[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77RHX2WMNEJWWBXTQ8GYV16":3,"skill-same:01M77RHX2WMNEJWWBXTQ8GYV16":54},{"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":32,"commitSha":33,"skillDirectory":34,"forksCount":35,"repositoryHtmlUrl":36,"repositoryHomepageUrl":6,"ownerHtmlUrl":37,"ownerBlogUrl":38,"ownerDescription":6,"updatedAt":39,"install":40,"resources":53},"01M77RHX2WMNEJWWBXTQ8GYV16","anvil-media",null,"Generate, inspect, or cancel images and videos through Anvil's bounded named-workflow MCP tools.","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\u002Fexamples\u002Fhermes\u002Fskills\u002Fanvil-media\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: anvil-media\ndescription: Generate, inspect, or cancel images and videos through Anvil's bounded named-workflow MCP tools.\nmetadata:\n  version: \"1.0.8\"\n  hermes:\n    tags: [media, image-generation, video-generation, mcp]\n    category: media\n---\n\n# Anvil media\n\nUse this skill automatically when the user asks Hermes to create, generate,\nrender, or revise an image or video through Anvil, or asks to inspect or cancel\nan earlier Anvil media job. A slash-command invocation is optional.\n\n## Procedure\n\n1. Call `mcp__anvil_media__media_capabilities` and\n   `mcp__anvil_media__media_workflow_list`. Select one exact returned workflow\n   whose kind matches the request. If it is unavailable, report its stated\n   reasons and stop.\n   Candidate recipe categories marked `static-reviewed` or `needs-qualification`\n   are discovery only. Execute only the exact available registered workflow.\n   If references are requested, inspect `assetUpload` and the exact artifact\n   roles\u002Fmedia limits. When intake is available, use the supported\n   `anvil-serving media asset upload --file \u003Clocal-file> --client-config \u003Cprotected-definition>`\n   preview\u002Fconfirmation path and bind the returned owned artifact ID. Never\n   send local paths, arbitrary URLs or base64 blobs as workflow parameters.\n   Missing reference\u002Faudio roles or intake remain explicit unsupported inputs.\n2. Call `mcp__anvil_media__media_workflow_show` for that exact workflow and\n   version. Treat its parameter schema, quality profiles, availability, and\n   limits as authoritative.\n3. Choose only a quality profile declared by the workflow. Use `draft` for an\n   explicitly quick or rough request, `high` for an explicitly high-detail or\n   largest-output request, and the returned `defaultQualityProfile` otherwise.\n   If the workflow declares no quality profiles, use the empty string as its\n   exact `quality_profile` value and preserve it through any resume.\n   These names select exact server-owned settings; they are not a promise of\n   subjective visual quality. Never include fields named by\n   `qualityProfileParameters` in `parameters`.\n4. Build the remaining parameters from `profiledParameterSchema` when it is\n   present, and otherwise from `schema`. Preserve the user's prompt faithfully.\n   If a seed is required and the user did not provide one, choose a fresh valid\n   integer without asking. Ask only when a genuinely required creative choice\n   cannot be inferred.\n5. Call `mcp__anvil_media__media_workflow_validate` for the selected workflow.\n   Pass only `workflow_id` and `version`; this tool does not accept parameters.\n   If an otherwise available workflow returns `backend_unavailable`, treat it\n   only as a cold-worker signal and continue to submission so the gateway can\n   create the bounded lifecycle approval. Do not diagnose, name, or offer to\n   inspect infrastructure. For any other validation failure, report the result\n   and stop.\n6. Create one opaque idempotency key for the user's intent using\n   `hermes-media-` followed by exactly 32 lowercase hexadecimal characters.\n   Never put an ellipsis in this key. Reuse it only when\n   retrying the identical workflow, version, quality profile, and parameters. A\n   changed request gets a new key. Preserve a resume bundle containing the exact\n   workflow ID and version, quality profile, full parameters, and idempotency\n   key; never reconstruct those fields from conversation or session history.\n   When advertised, call `mcp__anvil_media__media_workflow_prepare` with this\n   workflow\u002Fversion\u002Fkey\u002Fprofile and either literal parameters or the recipe's\n   supported structured brief. Preserve its exact `runArguments`, including\n   `plan_digest`. Unsupported reference roles and negative controls must fail\n   explicitly; never turn an identity image or source video into prompt prose.\n7. Call `mcp__anvil_media__media_workflow_run` with the exact\n   `quality_profile`. Preserve the returned job ID and state. Do not resubmit\n   merely because generation is long-running. If it returns\n   `awaiting_approval`, report the exact bounded operator action and transaction,\n   plus the complete resume bundle and returned job ID.