[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77RG5S7TKZ8RQP3GWSWKVJP":3,"skill-same:01M77RG5S7TKZ8RQP3GWSWKVJP":52},{"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},"01M77RG5S7TKZ8RQP3GWSWKVJP","anvil-serving-candidate-operations",null,"Prepare, recover, switch, test, and restore an Anvil Serving model candidate through managed recipes and serves. Use for RTX 5090 lane preparation, Windows\u002FWSL GPU-memory discrepancies, HF token handoff, unhealthy recipe-loaded containers, Q4\u002FMTP or other profile switches, live endpoint checks, and bounded candidate tests that must preserve the starting state.","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\u002F.agents\u002Fskills\u002Fanvil-serving-candidate-operations\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: anvil-serving-candidate-operations\ndescription: Prepare, recover, switch, test, and restore an Anvil Serving model candidate through managed recipes and serves. Use for RTX 5090 lane preparation, Windows\u002FWSL GPU-memory discrepancies, HF token handoff, unhealthy recipe-loaded containers, Q4\u002FMTP or other profile switches, live endpoint checks, and bounded candidate tests that must preserve the starting state.\n---\n\n# Anvil Serving Candidate Operations\n\nRun candidate work as one evidence-backed transaction. Container health,\nendpoint readiness, preflight, benchmark qualification, and route promotion are\nseparate states; never infer one from another.\n\n## Start state\n\n1. Read `README.md`, `CLAUDE.md`, the selected recipe, and\n   `.agents\u002Fskills\u002Fanvil-serving-llm-qualification\u002FSKILL.md` before mutation.\n2. Record repository revision and dirty state, command host\u002Fruntime, topology\n   target, router and serve state, reservations, loaded recipe containers, GPU\n   inventory, candidate recipe revision, and exact state to restore.\n3. Use `operation_contracts` and the structured controller\u002FMCP tools first.\n   Use the documented CLI commands below only when the wrapper is absent and\n   report that product gap.\n\n## Preflight the host boundary\n\n1. Resolve credentials by variable name only. If the general\n   `host-operations:credential-source-diagnostics` plugin skill is installed,\n   use it to trace the shell, Windows\u002FWSL, service, Compose, and container\n   boundary. Otherwise check presence without printing, hashing, or copying the\n   value. A `.env` file is not proof that the candidate process receives\n   `HF_TOKEN` or the recipe's declared credential variable.\n2. Run `anvil-serving host gpus --json` for product inventory. On Windows\u002FWSL,\n   use `host-operations:windows-gpu-lane-hygiene` when installed; otherwise\n   capture Windows and WSL NVIDIA views, exact PIDs, managed workloads, active\n   allocation, and residual WDDM\u002Fdriver reservation. Do not claim the lane is\n   empty from a utilization percentage or an empty CUDA process table alone.\n3. Stop for a human gate before terminating a process, changing Windows\n   graphics preferences, restarting Docker\u002FWSL, unloading a live serve, or\n   changing a deployment role. Prefer per-app iGPU placement and exact managed\n   workload lifecycle over global graphics changes.\n\n## Bound host memory before an unqualified load\n\nGPU fit is not host-RAM fit. Record available RAM, swap, desktop\u002FOS reserve,\nand the loader peak estimate before a new checkpoint\u002Fruntime startup. If the\nloader peak is unknown, require a managed container RAM and RAM-plus-swap\nceiling that preserves the reserve. If the managed lifecycle cannot enforce\nthat ceiling, stop the candidate trial and record the product gap; do not\nsubstitute raw Docker or rely on the global OOM killer. Preserve exact known-good\nrestoration separately from new candidate experimentation.\n\nFor every authorized candidate load or model swap, retain bounded, timestamped\nhost-memory snapshots in private attempt evidence: before unloading the\nincumbent, immediately before loading the candidate, after startup, and after\nrestoration. A refusal or startup failure requires an immediate snapshot before\ncleanup when safe; capture must never delay urgent restoration. Record missing\nphases and unavailable fields explicitly. A documentation-only change does not\nrequire a live capture or model operation.\n\nEach snapshot must include:\n\n- Total, free and available host RAM; swap total\u002Fused; memory pressure; file\n  cache, shared memory and reclaimable\u002Funreclaimable kernel memory. Keep these\n  accounting views separate rather than adding overlapping counters.\n- The largest processes and managed services\u002Fcontainers, with PID\u002Fparent,\n  identity, counts and proportional resident memory (PSS) when available.