[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7EJQJ75C550Y4JSSXHSJRTX":3,"skill-same:01M7EJQJ75C550Y4JSSXHSJRTX":66},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":27,"commitSha":28,"skillDirectory":29,"forksCount":30,"repositoryHtmlUrl":31,"repositoryHomepageUrl":6,"ownerHtmlUrl":32,"ownerBlogUrl":6,"ownerDescription":33,"updatedAt":34,"install":35,"resources":47},"01M7EJQJ75C550Y4JSSXHSJRTX","slam-runtime",null,"Run and troubleshoot the CUDA-only Gaussian-SLAM pipeline: validate the environment, construct safe run_slam.py invocations, understand tracking\u002Fmapping\u002Fsubmaps and outputs, control seeds and W&B, and prepare non-submitting SLURM plans.","lza6\u002FArmor-piercing-system-prompts","lza6","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F128880206?v=4","01M7EB3K7WPX5T28EJ2NVHE6EG","Chinese-tingfeng","01M7EB3K97E1KQ16QWKJMFNGX8",11,"MIT","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F04-安全渗透\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: slam-runtime\ndescription: \"Run and troubleshoot the CUDA-only Gaussian-SLAM pipeline:\n  validate the environment, construct safe run_slam.py invocations, understand\n  tracking\u002Fmapping\u002Fsubmaps and outputs, control seeds and W&B, and prepare\n  non-submitting SLURM plans.\"\ndisable-model-invocation: true\nmetadata:\n  disco-role: operating\nlicense: MIT\n---\n\n# Gaussian-SLAM runtime\n\nUse this skill when a user wants to execute, configure, reproduce, or recover a\nGaussian-SLAM run from this repository. The entry point is `run_slam.py`; the\nrepository's execution path is **CUDA-only**. Do not offer a CPU fallback for\ntracking, mapping, rendering, or evaluation. Dataset acquisition and dataset\nfile-layout instructions belong to `datasets-and-configuration`; evaluation\nmetric internals belong to `evaluation-and-mapping`.\n\n## Input\u002Foutput contract\n\n### Required inputs\n\n- A scene YAML path passed as the positional `config_path` argument.\n- A config resolving to `dataset_name`, `data`, `cam`, `tracking`, `mapping`,\n  `seed`, and `use_wandb`. A scene config normally inherits a dataset base\n  config and supplies `data.scene_name`, `data.input_path`, and\n  `data.output_path`.\n- A prepared RGB-D scene and camera metadata accepted by the selected dataset\n  adapter. The dataset's layout is intentionally not specified here; ask the\n  dataset skill for that contract.\n- A CUDA-capable PyTorch installation and both compiled extensions:\n  `simple_knn._C` and `gaussian_rasterizer._C`.\n\nUse explicit `--input_path` and `--output_path` values for portable runs. The\nconfig loader resolves `inherit_from` using the path as written, relative to\nthe process working directory, so launch from the repository root (or use\npaths that resolve from the chosen working directory).\n\n### Produced artifacts\n\n`run_slam.py` creates the output directory, saves the effective `config.yaml`,\nthen runs tracking\u002Fmapping followed by the repository evaluator. Common\nartifacts are:\n\n- `config.yaml`: effective merged configuration after CLI overrides;\n- `estimated_c2w.ckpt`: estimated camera-to-world poses;\n- `submaps\u002F000000.ckpt`, etc.: checkpoints containing Gaussian parameters and\n  the keyframe IDs belonging to the completed prior submap;\n- `mapping_vis\u002F*.jpg` and the created `tracking_vis\u002F` directory;\n- evaluator artifacts such as `rendering_metrics.json`,\n  `reconstruction_metrics.json`, `ate_aligned.json`, PNG summaries, a Replica\n  mesh, and a refined global-map PLY when the corresponding evaluation phase\n  succeeds.\n\nThe evaluator catches failures per evaluation phase and prints a traceback, so\nan `All done.✨` line is not sufficient evidence that every evaluation artifact\nexists. Check the expected files and the log. The active final Gaussian model\nis not explicitly finalized after the main loop; verify submap completeness\nbefore consuming checkpoints as a final map.