\n   Copy the server-returned `resumeBundle` object exactly into a literal,\n   copyable `Resume bundle` mapping. It must contain `workflow_id`, `version`,\n   `quality_profile`, complete `parameters`, the full `idempotency_key`,\n   `job_id`, and `approval_transaction_id`. Do not hand-build, abbreviate,\n   omit, reorder, or replace any value with prose or an ellipsis. The job ID\n   must appear inside the mapping even when it also appears elsewhere. Before\n   sending, compare all seven fields to the returned object and rewrite the\n   reply if any field differs. Never say that the bundle is stored, remembered,\n   or preserved unless every field is also present in the current reply. If the\n   server omits `resumeBundle` or you cannot reproduce it exactly, cancel that\n   job with `mcp__anvil_media__media_job_cancel`, report\n   `resume_bundle_incomplete`, and do not request approval. Treat the emitted\n   bundle as a user-facing result, not internal bookkeeping.\n   Hermes has no worker-lifecycle authority; stop without changing the worker.\n   Do not omit the caller-generated idempotency key.\n   If preparation returned `runArguments`, submit that object unchanged.\n   Preserve any returned `plan_digest` in the exact resume mapping and compare\n   it too; include it when replaying the five submission fields below.\n8. Poll `mcp__anvil_media__media_job_status` at a bounded cadence until the job\n   is terminal or the interaction's wait budget ends. Report\n   `awaiting_approval`, `preparing`, `queued`, and `running` as real\n   states; never invent completion. After an operator applies a reported cold\n   lifecycle approval, retry from the complete resume bundle so the reserved job\n   resumes exactly once. Pass its five submission fields to\n   `media_workflow_run`: `workflow_id`, `version`, `parameters`,\n   `quality_profile`, and `idempotency_key`, plus `plan_digest` when present.\n   Do not pass `job_id` or\n   `approval_transaction_id` to that tool: use `job_id` only for the same-job\n   equality check, and use `approval_transaction_id` only to correlate the\n   reported operator approval. Require `created: false` and the same job ID\n   returned by the original submission; any mismatch is a hard stop. Never use\n   `session_search`, session history, or another job to reconstruct or resume a\n   request. After a successful reattachment, `preparing`, `queued`, and `running`\n   are tool-loop states: call job status again instead of ending the response\n   with narration such as \"now polling.\"\n9. For a completed job, call\n   `mcp__anvil_media__media_artifact_inspect` for every returned artifact ID.\n   For an image, present the native image content returned by that tool to the\n   user; do not echo its encoded bytes as text. Also report the authenticated\n   `resource`, media type, byte length, digest, expiry, selected quality\n   profile, and the job's `latency` fields. Video remains resource-only.\n10. Except for the mandatory fail-closed cancellation in step 7 when the\n    server-issued resume bundle is missing or cannot be copied exactly, call\n    `mcp__anvil_media__media_job_cancel` only when the user asks to cancel that\n    job. Report the returned cancellation state exactly.\n\n## Envelope and content parsing (observed live 2026-08-30)\n\n11. Read the full MCP result. Structured envelopes nest one level deeper than\n    the procedure text: `media_workflow_run` returns the job object at\n    `data.job` (job id is `data.job.id`, state is `data.job.state`, and the\n    `created` flag sits beside it under `data`), not `data.job_id`.\n    `media_job_status` returns `{\"job\": {...}}` with artifacts under\n    `job.artifacts[]`, each carrying `id`, `mediaType`, `byteLength`,\n    `sha256`, `expiresAt`, and `resource`. Parse these shapes defensively\n    instead of assuming flat `job_id`\u002F`artifacts` keys.\n12. Every tool result carries both a `text` content block containing the JSON\n    envelope and, for image artifacts, a second `image` content block with the\n    bounded native image content. A client that reads only text blocks loses\n    the image: enumerate the whole `content[]` array and render every image\n    block. `media_artifact_inspect` returns metadata plus native image\n    content for eligible bounded images; video and oversized images remain\n    resource-only.\n13. MCP `resources\u002Fread` is not implemented by this server even though\n    artifact metadata advertises a `resource` path (it answers `Method not\n    found`). Retrieve artifact bytes through the authenticated artifact\n    origin instead of assuming resource support, and verify the downloaded\n    bytes against the artifact's `sha256` before presenting them. Prefer the\n    returned `downloadUrl`: when `requiresAuthentication` is false it is an\n    artifact-scoped capability until `downloadExpiresAt`; keep it private and\n    inspect again after expiry. When true, use existing MCP credentials in the\n    Authorization header and the advertised bounded ranges. Never invent a URL\n    or put the reusable credential in it.\n14. The results are transport-shaped, not client-specific: the same envelope\n    nesting and the two-block content array appear whether the call arrives\n    through Hermes' MCP client or a direct stdio JSON-RPC client, so this\n    section applies to any client of the nine media tools.