[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-huggingface-harbor-hf":3,"mdc-n7cqnx-key":35,"related-repo-huggingface-harbor-hf":1183,"related-org-huggingface-harbor-hf":1192},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":24,"repoUrl":25,"updatedAt":26,"license":27,"forks":28,"topics":29,"repo":30,"sourceUrl":33,"mdContent":34},"harbor-hf","manage Harbor benchmark campaigns on Hugging Face","Plan, profile, validate, launch, monitor, reconcile, recover, verify and score Harbor benchmark campaigns, then publish them through Hugging Face Jobs, Inference Providers, and Inference Endpoints.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},"huggingface","Hugging Face","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fhuggingface.png",[12,15,18,21],{"name":9,"slug":13,"type":14},"hugging-face","tag",{"name":16,"slug":17,"type":14},"Benchmarking","benchmarking",{"name":19,"slug":20,"type":14},"Monitoring","monitoring",{"name":22,"slug":23,"type":14},"Deployment","deployment",0,"https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fharbor-hf","2026-08-13T05:04:13.527907",null,2,[],{"repoUrl":25,"stars":24,"forks":28,"topics":31,"description":32},[],"Run reproducible Harbor benchmarks on Hugging Face infrastructure.","https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fharbor-hf\u002Ftree\u002FHEAD\u002Fskills\u002Fharbor-hf","---\nname: harbor-hf\ndescription: \"Plan, profile, validate, launch, monitor, reconcile, recover, verify and score Harbor benchmark campaigns, then publish them through Hugging Face Jobs, Inference Providers, and Inference Endpoints.\"\n---\n\n# Harbor HF operations\n\nUse this skill for operational work with `harbor-hf`. It covers single runs,\nfull campaigns, provider-backed agents, deployment profiles, paid canaries,\nHugging Face Jobs, recovery, private evidence, and result publication.\n\nA Harbor HF campaign is a durable control-plane object. Treat HF Jobs as\nreplaceable workers. Read campaign state from the coordination Dataset and\ncanonical evidence from the private Bucket.\n\n## Source documents\n\nRead the complete source document before acting in that area:\n\n- New campaign or live operation: `docs\u002Fharbor-cookbook.md`,\n  `docs\u002Frun-spec.md`, `docs\u002Fbenchmark-sources.md`, and\n  `docs\u002Fsingle-job-campaign-controller.md`.\n- New deployment or concurrency change: `docs\u002Fdeployment-profiling.md`.\n- Provider-backed agent: `docs\u002Fprovider-agent-architecture.md` and\n  `docs\u002Fharbor-integration-contract.md`.\n- Evidence, audit, restoration, or repair: `docs\u002Ftrial-evidence-bundle.md`.\n- Publication or catalog change: `docs\u002Fresult-publication.md` and the current\n  `results publish` and `results catalog` CLI help.\n\nThe checked-out CLI is authoritative for command syntax. Run\n`uv run harbor-hf \u003Cgroup> \u003Ccommand> --help` before using a mutating command.\nDo not copy a stale command from a report or an older campaign checkout.\n\nLoad the focused references in this skill as needed:\n\n- `references\u002Fplanning-and-capacity.md` for immutable inputs, profiling,\n  canaries and wave sizing, including timeouts and spend.\n- `references\u002Flaunch-and-monitoring.md` for submission, reconciliation, Jobs,\n  live observations and cancellation plus operator handoff.\n- `references\u002Frecovery.md` for failure classification, retries, interrupted\n  finalization, immutable replacements, and duplicate prevention.\n- `references\u002Fevidence-and-publication.md` for evidence, checksums, secret\n  scans and scoring plus publication and catalog decisions.\n- `references\u002Fprovider-agents-and-security.md` for custom agents, ingress,\n  credentials, identity checks, and provider-specific canaries.\n- `references\u002Foperator-checklists.md` for the final go\u002Fno-go and reporting\n  checklists.\n\n## Operating invariants\n\nKeep these rules in force throughout the session:\n\n- Run models and benchmark tasks only on remote Hugging Face infrastructure.\n- Pin every executable source and model reference. Pin tasks and images as well\n  as agents and workers.\n- Keep the control Dataset, input Bucket, evidence Bucket, and unpublished\n  results private.\n- Serialize campaign control mutations. A provider controller runs one internal\n  wave at a time; trial requests may overlap only within locked provider limits.\n- Treat `execution.concurrent_trials` and provider request concurrency as\n  separate limits.\n- Admit a wave only after measured end-to-end duration fits its deadline with\n  explicit reserve for drain and evidence publication.\n- Retry infrastructure failures only. Agent exits, time limits, refusals,\n  incomplete tasks, and verifier rejections remain terminal outcomes.\n- Never rerun an agent or judge to repair frozen historical evidence.\n- Preserve physical executions and terminal evidence together with hashes and\n  recovery provenance. Write terminal markers last.\n- Public Git benchmark sources are anonymous. Local or private benchmark files\n  use immutable bundles. Never forward a Git credential, SSH key, SSH agent, or\n  local credential helper into remote infrastructure.\n- Never copy an ambient local login into a remote secret. Every runtime\n  credential must be purpose-scoped and explicitly approved for its exact\n  source and destination. `harbor-hf auth add-job-token` records that approval,\n  stores the value in Harbor HF's private token file, and stores only the\n  selected name in its JSON config.\n- Never put credentials, route capabilities, authorization headers, cookies,\n  secret query parameters, or environment values in durable artifacts.\n- For endpoint work, success requires a verified paused endpoint with zero\n  ready replicas.\n- A reward of zero is a valid benchmark result when the execution evidence is\n  complete.\n- Missing, conflicting, unsafe, or ambiguous evidence fails closed.\n\n## Authorization boundary\n\nPlanning, validation, local inspection, dry runs, and artifact verification do\nnot authorize paid remote work. Obtain explicit user authorization before a\ncommand that can submit a Job, resume an endpoint, forward paid provider\nrequests, install automation, publish results, or change the primary catalog.\n\nBefore paid work, state the exact manifest, campaign identity policy, namespace,\nprovider or endpoint, model, trial count, wave count, maximum active waves,\nwave deadline, Job deadline, spend cap, estimated wave cost, and recovery\npolicy. Ask for missing choices. Never ask the user to paste a secret into chat.\n\n## Required workflow\n\n### Repository and state inspection\n\n1. Confirm the checkout and branch. Record the commit and working tree.\n2. Read repository instructions and the source documents listed above.\n3. Inspect the manifest, previous plans, launch records, active campaigns,\n   running Jobs, and owned endpoints.