\n\n## Required preflight\n\nRun these checks before a full scene. They are safe and do not download data,\nsubmit jobs, or contact W&B:\n\n```bash\nnvidia-smi\npython - \u003C\u003C'PY'\nimport torch\nprint(\"torch\", torch.__version__, \"cuda\", torch.version.cuda)\nassert torch.cuda.is_available(), \"Gaussian-SLAM requires CUDA\"\nprint(torch.cuda.get_device_name(0), torch.cuda.get_device_capability(0))\nPY\npython - \u003C\u003C'PY'\nimport simple_knn._C\nimport gaussian_rasterizer._C\nprint(\"CUDA extensions import\")\nPY\npython run_slam.py --help\npython skills\u002Fdisco\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002Fscripts\u002Fcheck_cli.py \\\n  configs\u002FReplica\u002Froom0.yaml\n```\n\nThe repository environment specifies Python 3.10, PyTorch 2.1.2 with CUDA\n12.1, torchvision 0.16.2, FAISS-GPU, Open3D 0.18.0, and pinned builds of the\ntwo CUDA extensions. Building extensions needs a compatible host compiler; the\nREADME requires exported `CC` and `CXX`. A verified inspection environment\npassed CUDA allocation and both extension imports on an NVIDIA A100-SXM4-40GB\nwith compute capability 8.0 (SM80). Treat that as evidence for A100\u002FSM80\ncompatibility, not as a promise for arbitrary GPUs. Match the PyTorch\u002FCUDA\nruntime, host compiler, toolkit, and extension build target to the actual\nmachine.\n\n`GaussianModel` imports `simple_knn._C.distCUDA2` at module import and uses it\nwhen adding points. `src\u002Futils\u002Futils.py` imports the custom\n`gaussian_rasterizer`; render settings, tensors, and camera optimization are\nalso hard-coded to CUDA. Missing either extension, a CPU-only Torch build, or a\nvisible device mismatch is a prerequisite failure—not something to work around\nby changing the device to CPU.\n\n## Standard run\n\nFrom the repository root, start with a small bounded scene\u002Fconfig if one is\navailable, then use the full run:\n\n```bash\npython run_slam.py configs\u002F\u003Cdataset>\u002F\u003Cscene>.yaml \\\n  --input_path \u003Cscene-input> \\\n  --output_path \u003Crun-output>\n```\n\nThe process is sequential: it loads and recursively merges YAML inheritance,\nupdates selected fields from CLI arguments, optionally initializes W&B, seeds\nPython\u002FNumPy\u002FPyTorch\u002FCUDA, runs `GaussianSLAM.run()`, and then constructs an\n`Evaluator` from the saved output config. The first two frames use dataset poses;\nsubsequent frames go through the configured tracker. Mapping runs on the\nconfigured keyframe cadence and always appends the last frame to the mapping\nIDs. Do not infer that `--gt_camera` changes this flow: the flag is parsed but\nis not applied by `update_config_with_args`; select\n`tracking.odometry_type: gt` in YAML when ground-truth tracking is intended.\n\nFor a deterministic, non-W&B smoke invocation, make the output unique and\nexplicit:\n\n```bash\nDISABLE_WANDB=true python run_slam.py configs\u002F\u003Cdataset>\u002F\u003Cscene>.yaml \\\n  --input_path \u003Cscene-input> \\\n  --output_path \u003Crun-output> \\\n  --seed 0\n```\n\nThis controls the software RNGs (`torch`, CUDA, NumPy, Python, and hash seed)\nand disables cuDNN benchmarking, but the README explicitly warns that the\nDifferential Gaussian rasterizer is nondeterministic. Compare runs with a\nsmall tolerance and record GPU, driver, extension build, config, and seed.\n\n## CLI and configuration\n\nRead [references\u002Fcli-reference.md](references\u002Fcli-reference.md) before changing\nhyperparameters. It records every parser option, its target config field, and\nsource-level caveats (including parsed-but-ineffective options and truthiness\nchecks that ignore zero). Use the checker to catch missing inherited fields and\ncontradictory filter flags before starting a GPU job.\n\nThe highest-risk runtime changes are tracking iterations\u002Floss weighting and\nmapping cadence\u002Fiterations. Large new-submap point counts, high-resolution\nframes, and long optimization schedules increase VRAM and wall time. Lower\ncadence or iteration counts only for a smoke test; do not call a shortened run\npaper reproduction.\n\n## Tracking, mapping, and submaps\n\n- **Tracking.