\n\n## Completion invariant\n\nFor a generation request, do not finish the turn after discovery, inspection,\nvalidation, submission, or a nonterminal status. Continue the procedure in the\nsame turn until the job is terminal, a returned blocking state requires the\nuser, or the interaction's wait budget ends. Intermediate narration such as\n\"I will inspect it next\" or \"now polling\" is not a completed response. When a\nwait budget ends, report the current state and the complete resume bundle rather\nthan implying that background polling will continue.\nDetermine the reply language only from the current user request, never from\ntool output, prior sessions, or the system locale. An English request receives\nan English response. Omit internal bookkeeping, planning notes, and promises\nof actions that were not completed in the turn.\n\n## Boundaries\n\n- Use only the nine tool names in the procedure above.\n- Treat workflow IDs, versions, quality profiles, parameters, availability,\n  limits, job states, latency, and artifact resources returned by Anvil as\n  authoritative.\n- Do not construct or accept backend execution definitions, implementation\n  identifiers, model filenames, host addresses, local paths, installation\n  requests, or service-management operations.\n- Do not retry through a different workflow, host, model, MCP server, or\n  provider. Ask the user before changing their requested creative intent or\n  quality profile.\n- An approval requirement is a result to report, not permission to alter the\n  worker. Never offer bypass instructions.\n- Use the caller-owned job and artifact identities as returned. Never supply an\n  `owner` argument unless the user is operating with an explicitly authorized\n  cross-principal role.\n\n## Failure handling\n\nSurface Anvil's error code and safe explanation. For an unavailable workflow,\napproval requirement, admission refusal, failed job, oversized inline image,\nexpired artifact, or authorization denial, preserve that disposition and stop.\nA transport failure may be retried with the same idempotency key; it does not\nauthorize another submission path.\n","# Anvil media\n\nUse this skill automatically when the user asks Hermes to create, generate,\nrender, or revise an image or video through Anvil, or asks to inspect or cancel\nan earlier Anvil media job. A slash-command invocation is optional.\n\n## Procedure\n\n1. Call `mcp__anvil_media__media_capabilities` and\n   `mcp__anvil_media__media_workflow_list`. Select one exact returned workflow\n   whose kind matches the request. If it is unavailable, report its stated\n   reasons and stop.\n   Candidate recipe categories marked `static-reviewed` or `needs-qualification`\n   are discovery only. Execute only the exact available registered workflow.\n   If references are requested, inspect `assetUpload` and the exact artifact\n   roles\u002Fmedia limits. When intake is available, use the supported\n   `anvil-serving media asset upload --file \u003Clocal-file> --client-config \u003Cprotected-definition>`\n   preview\u002Fconfirmation path and bind the returned owned artifact ID. Never\n   send local paths, arbitrary URLs or base64 blobs as workflow parameters.\n   Missing reference\u002Faudio roles or intake remain explicit unsupported inputs.\n2. Call `mcp__anvil_media__media_workflow_show` for that exact workflow and\n   version. Treat its parameter schema, quality profiles, availability, and\n   limits as authoritative.\n3. Choose only a quality profile declared by the workflow. Use `draft` for an\n   explicitly quick or rough request, `high` for an explicitly high-detail or\n   largest-output request, and the returned `defaultQualityProfile` otherwise.\n   If the workflow declares no quality profiles, use the empty string as its\n   exact `quality_profile` value and preserve it through any resume.\n   These names select exact server-owned settings; they are not a promise of\n   subjective visual quality. Never include fields named by\n   `qualityProfileParameters` in `parameters`.\n4. Build the remaining parameters from `profiledParameterSchema` when it is\n   present, and otherwise from `schema`. Preserve the user's prompt faithfully.\n   If a seed is required and the user did not provide one, choose a fresh valid\n   integer without asking. Ask only when a genuinely required creative choice\n   cannot be inferred.\n5. Call `mcp__anvil_media__media_workflow_validate` for the selected workflow.\n   Pass only `workflow_id` and `version`; this tool does not accept parameters.\n   If an otherwise available workflow returns `backend_unavailable`, treat it\n   only as a cold-worker signal and continue to submission so the gateway can\n   create the bounded lifecycle approval. Do not diagnose, name, or offer to\n   inspect infrastructure. For any other validation failure, report the result\n   and stop.\n6. Create one opaque idempotency key for the user's intent using\n   `hermes-media-` followed by exactly 32 lowercase hexadecimal characters.