\n4. Determine whether the request concerns a new campaign, an immutable retry,\n   an audit, publication, or a recovery operation.\n5. Check for an existing campaign or Job before creating anything. Repeated\n   prompts must remain idempotent.\n\n### Immutable planning\n\n1. Resolve full commits and SHA-256 digests for every behavior-affecting input.\n   Resolve public Git anonymously. Resolve a local directory into the exact\n   bundle source lock without uploading it during planning.\n2. Verify the complete task set and exact logical attempt count.\n3. Validate the manifest and write the campaign plan to a durable local file.\n4. Plan again from a clean checkout when introducing a new benchmark,\n   deployment, agent, or source revision. Compare plan digests and identities.\n5. Inspect the plan directly. Record run and shard counts plus trial and wave\n   counts instead of trusting a filename or manifest label.\n6. Confirm that infrastructure retries do not consume logical attempt numbers.\n\nUse:\n\n```bash\nuv run harbor-hf validate MANIFEST\nuv run harbor-hf campaign plan MANIFEST --format json > PLAN.json\n```\n\n### Capacity gate\n\nA valid schema does not prove that a wave can finish. Profile the exact workload\nor run a representative paid canary before the full campaign. Record complete\ntrial wall time, p50, p95, maximum duration, provider calls per trial, request\nspacing, queueing, evidence finalization time, and cleanup time.\n\nRun the bundled duration check with a conservative planning duration:\n\n```bash\nuv run python skills\u002Fharbor-hf\u002Fscripts\u002Fcheck_wave_budget.py \\\n  --manifest MANIFEST \\\n  --plan PLAN.json \\\n  --planning-trial-seconds SECONDS \\\n  --reserve-seconds SECONDS \\\n  --headroom-factor FACTOR\n```\n\nThe command must exit successfully. Review every deployment group and every\nwave in its JSON output. A failing report blocks submission. Reduce\n`max_shards_per_wave`, reduce `max_trials_per_shard`, or create a new profile.\nIncreasing a deadline without checking the enclosing HF Job and Sandbox limits\nis not an acceptable fix.\n\n### Canary and pilot gate\n\nUse separate stages:\n\n1. A transport canary proves source checkout, agent installation, ingress,\n   provider routing, judge routing, session capture, trajectory identity,\n   workspace output, checksums, and secret isolation.\n2. A representative pilot wave proves throughput and deadline sizing under the\n   selected concurrency and request pacing.\n3. The full campaign starts only after both stages pass and their evidence is\n   verified.\n\nA canary that scores zero can pass the transport gate. A canary with an\nexception, missing evidence, wrong identity, unsafe credential exposure, or an\nunexplained provider error fails the gate.\n\n### Submission gate\n\nBefore submission:\n\n1. Re-run validation and planning from the pinned worker revision.\n2. Compare the new plan digest with the approved plan.\n3. Run `campaign submit --dry-run` and inspect every remote write.\n4. Confirm Bucket privacy, source lock, bundle upload or reuse action, secret\n   names, namespace, Job image digest, worker revision, Harbor revision, agent\n   revision, provider route, and judge policy. Prove that no Git credential is\n   included.\n5. Confirm duration arithmetic and budget arithmetic from the same manifest.\n6. Save the approved manifest, plan, duration report, and launch decision.\n7. Submit once and capture the returned campaign ID, controller Job ID, input\n   digest, plan digest, and launch receipt.\n\n### Live operation\n\nThe campaign projection and controller status are the primary status surfaces:\n\n```bash\nuv run harbor-hf campaign status CAMPAIGN_ID --namespace NAMESPACE\nuv run harbor-hf campaign reconcile CAMPAIGN_ID --namespace NAMESPACE --dry-run\n```\n\nA provider campaign advances inside its detached controller Job. Do not run an\napplied local reconciliation loop. Check the controller attempt, claim,\nheartbeat age, current wave, remaining-time admission, and block reason. HF Job\nstate and logs are supporting evidence. While work runs, confirm that terminal\ntrial bundles and provider records appear in the private Bucket.\n\nInstall the shared watchdog only for an explicit campaign list. The watchdog\nmay launch a sequential replacement after a retryable infrastructure failure.\nIt cannot continue a capacity or policy pause.\n\nThe controller stops new work when observed throughput makes the locked duration\ninfeasible, provider errors exceed the profile boundary, evidence publication\nstalls, spend admission fails, or ownership becomes uncertain.\n\n### Recovery gate\n\nObserve before mutating. Load the campaign projection, control events, HF Job\nstate, wave evidence, trial markers, and checksums. Classify each affected\nlogical trial before requesting a retry.\n\nUse `campaign retry` only for trials already classified as retryable\ninfrastructure failures. An immutable manifest change requires a new campaign\nidentity and explicit linkage to the superseded campaign. Build a trial-level\nledger before any replacement so completed logical outcomes are not executed\nagain by accident.\n\nNever use `seal` merely to make a campaign terminal. Seal only a drained partial\ncampaign whose exhausted infrastructure outcomes are intentionally accepted as\nzero under the declared publication policy.\n\n### Evidence and publication gate\n\nBefore publication:\n\n1. Reconcile until all Jobs, waves, leases, retries, and cleanup actions are\n   terminal.\n2. Verify endpoint pause evidence when applicable.\n3. Run remote artifact verification.\n4. Deep-validate representative and repaired trial bundles locally.\n5. Verify every declared digest and terminal-marker ordering.\n6. Scan all retained private and candidate public files for known secret values\n   and high-confidence generic credential patterns.\n7. Confirm trial counts, attempt counts, task digests, selected executions,\n   rewards, judge identities, and result classification.\n8. Dry-run publication, publish once, and record the receipt and exact Dataset\n   revision.\n9. Change the primary catalog only with explicit authorization and a stated\n   reason.\n\n## Hard stops\n\nStop without launching or mutating when any of these conditions holds:\n\n- A required commit, digest, task identity, agent revision, or image pin is\n  mutable or unknown.\n- The task count, attempt count, shard count, or planned wave count differs from\n  the approved protocol.\n- No representative timing evidence exists for a paid full campaign.\n- The duration checker fails or depends on unexplained concurrency.\n- Estimated wave cost is missing, exceeds the campaign cap, or leaves no room\n  for the declared retry policy.\n- A provider or endpoint quota is unknown.\n- A secret value would enter a command, manifest, lock, log, or agent process.