[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77RHF79VM55HTBEKR2C14CY":3,"skill-same:01M77RHF79VM55HTBEKR2C14CY":46},{"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},"01M77RHF79VM55HTBEKR2C14CY","anvil-serving-workbench",null,"Operate anvil-serving from an agent harness using MCP\u002Fcontroller-backed playbooks, CLI fallbacks, and sub-agent evidence packets. Use when asked to inspect Anvil Serving status, choose models, run preflight or benchmarks, sync OpenClaw\u002FCodex\u002FClaude Code workbench config, collect promotion evidence, or coordinate small-model sub-agents for anvil-serving operations.","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.claude\u002Fskills\u002Fanvil-serving-workbench\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: anvil-serving-workbench\ndescription: Operate anvil-serving from an agent harness using MCP\u002Fcontroller-backed playbooks, CLI fallbacks, and sub-agent evidence packets. Use when asked to inspect Anvil Serving status, choose models, run preflight or benchmarks, sync OpenClaw\u002FCodex\u002FClaude Code workbench config, collect promotion evidence, or coordinate small-model sub-agents for anvil-serving operations.\n---\n\n# Anvil Serving Workbench\n\nUse this skill as the safe entry point for operating anvil-serving from a large\nlanguage model workbench. It chooses documented playbooks, prefers structured\nMCP\u002Fcontroller tools, and returns auditable evidence instead of ad hoc shell\nnarration.\n\n## Start Here\n\n1. Read `README.md`, `CLAUDE.md`, and `docs\u002FOPERATOR-SKILLS-AND-SUBAGENTS.md`\n   before changing behavior or running operational commands.\n2. List or inspect available MCP tools first. Use `operation_contracts` before\n   topology-aware or controller-backed work so the selected CLI operation,\n   target context, transport, and MCP wrapper agree. Use the grouped catalog\n   below rather than relying on a memorized subset. If the repo-scoped MCP\n   server is unavailable, report that loss of controller coverage explicitly.\n3. Use documented `anvil-serving` CLI verbs only when an MCP wrapper is missing.\n   Safe fallbacks are read-only or preview-first verbs such as `eval usage`,\n   `models sync`, `models recipes`, `models score`,\n   `harness sync openclaw --out -`,\n   and other render\u002Finspect commands. Return the command preview and mark the\n   missing wrapper as a product gap. Before using a CLI fallback, verify that\n   it resolves to this checkout or the intended installed version; do not imply\n   that an unverified fallback is controller-backed.\n4. Use `127.0.0.1` in local URLs. Do not introduce `localhost`.\n   For resource-owned commands, pass the deployed topology and declare the\n   actual command host\u002Fruntime; do not infer command identity from the target.\n5. For OpenClaw Talk \u002F Anvil Voice work, record the command host before\n   interpreting loopback. Reference deployment: Companion Node runs OpenClaw\n   Gateway plus Anvil Voice Realtime\u002Fproxy and reserves its 16 GB RAM for\n   OpenClaw, Claude Code, and Codex. Do not run STT\u002FTTS\u002FLLM model serves on\n   Mini for reference testing. Primary Node owns the router, candidate serves,\n   and STT\u002FTTS model endpoints. Prefer `dark-audio` or\n   `mini-dark-audio-proxy` for OpenClaw Talk; `mini-audio` is an explicit\n   optional same-host\u002Flocal-audio mode only. A non-gateway checkout failing to\n   reach Mini proxy loopback is topology evidence, not proof the live\n   Mini\u002FDark path is down.\n6. Pass credentials by environment variable name only. Never place literal keys\n   in configs, fixtures, packets, logs, or prompts.\n\n## MCP Tool Map\n\n- Gateway: `router_status`, `router_fleet_status`, `router_config_export`, `router_logs`, `router_manage`,\n  `router_transition`, `router_configuration`, `runtime_experiment`, and `decision_summary`.\n- Serves and residency: `serves_status`, `reservation_status`,\n  `serves_manage`, `serves_mode`, `serves_logs`, `serves_probe`, `serves_profile`,\n  and `serves_promote`.\n- Managed media worker: `media_worker_prepare`, `media_worker_status`,\n  `media_worker_logs`, and `media_worker_teardown`. Lifecycle mutations retain\n  the managed-serve dry-run, confirmation, and human-approval gates.\n- Caller media: `media_capabilities`, `media_workflow_list`,\n  `media_workflow_show`, `media_workflow_validate`, `media_workflow_prepare`, `media_workflow_run`,\n  `media_job_status`, `media_job_cancel`, and `media_artifact_inspect`.\n- Voice: `voice_manage` and `voice_proxy_manage`.