[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M753DM5Y1CF77260H6K2Y0PT":3,"skill-same:01M753DM5Y1CF77260H6K2Y0PT":60},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":31,"install":32,"resources":45},"01M753DM5Y1CF77260H6K2Y0PT","prepare-repo-skill-env",null,"Create, install, and prove the minimum backend-aware Python inspection environment set for a local repository before create-repo-skill. Use when a repo package must be installed into new or existing Conda, micromamba, uv, or venv environments; when selected native tests\u002Fexamples require CPU, CUDA, ROCm, MPS, or vendor accelerator dependencies; when a broken inspection environment needs diagnosis; or when hardware\u002Fbackend compatibility must be verified. Consume the confirmed extraction scope and backend verification plan, then run explicit terminal commands instead of delegating environment decisions to a bundled setup script.","VectorSpaceLab\u002FAREX-Skill","VectorSpaceLab","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F180637102?v=4","01M7538C5WPSJGWF7KHR4EFZ9F","01M7538C6R2GFP53NPXTCZNBRT",351,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fprepare-repo-skill-env\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: prepare-repo-skill-env\ndescription: \"Create, install, and prove the minimum backend-aware Python inspection environment set for a local repository before create-repo-skill. Use when a repo package must be installed into new or existing Conda, micromamba, uv, or venv environments; when selected native tests\u002Fexamples require CPU, CUDA, ROCm, MPS, or vendor accelerator dependencies; when a broken inspection environment needs diagnosis; or when hardware\u002Fbackend compatibility must be verified. Consume the confirmed extraction scope and backend verification plan, then run explicit terminal commands instead of delegating environment decisions to a bundled setup script.\"\nmetadata:\n  disco-role: meta\n---\n\n# Prepare Repo Skill Env\n\n## Purpose\n\nUse this skill before `create-repo-skill` when a local repository needs a\nverified Python environment for live API and runtime inspection. The caller\nshould provide or let this skill infer:\n\n- The repository path.\n- A new private environment prefix, or an existing environment that may be\n  inspected.\n- The confirmed extraction scope, backend-classified native candidate map, and\n  backend verification plan from `create-repo-skill`, when applicable.\n\nCreate the environment, install only what the confirmed scope requires, and\nprove that the package works from the environment's Python. \"Minimum\" means the\nsmallest environment set that covers every required backend in the selected\nscope, not CPU-only by default. Usually one GPU-capable environment can also\ncover CPU checks; create additional prefixes only for conflicting backend\nvariants or an independently selected CPU fallback claim. Run the relevant\nConda, micromamba, uv, Python, and pip commands directly in the terminal. Keep\neach material step visible and separately diagnosable; do not hide environment\ncreation, dependency selection, installation, or verification behind a bundled\nsetup\u002Fbootstrap script.\n\nPrefer Conda when it is already available because it handles Python versions\nand compiled dependencies well. Micromamba is an equivalent prefix-based path\nwhen it is already installed. If neither is available, use an existing\ncompatible Python with `venv`, or an installed `uv` when it can create the\nrequested environment. Do not mutate Conda `base`, the Python running DisCo, or\na user-owned environment without authorization.\n\nDo not run the repo-native candidate tests\u002Fexamples during environment\npreparation. Install their required variants and run the smallest backend smoke\nchecks needed to prove that final native verification is feasible; the native\ncases themselves run after the generated skill is integrated.\n\nThe default private prefix is\n`$DISCO_CODING_AGENT_DIR\u002Fenvs\u002F\u003Cskill-id>-inspection` when that variable is set,\notherwise `~\u002F.disco\u002Fagent\u002Fenvs\u002F\u003Cskill-id>-inspection`. Resolve it to a concrete\nabsolute path before executing commands.\n\nThis skill produces private setup evidence. Environment paths, checkout paths,\ncommands, and logs may be handed to `create-repo-skill`, but must not be copied\ninto the generated public repo skill.\n\n## Required Outputs\n\nDeliver an `ok`, `partial`, or `failed` environment handoff. In every case,\nwrite a private `repo_env_report.json` in the caller's artifact directory, or\nanother explicit private path when invoked directly.\n\nFor success, record:\n\n- Repository path, environment manager, prefix, and environment Python.\n- Confirmed extraction scope and the dependency groups selected or skipped.\n- Backend verification plan, required\u002Foptional\u002Falternative candidates, CPU\n  substitution decisions, hardware compatibility verdicts, and the minimum\n  environment set selected.\n- Exact environment creation, install, and verification commands with exit\n  status and concise output\u002Ferror evidence.\n- Installed distribution names and successfully imported module names.\n- Hardware\u002Fbackend verdict when relevant.\n- Any unsafe, credentialed, destructive, or long-running checks that were\n  deliberately skipped.\n\nUse `partial` only when a required backend cannot be prepared and the user\nexplicitly authorizes skill drafting with that exact limitation. A partial\nhandoff is not ready for full backend verification or auto-import. Without that\nauthorization, report `failed`.\n\nFor failure, record the failed phase, exact command, exit status, key stderr,\nrelevant platform\u002Fbackend facts, attempted remedies, and the next viable\naction. State explicitly that the environment is not ready for\n`create-repo-skill`.\n\n## Reference Map\n\nRead the relevant references before acting:\n\n- [references\u002Finstall-planning.md](references\u002Finstall-planning.md): packaging\n  discovery, backend verification plan interpretation, environment-manager\n  selection, direct command templates, extraction-scope dependency choices,\n  and install order.