** `Tracker.track` renders the current Gaussian model and optimizes\n  camera rotation\u002Ftranslation using RGB and depth losses. `odometry_type` is\n  `gt`, `const_speed`, or `odometer`; the latter uses the configured visual\n  odometer for initialization. Alpha and depth-outlier filters affect the\n  tracking mask. A high initial loss can double tracking iterations; with\n  `help_camera_initialization`, the tracker may retry initialization with an\n  odometer estimate.\n- **Mapping.** At mapping frames, the mapper creates a keyframe, seeds 3D\n  points from RGB-D and the estimated pose, adds non-duplicate points to the\n  current Gaussian model using `distCUDA2`, optimizes the current submap, and\n  writes a visualization. New submaps use edge\u002Fgradient seeding; an existing\n  submap uses alpha\u002Fdepth-error regions. Dataset-specific point-cloud filtering\n  is an implementation detail; do not substitute a missing dataset contract.\n- **Submap boundaries.** With `submap_using_motion_heuristic: true`, the source\n  starts a new submap when motion from the last submap start exceeds a hard-coded\n  50-degree rotation or 0.5 translation threshold. Otherwise it uses\n  `new_submap_every` frame IDs and also includes the last frame. At a boundary,\n  the prior model is written as a zero-padded checkpoint, keyframes are reset,\n  and a new model starts at that frame. The loop saves poses at the end but does\n  not perform a separate final active-model checkpoint; inspect the output\n  rather than assuming the last active model is serialized.\n\n## W&B and seed controls\n\nConfigs default to `use_wandb: False`. For an offline\u002Flocal run, prefer that\nsetting or set `DISABLE_WANDB=true`; the entry point treats exactly the string\n`true` as a forced disable. If W&B is enabled, `run_slam.py` calls `wandb.init`,\nlogs source code, and uploads tracking\u002Fmapping\u002Fevaluation records. Credentials\nand network access are not required for the default disabled path. If a site\nneeds W&B, decide explicitly whether online or offline mode is appropriate and\nensure its run directory is writable: the source contains a site-specific\nabsolute W&B directory that may not exist on another machine, and it is not a\nCLI setting. Patch it to a writable path or keep W&B disabled; do not expose\ncredentials in configs or SLURM scripts.\n\n## SLURM\n\nUse [references\u002Fcluster-reproduction.md](references\u002Fcluster-reproduction.md) or\nrun the safe planner:\n\n```bash\npython skills\u002Fdisco\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002Fscripts\u002Frender_sbatch_plan.py \\\n  --dataset Replica --scene room0 --scene room1 \\\n  --input-root \u003Cinput-root> --output-root \u003Coutput-root>\n```\n\nThe planner only prints a reviewable batch script; it never runs `sbatch`,\nstarts Python SLAM, downloads data, or authenticates W&B. Review the generated\narray bounds, config directory, scene names, input\u002Foutput roots, partition,\nGPU request, and environment activation before manually submitting it. The\nrepository example contains site placeholders, a dataset-directory case\nmismatch for ScanNet++, a fixed array example that must be changed per dataset,\nand a trailing continuation after `--group_name`; use the corrected recipe in\nthe reference instead of copying it verbatim.\n\n## Failure recovery\n\n1. Preserve the run log and output directory. Do not delete partial submaps or\n   overwrite a prior run while diagnosing it.\n2. Classify the first failure: environment\u002Fextension, config resolution,\n   dataset contract, CUDA\u002FVRAM, W&B, or post-SLAM evaluation.\n3. Re-run the corresponding preflight or checker. For config failures, use an\n   explicit scene config and absolute\u002Fworking-directory-correct input\u002Foutput\n   overrides. For extension failures, rebuild both pinned extensions for the\n   active Torch\u002FCUDA ABI and GPU architecture; never install a CPU substitute.\n4. For CUDA out-of-memory, reduce the smoke-test frame limit and mapping point\n   sample\u002Fiterations, close competing GPU jobs, and use a fresh output path.