blocking wait before the fix and passes afterwards. A matched timing run is still needed for release review. `HOS_TEST_AUTOLOAD` defaults off and is enabled by the disposable contributor launch. `RX_VK_TEST_AGENT` is unset by default. It enables the test HUD when provided. `RX_VK_TEST_LABEL` defaults to `Gameplay test`. ## Risks The opt-in test capture adds an overlay draw and, for bridge captures, a second readback and PNG. Timing coverage is still open. Periodic test captures wait for the next opportunity if the display is busy. Regular launches keep the existing settings and screenshot path. ## Screenshots These bridge captures show the same disposable Kokiri checkpoint and camera, with SDR and frame generation off. The second is from the exact rebased windowed agent session. ![Stock main at 89c3d21e](https://git.magicbean.dev/attachments/e9090ebb-3053-417f-a265-c3d411bc699d) ![Windows test HUD at b324a3cb with test banner, hardware and Dev metrics](https://git.magicbean.dev/attachments/0dc22b9d-a112-473e-b2f5-66b7acbf8d42) ## 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. - [ ] `mise run test` passes (any failure that also fails on `main` is named). - [ ] 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 check under `tests/` that reads changed source is updated, not weakened. - [x] Screenshots of the same spot on `main` and here, for a change that shows on screen. - [x] Docs updated. - [ ] Host code builds on Windows (llvm-mingw) and Linux, with UTF-8 paths. - [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.
feat(testing): show test context and Dev metrics in agent runs
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
Website / Website on Cloudflare (pull_request_target) Has been skipped
CI / Workflow lint (actionlint) (pull_request) Successful in 15s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 16s
Third-party parts / As their notes say (pull_request) Successful in 17s
PR description / Description follows the template (pull_request_target) Failing after 6s
c90d8f844a
Assisted-by: OpenAI Codex
Author
Member

Tested in game: Windows 11, NVIDIA RTX 4070, driver 621.01. The pre-rebase feature build loaded the disposable Kokiri Forest checkpoint in a visible window and offscreen. Both modes produced the test HUD and clean same-frame capture, then exited normally. Seven focused checks passed after rebasing.

Tested in game: Windows 11, NVIDIA RTX 4070, driver 621.01. The pre-rebase feature build loaded the disposable Kokiri Forest checkpoint in a visible window and offscreen. Both modes produced the test HUD and clean same-frame capture, then exited normally. Seven focused checks passed after rebasing.
HumanBean approved these changes 2026-11-20 07:14:26 +00:00
Dismissed
HumanBean added this to the WIP project 2026-11-20 07:14:30 +00:00
futondestroyer force-pushed codex/agent-test-hud from c90d8f844a
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
Website / Website on Cloudflare (pull_request_target) Has been skipped
CI / Workflow lint (actionlint) (pull_request) Successful in 15s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 16s
Third-party parts / As their notes say (pull_request) Successful in 17s
PR description / Description follows the template (pull_request_target) Failing after 6s
to b324a3cb4d
All checks were successful
PR description / Description follows the template (pull_request_target) Successful in 4s
2026-11-20 13:47:27 +00:00
Compare
futondestroyer dismissed HumanBean's review 2026-11-20 13:47:29 +00:00
Reason:

New commits pushed, approval review dismissed automatically according to repository settings

futondestroyer changed title from feat(testing): show test context and Dev metrics in agent runs to WIP: feat(testing): show test context and Dev metrics in agent runs 2026-11-20 13:47:31 +00:00
Author
Member

These SDR captures use the same disposable Kokiri checkpoint and camera. The second capture came from the rebased Windows agent session. I also verified both overlays in its live game window.

Stock main at 89c3d21e

Test HUD at b324a3cb with build, agent, test label, hardware and Dev metrics

These SDR captures use the same disposable Kokiri checkpoint and camera. The second capture came from the rebased Windows agent session. I also verified both overlays in its live game window. ![Stock main at 89c3d21e](https://git.magicbean.dev/attachments/e9090ebb-3053-417f-a265-c3d411bc699d) ![Test HUD at b324a3cb with build, agent, test label, hardware and Dev metrics](https://git.magicbean.dev/attachments/0dc22b9d-a112-473e-b2f5-66b7acbf8d42)
Author
Member

Rebased onto 89c3d21e at b324a3cb and fixed the test screenshot lock order found in independent review. On the RTX 4070, the disposable launch reached play with its HUD in the live window and offscreen capture. Expanded HDR10 with built-in 2x frame generation also produced an annotated generated frame. Ordinary-launch controls kept a single clean screenshot. The exact-head before/after pair is in the description. This remains WIP for timing and the remaining platform and CI checks.

Rebased onto 89c3d21e at b324a3cb and fixed the test screenshot lock order found in independent review. On the RTX 4070, the disposable launch reached play with its HUD in the live window and offscreen capture. Expanded HDR10 with built-in 2x frame generation also produced an annotated generated frame. Ordinary-launch controls kept a single clean screenshot. The exact-head before/after pair is in the description. This remains WIP for timing and the remaining platform and CI checks.
Owner
@rebase
Member

I rebased this pull request onto main (51f5f9da56): the new head is 5ce271b73e.

Conflicts resolved:

  • gpu/vk/vk_imgui.cpp: Kept colour-vision correction alongside screenshot handling and overlay recording.
  • gpu/vk/vk_imgui.h: Kept both the colour-vision and screenshot parameters.
  • tests/test_fps_overlay_order.py: Kept the colour-mode checks and test HUD recording coverage. No tests or builds were run.

Nothing was built or run: read the resolution before merging.


Automated rebase by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.

I rebased this pull request onto `main` (51f5f9da5641): the new head is 5ce271b73ebc. Conflicts resolved: - gpu/vk/vk_imgui.cpp: Kept colour-vision correction alongside screenshot handling and overlay recording. - gpu/vk/vk_imgui.h: Kept both the colour-vision and screenshot parameters. - tests/test_fps_overlay_order.py: Kept the colour-mode checks and test HUD recording coverage. No tests or builds were run. Nothing was built or run: read the resolution before merging. --- <sub>Automated rebase by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.</sub>
BeanBot force-pushed codex/agent-test-hud from b324a3cb4d
All checks were successful
PR description / Description follows the template (pull_request_target) Successful in 4s
to 5ce271b73e
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 11s
PR description / Description follows the template (pull_request_target) Successful in 15s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 17s
CI / Workflow lint (actionlint) (pull_request) Successful in 18s
Third-party parts / As their notes say (pull_request) Successful in 18s
CI / Website build and checks (pull_request) Has been cancelled
2026-11-20 22:02:48 +00:00
Compare
BeanBot left a comment

Recommendation: Request changes. The test HUD adds continuous Dev-metric collection and rendering, but its effect on the gameplay measurements it is meant to support has not been measured. The capture path also adds readback and PNG work whose cost is unknown.

Disposable agent runs now copy a selected checkpoint, enter play, display test context and Dev metrics, and produce annotated bridge screenshots with clean companions. The feature has been demonstrated on Windows, but its timing cost and behavior on other required hardware remain unverified.

Does it help: Unproven. The reported Windows runs show the feature working on an RTX 4070 at an earlier head, but there are no matched timing or capture-cost numbers against the base. The latest rebase was not run.
Other systems: The change affects windowed and headless rendering, bridge and periodic captures, frame generation, HDR, and title auto-continue. An earlier head was exercised on Windows NVIDIA, including HDR10 and generated frames; Linux RADV, Steam Deck, alternate GPU presentation, and the rebased head remain untested.
Switches: HOS_TEST_AUTOLOAD defaults off; RX_VK_TEST_AGENT defaults unset and enables the HUD; RX_VK_TEST_LABEL defaults to Gameplay test. They are documented in switches.def and used by the disposable launcher. Their opt-in defaults limit ordinary-launch exposure, but the cost of the enabled configuration is unmeasured.

Reviewed: 5ce271b73e against 51f5f9da56: Read the repository rules, contribution guide, PR template, standards, relevant control-bridge and architecture sections, commits, and changed renderer and launcher excerpts. Skimmed the test changes and recorded runs. The full changed code files and all dependent paths were not read; Tooling's adjacent AGENTS.md and docs/review-checklist.md were absent. Network access and overlapping PRs were unavailable.; nothing run (a code review).
Checked and fine:

  • The launcher requires an agent identifier and label together and requires a selected checkpoint.
  • The test profile uses its own user directory and overrides the relevant environment values.
  • Bridge captures retain the clean pixels before recording the test overlay.
  • The screenshot lock is attempted after inline frame generation, rather than blocking before it.

Blocking (to change before merge)

  1. gpu/vk/vk_menu.cpp:3147: Agent runs force Dev metrics continuously, yet the PR supplies no matched measurement of their CPU, GPU or frame-pacing cost. A contributor comparing a small performance change could measure the HUD's overhead or distortion instead. Measure warmed, interleaved runs with the HUD off and on, including frame times and 1% lows, and reduce or isolate the cost if it materially affects results.
  2. gpu/vk/vk_replay.cpp:5097: Each annotated bridge capture adds another GPU readback and PNG, but its timing and memory cost are not measured. Frequent captures can disturb the test being observed; measure capture cost and frame pacing under the intended cadence, and adjust the capture strategy if the disturbance is material.

Suggestions (the author may take or leave)

  • Complete the Linux RADV or Steam Deck headless run, alternate presentation coverage, current-head suite, and required Tier 1 checks recorded in Still to test.
  • Verify the rebase conflict resolution on the exact head, including colour-vision correction and screenshots.

Testing: No builds, tests, scripts, or game runs were executed, as requested. The author's Windows evidence is for b324a3cb; the subsequent rebase to 5ce271b73e reports no runs.
Overlaps and merge order: Unknown; the network is closed.


Automated code review by BeanBot (gpt-6.1-terra), asked for by BeanBot, after its rebase. Nothing was run; a maintainer decides.

**Recommendation: Request changes.** The test HUD adds continuous Dev-metric collection and rendering, but its effect on the gameplay measurements it is meant to support has not been measured. The capture path also adds readback and PNG work whose cost is unknown. Disposable agent runs now copy a selected checkpoint, enter play, display test context and Dev metrics, and produce annotated bridge screenshots with clean companions. The feature has been demonstrated on Windows, but its timing cost and behavior on other required hardware remain unverified. **Does it help:** Unproven. The reported Windows runs show the feature working on an RTX 4070 at an earlier head, but there are no matched timing or capture-cost numbers against the base. The latest rebase was not run. **Other systems:** The change affects windowed and headless rendering, bridge and periodic captures, frame generation, HDR, and title auto-continue. An earlier head was exercised on Windows NVIDIA, including HDR10 and generated frames; Linux RADV, Steam Deck, alternate GPU presentation, and the rebased head remain untested. **Switches:** HOS_TEST_AUTOLOAD defaults off; RX_VK_TEST_AGENT defaults unset and enables the HUD; RX_VK_TEST_LABEL defaults to Gameplay test. They are documented in switches.def and used by the disposable launcher. Their opt-in defaults limit ordinary-launch exposure, but the cost of the enabled configuration is unmeasured. **Reviewed:** 5ce271b73ebc against 51f5f9da5641: Read the repository rules, contribution guide, PR template, standards, relevant control-bridge and architecture sections, commits, and changed renderer and launcher excerpts. Skimmed the test changes and recorded runs. The full changed code files and all dependent paths were not read; Tooling's adjacent AGENTS.md and docs/review-checklist.md were absent. Network access and overlapping PRs were unavailable.; nothing run (a code review). **Checked and fine:** - The launcher requires an agent identifier and label together and requires a selected checkpoint. - The test profile uses its own user directory and overrides the relevant environment values. - Bridge captures retain the clean pixels before recording the test overlay. - The screenshot lock is attempted after inline frame generation, rather than blocking before it. **Blocking** (to change before merge) 1. gpu/vk/vk_menu.cpp:3147: Agent runs force Dev metrics continuously, yet the PR supplies no matched measurement of their CPU, GPU or frame-pacing cost. A contributor comparing a small performance change could measure the HUD's overhead or distortion instead. Measure warmed, interleaved runs with the HUD off and on, including frame times and 1% lows, and reduce or isolate the cost if it materially affects results. 2. gpu/vk/vk_replay.cpp:5097: Each annotated bridge capture adds another GPU readback and PNG, but its timing and memory cost are not measured. Frequent captures can disturb the test being observed; measure capture cost and frame pacing under the intended cadence, and adjust the capture strategy if the disturbance is material. **Suggestions** (the author may take or leave) - Complete the Linux RADV or Steam Deck headless run, alternate presentation coverage, current-head suite, and required Tier 1 checks recorded in Still to test. - Verify the rebase conflict resolution on the exact head, including colour-vision correction and screenshots. **Testing:** No builds, tests, scripts, or game runs were executed, as requested. The author's Windows evidence is for b324a3cb; the subsequent rebase to 5ce271b73ebc reports no runs. **Overlaps and merge order:** Unknown; the network is closed. --- <sub>Automated code review by BeanBot (gpt-6.1-terra), asked for by BeanBot, after its rebase. Nothing was run; a maintainer decides.</sub>
@ -3130,0 +3144,4 @@
return agent;
}
static int effective_fps_level() {
Member

Please measure warmed, matched runs with the test HUD off and on, including frame times and 1% lows. For small performance comparisons, continuous Dev metrics and overlay rendering could change the result being measured.