\n- [references\u002Fhardware-and-backends.md](references\u002Fhardware-and-backends.md):\n  accelerator probing, backend package choices, and compatibility failures.\n- [references\u002Fverification-and-failure-report.md](references\u002Fverification-and-failure-report.md):\n  direct verification commands, the report contract, and handoff templates.\n\n## Workflow\n\n### 1. Resolve Inputs and Mutation Boundaries\n\nResolve these before installation:\n\n- `repo_path`: canonical local checkout path.\n- `environment_prefix`: canonical target prefix\u002Fpath.\n- Confirmed included\u002Fexcluded directories, selected workflows, and user\n  requirements from `create-repo-skill` when available.\n- Backend-classified native candidates owned by included workflows. Require\n  backend requirement, criticality, CPU substitution, dependency variant,\n  hardware prerequisites, preparation smoke, and final native expectation.\n- The caller's backend verification plan and proposed minimum environment set.\n  If either is missing during a `create-repo-skill` call, derive it from repo\n  evidence before installation rather than defaulting silently to CPU.\n- Python version supported by repo metadata and required backend wheels.\n- Distribution name and import module names.\n- Required extras, requirements files, build tools, and backend packages.\n- Backend expectation: `auto`, `cpu`, `cuda`, `rocm`, `mps`, or another\n  documented backend.\n- Whether the prefix is new, private and reusable, or user-owned.\n- Whether potentially breaking mutation of an existing environment is already\n  authorized.\n\nWhen called automatically by `create-repo-skill`, use its resolved private\ndefault prefix without asking again. During a direct invocation, ask for a\nprefix or propose the private default before writing there. Never infer a path\noutside the stated default-prefix policy.\n\nIf the prefix exists, inspect it read-only first. Imports, version queries,\n`pip check`, metadata queries, and safe CLI help are read-only. Installing,\nupgrading, downgrading, uninstalling, or repairing packages is mutation. Ask\nbefore mutating a user-provided environment unless the user already authorized\nthat operation; otherwise create a new private prefix when allowed.\n\n### 2. Inspect the Repository Before Choosing Commands\n\nRead [references\u002Finstall-planning.md](references\u002Finstall-planning.md). Inspect\n`pyproject.toml`, `setup.cfg`, `setup.py`, requirements files, lockfiles,\n`environment.yml`, CI install steps, source roots, entry points, and documented\ninstall variants.\n\nBuild a concise install and backend map before running anything:\n\n```text\nSelected Python: 3.11 (repo supports >=3.10; required wheels available)\nIncluded scope: src\u002Fpackage, docs\u002Fgpu-inference.md, examples\u002Finfer.py\nExcluded scope: training\u002F, benchmarks\u002F\nRequired candidate: examples\u002Finfer.py | cuda | required | CPU substitute: none\nOptional candidate: benchmarks\u002Fthroughput.py | cuda | optional | excluded\nHost verdict: compatible NVIDIA GPU\u002Fdriver\u002Fwheel\nEnvironment set: one CUDA-capable prefix covering CUDA and CPU checks\nInstall: base package + [inference,cuda] + compatible torch CUDA wheel\nSkip: [train], [all], requirements-dev.txt, benchmark dependencies\nPreparation smoke: torch CUDA allocation + package backend import\nFinal native case: examples\u002Finfer.py with tiny fixture (run after integration)\n```\n\nDo not assume the repository name is the distribution or import name. Do not\ninstall all extras or every requirements file as a substitute for analysis.\n\n### 3. Probe Available Managers and Hardware\n\nRun discovery commands separately so their outcomes remain visible:\n\n```bash\ncommand -v conda\ncommand -v micromamba\ncommand -v uv\ncommand -v python3.11\ncommand -v python3\ncommand -v python\n```\n\nUse platform-equivalent commands on Windows. Query versions for tools that\nexist. Do not treat a missing command as a fatal error until all suitable paths\nhave been evaluated.\n\nRead [references\u002Fhardware-and-backends.md](references\u002Fhardware-and-backends.md)\nand probe only the hardware relevant to the requested or repo-required backend.\nA visible GPU does not by itself require a GPU install for package inspection.\n\nApply the backend plan after probing:\n\n- When an included capability has a `required` backend and `CPU substitute:\n  none`, prepare and verify that backend whenever compatible hardware and\n  packages are available. A CPU-only prefix is not ready for that scope.\n- When `CPU substitute` is `full` and the selected CPU candidate verifies the\n  same behavior, use the CPU path and record the evidence.\n- When the backend is `optional`, omit it from the minimum environment when it\n  is unavailable or outside the selected scope, but preserve an explicit\n  unverified-capability note.\n- When required hardware or a compatible dependency variant is unavailable,\n  do not install an arbitrary GPU wheel or call the CPU environment `ok`.\n  Report `failed`, ask to narrow scope, or produce `partial` only after the user\n  accepts drafting with the required-backend limitation.\n- When required backend variants conflict, create separate concrete prefixes\n  and verify each one. Do not mutate one prefix repeatedly between incompatible\n  CPU\u002FCUDA\u002FROCm\u002Fvendor states.\n\nIf no suitable Conda\u002Fmicromamba, uv, or Python runtime exists, do not invoke a\nbundled downloader. Explain the host change required and ask before installing\na runtime or environment manager unless the user already authorized that exact\nkind of host-level installation. Once authorized, use the platform's normal\npackage manager or the selected tool's documented installer with explicit,\nreviewable terminal commands, then resume this workflow. Do not silently curl\nand execute an installer.\n\n### 4. Create the Required Environment Set with Direct Commands\n\nSubstitute concrete absolute paths and the selected Python version in every\ncommand. Do not execute examples with placeholder paths.