[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77RHSS9PH197S2JYCK0YSJC":3,"skill-same:01M77RHSS9PH197S2JYCK0YSJC":56},{"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":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":6,"updatedAt":31,"install":32,"resources":45},"01M77RHSS9PH197S2JYCK0YSJC","anvil-serving-llm-qualification",null,"Qualify a pinned local LLM or VLM on Anvil Serving with reproducible text, image, video, capacity, router, and storage evidence. Use for model bakeoffs, failure diagnosis and researched configuration trials, quant comparisons, context\u002Fconcurrency qualification, multimodal enablement, or production-role recommendations that must remain human-gated.","fakoli\u002Fanvil-serving","fakoli","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F1631326?v=4","01M77RFVS9JSVF3PFEJ4114QDT","Sekou Doumbouya","01M77RFVT2YT7DTK045N6XFE5G",1,"MIT","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Fdocs\u002Ffindings\u002F2026-10-17-glm53-serving-performance-evidence\u002Fplans\u002Fsource\u002Fskills\u002Fanvil-serving-llm-qualification\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: anvil-serving-llm-qualification\ndescription: Qualify a pinned local LLM or VLM on Anvil Serving with reproducible text, image, video, capacity, router, and storage evidence. Use for model bakeoffs, failure diagnosis and researched configuration trials, quant comparisons, context\u002Fconcurrency qualification, multimodal enablement, or production-role recommendations that must remain human-gated.\n---\n\n# Anvil Serving LLM Qualification\n\nUse the product CLI for lifecycle, cache, preflight, benchmark, and router work.\nDo not create skill-local operational scripts or promote a route.\n\nThe maintained layout is `PRODUCT_ROOT\u002Fskills\u002Fanvil-serving-llm-qualification`.\nFor global discovery, symlink that directory rather than copying it. Resolve\nits real path and verify the product root contains `pyproject.toml` and\n`anvil_serving\u002F` before using repository-relative paths. For a copied install,\nuse the available workspace-resolution skill to locate the product checkout\nand read its canonical skill; never infer the checkout from the caller's cwd.\nRead `references\u002Fevidence-contract.md` and `references\u002Fconfiguration-search.md`\nbefore starting. The latter owns failure investigation and candidate stopping\nrules; preserve failed configurations while testing supported successors.\nFor runtime crashes, configuration improvement, or hardware-limit exploration,\nread [runtime-investigation.md](references\u002Fruntime-investigation.md) before live\nrequests. Validate its incident scenario offline first; finish the incident\nbranch separately from a replacement-model comparison.\nApply the shared `session-improvement-loop:research-synthesis` policy through its Anvil\nbindings without requiring the\nuser to ask for fusion: use one initial researcher and shared source registry,\nthen independent answers only for a named unresolved material decision.\nBudget one synthesis and keep local-only questions on retained local evidence.\n\n## Workflow\n\n1. Record the repository revision, dirty state, live serve\u002Frouter state, exact\n   model\u002Fruntime revisions, GPU identity, and cache inventory.\n2. Work in an isolated `codex\u002F` worktree. Preserve unrelated and untracked\n   files.\n   Before downloading or loading a candidate, validate the offline harness:\n\n   ```text\n   python scripts\u002Frun_tests.py tests\u002Ftest_agentic_benchmark.py tests\u002Ftest_benchmark_suite_runner.py -q\n   ```\n\n   Stop on a failed fixture\u002Foracle check. Strict tool protocol, fixture state,\n   final-answer formatting and independent coding quality remain separate\n   evidence. Record the oracle revision; never rewrite historical scores as\n   if they had used a newer scorer.\n3. Gate storage with `models cache inventory`. For any removal, use an exact\n   revision dry-run and explicit confirmation. Never broad-prune Docker or\n   delete volumes.\n4. Record every source with observed date, evidence class, hardware\u002Fruntime\n   relevance, and decision impact. Treat model cards and discussions as priors,\n   not local proof.\n5. Pull exact revisions with `models pull`; a shell\u002Fclient timeout does not\n   prove failure. Before retrying, inspect whether the one downloader container\n   remains active. Never start competing writers against the same cache.\n6. Load one isolated candidate at a time with a pinned recipe and discovered\n   GPU UUID. For a Windows\u002FWSL GPU lane, credential handoff, unhealthy loaded\n   recipe, managed switch, or restore transaction, use\n   `.agents\u002Fskills\u002Fanvil-serving-candidate-operations\u002FSKILL.md`. Do not change\n   a live alias.