\n   Treat a successful reduced run as a diagnostic only.\n5. For W&B errors, disable it with `DISABLE_WANDB=true` and rerun; separately\n   repair writable run-directory\u002Fnetwork\u002Fauth configuration if online logging\n   is required.\n6. If tracking diverges, inspect the initial-loss message and try the\n   config-supported odometry initialization or a bounded iteration adjustment.\n   Keep the change in the saved config and record it; do not silently switch to\n   ground-truth poses.\n7. If SLAM completes but evaluation reports a traceback, verify\n   `estimated_c2w.ckpt` and `submaps\u002F` first, then treat the affected evaluation\n   artifact as missing. Evaluation failures do not prove tracking\u002Fmapping\n   failed, and `All done.✨` does not prove evaluation succeeded.\n\nSee [references\u002Ftroubleshooting.md](references\u002Ftroubleshooting.md) for symptom,\nlikely cause, safe check, and recovery tables.\n\n## Difficult synthetic cases for verification\n\n- **Case A — ignored override and portable logging:** provide a valid scene\n  config with `--track_w_color_loss 0.1 --gt_camera`, `DISABLE_WANDB=true`, and\n  a writable output path. The checker\u002Fhandoff must warn that those two options\n  do not update the config, while the run plan preserves the YAML odometry and\n  W&B-disabled state rather than claiming ground-truth tracking.\n- **Case B — boundary plus GPU failure:** use a tiny synthetic RGB-D fixture and\n  a config with `map_every: 1`, a boundary at the last frame, and a deliberately\n  oversized `new_submap_points_num` for the available VRAM. The expected\n  recovery is a preserved partial output, an OOM classification, a reduced\n  diagnostic config in a new output directory, and an explicit check that only\n  completed prior submaps are consumed—not a CPU rerun or an assumed final\n  checkpoint.\n\n## Scope and evidence limits\n\nThis skill is runtime-focused. It does not define dataset directory layouts,\ndownload commands, metric formulas, mesh evaluation semantics, or global-map\nquality thresholds. No full-scene execution, dataset download, SLURM submission,\nor W&B network\u002Flogin test was performed during construction. The environment\nreport proves package\u002FCUDA allocation\u002Fextension-import readiness on an SM80 A100\ninspection host, not end-to-end scene correctness on every supported dataset\nor GPU.\n\nLoad the focused references progressively:\n\n- [references\u002Fworkflows.md](references\u002Fworkflows.md) for execution stages and\n  output validation;\n- [references\u002Fcli-reference.md](references\u002Fcli-reference.md) for exact flags;\n- [references\u002Fcluster-reproduction.md](references\u002Fcluster-reproduction.md) for\n  a corrected, non-submitting cluster recipe;\n- [references\u002Ftroubleshooting.md](references\u002Ftroubleshooting.md) for recovery.\n","# Gaussian-SLAM runtime\n\nUse this skill when a user wants to execute, configure, reproduce, or recover a\nGaussian-SLAM run from this repository. The entry point is `run_slam.py`; the\nrepository's execution path is **CUDA-only**. Do not offer a CPU fallback for\ntracking, mapping, rendering, or evaluation. Dataset acquisition and dataset\nfile-layout instructions belong to `datasets-and-configuration`; evaluation\nmetric internals belong to `evaluation-and-mapping`.\n\n## Input\u002Foutput contract\n\n### Required inputs\n\n- A scene YAML path passed as the positional `config_path` argument.\n- A config resolving to `dataset_name`, `data`, `cam`, `tracking`, `mapping`,\n  `seed`, and `use_wandb`. A scene config normally inherits a dataset base\n  config and supplies `data.scene_name`, `data.input_path`, and\n  `data.output_path`.\n- A prepared RGB-D scene and camera metadata accepted by the selected dataset\n  adapter. The dataset's layout is intentionally not specified here; ask the\n  dataset skill for that contract.\n- A CUDA-capable PyTorch installation and both compiled extensions:\n  `simple_knn._C` and `gaussian_rasterizer._C`.