English
-
OpenCode render workflow now defaults to one agent, script summaries and reuse of unchanged shader reviews/tests; additional agents require a user request.
-
Developer tooling: added reusable offline F1 capture preview, full-frame coverage, trace, and temporal-jitter candidate tools, a streaming private feedback-coverage report, a task-based tools index, the recovered private feedback-archive package, and an installable OpenCode render-investigation workflow; reviewed candidates remain separate from the production shader map.
-
Historical temporal jitter mapping batch: added 35 reviewed VS mappings (map 84→119) across slots 0/1/4/7/8, with 44 candidate VS held for unresolved clip-XY producer or mixed-camera evidence.
LoTemporalJitterTest --feedback-mapping-batchpassed 216,625 focused checks; later entries record the separately authorized follow-up trials and their bounded acceptance scope. -
Temporal jitter held-candidate follow-up: added one limited slot-8 mapping for the late-light candidate observed at
draw1206..1210, aligning the floor light pass with the scene depth/material jitter path. The map now contains 120 entries after the earlier 35-source batch. Captured-floor CPU checks passed 334,086 checks and the mapping batch passed 216,624 after removing this held candidate; the complete Windows SDK-on build succeeded, and the user accepted the bounded RTX 5080 floor-scene run. Broader runtime coverage remains open. -
Hybrid SR and bounded NGX sizing recovery: development builds include camera/depth Hybrid motion fallback for DLSS/FSR and bounded retry recovery for transient DLSS sizing failures; validation remains limited to the documented scope.
-
F1 capture diagnostics: full-frame jitter candidate triage is now the default, with strict geometry/depth pairing available as a separate report; successful draw records include jitter/uploaded-VP and tex0 binding metadata while preserving unknown states when evidence is incomplete. Capture command records do not imply GPU completion, and reviewed fixture export behavior is unchanged.
-
Late-pass e810 trial: added a guarded
e810cfacc107fd3c/fe31slot-7 path requiringconstantScreenSample, valid format-6 2D guest 1×1 tex0, repeat UV, no same-address resolved hit and no VS tex0 read. The SDK-on guarded build exited 0 and--captured-f6131-e810passed 66,252 checks across 32 phases at 1440p/4K. Independent capture review confirmed the mapping applied and bound on every accepted frame; the bounded RTX 5080 visual run was accepted by the user. Remaining candidates are tracked separately in the screen-sampling batch, and no broad candidate mapping is claimed. -
Screen-sampling candidate batch: authorized a separate manual trial adding exactly 40 VS mappings (35 slot 7, 5 slot 8) covering 101 observed PS pairings, excluding the two mixed-camera cases. The selector passed 78,316 CPU checks across 32 phases at 1440p/4K, with 0.487798 px old separation and 0.000279 px maximum projection error. The batch build passed, package
manual-screen-batch1is available, and the user accepted the manual run; no new capture evidence, whole-game or cross-hardware claim is made, and v0.6.20 was published on 2026-09-25. -
Capture CLI maintenance: fixed the direct
image_diff.pysubprocess path after aninspect.pyimport collision and kept the image comparison path Pillow-only; two affected checks, including the direct CLI regression, passed. -
Exploration bridge flicker investigation: aligned two observed material position paths with scene depth through the temporal-scene slot mappings
a027ab99fa3e3b0d -> slot 7andff769ec7b88e575f -> slot 8. The focused--captured-f2548-layerscheck passed 722,316 checks across 11 captured draws and 32 phases at 720p through 4K; the old-material negative control showed 0.487760 px separation. A later bounded floor-light follow-up was accepted on RTX 5080; broader same-scene coverage remains open; the fix is included in v0.6.20.
简体中文
-
OpenCode 渲染工作流默认单 agent，先运行脚本并读取摘要，复用未变化的 shader 审核及测试结果；仅在用户要求时启动其他 agent。
-
开发工具：新增可复用的 F1 捕获预览、全帧覆盖、寄存器状态和时序抖动候选离线工具、私有反馈覆盖流式报告、按任务组织的工具索引、恢复的私有反馈归档包，以及可安装的 OpenCode 渲染调查工作流；审核中的候选与生产 shader 映射分离。
-
历史时序抖动映射批次：新增 35 条已审阅 VS 映射（映射表 84→119），覆盖 slot 0/1/4/7/8；44 个候选 VS 因 clip-XY producer 或混合相机证据不足暂缓。
LoTemporalJitterTest --feedback-mapping-batch通过 216,625 项定向检查；后续条目记录了另行授权的跟进试验及其限定验收范围。 -
时序抖动暂缓候选后续：根据后期光照捕获（
draw1206..1210）增加一条限定范围的 slot 8 生产映射，使地板光照与场景深度／材质抖动路径对齐。此前 35-source 批次之后，映射表现为 120 条。捕获地板 CPU 检查通过 334,086 项，移除该暂缓候选后的映射批次通过 216,624 项；完整 Windows SDK-on build 已成功，用户已验收限定的 RTX 5080 地板场景，其他运行范围仍待完成。 -
Hybrid SR 与有界 NGX 尺寸恢复：开发构建包含 DLSS/FSR 相机／深度 Hybrid 运动回退，以及针对暂态 DLSS 尺寸失败的有界重试恢复；验证范围仍以文档记录为准。
-
F1 捕获诊断：时序抖动候选分析默认覆盖整帧，严格几何／深度配对改为单独报告；成功记录的绘制新增 jitter／已上传 VP 以及 tex0 绑定元数据，证据不足时保留 unknown。捕获命令被记录不等于 GPU 已完成，审核夹具导出行为保持不变。
-
后期 e810 trial：增加受守卫的
e810cfacc107fd3c／fe31slot 7 路径，要求constantScreenSample、有效 format-6 2D、guest 1×1 tex0、repeat UV、无同地址已解析命中且 VS 不读取 tex0。SDK-on 受守卫构建退出码为 0，--captured-f6131-e810在 1440p/4K 的 32 个相位通过 66,252 项检查；独立捕获复核确认映射在验收帧上实际应用并绑定，限定 RTX 5080 画面运行已获用户验收，剩余候选另见屏幕采样批次，不宣称批量候选映射。 -
屏幕采样候选批次：授权单独的实机试验，增加40条VS映射（slot 7为35条、slot 8为5条），覆盖101个已观察PS配对，并排除两个混合相机案例。selector通过78,316项CPU检查，覆盖32个相位和1440p/4K，旧分离为0.487798 px、最大投影误差为0.000279 px；批次构建已通过，
manual-screen-batch1包已准备，用户已验收手动运行；不宣称新的capture证据、画面全覆盖或跨硬件结果，v0.6.20 已于 2026-09-25 发布。 -
捕获 CLI 维护：修复
image_diff.py直接子进程调用与inspect.py的导入冲突，并保持图像比较路径仅依赖 Pillow；两个受影响检查（含直接 CLI 回归）均已通过。 -
探索场景桥面闪烁调查：通过时序场景 slot 映射
a027ab99fa3e3b0d -> slot 7、ff769ec7b88e575f -> slot 8，使两条已观察到的材质位置路径与场景深度对齐。定向--captured-f2548-layers检查通过 722,316 项检查，覆盖 11 条捕获绘制、720p 至 4K 的 32 个相位；旧材质负对照分离为 0.487760 px。后续限定的RTX 5080地板光照场景已获用户验收，更广场景覆盖仍待完成；v0.6.20 已于 2026-09-25 发布。