\n7. Diagnose startup and request failures down-stack: caller, product status\u002Flogs,\n   container exit\u002Fhealth, then engine\u002Fmodel-download error. Retain the earliest\n   actionable error, research it, and test a versioned configuration remedy\n   under `references\u002Fconfiguration-search.md` before rejecting a candidate.\n8. For a hardware-gated upstream kernel patch, pin the upstream commit, verify\n   the exact source and result hashes, and fail startup unless the required\n   symbol imports or the engine logs prove the intended path loaded. Treat the\n   patch as an exact recipe artifact, not generic runtime guidance.\n9. For missing hardware-specific MoE\u002FGEMM config warnings or measured kernel\n   bottlenecks, use `skills\u002Fanvil-serving-kernel-tuning\u002FSKILL.md`. Do not\n   recommend a generated tune without the identical untuned-versus-tuned A\u002FB.\n10. Declare and qualify the model-native reasoning and sampling policy intended\n    for the role. Use thinking-disabled probes only when supported and relevant;\n    they do not qualify a deployment that will use a different reasoning policy.\n11. Before requests for a serving comparison, retain the content-bound frozen\n    plan and matching `metrics_coverage` for its profile × workload × context ×\n    requested-concurrency matrix. Run capacity at the declared points with the campaign’s output\u002Freasoning\n    headroom. Keep performance measurement, quality, and promotion independent:\n    preserve quality misses, stop unsafe infrastructure\u002Fresource failures, and\n    never call the comparison complete with required non-measured metrics.\n12. For images\u002Fvideo, run direct endpoint preflight and\n    `eval benchmark multimodal` before router work. If direct `video_url` fails,\n    retain the failure and pause routed video qualification while investigating\n    the direct configuration. Resume only after its direct gates pass.\n13. If direct video passes, verify same-dialect preservation, fail-closed\n    unsupported translation, admission\u002Fcount\u002Ftoken limits, streaming, tools,\n    malformed media, and an isolated router configuration.\n14. A profile passes only with exact identity, all deterministic assertions,\n    visible answers, allowed finish reasons, matching media hashes, valid\n    tools, and no OOM\u002Fparser corruption.\n15. Explicitly classify the result as either isolated-only or client-facing.\n    If no public alias, router metadata, or selected backing identity changed,\n    prove that fact and do not rewrite client catalogs. If an authorized\n    promotion, rollback, or route switch changes any client-facing alias's\n    backing identity, `context_limit_tokens`, `max_output_tokens`, reasoning,\n    or modality contract, client propagation is a required qualification\n    closure gate even when the request did not mention clients.\n16. Every approved promotion must also pass the all-host provisioning and\n    rebuild-snapshot gate in `skills\u002Fanvil-serving-release-readiness\u002FSKILL.md`,\n    even when advertised limits are unchanged. Record source bindings,\n    inventory coverage and repeat convergence; missing coverage is\n    `harness_convergence_pending`, not completed promotion.\n    For client-facing changes on Companion Node, use\n    `converge-mini-vision-clients` when installed; otherwise enforce the same\n    closure directly. Reconcile Hermes, Pi, and OpenClaw from one authenticated\n    router snapshot. Enumerate\n    every retained Anvil-backed Hermes profile, preserve provider\u002Fauth and\n    compaction settings, validate real client behavior, and require an\n    idempotent final dry-run. If live client mutation\u002Frestart authority is not\n    present, report the qualification as client-convergence-pending rather than\n    complete or deployed.\n17. Restore the exact starting serve\u002Frouter state. Freeze the final candidate\n    source, run focused and full repository gates once, then delegate publication to\n    `skills\u002Fanvil-serving-benchmark-docs\u002FSKILL.md`. Apply its\n    `references\u002Fartifact-set-contract.md` and\n    `anvil-serving.benchmark-artifact-set\u002Fv1` template around the native raw\n    artifacts, dated finding, precision\u002Fmodality decision table, and required\n    publication-ready result card. Add a publication summary only when compact\n    public communication is requested; otherwise mark that role\n    `not-applicable` with a reason. For a format-only\n    refresh backed by complete retained artifacts, reconcile the published\n    values without restarting the serve or rerunning the benchmark.