\n\nUse explicit `--input_path` and `--output_path` values for portable runs. The\nconfig loader resolves `inherit_from` using the path as written, relative to\nthe process working directory, so launch from the repository root (or use\npaths that resolve from the chosen working directory).\n\n### Produced artifacts\n\n`run_slam.py` creates the output directory, saves the effective `config.yaml`,\nthen runs tracking\u002Fmapping followed by the repository evaluator. Common\nartifacts are:\n\n- `config.yaml`: effective merged configuration after CLI overrides;\n- `estimated_c2w.ckpt`: estimated camera-to-world poses;\n- `submaps\u002F000000.ckpt`, etc.: checkpoints containing Gaussian parameters and\n  the keyframe IDs belonging to the completed prior submap;\n- `mapping_vis\u002F*.jpg` and the created `tracking_vis\u002F` directory;\n- evaluator artifacts such as `rendering_metrics.json`,\n  `reconstruction_metrics.json`, `ate_aligned.json`, PNG summaries, a Replica\n  mesh, and a refined global-map PLY when the corresponding evaluation phase\n  succeeds.\n\nThe evaluator catches failures per evaluation phase and prints a traceback, so\nan `All done.✨` line is not sufficient evidence that every evaluation artifact\nexists. Check the expected files and the log. The active final Gaussian model\nis not explicitly finalized after the main loop; verify submap completeness\nbefore consuming checkpoints as a final map.\n\n## Required preflight\n\nRun these checks before a full scene. They are safe and do not download data,\nsubmit jobs, or contact W&B:\n\n```bash\nnvidia-smi\npython - \u003C\u003C'PY'\nimport torch\nprint(\"torch\", torch.__version__, \"cuda\", torch.version.cuda)\nassert torch.cuda.is_available(), \"Gaussian-SLAM requires CUDA\"\nprint(torch.cuda.get_device_name(0), torch.cuda.get_device_capability(0))\nPY\npython - \u003C\u003C'PY'\nimport simple_knn._C\nimport gaussian_rasterizer._C\nprint(\"CUDA extensions import\")\nPY\npython run_slam.py --help\npython skills\u002Fdisco\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002Fscripts\u002Fcheck_cli.py \\\n  configs\u002FReplica\u002Froom0.yaml\n```\n\nThe repository environment specifies Python 3.10, PyTorch 2.1.2 with CUDA\n12.1, torchvision 0.16.2, FAISS-GPU, Open3D 0.18.0, and pinned builds of the\ntwo CUDA extensions. Building extensions needs a compatible host compiler; the\nREADME requires exported `CC` and `CXX`. A verified inspection environment\npassed CUDA allocation and both extension imports on an NVIDIA A100-SXM4-40GB\nwith compute capability 8.0 (SM80). Treat that as evidence for A100\u002FSM80\ncompatibility, not as a promise for arbitrary GPUs. Match the PyTorch\u002FCUDA\nruntime, host compiler, toolkit, and extension build target to the actual\nmachine.\n\n`GaussianModel` imports `simple_knn._C.distCUDA2` at module import and uses it\nwhen adding points. `src\u002Futils\u002Futils.py` imports the custom\n`gaussian_rasterizer`; render settings, tensors, and camera optimization are\nalso hard-coded to CUDA. Missing either extension, a CPU-only Torch build, or a\nvisible device mismatch is a prerequisite failure—not something to work around\nby changing the device to CPU.\n\n## Standard run\n\nFrom the repository root, start with a small bounded scene\u002Fconfig if one is\navailable, then use the full run:\n\n```bash\npython run_slam.py configs\u002F\u003Cdataset>\u002F\u003Cscene>.yaml \\\n  --input_path \u003Cscene-input> \\\n  --output_path \u003Crun-output>\n```\n\nThe process is sequential: it loads and recursively merges YAML inheritance,\nupdates selected fields from CLI arguments, optionally initializes W&B, seeds\nPython\u002FNumPy\u002FPyTorch\u002FCUDA, runs `GaussianSLAM.run()`, and then constructs an\n`Evaluator` from the saved output config. The first two frames use dataset poses;\nsubsequent frames go through the configured tracker. Mapping runs on the\nconfigured keyframe cadence and always appends the last frame to the mapping\nIDs. Do not infer that `--gt_camera` changes this flow: the flag is parsed but\nis not applied by `update_config_with_args`; select\n`tracking.odometry_type: gt` in YAML when ground-truth tracking is intended.