\n   Never put an ellipsis in this key. Reuse it only when\n   retrying the identical workflow, version, quality profile, and parameters. A\n   changed request gets a new key. Preserve a resume bundle containing the exact\n   workflow ID and version, quality profile, full parameters, and idempotency\n   key; never reconstruct those fields from conversation or session history.\n   When advertised, call `mcp__anvil_media__media_workflow_prepare` with this\n   workflow\u002Fversion\u002Fkey\u002Fprofile and either literal parameters or the recipe's\n   supported structured brief. Preserve its exact `runArguments`, including\n   `plan_digest`. Unsupported reference roles and negative controls must fail\n   explicitly; never turn an identity image or source video into prompt prose.\n7. Call `mcp__anvil_media__media_workflow_run` with the exact\n   `quality_profile`. Preserve the returned job ID and state. Do not resubmit\n   merely because generation is long-running. If it returns\n   `awaiting_approval`, report the exact bounded operator action and transaction,\n   plus the complete resume bundle and returned job ID.\n   Copy the server-returned `resumeBundle` object exactly into a literal,\n   copyable `Resume bundle` mapping. It must contain `workflow_id`, `version`,\n   `quality_profile`, complete `parameters`, the full `idempotency_key`,\n   `job_id`, and `approval_transaction_id`. Do not hand-build, abbreviate,\n   omit, reorder, or replace any value with prose or an ellipsis. The job ID\n   must appear inside the mapping even when it also appears elsewhere. Before\n   sending, compare all seven fields to the returned object and rewrite the\n   reply if any field differs. Never say that the bundle is stored, remembered,\n   or preserved unless every field is also present in the current reply. If the\n   server omits `resumeBundle` or you cannot reproduce it exactly, cancel that\n   job with `mcp__anvil_media__media_job_cancel`, report\n   `resume_bundle_incomplete`, and do not request approval. Treat the emitted\n   bundle as a user-facing result, not internal bookkeeping.\n   Hermes has no worker-lifecycle authority; stop without changing the worker.\n   Do not omit the caller-generated idempotency key.\n   If preparation returned `runArguments`, submit that object unchanged.\n   Preserve any returned `plan_digest` in the exact resume mapping and compare\n   it too; include it when replaying the five submission fields below.\n8. Poll `mcp__anvil_media__media_job_status` at a bounded cadence until the job\n   is terminal or the interaction's wait budget ends. Report\n   `awaiting_approval`, `preparing`, `queued`, and `running` as real\n   states; never invent completion. After an operator applies a reported cold\n   lifecycle approval, retry from the complete resume bundle so the reserved job\n   resumes exactly once. Pass its five submission fields to\n   `media_workflow_run`: `workflow_id`, `version`, `parameters`,\n   `quality_profile`, and `idempotency_key`, plus `plan_digest` when present.\n   Do not pass `job_id` or\n   `approval_transaction_id` to that tool: use `job_id` only for the same-job\n   equality check, and use `approval_transaction_id` only to correlate the\n   reported operator approval. Require `created: false` and the same job ID\n   returned by the original submission; any mismatch is a hard stop. Never use\n   `session_search`, session history, or another job to reconstruct or resume a\n   request. After a successful reattachment, `preparing`, `queued`, and `running`\n   are tool-loop states: call job status again instead of ending the response\n   with narration such as \"now polling.\"\n9. For a completed job, call\n   `mcp__anvil_media__media_artifact_inspect` for every returned artifact ID.\n   For an image, present the native image content returned by that tool to the\n   user; do not echo its encoded bytes as text. Also report the authenticated\n   `resource`, media type, byte length, digest, expiry, selected quality\n   profile, and the job's `latency` fields. Video remains resource-only.\n10. Except for the mandatory fail-closed cancellation in step 7 when the\n    server-issued resume bundle is missing or cannot be copied exactly, call\n    `mcp__anvil_media__media_job_cancel` only when the user asks to cancel that\n    job. Report the returned cancellation state exactly.\n\n## Envelope and content parsing (observed live 2026-08-30)\n\n11. Read the full MCP result. Structured envelopes nest one level deeper than\n    the procedure text: `media_workflow_run` returns the job object at\n    `data.job` (job id is `data.job.id`, state is `data.job.state`, and the\n    `created` flag sits beside it under `data`), not `data.job_id`.\n    `media_job_status` returns `{\"job\": {...}}` with artifacts under\n    `job.artifacts[]`, each carrying `id`, `mediaType`, `byteLength`,\n    `sha256`, `expiresAt`, and `resource`. Parse these shapes defensively\n    instead of assuming flat `job_id`\u002F`artifacts` keys.\n12. Every tool result carries both a `text` content block containing the JSON\n    envelope and, for image artifacts, a second `image` content block with the\n    bounded native image content. A client that reads only text blocks loses\n    the image: enumerate the whole `content[]` array and render every image\n    block. `media_artifact_inspect` returns metadata plus native image\n    content for eligible bounded images; video and oversized images remain\n    resource-only.