\n\n## Improve the workflow from evidence\n\nAt campaign closure or after a material process failure, read\n`references\u002Fimprovement-loop.md`. Capture scoped lessons in existing evidence,\nthen evaluate authorized skill changes independently. Preserve the running\nqualification baseline and the budget for the requested model outcome.\n\n## Decisions\n\n- Separate official\u002Fcommunity claims from locally measured results.\n- Compare quantized profiles only after every hard gate passes for the exact\n  frozen profile. A failed profile returns to configuration search; it does\n  not automatically reject the checkpoint or engine.\n- Compare speculation to an otherwise identical no-speculation control. Keep\n  model, revision, image, patch, TP, context, concurrency, KV dtype, memory\n  fraction, backends, batching, graph capture, transport, parsers, and offload\n  policy fixed; do not inherit a KV dtype already known to fail that hardware\n  path.\n- Prefer a quant only when it meets the campaign’s declared memory or\n  throughput improvement threshold relative to the next higher precision.\n- Limit a known-crashing vision quant to a text-only role; do not infer future\n  support from shipped vision weights.\n- Keep promotion separately human-gated.\n","# Anvil Serving LLM Qualification\n\nUse the product CLI for lifecycle, cache, preflight, benchmark, and router work.\nDo not create skill-local operational scripts or promote a route.\n\nThe maintained layout is `PRODUCT_ROOT\u002Fskills\u002Fanvil-serving-llm-qualification`.\nFor global discovery, symlink that directory rather than copying it. Resolve\nits real path and verify the product root contains `pyproject.toml` and\n`anvil_serving\u002F` before using repository-relative paths. For a copied install,\nuse the available workspace-resolution skill to locate the product checkout\nand read its canonical skill; never infer the checkout from the caller's cwd.\nRead `references\u002Fevidence-contract.md` and `references\u002Fconfiguration-search.md`\nbefore starting. The latter owns failure investigation and candidate stopping\nrules; preserve failed configurations while testing supported successors.\nFor runtime crashes, configuration improvement, or hardware-limit exploration,\nread [runtime-investigation.md](references\u002Fruntime-investigation.md) before live\nrequests. Validate its incident scenario offline first; finish the incident\nbranch separately from a replacement-model comparison.\nApply the shared `session-improvement-loop:research-synthesis` policy through its Anvil\nbindings without requiring the\nuser to ask for fusion: use one initial researcher and shared source registry,\nthen independent answers only for a named unresolved material decision.\nBudget one synthesis and keep local-only questions on retained local evidence.\n\n## Workflow\n\n1. Record the repository revision, dirty state, live serve\u002Frouter state, exact\n   model\u002Fruntime revisions, GPU identity, and cache inventory.\n2. Work in an isolated `codex\u002F` worktree. Preserve unrelated and untracked\n   files.\n   Before downloading or loading a candidate, validate the offline harness:\n\n   ```text\n   python scripts\u002Frun_tests.py tests\u002Ftest_agentic_benchmark.py tests\u002Ftest_benchmark_suite_runner.py -q\n   ```\n\n   Stop on a failed fixture\u002Foracle check. Strict tool protocol, fixture state,\n   final-answer formatting and independent coding quality remain separate\n   evidence. Record the oracle revision; never rewrite historical scores as\n   if they had used a newer scorer.\n3. Gate storage with `models cache inventory`. For any removal, use an exact\n   revision dry-run and explicit confirmation. Never broad-prune Docker or\n   delete volumes.\n4. Record every source with observed date, evidence class, hardware\u002Fruntime\n   relevance, and decision impact. Treat model cards and discussions as priors,\n   not local proof.\n5. Pull exact revisions with `models pull`; a shell\u002Fclient timeout does not\n   prove failure. Before retrying, inspect whether the one downloader container\n   remains active. Never start competing writers against the same cache.\n6. Load one isolated candidate at a time with a pinned recipe and discovered\n   GPU UUID. For a Windows\u002FWSL GPU lane, credential handoff, unhealthy loaded\n   recipe, managed switch, or restore transaction, use\n   `.agents\u002Fskills\u002Fanvil-serving-candidate-operations\u002FSKILL.md`. Do not change\n   a live alias.\n7. Diagnose startup and request failures down-stack: caller, product status\u002Flogs,\n   container exit\u002Fhealth, then engine\u002Fmodel-download error. Retain the earliest\n   actionable error, research it, and test a versioned configuration remedy\n   under `references\u002Fconfiguration-search.md` before rejecting a candidate.