\n- Host, models, and telemetry: `doctor_summary`, `host_summary`,\n  `gpu_inventory`, `host_shared_memory`, `operator_config_inventory`,\n  `operator_config_export`, `observability_collect`, `host_manage`,\n  `host_services_status`, `host_services_discover`, `host_services_capabilities`,\n  `host_services_logs`, `host_services_manage`, `container_exec`,\n  `models_inventory`, `recipe_containers`, `recipe_settings`, `recipe_manage`,\n  `model_cache_inventory`, and\n  `cache_prune_plan`.\n- Harness: `openclaw_sync`, `client_catalog_sync`, `routed_eval`, `hermes_media_sync`,\n  `openclaw_gateway_status`, and `openclaw_gateway_restart`.\n- Propagation: `propagation_capabilities` reports scoped owner operations.\n  The default controller has no execution profile. A constructor-bound service\n  exposes approved admission, dispatch and bounded fleet operations; inspect\n  each operation's availability and required scope before use. See\n  `docs\u002FPROPAGATION-NATIVE-WRITES.md` for the controlled integration boundary.\n- Evaluation and evidence: `preflight_probe`, `benchmark_probe`,\n  `benchmark_artifact`, `benchmark_harness_prepare`, `benchmark_harness_status`,\n  `benchmark_harness_cleanup`, `benchmark_job_preflight`, `benchmark_job_submit`, `benchmark_job_status`, `benchmark_job_list`,\n  `benchmark_job_logs`, `benchmark_job_cancel`, `benchmark_job_artifact`,\n  `workflow_packet_validate`,\n  `external_bench_sources`, `external_bench_list`, `external_bench_report`,\n  and `external_bench_compare`.\n- Transport discovery: `operation_contracts`.\n- Controller diagnostics: `controller_inspect` and `controller_logs` return\n  bounded metadata for one explicitly selected Docker controller, never raw\n  log messages. The Docker log tail is 1..200 lines (default 100), and projection\n  returns at most 200 allowlisted events.\n\n`benchmark_probe` and `benchmark_artifact` are bounded capacity tools. The\n`benchmark_job_*` tools provide durable context, agentic, and SWE job control;\nthey do not make a job's grader independent or authorize model promotion.\n\n## Gates\n\nStop for a human gate before changing the active direct alias, applying a serve\npromotion, destructive cache pruning, host repair, Docker\u002FWSL restart, public\nor non-loopback bind, or any operation that would persist a new serving target.\n\n`serves_promote` and live `serves_mode` enter\u002Fleave have the same three-part\ngate: `confirm=true`, `dry_run=false`, and `human_approved=true`. `host_manage` and mutating\n`router_transition` actions remain human-gated even though their MCP schemas\nsupport preview and confirmation.\n\nNever self-verify. The model that generated a candidate output cannot be the\ncritic or judge that validates it.\n\nDo not write Anvil `state.db` directly. If Anvil task entry is requested, use\nthe supported Anvil CLI or MCP path; if it is unavailable, report the blocker.\n\n## Playbook Selection\n\n- Controller diagnosis: use `controller_inspect` and `controller_logs` through\n  the declared controller when reachable. When that transport is unavailable,\n  use the verified local `anvil-serving controller inspect --container NAME`\n  and `anvil-serving controller logs --container NAME --tail 20` on its owning\n  host; this read-only fallback also applies when a remote MCP wrapper exists\n  but cannot be reached. Do not substitute SSH or another controller. Native\n  services and macOS-local diagnostics are unsupported in this version.\n  Configured bindings are not observed publication, and diagnostic `state=ok`\n  proves neither endpoint node\u002Fbuild identity nor deployment acceptance.\n  Diagnostics grant no repair, restart, route or promotion authority.\n- Readiness: inspect `operation_contracts`, `router_status`,\n  `router_fleet_status`, `serves_status`,\n  `reservation_status`, `doctor_summary`, `host_summary`, `gpu_inventory`,\n  `models_inventory`, and configured endpoint status. Use\n  `observability_collect` only for bounded declared capabilities.\n- Model catalog: read or sync model inventory and use `external_bench_sources`,\n  `external_bench_list`, `external_bench_report`, or `external_bench_compare`\n  for benchmark priors. Keep those priors advisory-only. Use\n  `models recipes list\u002Fshow` to inspect recorded configurations and\n  `recipe_containers` or `models recipes running` to discover label-owned\n  loaded containers without registry dependence. Create or\n  revise candidates through `models recipes create\u002Fupdate`, review the rendered\n  recipe, and use\n  `models recipes load` only with an exact container plus its documented\n  confirmation gate. Use `models recipes delete` only for an exact reviewed\n  registry entry. Run `models pull` only after an explicit network, disk, and\n  target-volume gate.