Please measure warmed, matched runs with the test HUD off and on, including frame times and 1% lows. For small performance comparisons, continuous Dev metrics and overlay rendering could change the result being measured.
@ -5042,1 +5095,3 @@
shot_asked.on = true, shot_asked.value = value, shot_asked.present = presents, shot_asked.ask = std::move(a);
const bool annotated = a.test_hud && annotate_test_shot(cmd, ask_image);
if (annotated) {
c.bufferOffset = (VkDeviceSize)kShotW * kShotH * 4;
Member

Please measure the readback and PNG cost at the intended capture cadence, including its effect on frame pacing. Frequent bridge captures could disturb the gameplay test they observe.

Please measure the readback and PNG cost at the intended capture cadence, including its effect on frame pacing. Frequent bridge captures could disturb the gameplay test they observe.
Owner
@fix
Member

I didn't fix this pull request: it needs the author: Both points require measurements that cannot be obtained under the instruction not to run builds, tests, scripts, or the game. Whether either cost is material, and therefore needs a code change, remains for the author to determine. Nothing was pushed.


Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.

I didn't fix this pull request: it needs the author: Both points require measurements that cannot be obtained under the instruction not to run builds, tests, scripts, or the game. Whether either cost is material, and therefore needs a code change, remains for the author to determine. Nothing was pushed. --- <sub>Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.</sub>
Owner
@fix
Member

I didn't change this pull request: what the review asks needs the author.

Left for the author

  • HUD performance cost, gpu/vk/vk_menu.cpp:3147: Run warmed, interleaved gameplay comparisons with the HUD off and on, recording CPU and GPU cost, frame times, and 1% lows. If the difference materially affects results, reduce or isolate the cost. This requires game runs and measurements.
  • Capture performance cost, gpu/vk/vk_replay.cpp:5097: Measure readback time, PNG time, memory use, and frame pacing with captures at the intended cadence against gameplay without captures. Adjust the capture strategy if the disturbance is material. This requires game runs and measurements.
  • Hardware and current-head coverage: Run the current-head suite and outstanding Tier 1 checks, including a Linux RADV or Steam Deck headless run and alternate presentation coverage, and record the results. These require builds, tests, and access to the relevant hardware.
  • Rebase conflict verification: Run the exact head and verify colour-vision correction and screenshots, including the rebased conflict resolution. This requires a build and game run.
  • HUD measurement, gpu/vk/vk_menu.cpp:3147: Measure warmed, matched HUD-off and HUD-on gameplay runs, including frame times and 1% lows, to establish whether the overlay distorts performance comparisons. This requires game runs.
  • Readback and PNG measurement, gpu/vk/vk_replay.cpp:5097: Measure readback and PNG cost and frame pacing at the intended bridge capture cadence, comparing against gameplay without captures. This requires game runs.

Nothing was pushed.


Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.

I didn't change this pull request: what the review asks needs the author. **Left for the author** - **HUD performance cost, gpu/vk/vk_menu.cpp:3147**: Run warmed, interleaved gameplay comparisons with the HUD off and on, recording CPU and GPU cost, frame times, and 1% lows. If the difference materially affects results, reduce or isolate the cost. This requires game runs and measurements. - **Capture performance cost, gpu/vk/vk_replay.cpp:5097**: Measure readback time, PNG time, memory use, and frame pacing with captures at the intended cadence against gameplay without captures. Adjust the capture strategy if the disturbance is material. This requires game runs and measurements. - **Hardware and current-head coverage**: Run the current-head suite and outstanding Tier 1 checks, including a Linux RADV or Steam Deck headless run and alternate presentation coverage, and record the results. These require builds, tests, and access to the relevant hardware. - **Rebase conflict verification**: Run the exact head and verify colour-vision correction and screenshots, including the rebased conflict resolution. This requires a build and game run. - **HUD measurement, gpu/vk/vk_menu.cpp:3147**: Measure warmed, matched HUD-off and HUD-on gameplay runs, including frame times and 1% lows, to establish whether the overlay distorts performance comparisons. This requires game runs. - **Readback and PNG measurement, gpu/vk/vk_replay.cpp:5097**: Measure readback and PNG cost and frame pacing at the intended bridge capture cadence, comparing against gameplay without captures. This requires game runs. Nothing was pushed. --- <sub>Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.</sub>
Owner
@fix
Member

I didn't change this pull request: what the review asks needs the author.

Left for the author

  • gpu/vk/vk_menu.cpp:3147: Measure HUD overhead: Run warmed, interleaved gameplay comparisons with the HUD off and on using the same workload and hardware; record CPU and GPU costs, frame times, and 1% lows. Determine from those results whether to reduce or isolate the cost. Runs and measurements are prohibited here.
  • gpu/vk/vk_replay.cpp:5097: Measure capture overhead: Measure readback time, PNG time, memory use, and frame pacing with annotated bridge captures at the intended cadence against a run without those captures. Adjust the strategy if the disturbance is material. Runs and measurements are prohibited here.
  • Complete hardware and current-head coverage: Run the recorded current-head suite and required Tier 1 checks, including a Linux RADV or Steam Deck headless run and alternate presentation coverage. These require builds, tests, and hardware runs, which are prohibited here.
  • Verify rebase conflict resolution: Build and run exact head 5ce271b73e to verify the rebase resolution, including colour-vision correction and screenshots. This requires a build and game run, which are prohibited here.
  • gpu/vk/vk_menu.cpp:3147: Matched HUD timing runs: Measure warmed, matched HUD-off and HUD-on gameplay runs on the same hardware and workload, including frame times and 1% lows. Runs and measurements are prohibited here.
  • gpu/vk/vk_replay.cpp:5097: Capture timing and pacing: Measure readback and PNG costs and their effect on frame pacing at the intended bridge capture cadence, compared with gameplay without those captures. Runs and measurements are prohibited here.

Nothing was pushed.


Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.

I didn't change this pull request: what the review asks needs the author. **Left for the author** - **gpu/vk/vk_menu.cpp:3147: Measure HUD overhead**: Run warmed, interleaved gameplay comparisons with the HUD off and on using the same workload and hardware; record CPU and GPU costs, frame times, and 1% lows. Determine from those results whether to reduce or isolate the cost. Runs and measurements are prohibited here. - **gpu/vk/vk_replay.cpp:5097: Measure capture overhead**: Measure readback time, PNG time, memory use, and frame pacing with annotated bridge captures at the intended cadence against a run without those captures. Adjust the strategy if the disturbance is material. Runs and measurements are prohibited here. - **Complete hardware and current-head coverage**: Run the recorded current-head suite and required Tier 1 checks, including a Linux RADV or Steam Deck headless run and alternate presentation coverage. These require builds, tests, and hardware runs, which are prohibited here. - **Verify rebase conflict resolution**: Build and run exact head 5ce271b73ebc to verify the rebase resolution, including colour-vision correction and screenshots. This requires a build and game run, which are prohibited here. - **gpu/vk/vk_menu.cpp:3147: Matched HUD timing runs**: Measure warmed, matched HUD-off and HUD-on gameplay runs on the same hardware and workload, including frame times and 1% lows. Runs and measurements are prohibited here. - **gpu/vk/vk_replay.cpp:5097: Capture timing and pacing**: Measure readback and PNG costs and their effect on frame pacing at the intended bridge capture cadence, compared with gameplay without those captures. Runs and measurements are prohibited here. Nothing was pushed. --- <sub>Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.</sub>
Owner
@fix
Member

I changed this pull request by bringing it up to date: the new head is 6845d91d62. I rebased it onto main (3df5c13ed2).

Left for the author

  • HUD performance cost, gpu/vk/vk_menu.cpp:3147: Measure warmed, interleaved runs with the HUD off and on against the base, including CPU and GPU cost, frame times and 1% lows. Use the results to decide whether the cost needs reduction or isolation.
  • Capture cost, gpu/vk/vk_replay.cpp:5097: Measure readback and PNG timing, memory use and frame pacing at the intended capture cadence against runs without those captures. Adjust the strategy if the measured disturbance is material.
  • Remaining hardware and suite coverage: Run the current-head suite and required Tier 1 checks, including a Linux RADV or Steam Deck headless run and alternate presentation coverage, and record the results.
  • Verify conflict resolution: After the bot updates the branch and resolves conflicts, verify the exact resulting head, including colour-vision correction and screenshots, and record the results.

Nothing was built or run: check the change. It will be reviewed again.


Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.

I changed this pull request by bringing it up to date: the new head is 6845d91d62ea. I rebased it onto `main` (3df5c13ed2ec). **Left for the author** - **HUD performance cost, gpu/vk/vk_menu.cpp:3147**: Measure warmed, interleaved runs with the HUD off and on against the base, including CPU and GPU cost, frame times and 1% lows. Use the results to decide whether the cost needs reduction or isolation. - **Capture cost, gpu/vk/vk_replay.cpp:5097**: Measure readback and PNG timing, memory use and frame pacing at the intended capture cadence against runs without those captures. Adjust the strategy if the measured disturbance is material. - **Remaining hardware and suite coverage**: Run the current-head suite and required Tier 1 checks, including a Linux RADV or Steam Deck headless run and alternate presentation coverage, and record the results. - **Verify conflict resolution**: After the bot updates the branch and resolves conflicts, verify the exact resulting head, including colour-vision correction and screenshots, and record the results. Nothing was built or run: check the change. It will be reviewed again. --- <sub>Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.</sub>
BeanBot force-pushed codex/agent-test-hud from 5ce271b73e
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 11s
PR description / Description follows the template (pull_request_target) Successful in 15s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 17s
CI / Workflow lint (actionlint) (pull_request) Successful in 18s
Third-party parts / As their notes say (pull_request) Successful in 18s
CI / Website build and checks (pull_request) Has been cancelled
to 6845d91d62
Some checks failed
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
CI / Python source checks (ubuntu-24.04, part 1/2) (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 10s
PR description / Description follows the template (pull_request_target) Successful in 11s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 11s
CI / Workflow lint (actionlint) (pull_request) Successful in 12s
Third-party parts / As their notes say (pull_request) Successful in 13s
CI / Website build and checks (pull_request) Has been cancelled
2026-11-20 22:46:17 +00:00
Compare
BeanBot left a comment

Recommendation: Request changes. Disposable tests can fail to enter play on Windows when their run folder has a non-ASCII path. The previously identified HUD and capture costs also remain unmeasured.

The change adds disposable checkpoint launches, a test-context HUD and clean companion images for bridge captures. It makes test context easier to see, but the Windows path failure blocks some contributors and the cost of collecting and capturing metrics remains unknown. Since the previous review, the branch was rebased onto 3df5c13ed2; no runs of the resulting head are reported. The previous review’s requests for matched HUD and capture measurements remain outstanding.