\n- A public Git source needs authentication, or a local\u002Fprivate source cannot be\n  represented by a complete verified bundle.\n- A credential would be copied from local configuration into a remote secret\n  without approval naming the exact source and destination.\n- An endpoint is running without a verified owner and watchdog.\n- Existing terminal evidence is ambiguous, checksum-invalid, or duplicated.\n- A requested retry would rerun an agent or benchmark failure.\n- Full publication evidence has not passed checksum and secret validation.\n\n## Final report\n\nReport concrete evidence:\n\n- manifest and plan paths plus SHA-256 digests;\n- campaign, controller attempt, run, wave, shard, trial, execution, and HF Job IDs;\n- requested and observed concurrency, pacing, trial latency, wave duration, and\n  deadline headroom;\n- spend cap, reservation, observed or unreported spend, and retry capacity;\n- counts by queued, active, complete, retryable, exhausted, failed, cancelled,\n  and invalid state;\n- endpoint pause state or provider-route closure;\n- artifact verification, deep validation, and secret-scan results;\n- publication ID, Dataset revision, catalog state, and any remaining blocker.\n\nDo not call a campaign successful because its Job exited zero. Success comes\nfrom the durable campaign projection and verified canonical evidence.\n",{"data":36,"body":37},{"name":4,"description":6},{"type":38,"children":39},"root",[40,49,63,68,75,80,195,208,213,282,288,293,387,393,398,403,409,416,445,451,484,489,584,590,595,600,717,738,744,749,767,772,778,783,829,835,840,928,933,938,943,949,954,967,980,986,991,1039,1045,1050,1118,1124,1129,1172,1177],{"type":41,"tag":42,"props":43,"children":45},"element","h1",{"id":44},"harbor-hf-operations",[46],{"type":47,"value":48},"text","Harbor HF operations",{"type":41,"tag":50,"props":51,"children":52},"p",{},[53,55,61],{"type":47,"value":54},"Use this skill for operational work with ",{"type":41,"tag":56,"props":57,"children":59},"code",{"className":58},[],[60],{"type":47,"value":4},{"type":47,"value":62},". It covers single runs,\nfull campaigns, provider-backed agents, deployment profiles, paid canaries,\nHugging Face Jobs, recovery, private evidence, and result publication.",{"type":41,"tag":50,"props":64,"children":65},{},[66],{"type":47,"value":67},"A Harbor HF campaign is a durable control-plane object. Treat HF Jobs as\nreplaceable workers. Read campaign state from the coordination Dataset and\ncanonical evidence from the private Bucket.",{"type":41,"tag":69,"props":70,"children":72},"h2",{"id":71},"source-documents",[73],{"type":47,"value":74},"Source documents",{"type":41,"tag":50,"props":76,"children":77},{},[78],{"type":47,"value":79},"Read the complete source document before acting in that area:",{"type":41,"tag":81,"props":82,"children":83},"ul",{},[84,122,134,154,166],{"type":41,"tag":85,"props":86,"children":87},"li",{},[88,90,96,98,104,106,112,114,120],{"type":47,"value":89},"New campaign or live operation: ",{"type":41,"tag":56,"props":91,"children":93},{"className":92},[],[94],{"type":47,"value":95},"docs\u002Fharbor-cookbook.md",{"type":47,"value":97},",\n",{"type":41,"tag":56,"props":99,"children":101},{"className":100},[],[102],{"type":47,"value":103},"docs\u002Frun-spec.md",{"type":47,"value":105},", ",{"type":41,"tag":56,"props":107,"children":109},{"className":108},[],[110],{"type":47,"value":111},"docs\u002Fbenchmark-sources.md",{"type":47,"value":113},", and\n",{"type":41,"tag":56,"props":115,"children":117},{"className":116},[],[118],{"type":47,"value":119},"docs\u002Fsingle-job-campaign-controller.md",{"type":47,"value":121},".",{"type":41,"tag":85,"props":123,"children":124},{},[125,127,133],{"type":47,"value":126},"New deployment or concurrency change: ",{"type":41,"tag":56,"props":128,"children":130},{"className":129},[],[131],{"type":47,"value":132},"docs\u002Fdeployment-profiling.md",{"type":47,"value":121},{"type":41,"tag":85,"props":135,"children":136},{},[137,139,145,147,153],{"type":47,"value":138},"Provider-backed agent: ",{"type":41,"tag":56,"props":140,"children":142},{"className":141},[],[143],{"type":47,"value":144},"docs\u002Fprovider-agent-architecture.md",{"type":47,"value":146}," and\n",{"type":41,"tag":56,"props":148,"children":150},{"className":149},[],[151],{"type":47,"value":152},"docs\u002Fharbor-integration-contract.md",{"type":47,"value":121},{"type":41,"tag":85,"props":155,"children":156},{},[157,159,165],{"type":47,"value":158},"Evidence, audit, restoration, or repair: ",{"type":41,"tag":56,"props":160,"children":162},{"className":161},[],[163],{"type":47,"value":164},"docs\u002Ftrial-evidence-bundle.md",{"type":47,"value":121},{"type":41,"tag":85,"props":167,"children":168},{},[169,171,177,179,185,187,193],{"type":47,"value":170},"Publication or catalog change: ",{"type":41,"tag":56,"props":172,"children":174},{"className":173},[],[175],{"type":47,"value":176},"docs\u002Fresult-publication.md",{"type":47,"value":178}," and the current\n",{"type":41,"tag":56,"props":180,"children":182},{"className":181},[],[183],{"type":47,"value":184},"results publish",{"type":47,"value":186}," and ",{"type":41,"tag":56,"props":188,"children":190},{"className":189},[],[191],{"type":47,"value":192},"results catalog",{"type":47,"value":194}," CLI help.",{"type":41,"tag":50,"props":196,"children":197},{},[198,200,206],{"type":47,"value":199},"The checked-out CLI is authoritative for command syntax. Run\n",{"type":41,"tag":56,"props":201,"children":203},{"className":202},[],[204],{"type":47,"value":205},"uv run harbor-hf \u003Cgroup> \u003Ccommand> --help",{"type":47,"value":207}," before using a mutating command.\nDo not copy a stale command from a report or an older campaign checkout.",{"type":41,"tag":50,"props":209,"children":210},{},[211],{"type":47,"value":212},"Load the focused references in this skill as needed:",{"type":41,"tag":81,"props":214,"children":215},{},[216,227,238,249,260,271],{"type":41,"tag":85,"props":217,"children":218},{},[219,225],{"type":41,"tag":56,"props":220,"children":222},{"className":221},[],[223],{"type":47,"value":224},"references\u002Fplanning-and-capacity.md",{"type":47,"value":226}," for immutable inputs, profiling,\ncanaries and wave sizing, including timeouts and spend.",{"type":41,"tag":85,"props":228,"children":229},{},[230,236],{"type":41,"tag":56,"props":231,"children":233},{"className":232},[],[234],{"type":47,"value":235},"references\u002Flaunch-and-monitoring.md",{"type":47,"value":237}," for submission, reconciliation, Jobs,\nlive observations and cancellation plus operator handoff.",{"type":41,"tag":85,"props":239,"children":240},{},[241,247],{"type":41,"tag":56,"props":242,"children":244},{"className":243},[],[245],{"type":47,"value":246},"references\u002Frecovery.md",{"type":47,"value":248}," for failure classification, retries, interrupted\nfinalization, immutable replacements, and duplicate