\n- Serve swap: inspect `reservation_status`, then preview lifecycle work with\n  `serves_manage`, an exclusive TP=2 transaction with `serves_mode`, or a\n  named transaction with `serves_promote`. The newer\n  role-based recipe flow is `anvil-serving serves switch ROLE [MODEL]`; it is\n  CLI-only, so return its preview and name the missing MCP wrapper. Require an\n  exact target and the documented confirmation gate before apply, then run\n  preflight before either capacity or quality benchmarks.\n- Voice lifecycle: preview audio with `voice_manage` and the persistent\n  Realtime proxy with `voice_proxy_manage`; live lifecycle changes on the\n  owning host require `confirm=true` plus `dry_run=false`.\n  Profile selection is topology selection: for reference OpenClaw Talk and\n  candidate benchmarks, keep Mini model-free and select Dark-host audio or a\n  Mini-side proxy to Dark. Use `mini-audio` only when explicitly testing the\n  optional same-host\u002Flocal-audio mode.\n- Harness sync: preview OpenClaw provider, skill, and agent config with\n  `openclaw_sync`; apply only to an explicit `out`\u002F`gateway_host` target and\n  preserve operator-owned keys. Use `client_catalog_sync` to reconcile local\n  OpenClaw and Pi context\u002Foutput limits from the router's authenticated status\n  and capability endpoints; preview first, preserve compaction\u002Fauth keys, and\n  require confirmation before writing. Use `hermes_media_sync` to install the\n  narrow media MCP catalog and packaged Hermes skill; keep credentials as\n  environment references and verify a second preview is empty. Use\n  `routed_eval` on the client-owning host, locally or through its declared\n  controller, for fail-closed router identity plus real OpenClaw\u002FHermes\n  acceptance; do not substitute a response from another host's clients.\n- Router operations: use `router_status`, `router_fleet_status`, bounded `router_logs`,\n  `router_manage`, `router_transition`, and `decision_summary`; lifecycle and\n  tier-transition mutation is preview-first and live only with `confirm=true`\n  plus `dry_run=false`.\n- Evaluation: before model validation or comparison, load\n  `skills\u002Fanvil-serving-llm-qualification\u002Freferences\u002Fevidence-contract.md`\n  and `skills\u002Fanvil-serving-benchmark-docs\u002FSKILL.md`. Declare the test scope\n  and freeze the required serving-metrics matrix before requests. A focused\n  smoke check must identify its limited scope; a full comparison requires\n  verified coverage of the frozen matrix. Preserve quality failures separately\n  from safe performance measurements and report every unmeasured metric.\n  Use `preflight_probe` with the model family's declared thinking\n  mode, reasoning effort\u002Fevidence expectation, visible-answer budget,\n  reasoning headroom, and allowed finish reasons. Use `benchmark_probe` or\n  `benchmark_artifact` for capacity only. For quality, run\n  `anvil-serving eval benchmark quality` with repeated attempts and preserve\n  visible output, reasoning-channel evidence, finish reasons, separate visible\n  and reasoning budgets, provenance, and per-attempt failure classification.\n  Do not rank or promote from an older one-shot score that lacks that evidence.\n- Promotion evidence: assemble status, gateway probes, preflight, benchmark\n  artifacts, serve\u002Fdirect-alias config diffs, and reviewer recommendation with\n  `promoted=false`.\n- OpenClaw COLO smoke\u002Feval: when validating the Mini-to-Dark path or gathering\n  release\u002Fblog evidence, use `examples\u002Fopenclaw\u002Fcolo_smoke.py --live`; add\n  `--run-generations --run-interaction-benchmark` for repeatable intent stats.\n  Preserve the artifact's `interaction_benchmarks` and recipe fields. Treat\n  generation caps, exact\u002Fstream benchmark caps, benchmark reasoning effort, and\n  endpoint overrides as serve\u002Fmodel-recipe settings, not skill or plugin\n  constants.