\n\nFor a deterministic, non-W&B smoke invocation, make the output unique and\nexplicit:\n\n```bash\nDISABLE_WANDB=true python run_slam.py configs\u002F\u003Cdataset>\u002F\u003Cscene>.yaml \\\n  --input_path \u003Cscene-input> \\\n  --output_path \u003Crun-output> \\\n  --seed 0\n```\n\nThis controls the software RNGs (`torch`, CUDA, NumPy, Python, and hash seed)\nand disables cuDNN benchmarking, but the README explicitly warns that the\nDifferential Gaussian rasterizer is nondeterministic. Compare runs with a\nsmall tolerance and record GPU, driver, extension build, config, and seed.\n\n## CLI and configuration\n\nRead [references\u002Fcli-reference.md](references\u002Fcli-reference.md) before changing\nhyperparameters. It records every parser option, its target config field, and\nsource-level caveats (including parsed-but-ineffective options and truthiness\nchecks that ignore zero). Use the checker to catch missing inherited fields and\ncontradictory filter flags before starting a GPU job.\n\nThe highest-risk runtime changes are tracking iterations\u002Floss weighting and\nmapping cadence\u002Fiterations. Large new-submap point counts, high-resolution\nframes, and long optimization schedules increase VRAM and wall time. Lower\ncadence or iteration counts only for a smoke test; do not call a shortened run\npaper reproduction.\n\n## Tracking, mapping, and submaps\n\n- **Tracking.** `Tracker.track` renders the current Gaussian model and optimizes\n  camera rotation\u002Ftranslation using RGB and depth losses. `odometry_type` is\n  `gt`, `const_speed`, or `odometer`; the latter uses the configured visual\n  odometer for initialization. Alpha and depth-outlier filters affect the\n  tracking mask. A high initial loss can double tracking iterations; with\n  `help_camera_initialization`, the tracker may retry initialization with an\n  odometer estimate.\n- **Mapping.** At mapping frames, the mapper creates a keyframe, seeds 3D\n  points from RGB-D and the estimated pose, adds non-duplicate points to the\n  current Gaussian model using `distCUDA2`, optimizes the current submap, and\n  writes a visualization. New submaps use edge\u002Fgradient seeding; an existing\n  submap uses alpha\u002Fdepth-error regions. Dataset-specific point-cloud filtering\n  is an implementation detail; do not substitute a missing dataset contract.\n- **Submap boundaries.** With `submap_using_motion_heuristic: true`, the source\n  starts a new submap when motion from the last submap start exceeds a hard-coded\n  50-degree rotation or 0.5 translation threshold. Otherwise it uses\n  `new_submap_every` frame IDs and also includes the last frame. At a boundary,\n  the prior model is written as a zero-padded checkpoint, keyframes are reset,\n  and a new model starts at that frame. The loop saves poses at the end but does\n  not perform a separate final active-model checkpoint; inspect the output\n  rather than assuming the last active model is serialized.\n\n## W&B and seed controls\n\nConfigs default to `use_wandb: False`. For an offline\u002Flocal run, prefer that\nsetting or set `DISABLE_WANDB=true`; the entry point treats exactly the string\n`true` as a forced disable. If W&B is enabled, `run_slam.py` calls `wandb.init`,\nlogs source code, and uploads tracking\u002Fmapping\u002Fevaluation records. Credentials\nand network access are not required for the default disabled path. If a site\nneeds W&B, decide explicitly whether online or offline mode is appropriate and\nensure its run directory is writable: the source contains a site-specific\nabsolute W&B directory that may not exist on another machine, and it is not a\nCLI setting. Patch it to a writable path or keep W&B disabled; do not expose\ncredentials in configs or SLURM scripts.