Does it help: Unproven. The author reports successful visible and offscreen gameplay, clean companion captures, HDR10 and generated-frame captures on Windows 11 with an RTX 4070 at b324a3cb. There are no matched HUD-on/off or capture-cost numbers, and the current head has not been run.
Other systems: The changed paths cover windowed and offscreen rendering, periodic and bridge screenshots, frame generation, HDR, colour-vision correction and title auto-continue. Earlier Windows NVIDIA runs provide partial evidence; Linux RADV or Steam Deck, alternate GPU presentation, current-head conflict resolution and required Tier 1 coverage remain open.
Switches: HOS_TEST_AUTOLOAD defaults off and is enabled for agent launches; RX_VK_TEST_AGENT defaults unset and enables the HUD; RX_VK_TEST_LABEL defaults to Gameplay test. All have switch-table rows. Their opt-in defaults limit ordinary-launch exposure, but the enabled configuration’s cost is unmeasured.

Reviewed: 6845d91d62 against 3df5c13ed2: Read the base AGENTS.md, CONTRIBUTING.md and PR template; reviewed the complete diff, commits, changed host autoload file, launcher session setup and selected renderer callers and locking paths. Skimmed the test changes, recorded runs, standards and relevant architecture sections. The full changed renderer, launcher and test files and all dependent paths were not read. Tooling’s adjacent AGENTS.md and docs/review-checklist.md were absent.; nothing run (a code review).
Checked and fine:

  • Agent launches require both a public identifier and label and a selected checkpoint.
  • Bridge captures copy clean pixels before adding the test overlay.
  • The screenshot UI lock is attempted after inline frame generation.

Blocking (to change before merge)

  1. host/oot_autoload.c:193: The new agent launch relies on hos_oot_newest_save, whose Windows implementation uses FindFirstFileA and stat on UTF-8 paths. A run folder under a Windows user directory containing characters outside the active code page can fail to find its copied checkpoint, leaving launch waiting for play until timeout. Use UTF-8-aware Windows filesystem operations for the save search and add coverage for a non-ASCII run path.
  2. The previously reviewed HUD overhead remains unmeasured: measure warmed, matched runs including CPU and GPU cost, frame times and 1% lows, and reduce or isolate any material distortion.
  3. The previously reviewed second readback and PNG cost remains unmeasured: measure capture timing, memory and frame pacing at the intended cadence and adjust the strategy if material.

Suggestions (the author may take or leave)

  • Complete the current-head suite, Linux RADV or Steam Deck headless run, alternate presentation and Tier 1 checks.
  • Verify colour-vision correction and screenshots on the exact rebased head.

Testing: No builds, tests, scripts or game runs were executed. The author’s reported gameplay evidence is for an earlier head.
Overlaps and merge order: Unknown; network access is closed. The PR says Beanstalk runner #89 is merged.


Automated code review by BeanBot (gpt-6.1-terra), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.

**Recommendation: Request changes.** Disposable tests can fail to enter play on Windows when their run folder has a non-ASCII path. The previously identified HUD and capture costs also remain unmeasured. The change adds disposable checkpoint launches, a test-context HUD and clean companion images for bridge captures. It makes test context easier to see, but the Windows path failure blocks some contributors and the cost of collecting and capturing metrics remains unknown. Since the previous review, the branch was rebased onto 3df5c13ed2ec; no runs of the resulting head are reported. The previous review’s requests for matched HUD and capture measurements remain outstanding. **Does it help:** Unproven. The author reports successful visible and offscreen gameplay, clean companion captures, HDR10 and generated-frame captures on Windows 11 with an RTX 4070 at b324a3cb. There are no matched HUD-on/off or capture-cost numbers, and the current head has not been run. **Other systems:** The changed paths cover windowed and offscreen rendering, periodic and bridge screenshots, frame generation, HDR, colour-vision correction and title auto-continue. Earlier Windows NVIDIA runs provide partial evidence; Linux RADV or Steam Deck, alternate GPU presentation, current-head conflict resolution and required Tier 1 coverage remain open. **Switches:** HOS_TEST_AUTOLOAD defaults off and is enabled for agent launches; RX_VK_TEST_AGENT defaults unset and enables the HUD; RX_VK_TEST_LABEL defaults to Gameplay test. All have switch-table rows. Their opt-in defaults limit ordinary-launch exposure, but the enabled configuration’s cost is unmeasured. **Reviewed:** 6845d91d62ea against 3df5c13ed2ec: Read the base AGENTS.md, CONTRIBUTING.md and PR template; reviewed the complete diff, commits, changed host autoload file, launcher session setup and selected renderer callers and locking paths. Skimmed the test changes, recorded runs, standards and relevant architecture sections. The full changed renderer, launcher and test files and all dependent paths were not read. Tooling’s adjacent AGENTS.md and docs/review-checklist.md were absent.; nothing run (a code review). **Checked and fine:** - Agent launches require both a public identifier and label and a selected checkpoint. - Bridge captures copy clean pixels before adding the test overlay. - The screenshot UI lock is attempted after inline frame generation. **Blocking** (to change before merge) 1. host/oot_autoload.c:193: The new agent launch relies on hos_oot_newest_save, whose Windows implementation uses FindFirstFileA and stat on UTF-8 paths. A run folder under a Windows user directory containing characters outside the active code page can fail to find its copied checkpoint, leaving launch waiting for play until timeout. Use UTF-8-aware Windows filesystem operations for the save search and add coverage for a non-ASCII run path. 2. The previously reviewed HUD overhead remains unmeasured: measure warmed, matched runs including CPU and GPU cost, frame times and 1% lows, and reduce or isolate any material distortion. 3. The previously reviewed second readback and PNG cost remains unmeasured: measure capture timing, memory and frame pacing at the intended cadence and adjust the strategy if material. **Suggestions** (the author may take or leave) - Complete the current-head suite, Linux RADV or Steam Deck headless run, alternate presentation and Tier 1 checks. - Verify colour-vision correction and screenshots on the exact rebased head. **Testing:** No builds, tests, scripts or game runs were executed. The author’s reported gameplay evidence is for an earlier head. **Overlaps and merge order:** Unknown; network access is closed. The PR says Beanstalk runner #89 is merged. --- <sub>Automated code review by BeanBot (gpt-6.1-terra), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.</sub>
@ -185,3 +191,3 @@
if (!rx_option(HOS_OPT_SKIP_TITLE, 0)) return;
if (!test_autoload && !rx_option(HOS_OPT_SKIP_TITLE, 0)) return;
if (!hos_oot_newest_save(hos_user_dir, save_path, sizeof save_path)) {
hos_log("skip title: no saved game yet: the title screen stays");
Member

This now makes agent launches depend on the Windows save search, which uses narrow FindFirstFileA/stat calls on UTF-8 paths. A run folder containing characters outside the active code page can leave the checkpoint undiscovered and launch waiting for play. Could the search use UTF-8-aware Windows filesystem calls, with a non-ASCII path test?

This now makes agent launches depend on the Windows save search, which uses narrow FindFirstFileA/stat calls on UTF-8 paths. A run folder containing characters outside the active code page can leave the checkpoint undiscovered and launch waiting for play. Could the search use UTF-8-aware Windows filesystem calls, with a non-ASCII path test?
Owner
@fix
Member

I changed this pull request where the review's points could be done here (the rest needs runs, measurements or decisions): the new head is 30fc1bf23d. It was already up to date with main.

Changed

  • Windows non-ASCII save paths, host/oot_autoload.c:187: Changed the Windows save search to use wide filesystem calls and added non-ASCII path coverage in host/oot_autoload.c:39 and tests/test_skip_title.py:72. The author should run the coverage on Windows.

Left for the author

  • HUD overhead measurement, gpu/vk/vk_menu.cpp:3147: Measure warmed, interleaved HUD-off and HUD-on runs against the base, including CPU and GPU cost, frame times and 1% lows; reduce or isolate any material distortion.
  • Capture cost measurement, gpu/vk/vk_replay.cpp:5097: Measure readback and PNG timing, memory use and frame pacing at the intended capture cadence; adjust the capture strategy if disturbance is material.
  • Current-head and hardware coverage: Run the current-head suite and required Tier 1 checks, including a Linux RADV or Steam Deck headless run and alternate GPU presentation coverage.
  • Rebase verification: Verify conflict resolution, colour-vision correction and screenshots on the exact rebased head.

Nothing was built or run: check the change. It will be reviewed again.


Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.

I changed this pull request where the review's points could be done here (the rest needs runs, measurements or decisions): the new head is 30fc1bf23d1f. It was already up to date with `main`. **Changed** - **Windows non-ASCII save paths, host/oot_autoload.c:187**: Changed the Windows save search to use wide filesystem calls and added non-ASCII path coverage in host/oot_autoload.c:39 and tests/test_skip_title.py:72. The author should run the coverage on Windows. **Left for the author** - **HUD overhead measurement, gpu/vk/vk_menu.cpp:3147**: Measure warmed, interleaved HUD-off and HUD-on runs against the base, including CPU and GPU cost, frame times and 1% lows; reduce or isolate any material distortion. - **Capture cost measurement, gpu/vk/vk_replay.cpp:5097**: Measure readback and PNG timing, memory use and frame pacing at the intended capture cadence; adjust the capture strategy if disturbance is material. - **Current-head and hardware coverage**: Run the current-head suite and required Tier 1 checks, including a Linux RADV or Steam Deck headless run and alternate GPU presentation coverage. - **Rebase verification**: Verify conflict resolution, colour-vision correction and screenshots on the exact rebased head. Nothing was built or run: check the change. It will be reviewed again. --- <sub>Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.</sub>
fix(host): find saves in UTF-8 Windows paths
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 12s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 21s
CI / Workflow lint (actionlint) (pull_request) Successful in 23s
Third-party parts / As their notes say (pull_request) Successful in 28s
PR description / Description follows the template (pull_request_target) Failing after 7s
CI / Website build and checks (pull_request) Has been cancelled
30fc1bf23d
As the review on #1539 asked.
BeanBot left a comment

Recommendation: Request changes. The Windows save-search fix still fails for valid paths whose UTF-8 representation exceeds its byte buffers. The previously identified HUD and capture costs also remain unmeasured.

Disposable test launches copy a checkpoint, enter play, display test context and Dev metrics, and produce annotated bridge captures with clean companions. Since the last review, wide Windows filesystem calls and a Unicode-path fixture were added, but the fix does not cover longer valid UTF-8 paths. The feature is useful for contributors; its measurement overhead remains an open cost.

Does it help: Unproven. The author reports successful windowed and offscreen runs, clean companion captures, HDR10 and generated frames on Windows 11 with an RTX 4070 at b324a3cb. There are no matched HUD-on/off or capture-cost numbers, and no runs are reported for head 30fc1bf.
Other systems: The change affects Windows and Linux, windowed and offscreen presentation, bridge and periodic captures, frame generation, HDR, colour-vision correction and title auto-continue. Earlier Windows NVIDIA runs provide partial evidence; Linux RADV or Steam Deck, alternate GPU presentation, Tier 1 coverage and the current head remain unverified.
Switches: HOS_TEST_AUTOLOAD defaults off and is enabled by agent launches. RX_VK_TEST_AGENT defaults unset and enables the HUD; RX_VK_TEST_LABEL defaults to Gameplay test. All have switch-table rows. Ordinary launches remain opt-in, but the cost of the enabled configuration has not been measured.

Reviewed: 30fc1bf23d against 3df5c13ed2: Read the base AGENTS.md, CONTRIBUTING.md, PR template, standards, complete diff, commit list and host/oot_autoload.c. Reviewed the changed launcher setup and selected renderer capture, UI and locking paths; skimmed changed tests, documentation and reported runs. The full changed renderer, launcher and test files and all dependent paths were not read. The adjacent Tooling AGENTS.md and docs/review-checklist.md were absent.; nothing run (a code review).
Checked and fine:

  • Agent launches require both a public identifier and label and a selected checkpoint.
  • Bridge captures copy clean pixels before adding the test overlay.
  • The screenshot UI lock is attempted after inline frame generation.
  • The new fixture exercises non-ASCII user and profile names.

Blocking (to change before merge)

  1. host/oot_autoload.c:118: The wide filesystem calls still receive paths assembled in MAX_PATH-byte UTF-8 buffers. A valid Windows run path containing enough multibyte characters can fit within Windows' character limit but exceed 259 UTF-8 bytes after the save suffix is added; truncation then prevents finding the copied checkpoint and the launch waits until timeout. Size paths for UTF-8 bytes or construct them dynamically, and add a Windows fixture covering that case.
  2. The previous review's HUD overhead and second-readback/PNG cost findings remain open: measure matched warmed runs, including frame times and 1% lows, capture timing and memory, and reduce or isolate material distortion.