\n\nCreate the primary prefix for the strongest required selected backend. A\nGPU-capable prefix normally also runs CPU checks. When the backend plan requires\nincompatible variants, derive distinct suffixes such as `-cuda`, `-rocm`, or\n`-cpu` from the caller's private prefix and record every prefix in the report.\n\nFor Conda:\n\n```bash\nconda create --yes --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" \"python=3.11\" pip\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -c \"import sys; print(sys.executable); print(sys.version)\"\n```\n\nFor an already installed micromamba:\n\n```bash\nmicromamba create --yes --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" \"python=3.11\" pip\nmicromamba run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -c \"import sys; print(sys.executable); print(sys.version)\"\n```\n\nFor venv on Unix-like systems:\n\n```bash\npython3.11 -m venv \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\"\n\"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\" -c \"import sys; print(sys.executable); print(sys.version)\"\n```\n\nOn Windows, use the environment's `Scripts\u002Fpython.exe`. If installed `uv` is\nthe best available path, make its runtime acquisition explicit:\n\n```bash\nuv venv --seed --python 3.11 \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\"\n\"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\" -c \"import sys; print(sys.executable); print(sys.version)\"\n```\n\nIf `uv` would download a Python runtime, treat that as runtime installation and\napply the authorization rule from step 3.\n\nDo not rely on shell activation. For subsequent Conda\u002Fmicromamba commands use\n`run --prefix`; for venv\u002Fuv use the environment Python's absolute path. This\nprevents commands from accidentally using DisCo's or the user's current\nPython.\n\n### 5. Install the Minimum Required Package Set\n\nExecute one meaningful install phase at a time, with a finite terminal-tool\ntimeout and observable output. A typical new Conda environment uses commands\nlike:\n\n```bash\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -m pip install -e \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\"\n```\n\nFor venv\u002Fuv:\n\n```bash\n\"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\" -m pip install -e \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\"\n```\n\nAdd only evidence-backed requirements or extras:\n\n```bash\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -m pip install -r \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\u002Frequirements-inference.txt\"\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -m pip install -e \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo[inference]\"\n```\n\nFollow repo documentation when backend foundations such as torch, JAX,\nTensorFlow, or a compiler\u002Ftoolkit must be installed first. Do not reflexively\nupgrade pip\u002Fsetuptools\u002Fwheel or broad dependency sets, especially in an\nexisting environment; do so only when repo\u002Fbuild evidence requires it.\n\nFor every required backend candidate, install its evidence-backed dependency\nvariant into the prefix assigned by the backend plan. Do not install a CPU-only\ntorch\u002FJAX\u002FTensorFlow build into a required CUDA\u002FROCm environment and then count\npackage importability as backend coverage. Conversely, do not install GPU\nstacks for optional or excluded candidates merely because hardware is visible.\n\nDo not combine creation, multiple installs, and verification into one large\nshell command. After each phase, record the exact command, exit status, elapsed\ntime, and relevant output in the private report. On failure, diagnose the\nspecific phase and adjust the next visible command rather than rerunning an\nopaque all-in-one installer.\n\nIf metadata resolution or downloads stall, interrupt the command, distinguish\nnetwork delay from compilation, and use the network guidance in\n[references\u002Finstall-planning.md](references\u002Finstall-planning.md). Proxy or VPN\ndetails supplied by the user are local execution context only; never bake them\ninto the skill or generated package content.\n\n### 6. Verify from the Environment Python\n\nRead [references\u002Fverification-and-failure-report.md](references\u002Fverification-and-failure-report.md).\nAt minimum, run direct commands for:\n\n- Environment Python identity and version.\n- `python -m pip check`.\n- Distribution metadata from the environment.\n- Imports using the environment Python, preferably from outside the checkout\n  and with `-I` when compatible, so the current directory or user site cannot\n  make a broken install look healthy.\n- Relevant CLI\u002FAPI smoke checks.\n- A preparation smoke check for every required backend in the backend plan,\n  using the prefix assigned to that backend. For CUDA\u002FROCm\u002FMPS or a vendor\n  accelerator, prove framework\u002Fbackend availability and a minimal device\n  operation when safe.\n\nExamples for Conda:\n\n```bash\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -m pip check\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -I -c \"from importlib.metadata import version; print(version('distribution-name'))\"\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -I -c \"import package_name; print(package_name.__file__)\"\n```\n\nEquivalent venv commands must use the environment Python's absolute path. A\nsuccessful environment creation or `pip install` is not verification. Fix and\nrerun every applicable failed gate before an `ok` handoff. Do not run the\nrepo-native cases yet; record which final cases this prepared environment must\nsupport after skill integration.