\n8. For a hardware-gated upstream kernel patch, pin the upstream commit, verify\n   the exact source and result hashes, and fail startup unless the required\n   symbol imports or the engine logs prove the intended path loaded. Treat the\n   patch as an exact recipe artifact, not generic runtime guidance.\n9. For missing hardware-specific MoE\u002FGEMM config warnings or measured kernel\n   bottlenecks, use `skills\u002Fanvil-serving-kernel-tuning\u002FSKILL.md`. Do not\n   recommend a generated tune without the identical untuned-versus-tuned A\u002FB.\n10. Declare and qualify the model-native reasoning and sampling policy intended\n    for the role. Use thinking-disabled probes only when supported and relevant;\n    they do not qualify a deployment that will use a different reasoning policy.\n11. Before requests for a serving comparison, retain the content-bound frozen\n    plan and matching `metrics_coverage` for its profile × workload × context ×\n    requested-concurrency matrix. Run capacity at the declared points with the campaign’s output\u002Freasoning\n    headroom. Keep performance measurement, quality, and promotion independent:\n    preserve quality misses, stop unsafe infrastructure\u002Fresource failures, and\n    never call the comparison complete with required non-measured metrics.\n12. For images\u002Fvideo, run direct endpoint preflight and\n    `eval benchmark multimodal` before router work. If direct `video_url` fails,\n    retain the failure and pause routed video qualification while investigating\n    the direct configuration. Resume only after its direct gates pass.\n13. If direct video passes, verify same-dialect preservation, fail-closed\n    unsupported translation, admission\u002Fcount\u002Ftoken limits, streaming, tools,\n    malformed media, and an isolated router configuration.\n14. A profile passes only with exact identity, all deterministic assertions,\n    visible answers, allowed finish reasons, matching media hashes, valid\n    tools, and no OOM\u002Fparser corruption.\n15. Explicitly classify the result as either isolated-only or client-facing.\n    If no public alias, router metadata, or selected backing identity changed,\n    prove that fact and do not rewrite client catalogs. If an authorized\n    promotion, rollback, or route switch changes any client-facing alias's\n    backing identity, `context_limit_tokens`, `max_output_tokens`, reasoning,\n    or modality contract, client propagation is a required qualification\n    closure gate even when the request did not mention clients.\n16. Every approved promotion must also pass the all-host provisioning and\n    rebuild-snapshot gate in `skills\u002Fanvil-serving-release-readiness\u002FSKILL.md`,\n    even when advertised limits are unchanged. Record source bindings,\n    inventory coverage and repeat convergence; missing coverage is\n    `harness_convergence_pending`, not completed promotion.\n    For client-facing changes on Companion Node, use\n    `converge-mini-vision-clients` when installed; otherwise enforce the same\n    closure directly. Reconcile Hermes, Pi, and OpenClaw from one authenticated\n    router snapshot. Enumerate\n    every retained Anvil-backed Hermes profile, preserve provider\u002Fauth and\n    compaction settings, validate real client behavior, and require an\n    idempotent final dry-run. If live client mutation\u002Frestart authority is not\n    present, report the qualification as client-convergence-pending rather than\n    complete or deployed.\n17. Restore the exact starting serve\u002Frouter state. Freeze the final candidate\n    source, run focused and full repository gates once, then delegate publication to\n    `skills\u002Fanvil-serving-benchmark-docs\u002FSKILL.md`. Apply its\n    `references\u002Fartifact-set-contract.md` and\n    `anvil-serving.benchmark-artifact-set\u002Fv1` template around the native raw\n    artifacts, dated finding, precision\u002Fmodality decision table, and required\n    publication-ready result card. Add a publication summary only when compact\n    public communication is requested; otherwise mark that role\n    `not-applicable` with a reason. For a format-only\n    refresh backed by complete retained artifacts, reconcile the published\n    values without restarting the serve or rerunning the benchmark.\n\n## Improve the workflow from evidence\n\nAt campaign closure or after a material process failure, read\n`references\u002Fimprovement-loop.md`. Capture scoped lessons in existing evidence,\nthen evaluate authorized skill changes independently. Preserve the running\nqualification baseline and the budget for the requested model outcome.\n\n## Decisions\n\n- Separate official\u002Fcommunity claims from locally measured results.