\n  Group desktop\u002Fagent child tools so many small workers are not missed. Retain\n  cgroup `memory.current`, anonymous\u002Ffile\u002Fshared\u002Fkernel and swap accounting when\n  available; container totals include cache and are not process PSS.\n- Candidate RAM and RAM-plus-swap ceilings, host reserve, required available\n  RAM, observed available RAM and signed headroom\u002Fshortfall. A ceiling is not\n  measured consumption. Bind the recipe, command host, phase and timestamp.\n\nUse supported host\u002Fstatus\u002Fmonitoring commands first. If Linux attribution or\nnumeric refusal values are unavailable, retain bounded read-only `\u002Fproc` and\ncgroup evidence and record the product gap. Never read process environments or\nretain raw command lines, credentials or shared environment files. A nearby\nsnapshot is not the guard's exact sample: label refusal-time values unknown\nunless the guard retained them. An admission refusal is not an engine OOM.\nInvestigate co-resident owners before proposing a new memory budget; do not\nsilently lower the reserve, evict caches, terminate agent\u002Fsession processes or\nretry the same frozen attempt. Apply existing lifecycle authority without\nrequesting duplicate approval.\n\nA host-wide OOM or desktop-session loss blocks retries until containment and\nloader diagnosis are independently checked. See the\n[September 17 failure](..\u002F..\u002F..\u002Fdocs\u002Ffindings\u002F2026-09-17-glm53-mixed35-startup.md)\nand [containment ticket](..\u002F..\u002F..\u002F.tickets\u002F2026-09-17-recipe-host-memory-containment.md).\n\n## Select one lifecycle path\n\nUse recipe lifecycle for an isolated candidate:\n\n```text\nanvil-serving models recipes show MODEL\nanvil-serving models recipes load MODEL --container NAME --gpu-device GPU --dry-run\nanvil-serving models recipes load MODEL --container NAME --gpu-device GPU --confirm\nanvil-serving models recipes status MODEL --container NAME\nanvil-serving models recipes logs MODEL --container NAME --tail 500 --contains ERROR --contains \"KV cache\"\nanvil-serving models recipes unload MODEL --container NAME --dry-run\nanvil-serving models recipes unload MODEL --container NAME --confirm\n```\n\nUse `anvil-serving serves switch ROLE MODEL --dry-run` only for a declared\ndeployment-role change. Apply with `--confirm` only after the preview identifies\nthe exact source artifacts and rollback state. A recipe load does not change\nrouter policy, and an isolated test does not require an alias change.\n\n## Recover an unhealthy candidate\n\n1. Run recipe `status`, then bounded recipe `logs`. Capture the earliest\n   actionable startup error and classify authentication, authorization\u002Flicense,\n   missing dependency, incompatible engine\u002Fmodel, resource exhaustion, or\n   failed health contract. Start with repeatable `--contains` literals for the\n   expected failure signatures so progress bars and routine startup output do\n   not flood agent context. Widen the filter or remove it only when the cause\n   remains unclear; retain an unfiltered artifact outside model context when a\n   complete raw log is required for evidence.\n2. Do not \"fix unhealthy\" by immediately recreating the container. First prove\n   the owning cause and whether the recipe, image, environment handoff, health\n   check, or product status surface is defective.\n3. If replacement is justified, verify the exact model\u002Fcontainer identity,\n   preview `unload`, obtain the mutation gate, unload, re-check the GPU lane,\n   preview `load`, load, and follow bounded logs from first startup.\n4. Use raw Docker only for the narrowest read-only diagnosis when the Anvil\n   surface cannot expose required status or logs. Record that missing capability\n   as a product gap and return to managed commands for mutation.\n5. On Windows\u002FWSL, fix one startup layer per managed recipe: establish NCCL\n   transport first, then runner compatibility, then model load\u002Fcompile, and only\n   then KV\u002Fcontext capacity. Preserve each exact failure. A transport fix is not\n   evidence that the model fits, and a compile OOM before KV allocation is not a\n   measured context limit.\n\n## Qualify the candidate\n\n1. Require recipe status and the candidate's native health endpoint to pass.\n2. Run `anvil-serving eval preflight` against the exact direct endpoint\u002Fmodel,\n   starting with thinking disabled and the smallest deterministic checks.\n3. Expand to the profile's context, tools, streaming, modality, and concurrency\n   gates only after smoke passes. Use the LLM qualification skill for capacity,\n   repeated quality evidence, and publication.\n4. Keep external recipe claims advisory. If the test came from a gist or\n   runbook, use `workflow-intake:external-procedure-intake` when installed to\n   retain its exact revision and the local adaptation matrix.