\n\n### 7. Write the Report and Handoff\n\nWrite `repo_env_report.json` from the actual command results. Do not invent\nmissing output or mark skipped gates as passed. Include at least:\n\n```json\n{\n  \"schemaVersion\": 2,\n  \"status\": \"ok\",\n  \"readiness\": {\n    \"skillDrafting\": true,\n    \"fullBackendVerification\": true,\n    \"backendGateEligibleForAutoImport\": true\n  },\n  \"repositoryPath\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\",\n  \"environment\": {\n    \"manager\": \"conda\",\n    \"prefix\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\",\n    \"pythonExecutable\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\",\n    \"backends\": [\"cpu\", \"cuda\"]\n  },\n  \"workflowEnvironment\": {\n    \"executable\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\",\n    \"cwd\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\",\n    \"package\": \"distribution-name\",\n    \"version\": \"1.2.3\"\n  },\n  \"additionalEnvironments\": [],\n  \"backendPlan\": {\n    \"required\": [\n      {\n        \"backend\": \"cuda\",\n        \"capability\": \"GPU inference\",\n        \"cpuSubstitute\": \"none\",\n        \"hostVerdict\": \"compatible\",\n        \"dependencyVariant\": \"repo cuda extra plus compatible framework wheel\",\n        \"preparationSmoke\": \"passed\",\n        \"finalNativeCandidates\": [\"examples\u002Finfer.py\"]\n      }\n    ],\n    \"optional\": [],\n    \"alternative\": []\n  },\n  \"installPlan\": {\n    \"python\": \"3.11\",\n    \"includedScope\": [],\n    \"excludedScope\": [],\n    \"installedGroups\": [],\n    \"skippedGroups\": []\n  },\n  \"commands\": [],\n  \"verification\": {},\n  \"warnings\": [],\n  \"failures\": []\n}\n```\n\n`environment.pythonExecutable` remains private report evidence. It is not the\nobject to pass directly to workflow `agent()`. The caller must use the canonical\n`workflowEnvironment` object with `executable`, optional `cwd`, distribution\n`package`, and exact `version`.\n\nUse `status: \"ok\"` only when every required preparation\u002Fbackend gate passes.\nUse `status: \"partial\"` only after explicit user acceptance, with\n`skillDrafting: true`, `fullBackendVerification: false`, and\n`backendGateEligibleForAutoImport: false`. Use `status: \"failed\"` when a\nrequired gate remains unresolved without that acceptance. The `commands`\nentries should identify phase, command, exit status, and concise stdout\u002Fstderr\nevidence; do not put credentials or secret-bearing environment variables in\nthe report.\n\nFor every handoff, provide:\n\n```text\nEnvironment status: \u003Cok|partial|failed>\nReady for skill drafting: \u003Cyes|no>\nReady for full backend verification: \u003Cyes|no>\nBackend gate eligible for auto-import: \u003Cyes|no>\nRepository path: \u003Crepo_path>\nEnvironment manager: \u003Cconda|micromamba|venv|uv>\nEnvironment prefix: \u003Cenvironment_prefix>\nAdditional environment(s): \u003Cprefix\u002Fbackend or none>\nTemporary inspection Python: \u003Cpython_executable>\nWorkflow environment: {\"executable\":\"\u003Cpython_executable>\",\"cwd\":\"\u003Crepo_path>\",\"package\":\"\u003Cdistribution_name>\",\"version\":\"\u003Cverified_version>\"}\nInstalled package name(s): \u003Cdistribution names>\nVerified import(s): \u003Cmodules>\nVerification report: \u003Crepo_env_report.json>\nRequired backend verdict(s): \u003Cbackend: preparation smoke passed\u002Fblocked>\nFinal native backend case(s): \u003Ccandidate ids to run after skill integration>\nAccepted partial limitation: \u003Cexact user-accepted block or none>\nAdditional notes: \u003Cextras installed, skipped unsafe checks, limitations>\n```\n\n`create-repo-skill` may continue normally from `ok`. It may continue drafting\nfrom `partial` only after preserving the accepted backend block and disabling\nauto-import. It must not continue from `failed` without repair or an explicitly\nnarrowed extraction scope.\n\n## Non-Negotiables\n\n- Execute environment creation, package installation, and verification as\n  explicit terminal commands; do not route them through a bundled wrapper.\n- Never install into or mutate Conda `base`.\n- Never use DisCo's current Python as the target environment unless the user\n  explicitly requests and authorizes that mutation.\n- Never delete or recreate an existing prefix without explicit authorization.\n- Never mutate a user-provided environment in a potentially breaking way\n  without authorization; use a new private prefix when permitted.\n- Never install broad extras, dev requirements, or backend packages unrelated\n  to the confirmed extraction scope.\n- Never call a CPU-only environment `ok` for a required CUDA\u002FROCm\u002FMPS\u002Fvendor\n  capability whose CPU substitute is `partial` or `none`.\n- Never silently downgrade a required backend candidate to optional because\n  hardware, wheels, drivers, or toolkits are unavailable.\n- Never mark a partial environment eligible for auto-import. Preserve the\n  required-backend block for final native verification and informed user\n  acceptance.\n- Never rely only on shell activation or an unqualified `python`\u002F`pip` after\n  environment creation.\n- Never treat successful creation or installation as proof of usability.\n- Never hide a failed or skipped verification gate in an `ok` report.\n- Never wait indefinitely on a stalled resolver, download, or build. Use finite\n  timeouts, retain useful failure evidence, and diagnose before retrying.\n- Never copy private environment paths, proxy details, command logs, or setup\n  reports into the generated public repo skill.\n","# Prepare Repo Skill Env\n\n## Purpose\n\nUse this skill before `create-repo-skill` when a local repository needs a\nverified Python environment for live API and runtime inspection. The caller\nshould provide or let this skill infer:\n\n- The repository path.\n- A new private environment prefix, or an existing environment that may be\n  inspected.\n- The confirmed extraction scope, backend-classified native candidate map, and\n  backend verification plan from `create-repo-skill`, when applicable.