\n- Compare quantized profiles only after every hard gate passes for the exact\n  frozen profile. A failed profile returns to configuration search; it does\n  not automatically reject the checkpoint or engine.\n- Compare speculation to an otherwise identical no-speculation control. Keep\n  model, revision, image, patch, TP, context, concurrency, KV dtype, memory\n  fraction, backends, batching, graph capture, transport, parsers, and offload\n  policy fixed; do not inherit a KV dtype already known to fail that hardware\n  path.\n- Prefer a quant only when it meets the campaign’s declared memory or\n  throughput improvement threshold relative to the next higher precision.\n- Limit a known-crashing vision quant to a text-only role; do not infer future\n  support from shipped vision weights.\n- Keep promotion separately human-gated.\n",{"name":5,"description":7},"main","c02ba7c047f81b9e5fcc2a26434f6b5bfb82f5b0","docs\u002Ffindings\u002F2026-10-17-glm53-serving-performance-evidence\u002Fplans\u002Fsource\u002Fskills\u002Fanvil-serving-llm-qualification",0,"https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving","https:\u002F\u002Fgithub.com\u002Ffakoli","https:\u002F\u002Fsekoudoumbouya.com","2026-11-14 02:44:30",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add fakoli\u002Fanvil-serving --skill anvil-serving-llm-qualification",{"tool":38,"command":39},"bunx","bunx skills add fakoli\u002Fanvil-serving --skill anvil-serving-llm-qualification",{"tool":41,"command":42},"pnpm","pnpm dlx skills add fakoli\u002Fanvil-serving --skill anvil-serving-llm-qualification","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Fdocs\u002Ffindings\u002F2026-10-17-glm53-serving-performance-evidence\u002Fplans\u002Fsource\u002Fskills\u002Fanvil-serving-llm-qualification\u002FSKILL.md",true,[46,52],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Fevidence-contract.md","REFERENCE","参考文档",5435,"https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Fdocs\u002Ffindings\u002F2026-10-17-glm53-serving-performance-evidence\u002Fplans\u002Fsource\u002Fskills\u002Fanvil-serving-llm-qualification\u002Freferences\u002Fevidence-contract.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"references\u002Fruntime-investigation.md",10489,"https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Fdocs\u002Ffindings\u002F2026-10-17-glm53-serving-performance-evidence\u002Fplans\u002Fsource\u002Fskills\u002Fanvil-serving-llm-qualification\u002Freferences\u002Fruntime-investigation.md",[57,63,69,74,79,86],{"skillNo":58,"name":59,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M77RK15F1RY23K6YY54W7YQC","voice","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcli_reference_audit\u002Finput\u002Fskills\u002Fvoice\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":66,"categories":67,"keywords":68,"publishedAt":19},"01M77RK06RC7DDZQE1MEZHA9NQ","workbench","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcli_reference_audit\u002Finput\u002Fexamples\u002Fopenclaw\u002Fskills\u002Fworkbench\u002FSKILL.md",[],[],{"skillNo":70,"name":65,"nameZh":6,"description":6,"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":19},"01M77RJY4CSSN04B1MRGPMSFH8","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcli_reference_audit\u002Finput\u002F.claude\u002Fskills\u002Fworkbench\u002FSKILL.md",[],[],{"skillNo":75,"name":65,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":78,"publishedAt":19},"01M77RJX2C6VRHYS7F452WTFKX","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcli_reference_audit\u002Finput\u002F.agents\u002Fskills\u002Fworkbench\u002FSKILL.md",[],[],{"skillNo":80,"name":81,"nameZh":6,"description":82,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M77RJSYENH9ERF4RB3P8Y0DY","anvil-serving-voice-ops","Operate topology-owned anvil-serving voice audio and Realtime proxy lifecycle, Mini-to-Dark loopback forwarding, profiles, sidecars, and bounded voice benchmarks. Use for voice status, logs, start\u002Fstop\u002Frestart, OpenClaw Talk topology checks, and voice evidence that must remain separate from LLM serve qualification evidence.","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Fskills\u002Fanvil-serving-voice-ops\u002FSKILL.md",[],[],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":19},"01M77RJP56J64R6Y73WZCDJ03D","anvil-serving-stt-benchmark","Prepare, validate, execute, compare, and publish reproducible multi-sample speech-to-text qualifications in anvil-serving. Use for STT model bakeoffs, corpus-based WER\u002FCER and latency testing, candidate overlays, concurrency tests, restoration checks, and fail-closed stt-benchmark-evidence\u002Fv1 artifacts. Do not use for a single voice-pipeline turn or LLM qualification.","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Fskills\u002Fanvil-serving-stt-benchmark\u002FSKILL.md",[],[]]