\n\n## Close client context propagation\n\nAn isolated recipe load does not require client changes. A managed serve\nswitch, router-profile change, promotion, rollback, or fix-forward recovery\ndoes whenever it changes a client-facing alias's backing identity,\n`context_limit_tokens`, `max_output_tokens`, reasoning, or modality contract.\n\nEvery approved promotion also requires the infrastructure capture and all-host\nprovisioning gate in `skills\u002Fanvil-serving-release-readiness\u002FSKILL.md`, even when\nlimits or aliases are unchanged. Bind rebuilt server dependencies and every\nlocal\u002Fremote harness catalog to the same approved profile and effective router\nhash, then prove repeat convergence. Never mark a healthy route complete while\nprovisioning can reinstall a stale model or client contract. Return\n`harness_convergence_pending` with the accepted serve preserved for any gap.\n\nFor such changes on Companion Node, use `converge-mini-vision-clients` when\ninstalled; otherwise enforce the same closure directly before reporting the\noperation restored or complete. Reconcile Hermes, Pi, and OpenClaw from the\nsame authenticated router snapshot; enumerate every retained Anvil-backed\nHermes profile; preserve provider\u002Fauth, unrelated settings, and compaction;\nand require real-client acceptance plus an idempotent second dry-run. Also\nverify the durable scheduled reconciliation job covers\n`openclaw,hermes,pi`, uses `--hermes-profiles all`, and last exited zero.\n\nIf the operation lacks authority to update\u002Frestart live clients, do not stage\nconfig in live paths. Record exact candidate bindings and return\n`client_convergence=pending`; a healthy model or router is not deployment\nclosure.\n\n## Close an authorized promotion's publication\n\nWhen the user separately approves promotion, complete the\n[publication closure gate](..\u002F..\u002F..\u002Fskills\u002Fanvil-serving-benchmark-docs\u002FSKILL.md#promotion-publication-closure)\nafter live model\u002Fclient acceptance. Publish the Anvil Serving benchmark\nnarrative and update the declared Grafana and Workbench benchmark views through\nthe operator's dashboard infrastructure repository. Verify actual result data\nand retain the live publication receipt; updating model labels or a private\nrunbook is insufficient. Preserve the exact rollback and serving state while\npublishing. Report `promoted_publication_pending` if a required surface remains\nunpublished; do not represent the entire promotion workflow as complete.\n\n## Restore and report\n\nFor isolated or unpromoted trials, stop the candidate through its owning\nrecipe\u002Fserve lifecycle, restore the exact starting serve and router state, and\nprove the GPU returns to the recorded baseline. For an explicitly approved\npromotion, preserve the accepted promoted serve and retained rollback instead;\ndo not restore the starting serve. Complete client and publication closure\nwithout undoing the approved cutover. Do not remove model caches or Docker\nvolumes as cleanup.\n\nWhen restoration is required for an unpromoted trial or separately authorized\nrollback, treat it as a fresh managed deployment, not as an assumed container\nrestart. Use a preserved `--keep-container` baseline only when its startup\npreparation is proven restart-idempotent; an image that verifies and patches a\nfile in place may fail its checksum on restart. Otherwise unload the preserved\ncontainer and recreate the exact starting recipe. Load the router credential\nfrom its declared user-local environment source before a mode transaction that\nmust drain or readmit a tier. After direct health and model identity pass,\nrequire authenticated routed smoke and structured-output acceptance, exact mode\nownership, GPU residency, and shared-memory status. Preserve failed restore\nattempts and the root cause rather than overwriting them with the successful\nretry.\n\nReturn the start-state fingerprint, credential presence boundary, GPU lane\nclassification, recipe\u002Fprofile identity, preview and confirmation gates,\nhealth and log cause, preflight\u002Fbenchmark artifacts, restoration checks,\nwhether client-facing metadata changed, client convergence state, product\ngaps, and `promoted=false` for isolated\u002Funpromoted trials. For an approved\npromotion, report `promoted=true` and the retained rollback, with completion\nstatus `promoted_publication_pending` until the publication gate passes.\nPromotion remains a separate human gate.\n","# Anvil Serving Candidate Operations\n\nRun candidate work as one evidence-backed transaction. Container health,\nendpoint readiness, preflight, benchmark qualification, and route promotion are\nseparate states; never infer one from another.