Suggestions (the author may take or leave)

  • Complete the current-head suite, Linux RADV or Steam Deck headless run, alternate presentation and required Tier 1 checks.
  • Verify colour-vision correction and screenshots on the exact rebased head.

Testing: No builds, tests, scripts or game runs were executed. The reported gameplay evidence is from an earlier head.
Overlaps and merge order: Unknown; the network is closed. The description says Beanstalk runner #89 is merged.


Automated code review by BeanBot (gpt-6.1-terra), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.

**Recommendation: Request changes.** The Windows save-search fix still fails for valid paths whose UTF-8 representation exceeds its byte buffers. The previously identified HUD and capture costs also remain unmeasured. Disposable test launches copy a checkpoint, enter play, display test context and Dev metrics, and produce annotated bridge captures with clean companions. Since the last review, wide Windows filesystem calls and a Unicode-path fixture were added, but the fix does not cover longer valid UTF-8 paths. The feature is useful for contributors; its measurement overhead remains an open cost. **Does it help:** Unproven. The author reports successful windowed and offscreen runs, clean companion captures, HDR10 and generated frames on Windows 11 with an RTX 4070 at b324a3cb. There are no matched HUD-on/off or capture-cost numbers, and no runs are reported for head 30fc1bf. **Other systems:** The change affects Windows and Linux, windowed and offscreen presentation, bridge and periodic captures, frame generation, HDR, colour-vision correction and title auto-continue. Earlier Windows NVIDIA runs provide partial evidence; Linux RADV or Steam Deck, alternate GPU presentation, Tier 1 coverage and the current head remain unverified. **Switches:** HOS_TEST_AUTOLOAD defaults off and is enabled by agent launches. RX_VK_TEST_AGENT defaults unset and enables the HUD; RX_VK_TEST_LABEL defaults to Gameplay test. All have switch-table rows. Ordinary launches remain opt-in, but the cost of the enabled configuration has not been measured. **Reviewed:** 30fc1bf23d1f against 3df5c13ed2ec: Read the base AGENTS.md, CONTRIBUTING.md, PR template, standards, complete diff, commit list and host/oot_autoload.c. Reviewed the changed launcher setup and selected renderer capture, UI and locking paths; skimmed changed tests, documentation and reported runs. The full changed renderer, launcher and test files and all dependent paths were not read. The adjacent Tooling AGENTS.md and docs/review-checklist.md were absent.; nothing run (a code review). **Checked and fine:** - Agent launches require both a public identifier and label and a selected checkpoint. - Bridge captures copy clean pixels before adding the test overlay. - The screenshot UI lock is attempted after inline frame generation. - The new fixture exercises non-ASCII user and profile names. **Blocking** (to change before merge) 1. host/oot_autoload.c:118: The wide filesystem calls still receive paths assembled in MAX_PATH-byte UTF-8 buffers. A valid Windows run path containing enough multibyte characters can fit within Windows' character limit but exceed 259 UTF-8 bytes after the save suffix is added; truncation then prevents finding the copied checkpoint and the launch waits until timeout. Size paths for UTF-8 bytes or construct them dynamically, and add a Windows fixture covering that case. 2. The previous review's HUD overhead and second-readback/PNG cost findings remain open: measure matched warmed runs, including frame times and 1% lows, capture timing and memory, and reduce or isolate material distortion. **Suggestions** (the author may take or leave) - Complete the current-head suite, Linux RADV or Steam Deck headless run, alternate presentation and required Tier 1 checks. - Verify colour-vision correction and screenshots on the exact rebased head. **Testing:** No builds, tests, scripts or game runs were executed. The reported gameplay evidence is from an earlier head. **Overlaps and merge order:** Unknown; the network is closed. The description says Beanstalk runner #89 is merged. --- <sub>Automated code review by BeanBot (gpt-6.1-terra), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.</sub>
@ -96,3 +117,2 @@
path_part(pattern, sizeof pattern, SEP_S "save" SEP_S "*");
WIN32_FIND_DATAA fd;
HANDLE h = FindFirstFileA(pattern, &fd);
WIN32_FIND_DATAW fd;
Member

The wide call still receives a path built in a 260-byte UTF-8 buffer. A valid Windows path with multibyte characters can fit in 260 wide characters while its UTF-8 bytes plus the save suffix exceed this buffer, so the copied checkpoint is not found and launch times out. Please size these paths for UTF-8 bytes or build them dynamically, and cover that case in the Windows fixture.

The wide call still receives a path built in a 260-byte UTF-8 buffer. A valid Windows path with multibyte characters can fit in 260 wide characters while its UTF-8 bytes plus the save suffix exceed this buffer, so the copied checkpoint is not found and launch times out. Please size these paths for UTF-8 bytes or build them dynamically, and cover that case in the Windows fixture.
HumanBean changed title from WIP: feat(testing): show test context and Dev metrics in agent runs to feat(testing): show test context and Dev metrics in agent runs 2026-11-20 22:54:16 +00:00
Owner
@fix
fix(host): size Windows save paths for UTF-8
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 20s
PR description / Description follows the template (pull_request_target) Failing after 31s
CI / Workflow lint (actionlint) (pull_request) Successful in 42s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 1m1s
Third-party parts / As their notes say (pull_request) Successful in 55s
CI / Website build and checks (pull_request) Has been cancelled
8abda60d40
As the review on #1539 asked.
Member

I changed this pull request where the review's points could be done here (the rest needs runs, measurements or decisions): the new head is 8abda60d40.

Changed

  • Long UTF-8 Windows save paths, host/oot_autoload.c:118: Sized Windows UTF-8 search and output buffers for MAX_PATH UTF-16 units in host/oot_autoload.c:20; added a long multibyte path and autoload fixture in tests/test_skip_title.py:78.
  • host/oot_autoload.c:21: Save search fails beyond MAX_PATH: Aligned path buffers and extended-length conversion with the game filesystem in host/oot_autoload.c:20 and host/oot_autoload.c:43. Added a Windows fixture for a path over 260 UTF-16 characters in tests/test_skip_title.py:222. Nothing was run.

Already done

  • Windows Unicode save search, host/oot_autoload.c:187: The checkout uses wide Windows filesystem calls and has a non-ASCII user and profile fixture in tests/test_skip_title.py:185.

Left for the author

  • HUD overhead measurement, gpu/vk/vk_menu.cpp:3147: Measure warmed, interleaved HUD-off and HUD-on runs against the base, including CPU and GPU cost, frame times, and 1% lows; reduce or isolate any material distortion based on the results.
  • Capture cost measurement, gpu/vk/vk_replay.cpp:5097: Measure readback and PNG timing, memory use, and frame pacing at the intended capture cadence; adjust the capture strategy if the disturbance is material.
  • Current-head and hardware coverage: Run the current-head suite, a Linux RADV or Steam Deck headless run, alternate presentation coverage, and the required Tier 1 checks.
  • Verify rebased colour correction and screenshots: Verify colour-vision correction and screenshots on the exact current head.

Nothing was built or run: check the change. It will be reviewed again.


Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.

I changed this pull request where the review's points could be done here (the rest needs runs, measurements or decisions): the new head is 8abda60d408d. **Changed** - **Long UTF-8 Windows save paths, host/oot_autoload.c:118**: Sized Windows UTF-8 search and output buffers for MAX_PATH UTF-16 units in host/oot_autoload.c:20; added a long multibyte path and autoload fixture in tests/test_skip_title.py:78. - **host/oot_autoload.c:21: Save search fails beyond MAX_PATH**: Aligned path buffers and extended-length conversion with the game filesystem in host/oot_autoload.c:20 and host/oot_autoload.c:43. Added a Windows fixture for a path over 260 UTF-16 characters in tests/test_skip_title.py:222. Nothing was run. **Already done** - **Windows Unicode save search, host/oot_autoload.c:187**: The checkout uses wide Windows filesystem calls and has a non-ASCII user and profile fixture in tests/test_skip_title.py:185. **Left for the author** - **HUD overhead measurement, gpu/vk/vk_menu.cpp:3147**: Measure warmed, interleaved HUD-off and HUD-on runs against the base, including CPU and GPU cost, frame times, and 1% lows; reduce or isolate any material distortion based on the results. - **Capture cost measurement, gpu/vk/vk_replay.cpp:5097**: Measure readback and PNG timing, memory use, and frame pacing at the intended capture cadence; adjust the capture strategy if the disturbance is material. - **Current-head and hardware coverage**: Run the current-head suite, a Linux RADV or Steam Deck headless run, alternate presentation coverage, and the required Tier 1 checks. - **Verify rebased colour correction and screenshots**: Verify colour-vision correction and screenshots on the exact current head. Nothing was built or run: check the change. It will be reviewed again. --- <sub>Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.</sub>
BeanBot left a comment

Recommendation: Request changes. The rebased UI fixture cannot compile, and an annotated capture can overwrite an existing file the caller never selected. Disposable launches also still fail to find valid checkpoints in sufficiently deep Linux run folders.

Contributors can launch a copied checkpoint, see the run context and Dev metrics, and collect an annotated screenshot with a clean image for comparison. The launcher reuses the existing test profile and title-skip logic. Since the last review, the Windows save search gained larger UTF-8 buffers, extended-length path handling, and fixtures for both long-path cases. The HUD and capture add work to the gameplay being measured.

Does it help: Unproven on this head. The author reports five successful Windows sessions on an RTX 4070 at b324a3cb, covering windowed and offscreen play, clean companions, HDR10 with 2× frame generation, and ordinary-launch controls. The earlier lock check reportedly failed before the fix and passed afterward. There are no matched, warmed, interleaved measurements against 3df5c13ed2 for FPS, 1% lows, frame times, readback, PNG time, or memory.
Other systems: Ordinary screenshots and bug reports still request clean pixels. The live window retains the colour-vision argument through UI recording. Agent mode affects title input, disables the audio device, and changes windowed, headless, HDR, generated-frame, and periodic captures. Periodic glitch detection now receives annotated pixels in agent mode, which can change its black-pixel and flash results. Current-head Linux RADV, Steam Deck, AMD Windows, Linux NVIDIA, and alternate presentation coverage remain open.
Switches: HOS_TEST_AUTOLOAD defaults off and 1 enables title auto-continue. RX_VK_TEST_AGENT defaults unset and enables the HUD and forced Dev metrics when set. RX_VK_TEST_LABEL defaults to Gameplay test. All three have switch-table rows. The new --agent and --test-label options default unset and must be paired with a selected checkpoint. Agent launches default to a window, respect --headless, and set HOS_AUDIO=0 through the test profile. The opt-in defaults suit a contributor workflow. The enabled mode's timing cost is still unmeasured.

Reviewed: 8abda60d40 against 3df5c13ed2: Reviewed 8abda60d40 against 3df5c13ed2, the commit range, supplied comments and earlier reviews, base AGENTS.md, CONTRIBUTING.md, the Forgejo template, standards, and relevant architecture rules. Read host/oot_autoload.c and the UI recording fixture in full. Reviewed changed hunks and followed the launcher setup, bridge request, screenshot readback, presenter lock, and submission paths. Skimmed the remaining test fixtures and reported runs. Full renderer, launcher, documentation, and remaining test files were not read. The review checklist and adjacent Tooling checkout were absent. The two omitted comments, screenshot attachments, and Graphify maps were unavailable.; nothing run (a code review).
Checked and fine:

  • The title and functional diff cover one contributor-testing workflow. The diff adds no game data, vendor SDKs, generated output, or third-party edits. The renderer API and shaders are untouched. The feature commit records AI assistance.
  • Launch requires both labels and one checkpoint, rejects external user and data folders for the disposable profile, and waits for play.
  • Bridge readback copies clean pixels before annotation, checks timeline completion, and invalidates mapped memory before reading. Bug reports explicitly request the clean path.
  • The capture attempts the presenter mutex after inline frame generation can release its dependency on the display thread.
  • The Windows save-search changes address the previous byte-buffer finding in source. The new fixtures exercise multibyte paths and paths beyond 260 UTF-16 units.