\n- Host\u002Fcache work: use `host_summary`, `gpu_inventory`, `host_shared_memory`, and `cache_prune_plan`\n  for read-only checks and plans. Use `host_manage` only for an exact reviewed\n  repair after its human gate. Report Docker\u002FWSL restart, WSL config edits, and\n  cache deletion as `blocked` or `human_required` unless the human approves the\n  existing gate; MCP cache pruning is plan-only. `host_summary` also reports the\n  persistent `host.toml` WSL cache-reclaim policy. When enabled, confirmed\n  model pulls and managed model-load boundaries inherit one disclosed,\n  readiness-aware, warning-only page-cache postcondition; do not add a second\n  manual reclaim after a `reclaimed` result. Ad-hoc Compose, voice, request-time\n  ComfyUI loads, and the request-triggered multiplexer remain outside that\n  policy. Treat changing the persistent opt-in as a reviewed host mutation.\n- Host-supervised service work: use `host_services_status`,\n  `host_services_discover`, `host_services_capabilities`, and bounded\n  `host_services_logs` before `host_services_manage`. Adopt only an exact\n  launchd or Anvil-owned Docker identity. Mutating service actions stay previews\n  until `dry_run=false` and `confirm=true`; never use raw supervisor commands.\n  Keep manager and engine separate, retain Parakeet\u002FKokoro LaunchAgent entries\n  as `legacy`, and use a qualified model-lifecycle or recipe workflow for any model\n  migration. Docker model bindings require their owning model declaration and\n  manifest path;\n  native model bindings require `memory_mib` and preserve the 4 GiB host reserve.\n- Public\u002Fprivate config work: use `operator_config_inventory` first for\n  metadata-only classification and dependency closure. Use\n  `operator_config_export` only for the resulting safe versionable files and\n  allowlisted, redacted Anvil-owned gateway fragment. Preserve its refusal on\n  symlinks, unsupported YAML export, outside-root or missing dependencies,\n  unsafe secret fields, and oversized files; never replace it with raw remote\n  filesystem access. When a whole-home export is blocked by YAML, use the\n  inventory and an explicit bounded `paths` selection for only the required\n  supported files; every selected file's direct dependency closure is still\n  mandatory.\n\n## Model Benchmark Source Freshness\n\nWhen researching Fast or Heavy tier candidates, include dates in the evidence.\nTreat serving posts, community recipes, benchmark tables, and issue threads as\ntime-sensitive.\n\n- Classify sources as `current` at 0-60 days, `aging` at 61-120 days, and\n  `stale` after 120 days, when undated, or when tied to a materially older\n  engine, driver, CUDA, quantization, or checkpoint generation.\n- Prefer current official model cards, vendor release notes, and vLLM, SGLang,\n  llama.cpp, NVIDIA, or model-owner documentation and issues.\n- Treat Reddit, forum posts, and community benchmarks as recipe discovery, not\n  promotion evidence. Label aging or stale hardware-matched posts as historical\n  priors and require a current official source or local benchmark.\n- Record candidate, URL, published or observed date, age class, evidence type,\n  hardware\u002Fengine relevance, and decision impact.\n- Never let stale community evidence justify a shortlist, serve swap, or\n  promotion by itself. Local preflight, benchmarks, independent quality eval,\n  and the human gate remain required.\n\n## Sub-Agents\n\nUse small\u002Flocal models for inventory scout, serve-config reviewer, preflight runner,\nbenchmark runner, and evidence reporter roles. Use a stronger independent model\nfor quality critic or adversarial review. Keep role outputs bounded: facts,\ntool calls, artifacts, recommendations, and blockers.\n\n## Result Packet\n\nReturn an `operator-workflow\u002Fv1` packet with `schema_version`, `request`,\n`gate_state`, `targets`, `tools_used`, `artifacts`, `advisory_priors`,\n`recommendation`, `human_gate_required`, and `promoted`. Each `tools_used`\nentry must include `source_class`, `ok`, `dry_run`, `confirmed`, `target`, and\n`error`. Allowed recommendations are `promote`, `do_not_promote`,\n`needs_more_data`, and `blocked`. Allowed `gate_state` values are\n`not_required`, `confirm_required`, `human_required`, and `blocked`. Validate\nfinal packets with `workflow_packet_validate` when the MCP\u002Fcontrol-plane tool is\navailable. That validator checks packet shape, gate consistency, evidence\nscope, and bounded paths; it does not prove evidence sufficiency or reviewer\nindependence. An independent critic must still reject promotion when the\nunderlying evidence is missing or self-generated.\n","# Anvil Serving Workbench\n\nUse this skill as the safe entry point for operating anvil-serving from a large\nlanguage model workbench. It chooses documented playbooks, prefers structured\nMCP\u002Fcontroller tools, and returns auditable evidence instead of ad hoc shell\nnarration.