\n\n## Start state\n\n1. Read `README.md`, `CLAUDE.md`, the selected recipe, and\n   `.agents\u002Fskills\u002Fanvil-serving-llm-qualification\u002FSKILL.md` before mutation.\n2. Record repository revision and dirty state, command host\u002Fruntime, topology\n   target, router and serve state, reservations, loaded recipe containers, GPU\n   inventory, candidate recipe revision, and exact state to restore.\n3. Use `operation_contracts` and the structured controller\u002FMCP tools first.\n   Use the documented CLI commands below only when the wrapper is absent and\n   report that product gap.\n\n## Preflight the host boundary\n\n1. Resolve credentials by variable name only. If the general\n   `host-operations:credential-source-diagnostics` plugin skill is installed,\n   use it to trace the shell, Windows\u002FWSL, service, Compose, and container\n   boundary. Otherwise check presence without printing, hashing, or copying the\n   value. A `.env` file is not proof that the candidate process receives\n   `HF_TOKEN` or the recipe's declared credential variable.\n2. Run `anvil-serving host gpus --json` for product inventory. On Windows\u002FWSL,\n   use `host-operations:windows-gpu-lane-hygiene` when installed; otherwise\n   capture Windows and WSL NVIDIA views, exact PIDs, managed workloads, active\n   allocation, and residual WDDM\u002Fdriver reservation. Do not claim the lane is\n   empty from a utilization percentage or an empty CUDA process table alone.\n3. Stop for a human gate before terminating a process, changing Windows\n   graphics preferences, restarting Docker\u002FWSL, unloading a live serve, or\n   changing a deployment role. Prefer per-app iGPU placement and exact managed\n   workload lifecycle over global graphics changes.\n\n## Bound host memory before an unqualified load\n\nGPU fit is not host-RAM fit. Record available RAM, swap, desktop\u002FOS reserve,\nand the loader peak estimate before a new checkpoint\u002Fruntime startup. If the\nloader peak is unknown, require a managed container RAM and RAM-plus-swap\nceiling that preserves the reserve. If the managed lifecycle cannot enforce\nthat ceiling, stop the candidate trial and record the product gap; do not\nsubstitute raw Docker or rely on the global OOM killer. Preserve exact known-good\nrestoration separately from new candidate experimentation.\n\nFor every authorized candidate load or model swap, retain bounded, timestamped\nhost-memory snapshots in private attempt evidence: before unloading the\nincumbent, immediately before loading the candidate, after startup, and after\nrestoration. A refusal or startup failure requires an immediate snapshot before\ncleanup when safe; capture must never delay urgent restoration. Record missing\nphases and unavailable fields explicitly. A documentation-only change does not\nrequire a live capture or model operation.\n\nEach snapshot must include:\n\n- Total, free and available host RAM; swap total\u002Fused; memory pressure; file\n  cache, shared memory and reclaimable\u002Funreclaimable kernel memory. Keep these\n  accounting views separate rather than adding overlapping counters.\n- The largest processes and managed services\u002Fcontainers, with PID\u002Fparent,\n  identity, counts and proportional resident memory (PSS) when available.\n  Group desktop\u002Fagent child tools so many small workers are not missed. Retain\n  cgroup `memory.current`, anonymous\u002Ffile\u002Fshared\u002Fkernel and swap accounting when\n  available; container totals include cache and are not process PSS.\n- Candidate RAM and RAM-plus-swap ceilings, host reserve, required available\n  RAM, observed available RAM and signed headroom\u002Fshortfall. A ceiling is not\n  measured consumption. Bind the recipe, command host, phase and timestamp.\n\nUse supported host\u002Fstatus\u002Fmonitoring commands first. If Linux attribution or\nnumeric refusal values are unavailable, retain bounded read-only `\u002Fproc` and\ncgroup evidence and record the product gap. Never read process environments or\nretain raw command lines, credentials or shared environment files. A nearby\nsnapshot is not the guard's exact sample: label refusal-time values unknown\nunless the guard retained them. An admission refusal is not an engine OOM.\nInvestigate co-resident owners before proposing a new memory budget; do not\nsilently lower the reserve, evict caches, terminate agent\u002Fsession processes or\nretry the same frozen attempt. Apply existing lifecycle authority without\nrequesting duplicate approval.\n\nA host-wide OOM or desktop-session loss blocks retries until containment and\nloader diagnosis are independently checked. See the\n[September 17 failure](..\u002F..\u002F..\u002Fdocs\u002Ffindings\u002F2026-09-17-glm53-mixed35-startup.md)\nand [containment ticket](..\u002F..\u002F..\u002F.tickets\u002F2026-09-17-recipe-host-memory-containment.md).