\n\nCreate the environment, install only what the confirmed scope requires, and\nprove that the package works from the environment's Python. \"Minimum\" means the\nsmallest environment set that covers every required backend in the selected\nscope, not CPU-only by default. Usually one GPU-capable environment can also\ncover CPU checks; create additional prefixes only for conflicting backend\nvariants or an independently selected CPU fallback claim. Run the relevant\nConda, micromamba, uv, Python, and pip commands directly in the terminal. Keep\neach material step visible and separately diagnosable; do not hide environment\ncreation, dependency selection, installation, or verification behind a bundled\nsetup\u002Fbootstrap script.\n\nPrefer Conda when it is already available because it handles Python versions\nand compiled dependencies well. Micromamba is an equivalent prefix-based path\nwhen it is already installed. If neither is available, use an existing\ncompatible Python with `venv`, or an installed `uv` when it can create the\nrequested environment. Do not mutate Conda `base`, the Python running DisCo, or\na user-owned environment without authorization.\n\nDo not run the repo-native candidate tests\u002Fexamples during environment\npreparation. Install their required variants and run the smallest backend smoke\nchecks needed to prove that final native verification is feasible; the native\ncases themselves run after the generated skill is integrated.\n\nThe default private prefix is\n`$DISCO_CODING_AGENT_DIR\u002Fenvs\u002F\u003Cskill-id>-inspection` when that variable is set,\notherwise `~\u002F.disco\u002Fagent\u002Fenvs\u002F\u003Cskill-id>-inspection`. Resolve it to a concrete\nabsolute path before executing commands.\n\nThis skill produces private setup evidence. Environment paths, checkout paths,\ncommands, and logs may be handed to `create-repo-skill`, but must not be copied\ninto the generated public repo skill.\n\n## Required Outputs\n\nDeliver an `ok`, `partial`, or `failed` environment handoff. In every case,\nwrite a private `repo_env_report.json` in the caller's artifact directory, or\nanother explicit private path when invoked directly.\n\nFor success, record:\n\n- Repository path, environment manager, prefix, and environment Python.\n- Confirmed extraction scope and the dependency groups selected or skipped.\n- Backend verification plan, required\u002Foptional\u002Falternative candidates, CPU\n  substitution decisions, hardware compatibility verdicts, and the minimum\n  environment set selected.\n- Exact environment creation, install, and verification commands with exit\n  status and concise output\u002Ferror evidence.\n- Installed distribution names and successfully imported module names.\n- Hardware\u002Fbackend verdict when relevant.\n- Any unsafe, credentialed, destructive, or long-running checks that were\n  deliberately skipped.\n\nUse `partial` only when a required backend cannot be prepared and the user\nexplicitly authorizes skill drafting with that exact limitation. A partial\nhandoff is not ready for full backend verification or auto-import. Without that\nauthorization, report `failed`.\n\nFor failure, record the failed phase, exact command, exit status, key stderr,\nrelevant platform\u002Fbackend facts, attempted remedies, and the next viable\naction. State explicitly that the environment is not ready for\n`create-repo-skill`.\n\n## Reference Map\n\nRead the relevant references before acting:\n\n- [references\u002Finstall-planning.md](references\u002Finstall-planning.md): packaging\n  discovery, backend verification plan interpretation, environment-manager\n  selection, direct command templates, extraction-scope dependency choices,\n  and install order.\n- [references\u002Fhardware-and-backends.md](references\u002Fhardware-and-backends.md):\n  accelerator probing, backend package choices, and compatibility failures.\n- [references\u002Fverification-and-failure-report.md](references\u002Fverification-and-failure-report.md):\n  direct verification commands, the report contract, and handoff templates.\n\n## Workflow\n\n### 1. Resolve Inputs and Mutation Boundaries\n\nResolve these before installation:\n\n- `repo_path`: canonical local checkout path.\n- `environment_prefix`: canonical target prefix\u002Fpath.\n- Confirmed included\u002Fexcluded directories, selected workflows, and user\n  requirements from `create-repo-skill` when available.\n- Backend-classified native candidates owned by included workflows. Require\n  backend requirement, criticality, CPU substitution, dependency variant,\n  hardware prerequisites, preparation smoke, and final native expectation.\n- The caller's backend verification plan and proposed minimum environment set.\n  If either is missing during a `create-repo-skill` call, derive it from repo\n  evidence before installation rather than defaulting silently to CPU.\n- Python version supported by repo metadata and required backend wheels.\n- Distribution name and import module names.\n- Required extras, requirements files, build tools, and backend packages.\n- Backend expectation: `auto`, `cpu`, `cuda`, `rocm`, `mps`, or another\n  documented backend.\n- Whether the prefix is new, private and reusable, or user-owned.\n- Whether potentially breaking mutation of an existing environment is already\n  authorized.\n\nWhen called automatically by `create-repo-skill`, use its resolved private\ndefault prefix without asking again. During a direct invocation, ask for a\nprefix or propose the private default before writing there. Never infer a path\noutside the stated default-prefix policy.\n\nIf the prefix exists, inspect it read-only first. Imports, version queries,\n`pip check`, metadata queries, and safe CLI help are read-only. Installing,\nupgrading, downgrading, uninstalling, or repairing packages is mutation. Ask\nbefore mutating a user-provided environment unless the user already authorized\nthat operation; otherwise create a new private prefix when allowed.\n\n### 2. Inspect the Repository Before Choosing Commands\n\nRead [references\u002Finstall-planning.md](references\u002Finstall-planning.md). Inspect\n`pyproject.toml`, `setup.cfg`, `setup.py`, requirements files, lockfiles,\n`environment.yml`, CI install steps, source roots, entry points, and documented\ninstall variants.