prevention.",{"type":41,"tag":85,"props":250,"children":251},{},[252,258],{"type":41,"tag":56,"props":253,"children":255},{"className":254},[],[256],{"type":47,"value":257},"references\u002Fevidence-and-publication.md",{"type":47,"value":259}," for evidence, checksums, secret\nscans and scoring plus publication and catalog decisions.",{"type":41,"tag":85,"props":261,"children":262},{},[263,269],{"type":41,"tag":56,"props":264,"children":266},{"className":265},[],[267],{"type":47,"value":268},"references\u002Fprovider-agents-and-security.md",{"type":47,"value":270}," for custom agents, ingress,\ncredentials, identity checks, and provider-specific canaries.",{"type":41,"tag":85,"props":272,"children":273},{},[274,280],{"type":41,"tag":56,"props":275,"children":277},{"className":276},[],[278],{"type":47,"value":279},"references\u002Foperator-checklists.md",{"type":47,"value":281}," for the final go\u002Fno-go and reporting\nchecklists.",{"type":41,"tag":69,"props":283,"children":285},{"id":284},"operating-invariants",[286],{"type":47,"value":287},"Operating invariants",{"type":41,"tag":50,"props":289,"children":290},{},[291],{"type":47,"value":292},"Keep these rules in force throughout the session:",{"type":41,"tag":81,"props":294,"children":295},{},[296,301,306,311,316,329,334,339,344,349,354,367,372,377,382],{"type":41,"tag":85,"props":297,"children":298},{},[299],{"type":47,"value":300},"Run models and benchmark tasks only on remote Hugging Face infrastructure.",{"type":41,"tag":85,"props":302,"children":303},{},[304],{"type":47,"value":305},"Pin every executable source and model reference. Pin tasks and images as well\nas agents and workers.",{"type":41,"tag":85,"props":307,"children":308},{},[309],{"type":47,"value":310},"Keep the control Dataset, input Bucket, evidence Bucket, and unpublished\nresults private.",{"type":41,"tag":85,"props":312,"children":313},{},[314],{"type":47,"value":315},"Serialize campaign control mutations. A provider controller runs one internal\nwave at a time; trial requests may overlap only within locked provider limits.",{"type":41,"tag":85,"props":317,"children":318},{},[319,321,327],{"type":47,"value":320},"Treat ",{"type":41,"tag":56,"props":322,"children":324},{"className":323},[],[325],{"type":47,"value":326},"execution.concurrent_trials",{"type":47,"value":328}," and provider request concurrency as\nseparate limits.",{"type":41,"tag":85,"props":330,"children":331},{},[332],{"type":47,"value":333},"Admit a wave only after measured end-to-end duration fits its deadline with\nexplicit reserve for drain and evidence publication.",{"type":41,"tag":85,"props":335,"children":336},{},[337],{"type":47,"value":338},"Retry infrastructure failures only. Agent exits, time limits, refusals,\nincomplete tasks, and verifier rejections remain terminal outcomes.",{"type":41,"tag":85,"props":340,"children":341},{},[342],{"type":47,"value":343},"Never rerun an agent or judge to repair frozen historical evidence.",{"type":41,"tag":85,"props":345,"children":346},{},[347],{"type":47,"value":348},"Preserve physical executions and terminal evidence together with hashes and\nrecovery provenance. Write terminal markers last.",{"type":41,"tag":85,"props":350,"children":351},{},[352],{"type":47,"value":353},"Public Git benchmark sources are anonymous. Local or private benchmark files\nuse immutable bundles. Never forward a Git credential, SSH key, SSH agent, or\nlocal credential helper into remote infrastructure.",{"type":41,"tag":85,"props":355,"children":356},{},[357,359,365],{"type":47,"value":358},"Never copy an ambient local login into a remote secret. Every runtime\ncredential must be purpose-scoped and explicitly approved for its exact\nsource and destination. ",{"type":41,"tag":56,"props":360,"children":362},{"className":361},[],[363],{"type":47,"value":364},"harbor-hf auth add-job-token",{"type":47,"value":366}," records that approval,\nstores the value in Harbor HF's private token file, and stores only the\nselected name in its JSON config.",{"type":41,"tag":85,"props":368,"children":369},{},[370],{"type":47,"value":371},"Never put credentials, route capabilities, authorization headers, cookies,\nsecret query parameters, or environment values in durable artifacts.",{"type":41,"tag":85,"props":373,"children":374},{},[375],{"type":47,"value":376},"For endpoint work, success requires a verified paused endpoint with zero\nready replicas.",{"type":41,"tag":85,"props":378,"children":379},{},[380],{"type":47,"value":381},"A reward of zero is a valid benchmark result when the execution evidence is\ncomplete.",{"type":41,"tag":85,"props":383,"children":384},{},[385],{"type":47,"value":386},"Missing, conflicting, unsafe, or ambiguous evidence fails closed.",{"type":41,"tag":69,"props":388,"children":390},{"id":389},"authorization-boundary",[391],{"type":47,"value":392},"Authorization boundary",{"type":41,"tag":50,"props":394,"children":395},{},[396],{"type":47,"value":397},"Planning, validation, local inspection, dry runs, and artifact verification do\nnot authorize paid remote work. Obtain explicit user authorization before a\ncommand that can submit a Job, resume an endpoint, forward paid provider\nrequests, install automation, publish results, or change the primary catalog.",{"type":41,"tag":50,"props":399,"children":400},{},[401],{"type":47,"value":402},"Before paid work, state the exact manifest, campaign identity policy, namespace,\nprovider or endpoint, model, trial count, wave count, maximum active waves,\nwave deadline, Job deadline, spend cap, estimated wave cost, and recovery\npolicy. Ask for missing choices. Never ask the user to paste a secret into chat.",{"type":41,"tag":69,"props":404,"children":406},{"id":405},"required-workflow",[407],{"type":47,"value":408},"Required workflow",{"type":41,"tag":410,"props":411,"children":413},"h3",{"id":412},"repository-and-state-inspection",[414],{"type":47,"value":415},"Repository and state inspection",{"type":41,"tag":417,"props":418,"children":419},"ol",{},[420,425,430,435,440],{"type":41,"tag":85,"props":421,"children":422},{},[423],{"type":47,"value":424},"Confirm the checkout and branch. Record the commit and working tree.",{"type":41,"tag":85,"props":426,"children":427},{},[428],{"type":47,"value":429},"Read repository instructions and the source documents listed above.",{"type":41,"tag":85,"props":431,"children":432},{},[433],{"type":47,"value":434},"Inspect the manifest, previous plans, launch records, active campaigns,\nrunning Jobs, and owned endpoints.",{"type":41,"tag":85,"props":436,"children":437},{},[438],{"type":47,"value":439},"Determine whether the request concerns a new campaign, an immutable retry,\nan audit, publication, or a recovery operation.",{"type":41,"tag":85,"props":441,"children":442},{},[443],{"type":47,"value":444},"Check for an existing campaign or Job before creating anything. Repeated\nprompts must remain idempotent.",{"type":41,"tag":410,"props":446,"children":448},{"id":447},"immutable-planning",[449],{"type":47,"value":450},"Immutable planning",{"type":41,"tag":417,"props":452,"children":453},{},[454,459,464,469,474,479],{"type":41,"tag":85,"props":455,"children":456},{},[457],{"type":47,"value":458},"Resolve full commits and SHA-256 digests for every behavior-affecting input.