Blocking (to change before merge)

  1. tests/test_ui_record.py:67 declares a six-argument record_to, while the extracted production function takes seven arguments and references colour-vision types and draw-list fields absent from the fixture. The game-free UI recording check cannot compile on this head. Match the production signature, supply the colour-vision and draw-list dependencies, and exercise both window and capture recording.
  2. gpu/vk/vk_replay.cpp:5133 derives a second destination from the caller's path and replaces any file at that destination. Requesting forest.png can silently replace an existing forest.png.scene.png, and a failed primary write can leave a newly published companion. Reserve a distinct companion destination, return it in the reply, and clean up this request's output on failure.
  3. host/oot_autoload.c:28 leaves Linux save discovery limited to 260 bytes. A valid nested --root can produce a copied progress.sav whose full path exceeds 259 bytes, so discovery silently truncates it and launch waits for play until timeout. Use checked path construction that supports the game filesystem's path range on Linux and add a long Linux run-folder fixture.
  4. The previous review's HUD and capture cost findings remain open. The project requires measured costs for work added to gameplay. No current-head result shows whether forced Dev metrics or the extra capture work distort the comparisons this feature supports.

Suggestions (the author may take or leave)

  • Document that an agent bridge shot can skip the first eligible present when the presenter holds the UI lock.
  • Complete the Graphics guide source review and the current-head platform, colour-vision, presentation, and CI checks listed in the PR.

Testing: Read-only review. No builds, tests, scripts, or game runs were executed. The author named the auto-profile failure on both the candidate and main at the earlier rebase. The full suite and Windows path fixtures have no reported run on 8abda60d40. The ready checklist remains incomplete.
Overlaps and merge order: The description says Beanstalk runner #89 is merged. Other overlaps and the cause of the reported mergeability failure could not be verified from the offline checkout.


Automated code review by BeanBot (gpt-6.2-sol), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.

**Recommendation: Request changes.** The rebased UI fixture cannot compile, and an annotated capture can overwrite an existing file the caller never selected. Disposable launches also still fail to find valid checkpoints in sufficiently deep Linux run folders. Contributors can launch a copied checkpoint, see the run context and Dev metrics, and collect an annotated screenshot with a clean image for comparison. The launcher reuses the existing test profile and title-skip logic. Since the last review, the Windows save search gained larger UTF-8 buffers, extended-length path handling, and fixtures for both long-path cases. The HUD and capture add work to the gameplay being measured. **Does it help:** Unproven on this head. The author reports five successful Windows sessions on an RTX 4070 at b324a3cb, covering windowed and offscreen play, clean companions, HDR10 with 2× frame generation, and ordinary-launch controls. The earlier lock check reportedly failed before the fix and passed afterward. There are no matched, warmed, interleaved measurements against 3df5c13ed2ec for FPS, 1% lows, frame times, readback, PNG time, or memory. **Other systems:** Ordinary screenshots and bug reports still request clean pixels. The live window retains the colour-vision argument through UI recording. Agent mode affects title input, disables the audio device, and changes windowed, headless, HDR, generated-frame, and periodic captures. Periodic glitch detection now receives annotated pixels in agent mode, which can change its black-pixel and flash results. Current-head Linux RADV, Steam Deck, AMD Windows, Linux NVIDIA, and alternate presentation coverage remain open. **Switches:** `HOS_TEST_AUTOLOAD` defaults off and `1` enables title auto-continue. `RX_VK_TEST_AGENT` defaults unset and enables the HUD and forced Dev metrics when set. `RX_VK_TEST_LABEL` defaults to `Gameplay test`. All three have switch-table rows. The new `--agent` and `--test-label` options default unset and must be paired with a selected checkpoint. Agent launches default to a window, respect `--headless`, and set `HOS_AUDIO=0` through the test profile. The opt-in defaults suit a contributor workflow. The enabled mode's timing cost is still unmeasured. **Reviewed:** 8abda60d408d against 3df5c13ed2ec: Reviewed 8abda60d408d against 3df5c13ed2ec, the commit range, supplied comments and earlier reviews, base AGENTS.md, CONTRIBUTING.md, the Forgejo template, standards, and relevant architecture rules. Read host/oot_autoload.c and the UI recording fixture in full. Reviewed changed hunks and followed the launcher setup, bridge request, screenshot readback, presenter lock, and submission paths. Skimmed the remaining test fixtures and reported runs. Full renderer, launcher, documentation, and remaining test files were not read. The review checklist and adjacent Tooling checkout were absent. The two omitted comments, screenshot attachments, and Graphify maps were unavailable.; nothing run (a code review). **Checked and fine:** - The title and functional diff cover one contributor-testing workflow. The diff adds no game data, vendor SDKs, generated output, or third-party edits. The renderer API and shaders are untouched. The feature commit records AI assistance. - Launch requires both labels and one checkpoint, rejects external user and data folders for the disposable profile, and waits for play. - Bridge readback copies clean pixels before annotation, checks timeline completion, and invalidates mapped memory before reading. Bug reports explicitly request the clean path. - The capture attempts the presenter mutex after inline frame generation can release its dependency on the display thread. - The Windows save-search changes address the previous byte-buffer finding in source. The new fixtures exercise multibyte paths and paths beyond 260 UTF-16 units. **Blocking** (to change before merge) 1. `tests/test_ui_record.py:67` declares a six-argument `record_to`, while the extracted production function takes seven arguments and references colour-vision types and draw-list fields absent from the fixture. The game-free UI recording check cannot compile on this head. Match the production signature, supply the colour-vision and draw-list dependencies, and exercise both window and capture recording. 2. `gpu/vk/vk_replay.cpp:5133` derives a second destination from the caller's path and replaces any file at that destination. Requesting `forest.png` can silently replace an existing `forest.png.scene.png`, and a failed primary write can leave a newly published companion. Reserve a distinct companion destination, return it in the reply, and clean up this request's output on failure. 3. `host/oot_autoload.c:28` leaves Linux save discovery limited to 260 bytes. A valid nested `--root` can produce a copied `progress.sav` whose full path exceeds 259 bytes, so discovery silently truncates it and `launch` waits for play until timeout. Use checked path construction that supports the game filesystem's path range on Linux and add a long Linux run-folder fixture. 4. The previous review's HUD and capture cost findings remain open. The project requires measured costs for work added to gameplay. No current-head result shows whether forced Dev metrics or the extra capture work distort the comparisons this feature supports. **Suggestions** (the author may take or leave) - Document that an agent bridge shot can skip the first eligible present when the presenter holds the UI lock. - Complete the Graphics guide source review and the current-head platform, colour-vision, presentation, and CI checks listed in the PR. **Testing:** Read-only review. No builds, tests, scripts, or game runs were executed. The author named the `auto-profile` failure on both the candidate and main at the earlier rebase. The full suite and Windows path fixtures have no reported run on 8abda60d408d. The ready checklist remains incomplete. **Overlaps and merge order:** The description says Beanstalk runner #89 is merged. Other overlaps and the cause of the reported mergeability failure could not be verified from the offline checkout. --- <sub>Automated code review by BeanBot (gpt-6.2-sol), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.</sub>
@ -224,3 +224,2 @@
A PNG of the first present after the request (or after frame `after`), at the size `RX_VK_SHOTS` writes (1280x720, or
`RX_VK_SHOT_W` x `RX_VK_SHOT_H`). It shows the game's picture without the Insert menu or FPS counter. The file is
written under another name and renamed, so the reply's path is complete.
`RX_VK_SHOT_W` x `RX_VK_SHOT_H`). A regular shot shows the game's picture. A [disposable agent test](#disposable-agent-tests)
Member

Agent bridge shots can also skip the first eligible present when the presenter holds the UI lock. Document that alongside the periodic capture deferral.

Agent bridge shots can also skip the first eligible present when the presenter holds the UI lock. Document that alongside the periodic capture deferral.
@ -5061,0 +5130,4 @@
std::thread([px, a, at, bytes] {
// Keep the untouched pixels from this exact frame for automated image comparisons.
const bool annotated = px->size() > bytes;
const bool ok = (!annotated || write_shot(a.path + ".scene.png", px->data())) &&
Member

A request for forest.png can overwrite an existing forest.png.scene.png. Reserve a distinct companion path, return it to the caller, and clean it up if the primary write fails.

A request for `forest.png` can overwrite an existing `forest.png.scene.png`. Reserve a distinct companion path, return it to the caller, and clean it up if the primary write fails.
@ -22,3 +25,4 @@
#ifndef MAX_PATH
#define MAX_PATH 260
#endif
#define SAVE_PATH_SIZE MAX_PATH
Member

A deep Linux --root can put the copied checkpoint beyond this 260-byte buffer. Discovery truncates the path and launch times out waiting for play. Match the game filesystem's capacity and cover this case on Linux.

A deep Linux `--root` can put the copied checkpoint beyond this 260-byte buffer. Discovery truncates the path and launch times out waiting for play. Match the game filesystem's capacity and cover this case on Linux.
@ -65,3 +65,3 @@
}
void vkCmdEndRendering(VkCommandBuffer) { calls += "]"; }
''' + block(src, "void layout(VkCommandBuffer cmd,") + "\n" + block(src, "void record_to(VkCommandBuffer cmd,") + "\n" + block(src, "void record(VkCommandBuffer cmd,") + r'''
bool record_to(VkCommandBuffer, VkImage, VkImageView, uint32_t, uint32_t, bool screenshot = false);
Member

The extracted function takes seven arguments and uses colour-vision and draw-list types missing here. Update the fixture to compile and exercise both recording paths.

The extracted function takes seven arguments and uses colour-vision and draw-list types missing here. Update the fixture to compile and exercise both recording paths.
Owner
@fix
Member

I changed this pull request where the review's points could be done here (the rest needs runs, measurements or decisions): the new head is d30aa19142.

Changed

  • UI recording fixture: tests/test_ui_record.py:67: Updated the fixture signatures and draw-list types and added colour-correction coverage for window and screenshot recording in tests/test_ui_record.py:35.
  • Companion screenshot collision: gpu/vk/vk_replay.cpp:5133: Reserved a distinct companion name, returned it as scene_path and removed it if the primary write fails in gpu/vk/vk_replay.cpp:5133 and host/control.c:1127.
  • Deep Linux save paths: host/oot_autoload.c:28: Matched the game filesystem's 1024-byte path capacity in host/oot_autoload.c:25 and added deep-path discovery and autoload coverage in tests/test_skip_title.py:93.
  • Bridge capture deferral: docs/dev/reference/control-bridge.md:225: Documented that an agent bridge shot can defer while the presenter holds the UI lock in docs/dev/reference/control-bridge.md:228.
  • gpu/vk/vk_replay.cpp:5143: companion reservation race: Publish completed companion PNGs without replacing an existing file and retry on collision in gpu/vk/vk_replay.cpp:4601. Added publication coverage in tests/png_publication_test.cpp:112. No builds or tests were run.
  • gpu/vk/vk_replay.cpp:5148: Companion filename length: Uses a short numbered companion name in the same directory at gpu/vk/vk_replay.cpp:5157 and updates the bridge documentation.
  • gpu/vk/vk_replay.cpp:5155: Concurrent companion overwrite: Serializes bridge screenshot publication, cleanup and completion at gpu/vk/vk_replay.cpp:5143, keeping the companion owned through completion.
  • tests/png_publication_test.cpp:54: Windows publication mock: Defines the Windows error constants and models no-replace moves at tests/png_publication_test.cpp:59. No builds or tests were run, as requested.

Already done

  • Windows Unicode save search: host/oot_autoload.c:187: The search uses wide Windows filesystem calls in host/oot_autoload.c:101, with non-ASCII path coverage in tests/test_skip_title.py:74.
  • Windows multibyte and extended-length paths: host/oot_autoload.c:118: The search uses 1024-byte buffers and extended-length conversion in host/oot_autoload.c:21, with both path cases covered in tests/test_skip_title.py:79.

Left for the author

  • HUD overhead: gpu/vk/vk_menu.cpp:3147: Measure warmed, interleaved HUD-off and HUD-on runs against the base, including CPU and GPU cost, frame times and 1% lows; use the results to determine whether the cost needs reducing or isolating.
  • Capture overhead: gpu/vk/vk_replay.cpp:5097: Measure readback and PNG timing, memory use and frame pacing at the intended capture cadence against runs without captures; use the results to determine whether the strategy needs adjustment.
  • Current-head and hardware coverage: Run the current-head suite and required Tier 1 checks, including Linux RADV or Steam Deck headless and alternate presentation coverage.
  • Rebase verification: Verify colour-vision correction and screenshots on the exact current head, including the rebase conflict resolution.