\n\n## Start Here\n\n1. Read `README.md`, `CLAUDE.md`, and `docs\u002FOPERATOR-SKILLS-AND-SUBAGENTS.md`\n   before changing behavior or running operational commands.\n2. List or inspect available MCP tools first. Use `operation_contracts` before\n   topology-aware or controller-backed work so the selected CLI operation,\n   target context, transport, and MCP wrapper agree. Use the grouped catalog\n   below rather than relying on a memorized subset. If the repo-scoped MCP\n   server is unavailable, report that loss of controller coverage explicitly.\n3. Use documented `anvil-serving` CLI verbs only when an MCP wrapper is missing.\n   Safe fallbacks are read-only or preview-first verbs such as `eval usage`,\n   `models sync`, `models recipes`, `models score`,\n   `harness sync openclaw --out -`,\n   and other render\u002Finspect commands. Return the command preview and mark the\n   missing wrapper as a product gap. Before using a CLI fallback, verify that\n   it resolves to this checkout or the intended installed version; do not imply\n   that an unverified fallback is controller-backed.\n4. Use `127.0.0.1` in local URLs. Do not introduce `localhost`.\n   For resource-owned commands, pass the deployed topology and declare the\n   actual command host\u002Fruntime; do not infer command identity from the target.\n5. For OpenClaw Talk \u002F Anvil Voice work, record the command host before\n   interpreting loopback. Reference deployment: Companion Node runs OpenClaw\n   Gateway plus Anvil Voice Realtime\u002Fproxy and reserves its 16 GB RAM for\n   OpenClaw, Claude Code, and Codex. Do not run STT\u002FTTS\u002FLLM model serves on\n   Mini for reference testing. Primary Node owns the router, candidate serves,\n   and STT\u002FTTS model endpoints. Prefer `dark-audio` or\n   `mini-dark-audio-proxy` for OpenClaw Talk; `mini-audio` is an explicit\n   optional same-host\u002Flocal-audio mode only. A non-gateway checkout failing to\n   reach Mini proxy loopback is topology evidence, not proof the live\n   Mini\u002FDark path is down.\n6. Pass credentials by environment variable name only. Never place literal keys\n   in configs, fixtures, packets, logs, or prompts.\n\n## MCP Tool Map\n\n- Gateway: `router_status`, `router_fleet_status`, `router_config_export`, `router_logs`, `router_manage`,\n  `router_transition`, `router_configuration`, `runtime_experiment`, and `decision_summary`.\n- Serves and residency: `serves_status`, `reservation_status`,\n  `serves_manage`, `serves_mode`, `serves_logs`, `serves_probe`, `serves_profile`,\n  and `serves_promote`.\n- Managed media worker: `media_worker_prepare`, `media_worker_status`,\n  `media_worker_logs`, and `media_worker_teardown`. Lifecycle mutations retain\n  the managed-serve dry-run, confirmation, and human-approval gates.\n- Caller media: `media_capabilities`, `media_workflow_list`,\n  `media_workflow_show`, `media_workflow_validate`, `media_workflow_prepare`, `media_workflow_run`,\n  `media_job_status`, `media_job_cancel`, and `media_artifact_inspect`.\n- Voice: `voice_manage` and `voice_proxy_manage`.\n- Host, models, and telemetry: `doctor_summary`, `host_summary`,\n  `gpu_inventory`, `host_shared_memory`, `operator_config_inventory`,\n  `operator_config_export`, `observability_collect`, `host_manage`,\n  `host_services_status`, `host_services_discover`, `host_services_capabilities`,\n  `host_services_logs`, `host_services_manage`, `container_exec`,\n  `models_inventory`, `recipe_containers`, `recipe_settings`, `recipe_manage`,\n  `model_cache_inventory`, and\n  `cache_prune_plan`.\n- Harness: `openclaw_sync`, `client_catalog_sync`, `routed_eval`, `hermes_media_sync`,\n  `openclaw_gateway_status`, and `openclaw_gateway_restart`.\n- Propagation: `propagation_capabilities` reports scoped owner operations.\n  The default controller has no execution profile. A constructor-bound service\n  exposes approved admission, dispatch and bounded fleet operations; inspect\n  each operation's availability and required scope before use. See\n  `docs\u002FPROPAGATION-NATIVE-WRITES.md` for the controlled integration boundary.\n- Evaluation and evidence: `preflight_probe`, `benchmark_probe`,\n  `benchmark_artifact`, `benchmark_harness_prepare`, `benchmark_harness_status`,\n  `benchmark_harness_cleanup`, `benchmark_job_preflight`, `benchmark_job_submit`, `benchmark_job_status`, `benchmark_job_list`,\n  `benchmark_job_logs`, `benchmark_job_cancel`, `benchmark_job_artifact`,\n  `workflow_packet_validate`,\n  `external_bench_sources`, `external_bench_list`, `external_bench_report`,\n  and `external_bench_compare`.