\nResolve public Git anonymously. Resolve a local directory into the exact\nbundle source lock without uploading it during planning.",{"type":41,"tag":85,"props":460,"children":461},{},[462],{"type":47,"value":463},"Verify the complete task set and exact logical attempt count.",{"type":41,"tag":85,"props":465,"children":466},{},[467],{"type":47,"value":468},"Validate the manifest and write the campaign plan to a durable local file.",{"type":41,"tag":85,"props":470,"children":471},{},[472],{"type":47,"value":473},"Plan again from a clean checkout when introducing a new benchmark,\ndeployment, agent, or source revision. Compare plan digests and identities.",{"type":41,"tag":85,"props":475,"children":476},{},[477],{"type":47,"value":478},"Inspect the plan directly. Record run and shard counts plus trial and wave\ncounts instead of trusting a filename or manifest label.",{"type":41,"tag":85,"props":480,"children":481},{},[482],{"type":47,"value":483},"Confirm that infrastructure retries do not consume logical attempt numbers.",{"type":41,"tag":50,"props":485,"children":486},{},[487],{"type":47,"value":488},"Use:",{"type":41,"tag":490,"props":491,"children":496},"pre",{"className":492,"code":493,"language":494,"meta":495,"style":495},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","uv run harbor-hf validate MANIFEST\nuv run harbor-hf campaign plan MANIFEST --format json > PLAN.json\n","bash","",[497],{"type":41,"tag":56,"props":498,"children":499},{"__ignoreMap":495},[500,533],{"type":41,"tag":501,"props":502,"children":505},"span",{"class":503,"line":504},"line",1,[506,512,518,523,528],{"type":41,"tag":501,"props":507,"children":509},{"style":508},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[510],{"type":47,"value":511},"uv",{"type":41,"tag":501,"props":513,"children":515},{"style":514},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[516],{"type":47,"value":517}," run",{"type":41,"tag":501,"props":519,"children":520},{"style":514},[521],{"type":47,"value":522}," harbor-hf",{"type":41,"tag":501,"props":524,"children":525},{"style":514},[526],{"type":47,"value":527}," validate",{"type":41,"tag":501,"props":529,"children":530},{"style":514},[531],{"type":47,"value":532}," MANIFEST\n",{"type":41,"tag":501,"props":534,"children":535},{"class":503,"line":28},[536,540,544,548,553,558,563,568,573,579],{"type":41,"tag":501,"props":537,"children":538},{"style":508},[539],{"type":47,"value":511},{"type":41,"tag":501,"props":541,"children":542},{"style":514},[543],{"type":47,"value":517},{"type":41,"tag":501,"props":545,"children":546},{"style":514},[547],{"type":47,"value":522},{"type":41,"tag":501,"props":549,"children":550},{"style":514},[551],{"type":47,"value":552}," campaign",{"type":41,"tag":501,"props":554,"children":555},{"style":514},[556],{"type":47,"value":557}," plan",{"type":41,"tag":501,"props":559,"children":560},{"style":514},[561],{"type":47,"value":562}," MANIFEST",{"type":41,"tag":501,"props":564,"children":565},{"style":514},[566],{"type":47,"value":567}," --format",{"type":41,"tag":501,"props":569,"children":570},{"style":514},[571],{"type":47,"value":572}," json",{"type":41,"tag":501,"props":574,"children":576},{"style":575},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[577],{"type":47,"value":578}," >",{"type":41,"tag":501,"props":580,"children":581},{"style":514},[582],{"type":47,"value":583}," PLAN.json\n",{"type":41,"tag":410,"props":585,"children":587},{"id":586},"capacity-gate",[588],{"type":47,"value":589},"Capacity gate",{"type":41,"tag":50,"props":591,"children":592},{},[593],{"type":47,"value":594},"A valid schema does not prove that a wave can finish. Profile the exact workload\nor run a representative paid canary before the full campaign. Record complete\ntrial wall time, p50, p95, maximum duration, provider calls per trial, request\nspacing, queueing, evidence finalization time, and cleanup time.",{"type":41,"tag":50,"props":596,"children":597},{},[598],{"type":47,"value":599},"Run the bundled duration check with a conservative planning duration:",{"type":41,"tag":490,"props":601,"children":603},{"className":492,"code":602,"language":494,"meta":495,"style":495},"uv run python skills\u002Fharbor-hf\u002Fscripts\u002Fcheck_wave_budget.py \\\n  --manifest MANIFEST \\\n  --plan PLAN.json \\\n  --planning-trial-seconds SECONDS \\\n  --reserve-seconds SECONDS \\\n  --headroom-factor FACTOR\n",[604],{"type":41,"tag":56,"props":605,"children":606},{"__ignoreMap":495},[607,634,650,668,686,703],{"type":41,"tag":501,"props":608,"children":609},{"class":503,"line":504},[610,614,618,623,628],{"type":41,"tag":501,"props":611,"children":612},{"style":508},[613],{"type":47,"value":511},{"type":41,"tag":501,"props":615,"children":616},{"style":514},[617],{"type":47,"value":517},{"type":41,"tag":501,"props":619,"children":620},{"style":514},[621],{"type":47,"value":622}," python",{"type":41,"tag":501,"props":624,"children":625},{"style":514},[626],{"type":47,"value":627}," skills\u002Fharbor-hf\u002Fscripts\u002Fcheck_wave_budget.py",{"type":41,"tag":501,"props":629,"children":631},{"style":630},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[632],{"type":47,"value":633}," \\\n",{"type":41,"tag":501,"props":635,"children":636},{"class":503,"line":28},[637,642,646],{"type":41,"tag":501,"props":638,"children":639},{"style":514},[640],{"type":47,"value":641},"  --manifest",{"type":41,"tag":501,"props":643,"children":644},{"style":514},[645],{"type":47,"value":562},{"type":41,"tag":501,"props":647,"children":648},{"style":630},[649],{"type":47,"value":633},{"type":41,"tag":501,"props":651,"children":653},{"class":503,"line":652},3,[654,659,664],{"type":41,"tag":501,"props":655,"children":656},{"style":514},[657],{"type":47,"value":658},"  --plan",{"type":41,"tag":501,"props":660,"children":661},{"style":514},[662],{"type":47,"value":663}," PLAN.json",{"type":41,"tag":501,"props":665,"children":666},{"style":630},[667],{"type":47,"value":633},{"type":41,"tag":501,"props":669,"children":671},{"class":503,"line":670},4,[672,677,682],{"type":41,"tag":501,"props":673,"children":674},{"style":514},[675],{"type":47,"value":676},"  --planning-trial-seconds",{"type":41,"tag":501,"props":678,"children":679},{"style":514},[680],{"type":47,"value":681}," SECONDS",{"type":41,"tag":501,"props":683,"children":684},{"style":630},[685],{"type":47,"value":633},{"type":41,"tag":501,"props":687,"children":689},{"class":503,"line":688},5,[690,695,699],{"type":41,"tag":501,"props":691,"children":692},{"style":514},[693],{"type":47,"value":694},"  --reserve-seconds",{"type":41,"tag":501,"props":696,"children":697},{"style":514},[698],{"type":47,"value":681},{"type":41,"tag":501,"props":700,"children":701},{"style":630},[702],{"type":47,"value":633},{"type":41,"tag":501,"props":704,"children":706},{"class":503,"line":705},6,[707,712],{"type":41,"tag":501,"props":708,"children":709},{"style":514},[710],{"type":47,"value":711},"  --headroom-factor",{"type":41,"tag":501,"props":713,"children":714},{"style":514},[715],{"type":47,"value":716}," FACTOR\n",{"type":41,"tag":50,"props":718,"children":719},{},[720,722,728,730,736],{"type":47,"value":721},"The command must exit successfully. Review every deployment group and every\nwave in its JSON output. A failing report blocks submission. Reduce\n",{"type":41,"tag":56,"props":723,"children":725},{"className":724},[],[726],{"type":47,"value":727},"max_shards_per_wave",{"type":47,"value":729},", reduce ",{"type":41,"tag":56,"props":731,"children":733},{"className":732},[],[734],{"type":47,"value":735},"max_trials_per_shard",{"type":47,"value":737},", or create a new profile.\nIncreasing a deadline without checking the enclosing HF Job and Sandbox limits\nis not an acceptable fix.",{"type":41,"tag":410,"props":739,"children":741},{"id":740},"canary-and-pilot-gate",[742],{"type":47,"value":743},"Canary and pilot gate",{"type":41,"tag":50,"props":745,"children":746},{},[747],{"type":47,"value":748},"Use separate stages:",{"type":41,"tag":417,"props":750,"children":751},{},[752,757,762],{"type":41,"tag":85,"props":753,"children":754},{},[755],{"type":47,"value":756},"A transport canary proves source checkout, agent installation, ingress,\nprovider routing, judge routing, session capture, trajectory identity,\nworkspace output, checksums, and secret isolation.",{"type":41,"tag":85,"props":758,"children":759},{},[760],{"type":47,"value":761},"A representative pilot wave proves throughput and deadline sizing under the\nselected concurrency and request pacing.",{"type":41,"tag":85,"props":763,"children":764},{},[765],{"type":47,"value":766},"The full campaign starts only after both stages pass and their evidence is\nverified.",{"type":41,"tag":50,"props":768,"children":769},{},[770],{"type":47,"value":771},"A canary that scores zero can pass the transport gate. A canary with an\nexception, missing evidence, wrong identity, unsafe credential exposure, or an\nunexplained provider error fails the gate.",{"type":41,"tag":410,"props":773,"children":775},{"id":774},"submission-gate",[776],{"type":47,"value":777},"Submission gate",{"type":41,"tag":50,"props":779,"children":780},{},[781],{"type":47,"value":782},"Before submission:",{"type":41,"tag":417,"props":784,"children":785},{},[786,791,796,809,814,819,824],{"type":41,"tag":85,"props":787,"children":788},{},[789],{"type":47,"value":790},"Re-run validation and planning from the pinned worker revision.",{"type":41,"tag":85,"props":792,"children":793},{},[794],{"type":47,"value":795},"Compare the new plan digest with the approved plan.",{"type":41,"tag":85,"props":797,"children":798},{},[799,801,807],{"type":47,"value":800},"Run ",{"type":41,"tag":56,"props":802,"children":804},{"className":803},[],[805],{"type":47,"value":806},"campaign submit --dry-run",{"type":47,"value":808}," and inspect every remote write.",{"type":41,"tag":85,"props":810,"children":811},{},[812],{"type":47,"value":813},"Confirm Bucket privacy, source lock, bundle upload or reuse action, secret\nnames, namespace, Job image digest, worker revision, Harbor revision, agent\nrevision, provider route, and judge policy. Prove that no Git credential is\nincluded.",{"type":41,"tag":85,"props":815,"children":816},{},[817],{"type":47,"value":818},"Confirm duration arithmetic and budget arithmetic from the same manifest.",{"type":41,"tag":85,"props":820,"children":821},{},[822],{"type":47,"value":823},"Save the approved manifest, plan, duration report, and launch decision.",{"type":41,"tag":85,"props":825,"children":826},{},[827],{"type":47,"value":828},"Submit once and capture the returned campaign ID, controller Job ID, input\ndigest, plan digest, and launch receipt.",{"type":41,"tag":410,"props":830,"children":832},{"id":831},"live-operation",[833],{"type":47,"value":834},"Live operation",{"type":41,"tag":50,"props":836,"children":837},{},[838],{"type":47,"value":839},"The campaign projection and controller status are the primary status surfaces:",{"type":41,"tag":490,"props":841,"children":843},{"className":492,"code":842,"language":494,"meta":495,"style":495},"uv run harbor-hf campaign status CAMPAIGN_ID --namespace NAMESPACE\nuv run harbor-hf campaign reconcile CAMPAIGN_ID --namespace NAMESPACE --dry-run\n",[844],{"type":41,"tag":56,"props":845,"children":846},{"__ignoreMap":495},[847,886],{"type":41,"tag":501,"props":848,"children":849},{"class":503,"line":504},[850,854,858,862,866,871,876,881],{"type":41,"tag":501,"props":851,"children":852},{"style":508},[853],{"type":47,"value":511},{"type":41,"tag":501,"props":855,"children":856},{"style":514},[857],{"type":47,"value":517},{"type":41,"tag":501,"props":859,"children":860},{"style":514},[861],{"type":47,"value":522},{"type":41,"tag":501,"props":863,"children":864},{"style":514},[865],{"type":47,"value":552},{"type":41,"tag":501,"props":867,"children":868},{"style":514},[869],{"type":47,"value":870}," status",{"type":41,"tag":501,"props":872,"children":873},{"style":514},[874],{"type":47,"value":875}," CAMPAIGN_ID",{"type":41,"tag":501,"props":877,"children":878},{"style":514},[879],{"type":47,"value":880}," --namespace",{"type":41,"tag":501,"props":882,"children":883},{"style":514},[884],{"type":47,"value":885}," NAMESPACE\n",{"type":41,"tag":501,"props":887,"children":888},{"class":503,"line":28},[889,893,897,901,905,910,914,918,923],{"type":41,"tag":501,"props":890,"children":891},{"style":508},[892],{"type":47,"value":511},{"type":41,"tag":501,"props":894,"children":895},{"style":514},[896],{"type":47,"value":517},{"type":41,"tag":501,"props":898,"children":899},{"style":514},[900],{"type":47,"value":522},{"type":41,"tag":501,"props":902,"children":903},{"style":514},[904],{"type":47,"value":552},{"type":41,"tag":501,"props":906,"children":907},{"style":514},[908],{"type":47,"value":909}," reconcile",{"type":41,"tag":501,"props":911,"children":912},{"style":514},[913],{"type":47,"value":875},{"type":41,"tag":501,"props":915,"children":916},{"style":514},[917],{"type":47,"value":880},{"type":41,"tag":501,"props":919,"children":920},{"style":514},[921],{"type":47,"value":922}," NAMESPACE",{"type":41,"tag":501,"props":924,"children":925},{"style":514},[926],{"type":47,"value":927}," --dry-run\n",{"type":41,"tag":50,"props":929,"children":930},{},[931],{"type":47,"value":932},"A provider campaign advances inside its detached controller Job. Do not run an\napplied local reconciliation loop. Check the controller attempt, claim,\nheartbeat age, current wave, remaining-time admission, and block reason. HF Job\nstate and logs are supporting evidence. While work runs, confirm that terminal\ntrial bundles and provider records appear in the private Bucket.",{"type":41,"tag":50,"props":934,"children":935},{},[936],{"type":47,"value":937},"Install the shared watchdog only for an explicit campaign list. The watchdog\nmay launch a sequential replacement after a retryable infrastructure failure.