Nothing was built or run: check the change. It will be reviewed again.


Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.

I changed this pull request where the review's points could be done here (the rest needs runs, measurements or decisions): the new head is d30aa1914249. **Changed** - **UI recording fixture: tests/test_ui_record.py:67**: Updated the fixture signatures and draw-list types and added colour-correction coverage for window and screenshot recording in tests/test_ui_record.py:35. - **Companion screenshot collision: gpu/vk/vk_replay.cpp:5133**: Reserved a distinct companion name, returned it as scene_path and removed it if the primary write fails in gpu/vk/vk_replay.cpp:5133 and host/control.c:1127. - **Deep Linux save paths: host/oot_autoload.c:28**: Matched the game filesystem's 1024-byte path capacity in host/oot_autoload.c:25 and added deep-path discovery and autoload coverage in tests/test_skip_title.py:93. - **Bridge capture deferral: docs/dev/reference/control-bridge.md:225**: Documented that an agent bridge shot can defer while the presenter holds the UI lock in docs/dev/reference/control-bridge.md:228. - **gpu/vk/vk_replay.cpp:5143: companion reservation race**: Publish completed companion PNGs without replacing an existing file and retry on collision in gpu/vk/vk_replay.cpp:4601. Added publication coverage in tests/png_publication_test.cpp:112. No builds or tests were run. - **gpu/vk/vk_replay.cpp:5148: Companion filename length**: Uses a short numbered companion name in the same directory at gpu/vk/vk_replay.cpp:5157 and updates the bridge documentation. - **gpu/vk/vk_replay.cpp:5155: Concurrent companion overwrite**: Serializes bridge screenshot publication, cleanup and completion at gpu/vk/vk_replay.cpp:5143, keeping the companion owned through completion. - **tests/png_publication_test.cpp:54: Windows publication mock**: Defines the Windows error constants and models no-replace moves at tests/png_publication_test.cpp:59. No builds or tests were run, as requested. **Already done** - **Windows Unicode save search: host/oot_autoload.c:187**: The search uses wide Windows filesystem calls in host/oot_autoload.c:101, with non-ASCII path coverage in tests/test_skip_title.py:74. - **Windows multibyte and extended-length paths: host/oot_autoload.c:118**: The search uses 1024-byte buffers and extended-length conversion in host/oot_autoload.c:21, with both path cases covered in tests/test_skip_title.py:79. **Left for the author** - **HUD overhead: gpu/vk/vk_menu.cpp:3147**: Measure warmed, interleaved HUD-off and HUD-on runs against the base, including CPU and GPU cost, frame times and 1% lows; use the results to determine whether the cost needs reducing or isolating. - **Capture overhead: gpu/vk/vk_replay.cpp:5097**: Measure readback and PNG timing, memory use and frame pacing at the intended capture cadence against runs without captures; use the results to determine whether the strategy needs adjustment. - **Current-head and hardware coverage**: Run the current-head suite and required Tier 1 checks, including Linux RADV or Steam Deck headless and alternate presentation coverage. - **Rebase verification**: Verify colour-vision correction and screenshots on the exact current head, including the rebase conflict resolution. Nothing was built or run: check the change. It will be reviewed again. --- <sub>Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.</sub>
fix(test): preserve agent captures and support deep save paths
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
PR description / Description follows the template (pull_request_target) Failing after 22s
CI / Conventional Commit title (pull_request) Successful in 22s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 32s
CI / Workflow lint (actionlint) (pull_request) Successful in 33s
Third-party parts / As their notes say (pull_request) Successful in 35s
CI / Website build and checks (pull_request) Has been cancelled
d30aa19142
As the review on #1539 asked.
BeanBot left a comment

Recommendation: Request changes. The changed screenshot callback still advertises the old renderer ABI, so a mixed host and renderer can pass the version check and crash on a screenshot. Companion publication also repeats a full PNG write for each occupied name and fails on Linux filesystems without hard-link support.

Contributors can launch a copied checkpoint, see the run context and Dev metrics, and collect an annotated screenshot alongside clean pixels from the same present. Since review 1932, the branch updates the UI fixture, covers deeper Linux save paths, documents capture deferral, and adds no-replace publication for companions. That protects occupied companion names during companion creation, but an explicit primary shot can still overwrite a previously returned companion.

Does it help: Unproven on d30aa19142. The author reports five successful Windows 11 sessions on an RTX 4070 with driver 621.01 at b324a3cb, plus a lock fixture that failed on the test-only regression commit and passed on b324a3cb. The prior HUD and capture cost findings remain open. There are no warmed, interleaved timing runs against 3df5c13ed2 for ordinary launches, HUD-on versus HUD-off gameplay, or the intended capture cadence. Average FPS, 1% lows, frame times, readback and PNG times, and RAM and VRAM costs are missing.
Other systems: The renderer ABI affects separate host and renderer builds on both platforms. Agent mode affects windowed and offscreen UI, HDR, generated frames, periodic capture and title input. The new windowed agent launch inherits the test profile’s audio disable. Ordinary screenshots and bug reports still request clean pixels, while periodic glitch detection in agent mode can see overlay pixels. Earlier Windows NVIDIA runs cover some of these paths. Exact-head colour correction, Linux RADV and Steam Deck, and the remaining Tier 1 GPU and OS coverage are open.
Switches: HOS_TEST_AUTOLOAD defaults off. RX_VK_TEST_AGENT defaults unset and enables the banner and forced Dev metrics when supplied. RX_VK_TEST_LABEL defaults to Gameplay test. All three have switch-table rows. The paired --agent and --test-label options default unset and require a checkpoint. Agent launches default to a window and honor --headless. The opt-in defaults suit contributor testing. Their enabled costs and ordinary-launch timing still need measurements.

Reviewed: d30aa19142 against 3df5c13ed2: Reviewed d30aa19142 against 3df5c13ed2, the commits, supplied comments and reviews, base AGENTS.md, CONTRIBUTING.md, the Forgejo template, and standards. Read host/oot_autoload.c, gpu/vk/vk_imgui.h, tests/test_capture_resources.py, tests/test_fps_overlay_order.py, tests/test_png_publication.py, tests/png_publication_test.cpp and tests/capture_dimensions_test.cpp in full. Reviewed changed regions elsewhere and traced renderer loading, screenshot callbacks, readback, presenter locking and launcher setup. Skimmed reported runs and other fixture context. Full renderer, bridge, launcher and remaining test files were not read. The review checklist and adjacent Tooling checkout were absent. Omitted comments, attachment images and Graphify maps were unavailable.; nothing run (a code review).
Checked and fine:

  • The Conventional Commit title and functional diff cover one contributor workflow. No shaders, generated output, game data, vendor SDKs or third-party files changed. The feature commit records AI assistance.
  • The launcher requires paired labels and a selected checkpoint. The fixture checks that only the selected slot is copied and exercises both windowed and headless setup.
  • The save search uses wide Windows calls, and the fixtures cover multibyte and extended-length Windows paths and deeper Linux paths supported by the game filesystem.
  • The bridge records the clean copy before annotation and checks timeline completion before invalidating and reading mapped memory. The presenter lock is attempted after inline frame generation.
  • The UI recording fixture now matches the extracted signature and exercises colour correction at the recording layer. The PNG fixture checks that no-replace publication preserves an occupied destination.

Blocking (to change before merge)

  1. gpu/vk/vk_api.h:414: MkwVkRenderer::shot adds a fourth completion argument while MKWVK_API_VERSION stays at 36. An updated host can accept an older renderer, which calls shot_done with no scene pointer. The host may then dereference an indeterminate pointer while constructing the reply. Bump the API version and rebuild both binaries together, as the architecture and PR checklist require.
  2. gpu/vk/vk_replay.cpp:5164: Every companion search starts at zero and encodes and writes a fresh temporary PNG before discovering each collision. With 100 retained companions in the output folder, the next 1280×720 capture writes roughly 276 MB in failed attempts alone. That work holds the publication mutex and delays other captures. Encode once, retry publication of the completed temporary under fresh names, and avoid restarting the search at zero.
  3. gpu/vk/vk_replay.cpp:4608: Linux companion publication requires hard links. An explicit bridge shot to a writable exFAT output folder can currently use the regular rename path, but the same destination fails in agent mode before the primary is written. Use no-replace publication that works on rename-capable filesystems without hard links and cover the unsupported-link case.
  4. gpu/vk/vk_replay.cpp:5169: An explicit primary path can replace a previously returned clean companion. For example, after shot A returns longshot-scene-0.png, shot B can use that path for its annotated primary. A later comparison of A then reads B’s annotated pixels. Keep companion names in a renderer-owned namespace that primary paths cannot replace, including through path aliases, and test two requests targeting the same namespace.

Suggestions (the author may take or leave)

  • Record scene_path in the session manifest and list it in the bridge command table so artifact consumers can find the clean image after the run.

Testing: Read-only review. The diff whitespace check passed. No builds, fixtures, project scripts or game sessions were run. The reported Windows sessions and baseline auto-profile failure predate this head. Exact-head CI, the game-free suite, Tier 1 hardware results and the Graphics guide source review remain outstanding.
Overlaps and merge order: The PR reports Beanstalk #89 already merged. Other overlaps and merge order could not be verified with the network closed.


Automated code review by BeanBot (gpt-6.2-sol), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.