\n\n## SLURM\n\nUse [references\u002Fcluster-reproduction.md](references\u002Fcluster-reproduction.md) or\nrun the safe planner:\n\n```bash\npython skills\u002Fdisco\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002Fscripts\u002Frender_sbatch_plan.py \\\n  --dataset Replica --scene room0 --scene room1 \\\n  --input-root \u003Cinput-root> --output-root \u003Coutput-root>\n```\n\nThe planner only prints a reviewable batch script; it never runs `sbatch`,\nstarts Python SLAM, downloads data, or authenticates W&B. Review the generated\narray bounds, config directory, scene names, input\u002Foutput roots, partition,\nGPU request, and environment activation before manually submitting it. The\nrepository example contains site placeholders, a dataset-directory case\nmismatch for ScanNet++, a fixed array example that must be changed per dataset,\nand a trailing continuation after `--group_name`; use the corrected recipe in\nthe reference instead of copying it verbatim.\n\n## Failure recovery\n\n1. Preserve the run log and output directory. Do not delete partial submaps or\n   overwrite a prior run while diagnosing it.\n2. Classify the first failure: environment\u002Fextension, config resolution,\n   dataset contract, CUDA\u002FVRAM, W&B, or post-SLAM evaluation.\n3. Re-run the corresponding preflight or checker. For config failures, use an\n   explicit scene config and absolute\u002Fworking-directory-correct input\u002Foutput\n   overrides. For extension failures, rebuild both pinned extensions for the\n   active Torch\u002FCUDA ABI and GPU architecture; never install a CPU substitute.\n4. For CUDA out-of-memory, reduce the smoke-test frame limit and mapping point\n   sample\u002Fiterations, close competing GPU jobs, and use a fresh output path.\n   Treat a successful reduced run as a diagnostic only.\n5. For W&B errors, disable it with `DISABLE_WANDB=true` and rerun; separately\n   repair writable run-directory\u002Fnetwork\u002Fauth configuration if online logging\n   is required.\n6. If tracking diverges, inspect the initial-loss message and try the\n   config-supported odometry initialization or a bounded iteration adjustment.\n   Keep the change in the saved config and record it; do not silently switch to\n   ground-truth poses.\n7. If SLAM completes but evaluation reports a traceback, verify\n   `estimated_c2w.ckpt` and `submaps\u002F` first, then treat the affected evaluation\n   artifact as missing. Evaluation failures do not prove tracking\u002Fmapping\n   failed, and `All done.✨` does not prove evaluation succeeded.\n\nSee [references\u002Ftroubleshooting.md](references\u002Ftroubleshooting.md) for symptom,\nlikely cause, safe check, and recovery tables.\n\n## Difficult synthetic cases for verification\n\n- **Case A — ignored override and portable logging:** provide a valid scene\n  config with `--track_w_color_loss 0.1 --gt_camera`, `DISABLE_WANDB=true`, and\n  a writable output path. The checker\u002Fhandoff must warn that those two options\n  do not update the config, while the run plan preserves the YAML odometry and\n  W&B-disabled state rather than claiming ground-truth tracking.\n- **Case B — boundary plus GPU failure:** use a tiny synthetic RGB-D fixture and\n  a config with `map_every: 1`, a boundary at the last frame, and a deliberately\n  oversized `new_submap_points_num` for the available VRAM. The expected\n  recovery is a preserved partial output, an OOM classification, a reduced\n  diagnostic config in a new output directory, and an explicit check that only\n  completed prior submaps are consumed—not a CPU rerun or an assumed final\n  checkpoint.\n\n## Scope and evidence limits\n\nThis skill is runtime-focused. It does not define dataset directory layouts,\ndownload commands, metric formulas, mesh evaluation semantics, or global-map\nquality thresholds. No full-scene execution, dataset download, SLURM submission,\nor W&B network\u002Flogin test was performed during construction. The environment\nreport proves package\u002FCUDA allocation\u002Fextension-import readiness on an SM80 A100\ninspection host, not end-to-end scene correctness on every supported dataset\nor GPU.