\n\nBuild a concise install and backend map before running anything:\n\n```text\nSelected Python: 3.11 (repo supports >=3.10; required wheels available)\nIncluded scope: src\u002Fpackage, docs\u002Fgpu-inference.md, examples\u002Finfer.py\nExcluded scope: training\u002F, benchmarks\u002F\nRequired candidate: examples\u002Finfer.py | cuda | required | CPU substitute: none\nOptional candidate: benchmarks\u002Fthroughput.py | cuda | optional | excluded\nHost verdict: compatible NVIDIA GPU\u002Fdriver\u002Fwheel\nEnvironment set: one CUDA-capable prefix covering CUDA and CPU checks\nInstall: base package + [inference,cuda] + compatible torch CUDA wheel\nSkip: [train], [all], requirements-dev.txt, benchmark dependencies\nPreparation smoke: torch CUDA allocation + package backend import\nFinal native case: examples\u002Finfer.py with tiny fixture (run after integration)\n```\n\nDo not assume the repository name is the distribution or import name. Do not\ninstall all extras or every requirements file as a substitute for analysis.\n\n### 3. Probe Available Managers and Hardware\n\nRun discovery commands separately so their outcomes remain visible:\n\n```bash\ncommand -v conda\ncommand -v micromamba\ncommand -v uv\ncommand -v python3.11\ncommand -v python3\ncommand -v python\n```\n\nUse platform-equivalent commands on Windows. Query versions for tools that\nexist. Do not treat a missing command as a fatal error until all suitable paths\nhave been evaluated.\n\nRead [references\u002Fhardware-and-backends.md](references\u002Fhardware-and-backends.md)\nand probe only the hardware relevant to the requested or repo-required backend.\nA visible GPU does not by itself require a GPU install for package inspection.\n\nApply the backend plan after probing:\n\n- When an included capability has a `required` backend and `CPU substitute:\n  none`, prepare and verify that backend whenever compatible hardware and\n  packages are available. A CPU-only prefix is not ready for that scope.\n- When `CPU substitute` is `full` and the selected CPU candidate verifies the\n  same behavior, use the CPU path and record the evidence.\n- When the backend is `optional`, omit it from the minimum environment when it\n  is unavailable or outside the selected scope, but preserve an explicit\n  unverified-capability note.\n- When required hardware or a compatible dependency variant is unavailable,\n  do not install an arbitrary GPU wheel or call the CPU environment `ok`.\n  Report `failed`, ask to narrow scope, or produce `partial` only after the user\n  accepts drafting with the required-backend limitation.\n- When required backend variants conflict, create separate concrete prefixes\n  and verify each one. Do not mutate one prefix repeatedly between incompatible\n  CPU\u002FCUDA\u002FROCm\u002Fvendor states.\n\nIf no suitable Conda\u002Fmicromamba, uv, or Python runtime exists, do not invoke a\nbundled downloader. Explain the host change required and ask before installing\na runtime or environment manager unless the user already authorized that exact\nkind of host-level installation. Once authorized, use the platform's normal\npackage manager or the selected tool's documented installer with explicit,\nreviewable terminal commands, then resume this workflow. Do not silently curl\nand execute an installer.\n\n### 4. Create the Required Environment Set with Direct Commands\n\nSubstitute concrete absolute paths and the selected Python version in every\ncommand. Do not execute examples with placeholder paths.\n\nCreate the primary prefix for the strongest required selected backend. A\nGPU-capable prefix normally also runs CPU checks. When the backend plan requires\nincompatible variants, derive distinct suffixes such as `-cuda`, `-rocm`, or\n`-cpu` from the caller's private prefix and record every prefix in the report.\n\nFor Conda:\n\n```bash\nconda create --yes --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" \"python=3.11\" pip\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -c \"import sys; print(sys.executable); print(sys.version)\"\n```\n\nFor an already installed micromamba:\n\n```bash\nmicromamba create --yes --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" \"python=3.11\" pip\nmicromamba run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -c \"import sys; print(sys.executable); print(sys.version)\"\n```\n\nFor venv on Unix-like systems:\n\n```bash\npython3.11 -m venv \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\"\n\"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\" -c \"import sys; print(sys.executable); print(sys.version)\"\n```\n\nOn Windows, use the environment's `Scripts\u002Fpython.exe`. If installed `uv` is\nthe best available path, make its runtime acquisition explicit:\n\n```bash\nuv venv --seed --python 3.11 \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\"\n\"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\" -c \"import sys; print(sys.executable); print(sys.version)\"\n```\n\nIf `uv` would download a Python runtime, treat that as runtime installation and\napply the authorization rule from step 3.\n\nDo not rely on shell activation. For subsequent Conda\u002Fmicromamba commands use\n`run --prefix`; for venv\u002Fuv use the environment Python's absolute path. This\nprevents commands from accidentally using DisCo's or the user's current\nPython.\n\n### 5. Install the Minimum Required Package Set\n\nExecute one meaningful install phase at a time, with a finite terminal-tool\ntimeout and observable output. A typical new Conda environment uses commands\nlike:\n\n```bash\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -m pip install -e \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\"\n```\n\nFor venv\u002Fuv:\n\n```bash\n\"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\" -m pip install -e \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\"\n```\n\nAdd only evidence-backed requirements or extras:\n\n```bash\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -m pip install -r \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\u002Frequirements-inference.txt\"\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -m pip install -e \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo[inference]\"\n```\n\nFollow repo documentation when backend foundations such as torch, JAX,\nTensorFlow, or a compiler\u002Ftoolkit must be installed first. Do not reflexively\nupgrade pip\u002Fsetuptools\u002Fwheel or broad dependency sets, especially in an\nexisting environment; do so only when repo\u002Fbuild evidence requires it.