\n\n## Select one lifecycle path\n\nUse recipe lifecycle for an isolated candidate:\n\n```text\nanvil-serving models recipes show MODEL\nanvil-serving models recipes load MODEL --container NAME --gpu-device GPU --dry-run\nanvil-serving models recipes load MODEL --container NAME --gpu-device GPU --confirm\nanvil-serving models recipes status MODEL --container NAME\nanvil-serving models recipes logs MODEL --container NAME --tail 500 --contains ERROR --contains \"KV cache\"\nanvil-serving models recipes unload MODEL --container NAME --dry-run\nanvil-serving models recipes unload MODEL --container NAME --confirm\n```\n\nUse `anvil-serving serves switch ROLE MODEL --dry-run` only for a declared\ndeployment-role change. Apply with `--confirm` only after the preview identifies\nthe exact source artifacts and rollback state. A recipe load does not change\nrouter policy, and an isolated test does not require an alias change.\n\n## Recover an unhealthy candidate\n\n1. Run recipe `status`, then bounded recipe `logs`. Capture the earliest\n   actionable startup error and classify authentication, authorization\u002Flicense,\n   missing dependency, incompatible engine\u002Fmodel, resource exhaustion, or\n   failed health contract. Start with repeatable `--contains` literals for the\n   expected failure signatures so progress bars and routine startup output do\n   not flood agent context. Widen the filter or remove it only when the cause\n   remains unclear; retain an unfiltered artifact outside model context when a\n   complete raw log is required for evidence.\n2. Do not \"fix unhealthy\" by immediately recreating the container. First prove\n   the owning cause and whether the recipe, image, environment handoff, health\n   check, or product status surface is defective.\n3. If replacement is justified, verify the exact model\u002Fcontainer identity,\n   preview `unload`, obtain the mutation gate, unload, re-check the GPU lane,\n   preview `load`, load, and follow bounded logs from first startup.\n4. Use raw Docker only for the narrowest read-only diagnosis when the Anvil\n   surface cannot expose required status or logs. Record that missing capability\n   as a product gap and return to managed commands for mutation.\n5. On Windows\u002FWSL, fix one startup layer per managed recipe: establish NCCL\n   transport first, then runner compatibility, then model load\u002Fcompile, and only\n   then KV\u002Fcontext capacity. Preserve each exact failure. A transport fix is not\n   evidence that the model fits, and a compile OOM before KV allocation is not a\n   measured context limit.\n\n## Qualify the candidate\n\n1. Require recipe status and the candidate's native health endpoint to pass.\n2. Run `anvil-serving eval preflight` against the exact direct endpoint\u002Fmodel,\n   starting with thinking disabled and the smallest deterministic checks.\n3. Expand to the profile's context, tools, streaming, modality, and concurrency\n   gates only after smoke passes. Use the LLM qualification skill for capacity,\n   repeated quality evidence, and publication.\n4. Keep external recipe claims advisory. If the test came from a gist or\n   runbook, use `workflow-intake:external-procedure-intake` when installed to\n   retain its exact revision and the local adaptation matrix.\n\n## Close client context propagation\n\nAn isolated recipe load does not require client changes. A managed serve\nswitch, router-profile change, promotion, rollback, or fix-forward recovery\ndoes whenever it changes a client-facing alias's backing identity,\n`context_limit_tokens`, `max_output_tokens`, reasoning, or modality contract.\n\nEvery approved promotion also requires the infrastructure capture and all-host\nprovisioning gate in `skills\u002Fanvil-serving-release-readiness\u002FSKILL.md`, even when\nlimits or aliases are unchanged. Bind rebuilt server dependencies and every\nlocal\u002Fremote harness catalog to the same approved profile and effective router\nhash, then prove repeat convergence. Never mark a healthy route complete while\nprovisioning can reinstall a stale model or client contract. Return\n`harness_convergence_pending` with the accepted serve preserved for any gap.\n\nFor such changes on Companion Node, use `converge-mini-vision-clients` when\ninstalled; otherwise enforce the same closure directly before reporting the\noperation restored or complete. Reconcile Hermes, Pi, and OpenClaw from the\nsame authenticated router snapshot; enumerate every retained Anvil-backed\nHermes profile; preserve provider\u002Fauth, unrelated settings, and compaction;\nand require real-client acceptance plus an idempotent second dry-run. Also\nverify the durable scheduled reconciliation job covers\n`openclaw,hermes,pi`, uses `--hermes-profiles all`, and last exited zero.