\n- Transport discovery: `operation_contracts`.\n- Controller diagnostics: `controller_inspect` and `controller_logs` return\n  bounded metadata for one explicitly selected Docker controller, never raw\n  log messages. The Docker log tail is 1..200 lines (default 100), and projection\n  returns at most 200 allowlisted events.\n\n`benchmark_probe` and `benchmark_artifact` are bounded capacity tools. The\n`benchmark_job_*` tools provide durable context, agentic, and SWE job control;\nthey do not make a job's grader independent or authorize model promotion.\n\n## Gates\n\nStop for a human gate before changing the active direct alias, applying a serve\npromotion, destructive cache pruning, host repair, Docker\u002FWSL restart, public\nor non-loopback bind, or any operation that would persist a new serving target.\n\n`serves_promote` and live `serves_mode` enter\u002Fleave have the same three-part\ngate: `confirm=true`, `dry_run=false`, and `human_approved=true`. `host_manage` and mutating\n`router_transition` actions remain human-gated even though their MCP schemas\nsupport preview and confirmation.\n\nNever self-verify. The model that generated a candidate output cannot be the\ncritic or judge that validates it.\n\nDo not write Anvil `state.db` directly. If Anvil task entry is requested, use\nthe supported Anvil CLI or MCP path; if it is unavailable, report the blocker.\n\n## Playbook Selection\n\n- Controller diagnosis: use `controller_inspect` and `controller_logs` through\n  the declared controller when reachable. When that transport is unavailable,\n  use the verified local `anvil-serving controller inspect --container NAME`\n  and `anvil-serving controller logs --container NAME --tail 20` on its owning\n  host; this read-only fallback also applies when a remote MCP wrapper exists\n  but cannot be reached. Do not substitute SSH or another controller. Native\n  services and macOS-local diagnostics are unsupported in this version.\n  Configured bindings are not observed publication, and diagnostic `state=ok`\n  proves neither endpoint node\u002Fbuild identity nor deployment acceptance.\n  Diagnostics grant no repair, restart, route or promotion authority.\n- Readiness: inspect `operation_contracts`, `router_status`,\n  `router_fleet_status`, `serves_status`,\n  `reservation_status`, `doctor_summary`, `host_summary`, `gpu_inventory`,\n  `models_inventory`, and configured endpoint status. Use\n  `observability_collect` only for bounded declared capabilities.\n- Model catalog: read or sync model inventory and use `external_bench_sources`,\n  `external_bench_list`, `external_bench_report`, or `external_bench_compare`\n  for benchmark priors. Keep those priors advisory-only. Use\n  `models recipes list\u002Fshow` to inspect recorded configurations and\n  `recipe_containers` or `models recipes running` to discover label-owned\n  loaded containers without registry dependence. Create or\n  revise candidates through `models recipes create\u002Fupdate`, review the rendered\n  recipe, and use\n  `models recipes load` only with an exact container plus its documented\n  confirmation gate. Use `models recipes delete` only for an exact reviewed\n  registry entry. Run `models pull` only after an explicit network, disk, and\n  target-volume gate.\n- Serve swap: inspect `reservation_status`, then preview lifecycle work with\n  `serves_manage`, an exclusive TP=2 transaction with `serves_mode`, or a\n  named transaction with `serves_promote`. The newer\n  role-based recipe flow is `anvil-serving serves switch ROLE [MODEL]`; it is\n  CLI-only, so return its preview and name the missing MCP wrapper. Require an\n  exact target and the documented confirmation gate before apply, then run\n  preflight before either capacity or quality benchmarks.\n- Voice lifecycle: preview audio with `voice_manage` and the persistent\n  Realtime proxy with `voice_proxy_manage`; live lifecycle changes on the\n  owning host require `confirm=true` plus `dry_run=false`.\n  Profile selection is topology selection: for reference OpenClaw Talk and\n  candidate benchmarks, keep Mini model-free and select Dark-host audio or a\n  Mini-side proxy to Dark. Use `mini-audio` only when explicitly testing the\n  optional same-host\u002Flocal-audio mode.