\n\nFor every required backend candidate, install its evidence-backed dependency\nvariant into the prefix assigned by the backend plan. Do not install a CPU-only\ntorch\u002FJAX\u002FTensorFlow build into a required CUDA\u002FROCm environment and then count\npackage importability as backend coverage. Conversely, do not install GPU\nstacks for optional or excluded candidates merely because hardware is visible.\n\nDo not combine creation, multiple installs, and verification into one large\nshell command. After each phase, record the exact command, exit status, elapsed\ntime, and relevant output in the private report. On failure, diagnose the\nspecific phase and adjust the next visible command rather than rerunning an\nopaque all-in-one installer.\n\nIf metadata resolution or downloads stall, interrupt the command, distinguish\nnetwork delay from compilation, and use the network guidance in\n[references\u002Finstall-planning.md](references\u002Finstall-planning.md). Proxy or VPN\ndetails supplied by the user are local execution context only; never bake them\ninto the skill or generated package content.\n\n### 6. Verify from the Environment Python\n\nRead [references\u002Fverification-and-failure-report.md](references\u002Fverification-and-failure-report.md).\nAt minimum, run direct commands for:\n\n- Environment Python identity and version.\n- `python -m pip check`.\n- Distribution metadata from the environment.\n- Imports using the environment Python, preferably from outside the checkout\n  and with `-I` when compatible, so the current directory or user site cannot\n  make a broken install look healthy.\n- Relevant CLI\u002FAPI smoke checks.\n- A preparation smoke check for every required backend in the backend plan,\n  using the prefix assigned to that backend. For CUDA\u002FROCm\u002FMPS or a vendor\n  accelerator, prove framework\u002Fbackend availability and a minimal device\n  operation when safe.\n\nExamples for Conda:\n\n```bash\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -m pip check\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -I -c \"from importlib.metadata import version; print(version('distribution-name'))\"\nconda run --prefix \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\" python -I -c \"import package_name; print(package_name.__file__)\"\n```\n\nEquivalent venv commands must use the environment Python's absolute path. A\nsuccessful environment creation or `pip install` is not verification. Fix and\nrerun every applicable failed gate before an `ok` handoff. Do not run the\nrepo-native cases yet; record which final cases this prepared environment must\nsupport after skill integration.\n\n### 7. Write the Report and Handoff\n\nWrite `repo_env_report.json` from the actual command results. Do not invent\nmissing output or mark skipped gates as passed. Include at least:\n\n```json\n{\n  \"schemaVersion\": 2,\n  \"status\": \"ok\",\n  \"readiness\": {\n    \"skillDrafting\": true,\n    \"fullBackendVerification\": true,\n    \"backendGateEligibleForAutoImport\": true\n  },\n  \"repositoryPath\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\",\n  \"environment\": {\n    \"manager\": \"conda\",\n    \"prefix\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\",\n    \"pythonExecutable\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\",\n    \"backends\": [\"cpu\", \"cuda\"]\n  },\n  \"workflowEnvironment\": {\n    \"executable\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Finspection-env\u002Fbin\u002Fpython\",\n    \"cwd\": \"\u002Fabsolute\u002Fpath\u002Fto\u002Frepo\",\n    \"package\": \"distribution-name\",\n    \"version\": \"1.2.3\"\n  },\n  \"additionalEnvironments\": [],\n  \"backendPlan\": {\n    \"required\": [\n      {\n        \"backend\": \"cuda\",\n        \"capability\": \"GPU inference\",\n        \"cpuSubstitute\": \"none\",\n        \"hostVerdict\": \"compatible\",\n        \"dependencyVariant\": \"repo cuda extra plus compatible framework wheel\",\n        \"preparationSmoke\": \"passed\",\n        \"finalNativeCandidates\": [\"examples\u002Finfer.py\"]\n      }\n    ],\n    \"optional\": [],\n    \"alternative\": []\n  },\n  \"installPlan\": {\n    \"python\": \"3.11\",\n    \"includedScope\": [],\n    \"excludedScope\": [],\n    \"installedGroups\": [],\n    \"skippedGroups\": []\n  },\n  \"commands\": [],\n  \"verification\": {},\n  \"warnings\": [],\n  \"failures\": []\n}\n```\n\n`environment.pythonExecutable` remains private report evidence. It is not the\nobject to pass directly to workflow `agent()`. The caller must use the canonical\n`workflowEnvironment` object with `executable`, optional `cwd`, distribution\n`package`, and exact `version`.\n\nUse `status: \"ok\"` only when every required preparation\u002Fbackend gate passes.\nUse `status: \"partial\"` only after explicit user acceptance, with\n`skillDrafting: true`, `fullBackendVerification: false`, and\n`backendGateEligibleForAutoImport: false`. Use `status: \"failed\"` when a\nrequired gate remains unresolved without that acceptance. The `commands`\nentries should identify phase, command, exit status, and concise stdout\u002Fstderr\nevidence; do not put credentials or secret-bearing environment variables in\nthe report.