\nIt cannot continue a capacity or policy pause.",{"type":41,"tag":50,"props":939,"children":940},{},[941],{"type":47,"value":942},"The controller stops new work when observed throughput makes the locked duration\ninfeasible, provider errors exceed the profile boundary, evidence publication\nstalls, spend admission fails, or ownership becomes uncertain.",{"type":41,"tag":410,"props":944,"children":946},{"id":945},"recovery-gate",[947],{"type":47,"value":948},"Recovery gate",{"type":41,"tag":50,"props":950,"children":951},{},[952],{"type":47,"value":953},"Observe before mutating. Load the campaign projection, control events, HF Job\nstate, wave evidence, trial markers, and checksums. Classify each affected\nlogical trial before requesting a retry.",{"type":41,"tag":50,"props":955,"children":956},{},[957,959,965],{"type":47,"value":958},"Use ",{"type":41,"tag":56,"props":960,"children":962},{"className":961},[],[963],{"type":47,"value":964},"campaign retry",{"type":47,"value":966}," only for trials already classified as retryable\ninfrastructure failures. An immutable manifest change requires a new campaign\nidentity and explicit linkage to the superseded campaign. Build a trial-level\nledger before any replacement so completed logical outcomes are not executed\nagain by accident.",{"type":41,"tag":50,"props":968,"children":969},{},[970,972,978],{"type":47,"value":971},"Never use ",{"type":41,"tag":56,"props":973,"children":975},{"className":974},[],[976],{"type":47,"value":977},"seal",{"type":47,"value":979}," merely to make a campaign terminal. Seal only a drained partial\ncampaign whose exhausted infrastructure outcomes are intentionally accepted as\nzero under the declared publication policy.",{"type":41,"tag":410,"props":981,"children":983},{"id":982},"evidence-and-publication-gate",[984],{"type":47,"value":985},"Evidence and publication gate",{"type":41,"tag":50,"props":987,"children":988},{},[989],{"type":47,"value":990},"Before publication:",{"type":41,"tag":417,"props":992,"children":993},{},[994,999,1004,1009,1014,1019,1024,1029,1034],{"type":41,"tag":85,"props":995,"children":996},{},[997],{"type":47,"value":998},"Reconcile until all Jobs, waves, leases, retries, and cleanup actions are\nterminal.",{"type":41,"tag":85,"props":1000,"children":1001},{},[1002],{"type":47,"value":1003},"Verify endpoint pause evidence when applicable.",{"type":41,"tag":85,"props":1005,"children":1006},{},[1007],{"type":47,"value":1008},"Run remote artifact verification.",{"type":41,"tag":85,"props":1010,"children":1011},{},[1012],{"type":47,"value":1013},"Deep-validate representative and repaired trial bundles locally.",{"type":41,"tag":85,"props":1015,"children":1016},{},[1017],{"type":47,"value":1018},"Verify every declared digest and terminal-marker ordering.",{"type":41,"tag":85,"props":1020,"children":1021},{},[1022],{"type":47,"value":1023},"Scan all retained private and candidate public files for known secret values\nand high-confidence generic credential patterns.",{"type":41,"tag":85,"props":1025,"children":1026},{},[1027],{"type":47,"value":1028},"Confirm trial counts, attempt counts, task digests, selected executions,\nrewards, judge identities, and result classification.",{"type":41,"tag":85,"props":1030,"children":1031},{},[1032],{"type":47,"value":1033},"Dry-run publication, publish once, and record the receipt and exact Dataset\nrevision.",{"type":41,"tag":85,"props":1035,"children":1036},{},[1037],{"type":47,"value":1038},"Change the primary catalog only with explicit authorization and a stated\nreason.",{"type":41,"tag":69,"props":1040,"children":1042},{"id":1041},"hard-stops",[1043],{"type":47,"value":1044},"Hard stops",{"type":41,"tag":50,"props":1046,"children":1047},{},[1048],{"type":47,"value":1049},"Stop without launching or mutating when any of these conditions holds:",{"type":41,"tag":81,"props":1051,"children":1052},{},[1053,1058,1063,1068,1073,1078,1083,1088,1093,1098,1103,1108,1113],{"type":41,"tag":85,"props":1054,"children":1055},{},[1056],{"type":47,"value":1057},"A required commit, digest, task identity, agent revision, or image pin is\nmutable or unknown.",{"type":41,"tag":85,"props":1059,"children":1060},{},[1061],{"type":47,"value":1062},"The task count, attempt count, shard count, or planned wave count differs from\nthe approved protocol.",{"type":41,"tag":85,"props":1064,"children":1065},{},[1066],{"type":47,"value":1067},"No representative timing evidence exists for a paid full campaign.",{"type":41,"tag":85,"props":1069,"children":1070},{},[1071],{"type":47,"value":1072},"The duration checker fails or depends on unexplained concurrency.",{"type":41,"tag":85,"props":1074,"children":1075},{},[1076],{"type":47,"value":1077},"Estimated wave cost is missing, exceeds the campaign cap, or leaves no room\nfor the declared retry policy.",{"type":41,"tag":85,"props":1079,"children":1080},{},[1081],{"type":47,"value":1082},"A provider or endpoint quota is unknown.",{"type":41,"tag":85,"props":1084,"children":1085},{},[1086],{"type":47,"value":1087},"A secret value would enter a command, manifest, lock, log, or agent process.",{"type":41,"tag":85,"props":1089,"children":1090},{},[1091],{"type":47,"value":1092},"A public Git source needs authentication, or a local\u002Fprivate source cannot be\nrepresented by a complete verified bundle.",{"type":41,"tag":85,"props":1094,"children":1095},{},[1096],{"type":47,"value":1097},"A credential would be copied from local configuration into a remote secret\nwithout approval naming the exact source and destination.",{"type":41,"tag":85,"props":1099,"children":1100},{},[1101],{"type":47,"value":1102},"An endpoint is running without a verified owner and watchdog.",{"type":41,"tag":85,"props":1104,"children":1105},{},[1106],{"type":47,"value":1107},"Existing terminal evidence is ambiguous, checksum-invalid, or duplicated.",{"type":41,"tag":85,"props":1109,"children":1110},{},[1111],{"type":47,"value":1112},"A requested retry would rerun an agent or benchmark failure.",{"type":41,"tag":85,"props":1114,"children":1115},{},[1116],{"type":47,"value":1117},"Full publication evidence has not passed checksum and secret validation.",{"type":41,"tag":69,"props":1119,"children":1121},{"id":1120},"final-report",[1122],{"type":47,"value":1123},"Final report",{"type":41,"tag":50,"props":1125,"children":1126},{},[1127],{"type":47,"value":1128},"Report concrete evidence:",{"type":41,"tag":81,"props":1130,"children":1131},{},[1132,1137,1142,1147,1152,1157,1162,1167],{"type":41,"tag":85,"props":1133,"children":1134},{},[1135],{"type":47,"value":1136},"manifest