\n13. MCP `resources\u002Fread` is not implemented by this server even though\n    artifact metadata advertises a `resource` path (it answers `Method not\n    found`). Retrieve artifact bytes through the authenticated artifact\n    origin instead of assuming resource support, and verify the downloaded\n    bytes against the artifact's `sha256` before presenting them. Prefer the\n    returned `downloadUrl`: when `requiresAuthentication` is false it is an\n    artifact-scoped capability until `downloadExpiresAt`; keep it private and\n    inspect again after expiry. When true, use existing MCP credentials in the\n    Authorization header and the advertised bounded ranges. Never invent a URL\n    or put the reusable credential in it.\n14. The results are transport-shaped, not client-specific: the same envelope\n    nesting and the two-block content array appear whether the call arrives\n    through Hermes' MCP client or a direct stdio JSON-RPC client, so this\n    section applies to any client of the nine media tools.\n\n## Completion invariant\n\nFor a generation request, do not finish the turn after discovery, inspection,\nvalidation, submission, or a nonterminal status. Continue the procedure in the\nsame turn until the job is terminal, a returned blocking state requires the\nuser, or the interaction's wait budget ends. Intermediate narration such as\n\"I will inspect it next\" or \"now polling\" is not a completed response. When a\nwait budget ends, report the current state and the complete resume bundle rather\nthan implying that background polling will continue.\nDetermine the reply language only from the current user request, never from\ntool output, prior sessions, or the system locale. An English request receives\nan English response. Omit internal bookkeeping, planning notes, and promises\nof actions that were not completed in the turn.\n\n## Boundaries\n\n- Use only the nine tool names in the procedure above.\n- Treat workflow IDs, versions, quality profiles, parameters, availability,\n  limits, job states, latency, and artifact resources returned by Anvil as\n  authoritative.\n- Do not construct or accept backend execution definitions, implementation\n  identifiers, model filenames, host addresses, local paths, installation\n  requests, or service-management operations.\n- Do not retry through a different workflow, host, model, MCP server, or\n  provider. Ask the user before changing their requested creative intent or\n  quality profile.\n- An approval requirement is a result to report, not permission to alter the\n  worker. Never offer bypass instructions.\n- Use the caller-owned job and artifact identities as returned. Never supply an\n  `owner` argument unless the user is operating with an explicitly authorized\n  cross-principal role.\n\n## Failure handling\n\nSurface Anvil's error code and safe explanation. For an unavailable workflow,\napproval requirement, admission refusal, failed job, oversized inline image,\nexpired artifact, or authorization denial, preserve that disposition and stop.\nA transport failure may be retried with the same idempotency key; it does not\nauthorize another submission path.\n",{"name":5,"metadata":24,"description":7},{"hermes":25,"version":31},{"tags":26,"category":27},[27,28,29,30],"media","image-generation","video-generation","mcp","1.0.8","main","c02ba7c047f81b9e5fcc2a26434f6b5bfb82f5b0","examples\u002Fhermes\u002Fskills\u002Fanvil-media",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":41,"prompt":51,"zipAvailable":52},[42,45,48],{"tool":43,"command":44},"npx","npx skills add fakoli\u002Fanvil-serving --skill anvil-media",{"tool":46,"command":47},"bunx","bunx skills add fakoli\u002Fanvil-serving --skill anvil-media",{"tool":49,"command":50},"pnpm","pnpm dlx skills add fakoli\u002Fanvil-serving --skill anvil-media","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Fexamples\u002Fhermes\u002Fskills\u002Fanvil-media\u002FSKILL.md",true,[],[55,61,67,72,77,84],{"skillNo":56,"name":57,"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":58,"categories":59,"keywords":60,"publishedAt":19},"01M77RK15F1RY23K6YY54W7YQC","voice","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcli_reference_audit\u002Finput\u002Fskills\u002Fvoice\u002FSKILL.md",[],[],{"skillNo":62,"name":63,"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":64,"categories":65,"keywords":66,"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":68,"name":63,"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":69,"categories":70,"keywords":71,"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":73,"name":63,"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":74,"categories":75,"keywords":76,"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":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"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":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"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",[],[]]