\n\nFor every handoff, provide:\n\n```text\nEnvironment status: \u003Cok|partial|failed>\nReady for skill drafting: \u003Cyes|no>\nReady for full backend verification: \u003Cyes|no>\nBackend gate eligible for auto-import: \u003Cyes|no>\nRepository path: \u003Crepo_path>\nEnvironment manager: \u003Cconda|micromamba|venv|uv>\nEnvironment prefix: \u003Cenvironment_prefix>\nAdditional environment(s): \u003Cprefix\u002Fbackend or none>\nTemporary inspection Python: \u003Cpython_executable>\nWorkflow environment: {\"executable\":\"\u003Cpython_executable>\",\"cwd\":\"\u003Crepo_path>\",\"package\":\"\u003Cdistribution_name>\",\"version\":\"\u003Cverified_version>\"}\nInstalled package name(s): \u003Cdistribution names>\nVerified import(s): \u003Cmodules>\nVerification report: \u003Crepo_env_report.json>\nRequired backend verdict(s): \u003Cbackend: preparation smoke passed\u002Fblocked>\nFinal native backend case(s): \u003Ccandidate ids to run after skill integration>\nAccepted partial limitation: \u003Cexact user-accepted block or none>\nAdditional notes: \u003Cextras installed, skipped unsafe checks, limitations>\n```\n\n`create-repo-skill` may continue normally from `ok`. It may continue drafting\nfrom `partial` only after preserving the accepted backend block and disabling\nauto-import. It must not continue from `failed` without repair or an explicitly\nnarrowed extraction scope.\n\n## Non-Negotiables\n\n- Execute environment creation, package installation, and verification as\n  explicit terminal commands; do not route them through a bundled wrapper.\n- Never install into or mutate Conda `base`.\n- Never use DisCo's current Python as the target environment unless the user\n  explicitly requests and authorizes that mutation.\n- Never delete or recreate an existing prefix without explicit authorization.\n- Never mutate a user-provided environment in a potentially breaking way\n  without authorization; use a new private prefix when permitted.\n- Never install broad extras, dev requirements, or backend packages unrelated\n  to the confirmed extraction scope.\n- Never call a CPU-only environment `ok` for a required CUDA\u002FROCm\u002FMPS\u002Fvendor\n  capability whose CPU substitute is `partial` or `none`.\n- Never silently downgrade a required backend candidate to optional because\n  hardware, wheels, drivers, or toolkits are unavailable.\n- Never mark a partial environment eligible for auto-import. Preserve the\n  required-backend block for final native verification and informed user\n  acceptance.\n- Never rely only on shell activation or an unqualified `python`\u002F`pip` after\n  environment creation.\n- Never treat successful creation or installation as proof of usability.\n- Never hide a failed or skipped verification gate in an `ok` report.\n- Never wait indefinitely on a stalled resolver, download, or build. Use finite\n  timeouts, retain useful failure evidence, and diagnose before retrying.\n- Never copy private environment paths, proxy details, command logs, or setup\n  reports into the generated public repo skill.\n",{"name":5,"metadata":23,"description":7},{"disco-role":24},"meta","main","ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6","cli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fprepare-repo-skill-env",27,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab","2026-09-03 10:25:21",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add VectorSpaceLab\u002FAREX-Skill --skill prepare-repo-skill-env",{"tool":38,"command":39},"bunx","bunx skills add VectorSpaceLab\u002FAREX-Skill --skill prepare-repo-skill-env",{"tool":41,"command":42},"pnpm","pnpm dlx skills add VectorSpaceLab\u002FAREX-Skill --skill prepare-repo-skill-env","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fprepare-repo-skill-env\u002FSKILL.md",true,[46,52,56],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Fhardware-and-backends.md","REFERENCE","参考文档",8166,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fprepare-repo-skill-env\u002Freferences\u002Fhardware-and-backends.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"references\u002Finstall-planning.md",12892,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fprepare-repo-skill-env\u002Freferences\u002Finstall-planning.md",{"path":57,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":58,"sourceHtmlUrl":59},"references\u002Fverification-and-failure-report.md",10803,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fcli\u002Fpackages\u002Fcoding-agent\u002Fsrc\u002Fdisco\u002Fskills\u002Fprepare-repo-skill-env\u002Freferences\u002Fverification-and-failure-report.md",[61,69,76,83,90,97],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":68},"01M799NYSH1R3RVAHSM2D88Y84","passnet-triton-opt","Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block\u002Fgrid\u002Fwarp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-triton-opt\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"publishedAt":68},"01M799NS5APEC0N1FZ7XZVW6BN","passnet-solve","END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-solve\u002FSKILL.md",[],[],{"skillNo":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"publishedAt":68},"01M799NNY9BKSM21V5R7QVTBJX","passnet-skill","PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under .\u002Fpass_dir\u002F, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-skill\u002FSKILL.md",[],[],{"skillNo":84,"name":85,"nameZh":6,"description":86,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"publishedAt":68},"01M799NHH6NP7Z1RK77KGXAN6S","passnet-pattern-fusion","Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()\u002Freplacement_args()\u002Freplacement_func() files, when a pass \"failed to match\", or when deciding how to fuse multiple kernels\u002Fops.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-pattern-fusion\u002FSKILL.md",[],[],{"skillNo":91,"name":92,"nameZh":6,"description":93,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":94,"categories":95,"keywords":96,"publishedAt":68},"01M799ND2J4WNC2PGAZ0A06NDF","passnet-orchestrate","Round\u002Fbatch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep\u002Frevert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-orchestrate\u002FSKILL.md",[],[],{"skillNo":98,"name":99,"nameZh":6,"description":100,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":101,"categories":102,"keywords":103,"publishedAt":68},"01M799N8SBTMY02ZDG2AZF5ZWE","passnet-feedback","Fast, deterministic feedback for PassNet work: pre-flight pattern\u002Fmatch verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-feedback\u002FSKILL.md",[],[]]