\n\nLoad the focused references progressively:\n\n- [references\u002Fworkflows.md](references\u002Fworkflows.md) for execution stages and\n  output validation;\n- [references\u002Fcli-reference.md](references\u002Fcli-reference.md) for exact flags;\n- [references\u002Fcluster-reproduction.md](references\u002Fcluster-reproduction.md) for\n  a corrected, non-submitting cluster recipe;\n- [references\u002Ftroubleshooting.md](references\u002Ftroubleshooting.md) for recovery.\n",{"name":5,"license":15,"metadata":24,"description":7,"disable-model-invocation":26},{"disco-role":25},"operating",true,"main","d69096e239f2a89aabc2cd078c47664a9d3b1e11","04-安全渗透\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime",3,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts","https:\u002F\u002Fgithub.com\u002Flza6","个人vx：Tf00798\r\n个人✈：@TFGY999","2026-10-08 14:29:22",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":26},[37,40,43],{"tool":38,"command":39},"npx","npx skills add lza6\u002FArmor-piercing-system-prompts --skill slam-runtime",{"tool":41,"command":42},"bunx","bunx skills add lza6\u002FArmor-piercing-system-prompts --skill slam-runtime",{"tool":44,"command":45},"pnpm","pnpm dlx skills add lza6\u002FArmor-piercing-system-prompts --skill slam-runtime","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F04-安全渗透\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002FSKILL.md",[48,54,58,62],{"path":49,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":52,"sourceHtmlUrl":53},"references\u002Fcli-reference.md","REFERENCE","参考文档",4757,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F04-安全渗透\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002Freferences\u002Fcli-reference.md",{"path":55,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":56,"sourceHtmlUrl":57},"references\u002Fcluster-reproduction.md",3988,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F04-安全渗透\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002Freferences\u002Fcluster-reproduction.md",{"path":59,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":60,"sourceHtmlUrl":61},"references\u002Ftroubleshooting.md",6666,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F04-安全渗透\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002Freferences\u002Ftroubleshooting.md",{"path":63,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":64,"sourceHtmlUrl":65},"references\u002Fworkflows.md",6128,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F04-安全渗透\u002Fgaussian-slam\u002Fsub-skills\u002Fslam-runtime\u002Freferences\u002Fworkflows.md",[67,75,80,87,94,101],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":74},"01M7FFG485KY83TD6HSC4YPCMC","obsidian-clipper-template-creator","Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator\u002FSKILL.md",[],[],"2026-11-16 10:05:34",{"skillNo":76,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":74},"01M7FFG0T6AZ0STEJPSB3CX84K","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator-antigravity-awesome-skills-main\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":74},"01M7FFFY0QEM14SKWPDYGMXM2P","python-api","Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP\u002FmAP semantics, and source-style import diagnostics.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Fpython-api\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":74},"01M7FFFVZQZ2GJ0C0TGP7JHD20","file-evaluation","Folder-based PASCAL VOC AP\u002FmAP evaluation from ground-truth and detection text files.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Ffile-evaluation\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":74},"01M7FFFSFNYSE77SJ6PE1CK2DA","object-detection-metrics","Guides legacy Object-Detection-Metrics PASCAL VOC AP\u002FmAP evaluation from detection text folders and direct Python API classes.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002FSKILL.md",[],[],{"skillNo":102,"name":103,"nameZh":6,"description":104,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":74},"01M7FFFP0GCK1RVXBCPQ8APBMT","nx-workspace-patterns","Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fnx-workspace-patterns\u002FSKILL.md",[],[]]