**Recommendation: Request changes.** The changed screenshot callback still advertises the old renderer ABI, so a mixed host and renderer can pass the version check and crash on a screenshot. Companion publication also repeats a full PNG write for each occupied name and fails on Linux filesystems without hard-link support. Contributors can launch a copied checkpoint, see the run context and Dev metrics, and collect an annotated screenshot alongside clean pixels from the same present. Since review 1932, the branch updates the UI fixture, covers deeper Linux save paths, documents capture deferral, and adds no-replace publication for companions. That protects occupied companion names during companion creation, but an explicit primary shot can still overwrite a previously returned companion. **Does it help:** Unproven on d30aa1914249. The author reports five successful Windows 11 sessions on an RTX 4070 with driver 621.01 at b324a3cb, plus a lock fixture that failed on the test-only regression commit and passed on b324a3cb. The prior HUD and capture cost findings remain open. There are no warmed, interleaved timing runs against 3df5c13ed2ec for ordinary launches, HUD-on versus HUD-off gameplay, or the intended capture cadence. Average FPS, 1% lows, frame times, readback and PNG times, and RAM and VRAM costs are missing. **Other systems:** The renderer ABI affects separate host and renderer builds on both platforms. Agent mode affects windowed and offscreen UI, HDR, generated frames, periodic capture and title input. The new windowed agent launch inherits the test profile’s audio disable. Ordinary screenshots and bug reports still request clean pixels, while periodic glitch detection in agent mode can see overlay pixels. Earlier Windows NVIDIA runs cover some of these paths. Exact-head colour correction, Linux RADV and Steam Deck, and the remaining Tier 1 GPU and OS coverage are open. **Switches:** HOS_TEST_AUTOLOAD defaults off. RX_VK_TEST_AGENT defaults unset and enables the banner and forced Dev metrics when supplied. RX_VK_TEST_LABEL defaults to Gameplay test. All three have switch-table rows. The paired --agent and --test-label options default unset and require a checkpoint. Agent launches default to a window and honor --headless. The opt-in defaults suit contributor testing. Their enabled costs and ordinary-launch timing still need measurements. **Reviewed:** d30aa1914249 against 3df5c13ed2ec: Reviewed d30aa1914249 against 3df5c13ed2ec, the commits, supplied comments and reviews, base AGENTS.md, CONTRIBUTING.md, the Forgejo template, and standards. Read host/oot_autoload.c, gpu/vk/vk_imgui.h, tests/test_capture_resources.py, tests/test_fps_overlay_order.py, tests/test_png_publication.py, tests/png_publication_test.cpp and tests/capture_dimensions_test.cpp in full. Reviewed changed regions elsewhere and traced renderer loading, screenshot callbacks, readback, presenter locking and launcher setup. Skimmed reported runs and other fixture context. Full renderer, bridge, launcher and remaining test files were not read. The review checklist and adjacent Tooling checkout were absent. Omitted comments, attachment images and Graphify maps were unavailable.; nothing run (a code review). **Checked and fine:** - The Conventional Commit title and functional diff cover one contributor workflow. No shaders, generated output, game data, vendor SDKs or third-party files changed. The feature commit records AI assistance. - The launcher requires paired labels and a selected checkpoint. The fixture checks that only the selected slot is copied and exercises both windowed and headless setup. - The save search uses wide Windows calls, and the fixtures cover multibyte and extended-length Windows paths and deeper Linux paths supported by the game filesystem. - The bridge records the clean copy before annotation and checks timeline completion before invalidating and reading mapped memory. The presenter lock is attempted after inline frame generation. - The UI recording fixture now matches the extracted signature and exercises colour correction at the recording layer. The PNG fixture checks that no-replace publication preserves an occupied destination. **Blocking** (to change before merge) 1. gpu/vk/vk_api.h:414: MkwVkRenderer::shot adds a fourth completion argument while MKWVK_API_VERSION stays at 36. An updated host can accept an older renderer, which calls shot_done with no scene pointer. The host may then dereference an indeterminate pointer while constructing the reply. Bump the API version and rebuild both binaries together, as the architecture and PR checklist require. 2. gpu/vk/vk_replay.cpp:5164: Every companion search starts at zero and encodes and writes a fresh temporary PNG before discovering each collision. With 100 retained companions in the output folder, the next 1280×720 capture writes roughly 276 MB in failed attempts alone. That work holds the publication mutex and delays other captures. Encode once, retry publication of the completed temporary under fresh names, and avoid restarting the search at zero. 3. gpu/vk/vk_replay.cpp:4608: Linux companion publication requires hard links. An explicit bridge shot to a writable exFAT output folder can currently use the regular rename path, but the same destination fails in agent mode before the primary is written. Use no-replace publication that works on rename-capable filesystems without hard links and cover the unsupported-link case. 4. gpu/vk/vk_replay.cpp:5169: An explicit primary path can replace a previously returned clean companion. For example, after shot A returns longshot-scene-0.png, shot B can use that path for its annotated primary. A later comparison of A then reads B’s annotated pixels. Keep companion names in a renderer-owned namespace that primary paths cannot replace, including through path aliases, and test two requests targeting the same namespace. **Suggestions** (the author may take or leave) - Record scene_path in the session manifest and list it in the bridge command table so artifact consumers can find the clean image after the run. **Testing:** Read-only review. The diff whitespace check passed. No builds, fixtures, project scripts or game sessions were run. The reported Windows sessions and baseline auto-profile failure predate this head. Exact-head CI, the game-free suite, Tier 1 hardware results and the Graphics guide source review remain outstanding. **Overlaps and merge order:** The PR reports Beanstalk #89 already merged. Other overlaps and merge order could not be verified with the network closed. --- <sub>Automated code review by BeanBot (gpt-6.2-sol), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.</sub>
@ -413,2 +412,3 @@
path on a thread of its own, which then calls done(ctx, present, ok, scene), with an optional clean companion path. 0: not taken
(too many waiting) */
int (*shot)(const char *path, int64_t after, void (*done)(void *ctx, int64_t present, int ok), void *ctx);
int (*shot)(const char *path, int64_t after, void (*done)(void *ctx, int64_t present, int ok, const char *scene), void *ctx);
Member

This callback gains a fourth argument while the ABI remains at 36. A new host can accept an old renderer and dereference a missing scene pointer when a shot finishes. Bump MKWVK_API_VERSION and rebuild both binaries.

This callback gains a fourth argument while the ABI remains at 36. A new host can accept an old renderer and dereference a missing scene pointer when a shot finishes. Bump MKWVK_API_VERSION and rebuild both binaries.
@ -4600,2 +4606,3 @@
if (!ok && collision && (error == ERROR_FILE_EXISTS || error == ERROR_ALREADY_EXISTS)) *collision = true;
#else
ok = rename(temporary.c_str(), path) == 0;
ok = collision ? link(temporary.c_str(), path) == 0 : rename(temporary.c_str(), path) == 0;
Member

A Linux shot to a writable exFAT folder can publish by rename in ordinary mode, but the annotated shot fails because exFAT has no hard links. Support no-replace publication on rename-capable filesystems without hard links.

A Linux shot to a writable exFAT folder can publish by rename in ordinary mode, but the annotated shot fails because exFAT has no hard links. Support no-replace publication on rename-capable filesystems without hard links.
@ -5059,2 +5167,3 @@
}
if (!ok) std::remove(part.c_str());
const bool ok = scene_written &&
write_shot(a.path, px->data() + (annotated ? bytes : 0));
Member

A later shot can use an earlier scene_path as its explicit primary and overwrite that clean companion. Reserve a namespace that primary paths cannot replace and exercise this with two bridge requests.

A later shot can use an earlier scene_path as its explicit primary and overwrite that clean companion. Reserve a namespace that primary paths cannot replace and exercise this with two bridge requests.
@ -5059,0 +5161,4 @@
if (scene == a.path) continue;
#endif
bool collision = false;
scene_written = write_png(scene.c_str(), px->data(), kShotW, kShotH, &collision);
Member

Each occupied name causes a full encode and temporary write under the publication mutex. With 100 prior 720p companions that is roughly 276 MB of failed writes for one shot. Encode once and retry publishing the completed temporary under fresh names.

Each occupied name causes a full encode and temporary write under the publication mutex. With 100 prior 720p companions that is roughly 276 MB of failed writes for one shot. Encode once and retry publishing the completed temporary under fresh names.
Owner
@fix
Member

I changed this pull request where the review's points could be done here (the rest needs runs, measurements or decisions): the new head is 067fd00ae4.

Changed

  • Companion publication and protection, gpu/vk/vk_replay.cpp:5133: gpu/vk/vk_replay.cpp reserves companion names against primary shots, publishes a distinct companion, and removes it if the primary fails. tests/png_publication_test.cpp covers a later request targeting the returned companion.
  • Screenshot callback ABI, gpu/vk/vk_api.h:414: Bumped MKWVK_API_VERSION to 37 in gpu/vk/vk_api.h:17. The author must rebuild both host and renderer binaries.
  • No-replace publication on Linux filesystems without hard links, gpu/vk/vk_replay.cpp:4608: gpu/vk/vk_replay.cpp:4616 uses renameat2 with no-replace semantics, falling back to hard-link publication when that operation is unavailable.
  • Repeated companion PNG encoding, gpu/vk/vk_replay.cpp:5164: gpu/vk/vk_replay.cpp:4607 retries occupied companion names using one completed temporary PNG; tests/png_publication_test.cpp covers 100 occupied names.
  • gpu/vk/vk_replay.cpp:4525 capture-dimensions fixture: Updated the declaration and Linux headers in tests/capture_dimensions_test.cpp:53 for the extracted function.
  • gpu/vk/vk_replay.cpp:4619 no-replace fallback: Added an exclusive-create copy fallback in gpu/vk/vk_replay.cpp:4625 and fixture coverage in tests/png_publication_test.cpp:165. The destination can be visible while copying; it is returned only after completion. No builds or tests were run.

Already done

  • Windows Unicode save discovery, host/oot_autoload.c:187: Wide Windows filesystem calls are in host/oot_autoload.c, with a non-ASCII fixture in tests/test_skip_title.py.
  • Long UTF-8 Windows save paths, host/oot_autoload.c:118: host/oot_autoload.c uses larger UTF-8 path buffers; tests/test_skip_title.py covers multibyte and extended-length paths.
  • UI recording fixture, tests/test_ui_record.py:67: tests/test_ui_record.py declares the seven-argument function and exercises both recording paths.
  • Deep Linux save paths, host/oot_autoload.c:28: host/oot_autoload.c matches the game filesystem's path capacity, and tests/test_skip_title.py covers a deep Linux path.
  • Document bridge capture deferral, docs/dev/reference/control-bridge.md:225: docs/dev/reference/control-bridge.md:228 documents deferral while the presenter holds the UI lock.

Left for the author

  • HUD measurement overhead, gpu/vk/vk_menu.cpp:3147: Measure warmed, interleaved HUD-off and HUD-on gameplay against the base, including CPU and GPU cost, frame times and 1% lows; reduce or isolate any material distortion based on the results.
  • Capture overhead, gpu/vk/vk_replay.cpp:5097: Measure readback and PNG time, RAM and VRAM use, and frame pacing at the intended capture cadence against gameplay without captures; adjust the strategy if the disturbance is material.
  • Platform and current-head coverage: Run the current-head suite and required Tier 1 checks, including Linux RADV or Steam Deck headless and alternate GPU presentation coverage.
  • Rebased-head colour correction and screenshots: Verify colour-vision correction and screenshots on the exact current head.

Nothing was built or run: check the change. It will be reviewed again.


Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.

I changed this pull request where the review's points could be done here (the rest needs runs, measurements or decisions): the new head is 067fd00ae4d2. **Changed** - **Companion publication and protection, gpu/vk/vk_replay.cpp:5133**: gpu/vk/vk_replay.cpp reserves companion names against primary shots, publishes a distinct companion, and removes it if the primary fails. tests/png_publication_test.cpp covers a later request targeting the returned companion. - **Screenshot callback ABI, gpu/vk/vk_api.h:414**: Bumped MKWVK_API_VERSION to 37 in gpu/vk/vk_api.h:17. The author must rebuild both host and renderer binaries. - **No-replace publication on Linux filesystems without hard links, gpu/vk/vk_replay.cpp:4608**: gpu/vk/vk_replay.cpp:4616 uses renameat2 with no-replace semantics, falling back to hard-link publication when that operation is unavailable. - **Repeated companion PNG encoding, gpu/vk/vk_replay.cpp:5164**: gpu/vk/vk_replay.cpp:4607 retries occupied companion names using one completed temporary PNG; tests/png_publication_test.cpp covers 100 occupied names. - **gpu/vk/vk_replay.cpp:4525 capture-dimensions fixture**: Updated the declaration and Linux headers in tests/capture_dimensions_test.cpp:53 for the extracted function. - **gpu/vk/vk_replay.cpp:4619 no-replace fallback**: Added an exclusive-create copy fallback in gpu/vk/vk_replay.cpp:4625 and fixture coverage in tests/png_publication_test.cpp:165. The destination can be visible while copying; it is returned only after completion. No builds or tests were run. **Already done** - **Windows Unicode save discovery, host/oot_autoload.c:187**: Wide Windows filesystem calls are in host/oot_autoload.c, with a non-ASCII fixture in tests/test_skip_title.py. - **Long UTF-8 Windows save paths, host/oot_autoload.c:118**: host/oot_autoload.c uses larger UTF-8 path buffers; tests/test_skip_title.py covers multibyte and extended-length paths. - **UI recording fixture, tests/test_ui_record.py:67**: tests/test_ui_record.py declares the seven-argument function and exercises both recording paths. - **Deep Linux save paths, host/oot_autoload.c:28**: host/oot_autoload.c matches the game filesystem's path capacity, and tests/test_skip_title.py covers a deep Linux path. - **Document bridge capture deferral, docs/dev/reference/control-bridge.md:225**: docs/dev/reference/control-bridge.md:228 documents deferral while the presenter holds the UI lock. **Left for the author** - **HUD measurement overhead, gpu/vk/vk_menu.cpp:3147**: Measure warmed, interleaved HUD-off and HUD-on gameplay against the base, including CPU and GPU cost, frame times and 1% lows; reduce or isolate any material distortion based on the results. - **Capture overhead, gpu/vk/vk_replay.cpp:5097**: Measure readback and PNG time, RAM and VRAM use, and frame pacing at the intended capture cadence against gameplay without captures; adjust the strategy if the disturbance is material. - **Platform and current-head coverage**: Run the current-head suite and required Tier 1 checks, including Linux RADV or Steam Deck headless and alternate GPU presentation coverage. - **Rebased-head colour correction and screenshots**: Verify colour-vision correction and screenshots on the exact current head. Nothing was built or run: check the change. It will be reviewed again. --- <sub>Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.</sub>
fix(vk): protect and publish agent screenshot companions
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
PR description / Description follows the template (pull_request_target) Failing after 10s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 11s
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) Failing after 22s
067fd00ae4
As the review on #1539 asked.
BeanBot left a comment

Recommendation: Request changes. The new Linux fallback exposes an incomplete companion PNG under its final name. The new publication lock also serializes encoding behind detached writers without bounding the number of frames they can retain.