\n- Harness sync: preview OpenClaw provider, skill, and agent config with\n  `openclaw_sync`; apply only to an explicit `out`\u002F`gateway_host` target and\n  preserve operator-owned keys. Use `client_catalog_sync` to reconcile local\n  OpenClaw and Pi context\u002Foutput limits from the router's authenticated status\n  and capability endpoints; preview first, preserve compaction\u002Fauth keys, and\n  require confirmation before writing. Use `hermes_media_sync` to install the\n  narrow media MCP catalog and packaged Hermes skill; keep credentials as\n  environment references and verify a second preview is empty. Use\n  `routed_eval` on the client-owning host, locally or through its declared\n  controller, for fail-closed router identity plus real OpenClaw\u002FHermes\n  acceptance; do not substitute a response from another host's clients.\n- Router operations: use `router_status`, `router_fleet_status`, bounded `router_logs`,\n  `router_manage`, `router_transition`, and `decision_summary`; lifecycle and\n  tier-transition mutation is preview-first and live only with `confirm=true`\n  plus `dry_run=false`.\n- Evaluation: before model validation or comparison, load\n  `skills\u002Fanvil-serving-llm-qualification\u002Freferences\u002Fevidence-contract.md`\n  and `skills\u002Fanvil-serving-benchmark-docs\u002FSKILL.md`. Declare the test scope\n  and freeze the required serving-metrics matrix before requests. A focused\n  smoke check must identify its limited scope; a full comparison requires\n  verified coverage of the frozen matrix. Preserve quality failures separately\n  from safe performance measurements and report every unmeasured metric.\n  Use `preflight_probe` with the model family's declared thinking\n  mode, reasoning effort\u002Fevidence expectation, visible-answer budget,\n  reasoning headroom, and allowed finish reasons. Use `benchmark_probe` or\n  `benchmark_artifact` for capacity only. For quality, run\n  `anvil-serving eval benchmark quality` with repeated attempts and preserve\n  visible output, reasoning-channel evidence, finish reasons, separate visible\n  and reasoning budgets, provenance, and per-attempt failure classification.\n  Do not rank or promote from an older one-shot score that lacks that evidence.\n- Promotion evidence: assemble status, gateway probes, preflight, benchmark\n  artifacts, serve\u002Fdirect-alias config diffs, and reviewer recommendation with\n  `promoted=false`.\n- OpenClaw COLO smoke\u002Feval: when validating the Mini-to-Dark path or gathering\n  release\u002Fblog evidence, use `examples\u002Fopenclaw\u002Fcolo_smoke.py --live`; add\n  `--run-generations --run-interaction-benchmark` for repeatable intent stats.\n  Preserve the artifact's `interaction_benchmarks` and recipe fields. Treat\n  generation caps, exact\u002Fstream benchmark caps, benchmark reasoning effort, and\n  endpoint overrides as serve\u002Fmodel-recipe settings, not skill or plugin\n  constants.\n- Host\u002Fcache work: use `host_summary`, `gpu_inventory`, `host_shared_memory`, and `cache_prune_plan`\n  for read-only checks and plans. Use `host_manage` only for an exact reviewed\n  repair after its human gate. Report Docker\u002FWSL restart, WSL config edits, and\n  cache deletion as `blocked` or `human_required` unless the human approves the\n  existing gate; MCP cache pruning is plan-only. `host_summary` also reports the\n  persistent `host.toml` WSL cache-reclaim policy. When enabled, confirmed\n  model pulls and managed model-load boundaries inherit one disclosed,\n  readiness-aware, warning-only page-cache postcondition; do not add a second\n  manual reclaim after a `reclaimed` result. Ad-hoc Compose, voice, request-time\n  ComfyUI loads, and the request-triggered multiplexer remain outside that\n  policy. Treat changing the persistent opt-in as a reviewed host mutation.\n- Host-supervised service work: use `host_services_status`,\n  `host_services_discover`, `host_services_capabilities`, and bounded\n  `host_services_logs` before `host_services_manage`. Adopt only an exact\n  launchd or Anvil-owned Docker identity. Mutating service actions stay previews\n  until `dry_run=false` and `confirm=true`; never use raw supervisor commands.\n  Keep manager and engine separate, retain Parakeet\u002FKokoro LaunchAgent entries\n  as `legacy`, and use a qualified model-lifecycle or recipe workflow for any model\n  migration. Docker model bindings require their owning model declaration and\n  manifest path;\n  native model bindings require `memory_mib` and preserve the 4 GiB host reserve.