\n\nIf the operation lacks authority to update\u002Frestart live clients, do not stage\nconfig in live paths. Record exact candidate bindings and return\n`client_convergence=pending`; a healthy model or router is not deployment\nclosure.\n\n## Close an authorized promotion's publication\n\nWhen the user separately approves promotion, complete the\n[publication closure gate](..\u002F..\u002F..\u002Fskills\u002Fanvil-serving-benchmark-docs\u002FSKILL.md#promotion-publication-closure)\nafter live model\u002Fclient acceptance. Publish the Anvil Serving benchmark\nnarrative and update the declared Grafana and Workbench benchmark views through\nthe operator's dashboard infrastructure repository. Verify actual result data\nand retain the live publication receipt; updating model labels or a private\nrunbook is insufficient. Preserve the exact rollback and serving state while\npublishing. Report `promoted_publication_pending` if a required surface remains\nunpublished; do not represent the entire promotion workflow as complete.\n\n## Restore and report\n\nFor isolated or unpromoted trials, stop the candidate through its owning\nrecipe\u002Fserve lifecycle, restore the exact starting serve and router state, and\nprove the GPU returns to the recorded baseline. For an explicitly approved\npromotion, preserve the accepted promoted serve and retained rollback instead;\ndo not restore the starting serve. Complete client and publication closure\nwithout undoing the approved cutover. Do not remove model caches or Docker\nvolumes as cleanup.\n\nWhen restoration is required for an unpromoted trial or separately authorized\nrollback, treat it as a fresh managed deployment, not as an assumed container\nrestart. Use a preserved `--keep-container` baseline only when its startup\npreparation is proven restart-idempotent; an image that verifies and patches a\nfile in place may fail its checksum on restart. Otherwise unload the preserved\ncontainer and recreate the exact starting recipe. Load the router credential\nfrom its declared user-local environment source before a mode transaction that\nmust drain or readmit a tier. After direct health and model identity pass,\nrequire authenticated routed smoke and structured-output acceptance, exact mode\nownership, GPU residency, and shared-memory status. Preserve failed restore\nattempts and the root cause rather than overwriting them with the successful\nretry.\n\nReturn the start-state fingerprint, credential presence boundary, GPU lane\nclassification, recipe\u002Fprofile identity, preview and confirmation gates,\nhealth and log cause, preflight\u002Fbenchmark artifacts, restoration checks,\nwhether client-facing metadata changed, client convergence state, product\ngaps, and `promoted=false` for isolated\u002Funpromoted trials. For an approved\npromotion, report `promoted=true` and the retained rollback, with completion\nstatus `promoted_publication_pending` until the publication gate passes.\nPromotion remains a separate human gate.\n",{"name":5,"description":7},"main","c02ba7c047f81b9e5fcc2a26434f6b5bfb82f5b0",".agents\u002Fskills\u002Fanvil-serving-candidate-operations",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-candidate-operations",{"tool":38,"command":39},"bunx","bunx skills add fakoli\u002Fanvil-serving --skill anvil-serving-candidate-operations",{"tool":41,"command":42},"pnpm","pnpm dlx skills add fakoli\u002Fanvil-serving --skill anvil-serving-candidate-operations","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fanvil-serving-candidate-operations\u002FSKILL.md",true,[46],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"agents\u002Fopenai.yaml","REFERENCE","参考文档",297,"https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fanvil-serving-candidate-operations\u002Fagents\u002Fopenai.yaml",[53,59,65,70,75,82],{"skillNo":54,"name":55,"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":56,"categories":57,"keywords":58,"publishedAt":19},"01M77RK15F1RY23K6YY54W7YQC","voice","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcli_reference_audit\u002Finput\u002Fskills\u002Fvoice\u002FSKILL.md",[],[],{"skillNo":60,"name":61,"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":62,"categories":63,"keywords":64,"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":66,"name":61,"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":67,"categories":68,"keywords":69,"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":71,"name":61,"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":72,"categories":73,"keywords":74,"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":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"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":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"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",[],[]]