Contributors can launch a copied checkpoint, see the build and test context beside Dev metrics, and get a clean comparison image from the same present. Since review 1936, the branch bumped the renderer ABI, stopped re-encoding on filename collisions, and added no-replace publication and companion-name protection. Those changes resolve the ABI and repeated-encoding findings. The fallback and writer backlog still need changes, and the earlier measurement requests remain open.

Does it help: Unproven on 067fd00ae4. The author reports five successful Windows 11 sessions on an RTX 4070 with driver 621.01 at b324a3cb, and a frame-generation lock fixture that failed on the regression commit and passed at b324a3cb. There are no matched, warmed, interleaved runs against 3df5c13ed2 showing ordinary-launch cost, HUD-on/off average FPS, 1% lows and frame times, or readback, PNG and memory costs at the intended capture cadence.
Other systems: The ABI bump requires matching host and renderer builds on Windows and Linux. Earlier Windows NVIDIA evidence covers live and offscreen play, ordinary screenshots, expanded HDR10 and built-in 2× frame generation. Exact-head colour correction, Linux RADV and Steam Deck, and the remaining Tier 1 GPU and OS combinations are unverified. Agent-mode periodic captures also feed annotated pixels into the existing black-pixel and flash detectors. Bug-report screenshots still explicitly request clean pixels.
Switches: HOS_TEST_AUTOLOAD defaults off and the agent launcher sets it to 1. RX_VK_TEST_AGENT defaults unset and enables the banner and Dev metrics. RX_VK_TEST_LABEL defaults to Gameplay test. All three have switch-table rows. The paired --agent and --test-label options default unset and require a checkpoint. Agent launches open a window by default, honor --headless, and inherit HOS_AUDIO=0 and an empty HOS_ENV_FILE from the test profile. The opt-in defaults fit contributor runs. The enabled timing cost is still open.

Reviewed: 067fd00ae4 against 3df5c13ed2: Reviewed 067fd00ae4 against 3df5c13ed2, the seven commits, the supplied comments and reviews, base AGENTS.md, CONTRIBUTING.md, the Forgejo template, standards, and relevant renderer architecture rules. Read host/oot_autoload.c, gpu/vk/vk_imgui.h, tests/capture_dimensions_test.cpp, tests/test_capture_dimensions.py and tests/test_png_publication.py in full. Reviewed changed regions in the remaining files and traced screenshot recording, readback, publication, callback wiring and launcher setup. Skimmed remaining fixture context and reported runs. Full renderer, launcher, bridge and remaining test files were outside the full-file coverage. The review checklist, adjacent Tooling checkout, omitted comments, screenshot attachments and Graphify maps were unavailable.; nothing run (a code review).
Checked and fine:

  • The conventional title and functional diff cover one contributor workflow. The diff includes no game data, vendor SDKs, shader changes, generated output or third-party edits. The feature commit records AI assistance.
  • The screenshot callback and ABI version agree. The host copies the companion path during the callback, and bug reports use the clean route.
  • The clean readback is recorded before annotation, and CPU access follows timeline completion and mapped-memory invalidation.
  • The UI lock is tried after inline frame generation, and the capture releases it after submission.
  • The launcher requires paired labels and a checkpoint, uses a disposable user folder, and waits for play. The save-search fixtures cover multibyte and extended-length Windows paths and deep Linux paths.

Blocking (to change before merge)

  1. gpu/vk/vk_replay.cpp:4629: The exclusive-create copy publishes the final name before the PNG is complete, violating the renderer's atomic PNG rule. A directory watcher on a filesystem without no-replace rename or hard links can open a partial image, and an interrupted process can leave it behind. Require an atomic no-replace publication primitive and return a clear error on filesystems that lack one.
  2. gpu/vk/vk_replay.cpp:5205: Each readback starts a detached thread that holds both pixel copies while it waits for the global publication lock. If a slow output directory delays writes while clients keep requesting shots, completed readbacks move past the 16-entry request queue into an unbounded writer backlog. At 4K each annotated request retains about 63 MiB of pixels. Use a bounded writer queue with admission that counts pending readbacks and writes, and return busy when full.
  3. gpu/vk/vk_replay.cpp:4607: Companion allocation starts at zero for every shot and gives up after 10,000 names. A persistent capture directory fails every subsequent annotated shot once those names are occupied, even when it has free space. Use unique short names with atomic collision checks so retries do not grow with the directory's capture history.

Suggestions (the author may take or leave)

  • Carry forward the earlier request for exact-head HUD and capture timings and complete the current-head suite, Tier 1 runs, colour-correction check, and Graphics guide source review. The reported auto-profile baseline failure was established on older commits.

Testing: Static review only. git diff --check passed. No builds, tests, scripts or game runs were executed. The reported Windows gameplay and lock fixture results predate the current head.
Overlaps and merge order: The description says Beanstalk runner #89 is already merged. Other overlaps could not be checked offline.


Automated code review by BeanBot (gpt-6.2-sol), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.

**Recommendation: Request changes.** The new Linux fallback exposes an incomplete companion PNG under its final name. The new publication lock also serializes encoding behind detached writers without bounding the number of frames they can retain. Contributors can launch a copied checkpoint, see the build and test context beside Dev metrics, and get a clean comparison image from the same present. Since review 1936, the branch bumped the renderer ABI, stopped re-encoding on filename collisions, and added no-replace publication and companion-name protection. Those changes resolve the ABI and repeated-encoding findings. The fallback and writer backlog still need changes, and the earlier measurement requests remain open. **Does it help:** Unproven on 067fd00ae4d2. The author reports five successful Windows 11 sessions on an RTX 4070 with driver 621.01 at b324a3cb, and a frame-generation lock fixture that failed on the regression commit and passed at b324a3cb. There are no matched, warmed, interleaved runs against 3df5c13ed2ec showing ordinary-launch cost, HUD-on/off average FPS, 1% lows and frame times, or readback, PNG and memory costs at the intended capture cadence. **Other systems:** The ABI bump requires matching host and renderer builds on Windows and Linux. Earlier Windows NVIDIA evidence covers live and offscreen play, ordinary screenshots, expanded HDR10 and built-in 2× frame generation. Exact-head colour correction, Linux RADV and Steam Deck, and the remaining Tier 1 GPU and OS combinations are unverified. Agent-mode periodic captures also feed annotated pixels into the existing black-pixel and flash detectors. Bug-report screenshots still explicitly request clean pixels. **Switches:** HOS_TEST_AUTOLOAD defaults off and the agent launcher sets it to 1. RX_VK_TEST_AGENT defaults unset and enables the banner and Dev metrics. RX_VK_TEST_LABEL defaults to Gameplay test. All three have switch-table rows. The paired --agent and --test-label options default unset and require a checkpoint. Agent launches open a window by default, honor --headless, and inherit HOS_AUDIO=0 and an empty HOS_ENV_FILE from the test profile. The opt-in defaults fit contributor runs. The enabled timing cost is still open. **Reviewed:** 067fd00ae4d2 against 3df5c13ed2ec: Reviewed 067fd00ae4d2 against 3df5c13ed2ec, the seven commits, the supplied comments and reviews, base AGENTS.md, CONTRIBUTING.md, the Forgejo template, standards, and relevant renderer architecture rules. Read host/oot_autoload.c, gpu/vk/vk_imgui.h, tests/capture_dimensions_test.cpp, tests/test_capture_dimensions.py and tests/test_png_publication.py in full. Reviewed changed regions in the remaining files and traced screenshot recording, readback, publication, callback wiring and launcher setup. Skimmed remaining fixture context and reported runs. Full renderer, launcher, bridge and remaining test files were outside the full-file coverage. The review checklist, adjacent Tooling checkout, omitted comments, screenshot attachments and Graphify maps were unavailable.; nothing run (a code review). **Checked and fine:** - The conventional title and functional diff cover one contributor workflow. The diff includes no game data, vendor SDKs, shader changes, generated output or third-party edits. The feature commit records AI assistance. - The screenshot callback and ABI version agree. The host copies the companion path during the callback, and bug reports use the clean route. - The clean readback is recorded before annotation, and CPU access follows timeline completion and mapped-memory invalidation. - The UI lock is tried after inline frame generation, and the capture releases it after submission. - The launcher requires paired labels and a checkpoint, uses a disposable user folder, and waits for play. The save-search fixtures cover multibyte and extended-length Windows paths and deep Linux paths. **Blocking** (to change before merge) 1. gpu/vk/vk_replay.cpp:4629: The exclusive-create copy publishes the final name before the PNG is complete, violating the renderer's atomic PNG rule. A directory watcher on a filesystem without no-replace rename or hard links can open a partial image, and an interrupted process can leave it behind. Require an atomic no-replace publication primitive and return a clear error on filesystems that lack one. 2. gpu/vk/vk_replay.cpp:5205: Each readback starts a detached thread that holds both pixel copies while it waits for the global publication lock. If a slow output directory delays writes while clients keep requesting shots, completed readbacks move past the 16-entry request queue into an unbounded writer backlog. At 4K each annotated request retains about 63 MiB of pixels. Use a bounded writer queue with admission that counts pending readbacks and writes, and return busy when full. 3. gpu/vk/vk_replay.cpp:4607: Companion allocation starts at zero for every shot and gives up after 10,000 names. A persistent capture directory fails every subsequent annotated shot once those names are occupied, even when it has free space. Use unique short names with atomic collision checks so retries do not grow with the directory's capture history. **Suggestions** (the author may take or leave) - Carry forward the earlier request for exact-head HUD and capture timings and complete the current-head suite, Tier 1 runs, colour-correction check, and Graphics guide source review. The reported auto-profile baseline failure was established on older commits. **Testing:** Static review only. git diff --check passed. No builds, tests, scripts or game runs were executed. The reported Windows gameplay and lock fixture results predate the current head. **Overlaps and merge order:** The description says Beanstalk runner #89 is already merged. Other overlaps could not be checked offline. --- <sub>Automated code review by BeanBot (gpt-6.2-sol), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.</sub>
@ -4594,17 +4604,62 @@ bool write_png(const char* path, const uint8_t* rgba, uint32_t w, uint32_t h) {
}
if (ok) {
phase = "publish";
for (unsigned i = 0; i < (companion ? 10000u : 1u); i++) {
Member

Every shot rescans from zero, and a directory with 10,000 companions can never accept another annotated shot. Use unique short names and keep the atomic collision check.

Every shot rescans from zero, and a directory with 10,000 companions can never accept another annotated shot. Use unique short names and keep the atomic collision check.
@ -4604,0 +4626,4 @@
// Claim the name exclusively, then copy the already encoded PNG.
FILE* in = fopen(temporary.c_str(), "rb");
if (in) {
FILE* out = fopen(candidate.c_str(), "wbx");
Member

This exposes the final companion name during the copy. A watcher can read a partial PNG, and interruption can leave it behind. Keep publication atomic and report unsupported filesystems explicitly.

This exposes the final companion name during the copy. A watcher can read a partial PNG, and interruption can leave it behind. Keep publication atomic and report unsupported filesystems explicitly.
@ -5056,3 +5205,1 @@
if (ok) {
std::remove(a.path.c_str());
ok = std::rename(part.c_str(), a.path.c_str()) == 0;
std::thread([px, a, at, bytes] {
Member

The request limit stops applying once this detached writer starts. Slow storage can pile up threads behind publication_mu, each retaining about 63 MiB at 4K. Bound outstanding writes and include them in admission.

The request limit stops applying once this detached writer starts. Slow storage can pile up threads behind publication_mu, each retaining about 63 MiB at 4K. Bound outstanding writes and include them in admission.
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
PR description / Description follows the template (pull_request_target) Failing after 10s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 11s
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) Failing after 22s
This pull request has changes conflicting with the target branch.
  • gpu/vk/vk_api.h
  • gpu/vk/vk_menu.cpp
  • gpu/vk/vk_replay.cpp
  • host/control.c
  • 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 'codex/agent-test-hud:futondestroyer-codex/agent-test-hud'
git switch 'futondestroyer-codex/agent-test-hud'
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
3 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!1539
No description provided.