\n- Public\u002Fprivate config work: use `operator_config_inventory` first for\n  metadata-only classification and dependency closure. Use\n  `operator_config_export` only for the resulting safe versionable files and\n  allowlisted, redacted Anvil-owned gateway fragment. Preserve its refusal on\n  symlinks, unsupported YAML export, outside-root or missing dependencies,\n  unsafe secret fields, and oversized files; never replace it with raw remote\n  filesystem access. When a whole-home export is blocked by YAML, use the\n  inventory and an explicit bounded `paths` selection for only the required\n  supported files; every selected file's direct dependency closure is still\n  mandatory.\n\n## Model Benchmark Source Freshness\n\nWhen researching Fast or Heavy tier candidates, include dates in the evidence.\nTreat serving posts, community recipes, benchmark tables, and issue threads as\ntime-sensitive.\n\n- Classify sources as `current` at 0-60 days, `aging` at 61-120 days, and\n  `stale` after 120 days, when undated, or when tied to a materially older\n  engine, driver, CUDA, quantization, or checkpoint generation.\n- Prefer current official model cards, vendor release notes, and vLLM, SGLang,\n  llama.cpp, NVIDIA, or model-owner documentation and issues.\n- Treat Reddit, forum posts, and community benchmarks as recipe discovery, not\n  promotion evidence. Label aging or stale hardware-matched posts as historical\n  priors and require a current official source or local benchmark.\n- Record candidate, URL, published or observed date, age class, evidence type,\n  hardware\u002Fengine relevance, and decision impact.\n- Never let stale community evidence justify a shortlist, serve swap, or\n  promotion by itself. Local preflight, benchmarks, independent quality eval,\n  and the human gate remain required.\n\n## Sub-Agents\n\nUse small\u002Flocal models for inventory scout, serve-config reviewer, preflight runner,\nbenchmark runner, and evidence reporter roles. Use a stronger independent model\nfor quality critic or adversarial review. Keep role outputs bounded: facts,\ntool calls, artifacts, recommendations, and blockers.\n\n## Result Packet\n\nReturn an `operator-workflow\u002Fv1` packet with `schema_version`, `request`,\n`gate_state`, `targets`, `tools_used`, `artifacts`, `advisory_priors`,\n`recommendation`, `human_gate_required`, and `promoted`. Each `tools_used`\nentry must include `source_class`, `ok`, `dry_run`, `confirmed`, `target`, and\n`error`. Allowed recommendations are `promote`, `do_not_promote`,\n`needs_more_data`, and `blocked`. Allowed `gate_state` values are\n`not_required`, `confirm_required`, `human_required`, and `blocked`. Validate\nfinal packets with `workflow_packet_validate` when the MCP\u002Fcontrol-plane tool is\navailable. That validator checks packet shape, gate consistency, evidence\nscope, and bounded paths; it does not prove evidence sufficiency or reviewer\nindependence. An independent critic must still reject promotion when the\nunderlying evidence is missing or self-generated.\n",{"name":5,"description":7},"main","c02ba7c047f81b9e5fcc2a26434f6b5bfb82f5b0",".claude\u002Fskills\u002Fanvil-serving-workbench",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-workbench",{"tool":38,"command":39},"bunx","bunx skills add fakoli\u002Fanvil-serving --skill anvil-serving-workbench",{"tool":41,"command":42},"pnpm","pnpm dlx skills add fakoli\u002Fanvil-serving --skill anvil-serving-workbench","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fanvil-serving-workbench\u002FSKILL.md",true,[],[47,53,59,64,69,76],{"skillNo":48,"name":49,"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":50,"categories":51,"keywords":52,"publishedAt":19},"01M77RK15F1RY23K6YY54W7YQC","voice","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcli_reference_audit\u002Finput\u002Fskills\u002Fvoice\u002FSKILL.md",[],[],{"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},"01M77RK06RC7DDZQE1MEZHA9NQ","workbench","https:\u002F\u002Fgithub.com\u002Ffakoli\u002Fanvil-serving\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fcli_reference_audit\u002Finput\u002Fexamples\u002Fopenclaw\u002Fskills\u002Fworkbench\u002FSKILL.md",[],[],{"skillNo":60,"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":61,"categories":62,"keywords":63,"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":65,"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":66,"categories":67,"keywords":68,"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":70,"name":71,"nameZh":6,"description":72,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":73,"categories":74,"keywords":75,"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":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"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",[],[]]