and plan paths plus SHA-256 digests;",{"type":41,"tag":85,"props":1138,"children":1139},{},[1140],{"type":47,"value":1141},"campaign, controller attempt, run, wave, shard, trial, execution, and HF Job IDs;",{"type":41,"tag":85,"props":1143,"children":1144},{},[1145],{"type":47,"value":1146},"requested and observed concurrency, pacing, trial latency, wave duration, and\ndeadline headroom;",{"type":41,"tag":85,"props":1148,"children":1149},{},[1150],{"type":47,"value":1151},"spend cap, reservation, observed or unreported spend, and retry capacity;",{"type":41,"tag":85,"props":1153,"children":1154},{},[1155],{"type":47,"value":1156},"counts by queued, active, complete, retryable, exhausted, failed, cancelled,\nand invalid state;",{"type":41,"tag":85,"props":1158,"children":1159},{},[1160],{"type":47,"value":1161},"endpoint pause state or provider-route closure;",{"type":41,"tag":85,"props":1163,"children":1164},{},[1165],{"type":47,"value":1166},"artifact verification, deep validation, and secret-scan results;",{"type":41,"tag":85,"props":1168,"children":1169},{},[1170],{"type":47,"value":1171},"publication ID, Dataset revision, catalog state, and any remaining blocker.",{"type":41,"tag":50,"props":1173,"children":1174},{},[1175],{"type":47,"value":1176},"Do not call a campaign successful because its Job exited zero. Success comes\nfrom the durable campaign projection and verified canonical evidence.",{"type":41,"tag":1178,"props":1179,"children":1180},"style",{},[1181],{"type":47,"value":1182},"html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"items":1184,"total":504},[1185],{"slug":4,"name":4,"fn":5,"description":6,"org":1186,"tags":1187,"stars":24,"repoUrl":25,"updatedAt":26},{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1188,1189,1190,1191],{"name":16,"slug":17,"type":14},{"name":22,"slug":23,"type":14},{"name":9,"slug":13,"type":14},{"name":19,"slug":20,"type":14},{"items":1193,"total":1366},[1194,1216,1232,1249,1263,1276,1287,1302,1314,1324,1337,1351],{"slug":1195,"name":1195,"fn":1196,"description":1197,"org":1198,"tags":1199,"stars":1213,"repoUrl":1214,"updatedAt":1215},"train-sentence-transformers","train sentence-transformers models","Train or fine-tune sentence-transformers models across `SentenceTransformer` (bi-encoder, dense or static embedding model for retrieval, similarity, clustering, classification, paraphrase mining, dedup, multimodal), `CrossEncoder` (reranker, pair scoring for two-stage retrieval \u002F pair classification), `SparseEncoder` (SPLADE, sparse embedding model for learned-sparse retrieval), and `MultiVectorEncoder` (ColBERT \u002F late-interaction, per-token embeddings scored with MaxSim). Covers loss selection, hard-negative mining, evaluators, distillation, LoRA, Matryoshka, and Hugging Face Hub publishing. Use for any sentence-transformers training task.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1200,1203,1204,1207,1210],{"name":1201,"slug":1202,"type":14},"Deep Learning","deep-learning",{"name":9,"slug":13,"type":14},{"name":1205,"slug":1206,"type":14},"LLM","llm",{"name":1208,"slug":1209,"type":14},"Python","python",{"name":1211,"slug":1212,"type":14},"Search","search",18914,"https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fsentence-transformers","2026-08-13T04:29:38.482788",{"slug":1217,"name":1217,"fn":1218,"description":1219,"org":1220,"tags":1221,"stars":1229,"repoUrl":1230,"updatedAt":1231},"trl-training","train and fine-tune LLMs with TRL","Train and fine-tune transformer language models using TRL (Transformers Reinforcement Learning). Supports SFT, DPO, GRPO, KTO, RLOO and Reward Model training via CLI commands.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1222,1225,1226,1227,1228],{"name":1223,"slug":1224,"type":14},"AI Infrastructure","ai-infrastructure",{"name":1201,"slug":1202,"type":14},{"name":9,"slug":13,"type":14},{"name":1205,"slug":1206,"type":14},{"name":1208,"slug":1209,"type":14},18850,"https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftrl","2026-04-06T18:25:32.746828",{"slug":1233,"name":1233,"fn":1234,"description":1235,"org":1236,"tags":1237,"stars":1246,"repoUrl":1247,"updatedAt":1248},"hf-cli","manage Hugging Face Hub resources via CLI","Hugging Face Hub CLI (`hf`) for downloading, uploading, and managing models, datasets, spaces, buckets, repos, papers, jobs, and more on the Hugging Face Hub. Use when: handling authentication; managing local cache; managing Hugging Face Buckets; running or scheduling jobs on Hugging Face infrastructure; managing Hugging Face repos; discussions and pull requests; browsing models, datasets and spaces; reading, searching, or browsing academic papers; managing collections; querying datasets; configuring spaces; setting up webhooks; or deploying and managing HF Inference Endpoints. Make sure to use this skill whenever the user mentions 'hf', 'huggingface', 'Hugging Face', 'huggingface-cli', or 'hugging face cli', or wants to do anything related to the Hugging Face ecosystem and to AI and ML in general. Also use for cloud storage needs like training checkpoints, data pipelines, or agent traces. Use even if the user doesn't explicitly ask for a CLI command. Replaces the deprecated `huggingface-cli`.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1238,1241,1244,1245],{"name":1239,"slug":1240,"type":14},"CLI","cli",{"name":1242,"slug":1243,"type":14},"Datasets","datasets",{"name":9,"slug":13,"type":14},{"name":1205,"slug":1206,"type":14},10861,"https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fskills","2026-08-09T04:08:11.055248",{"slug":1250,"name":1250,"fn":1251,"description":1252,"org":1253,"tags":1254,"stars":1246,"repoUrl":1247,"updatedAt":1262},"hf-cloud-aws-context-discovery","discover local AWS environment context","Discover the user's local AWS context (active profile, region, account ID, caller identity) at the start of any AWS task. Use this skill before any other AWS work — deploying to SageMaker, creating resources, calling AWS APIs, or anything that touches an AWS account. Use it especially when the user has not specified a region or profile explicitly, when they say things like \"use my AWS account\", \"deploy to AWS\", \"use my profile\", or when about to make any AWS CLI or SDK call. Never guess the region or account ID — always use this skill to read it from the local configuration first.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1255,1258,1259],{"name":1256,"slug":1257,"type":14},"AWS","aws",{"name":1239,"slug":1240,"type":14},{"name":1260,"slug":1261,"type":14},"Configuration","configuration","2026-07-08T05:55:33.716099",{"slug":1264,"name":1264,"fn":1265,"description":1266,"org":1267,"tags":1268,"stars":1246,"repoUrl":1247,"updatedAt":1275},"hf-cloud-python-env-setup","set up Python environments for AWS","Set up an isolated Python environment for SageMaker \u002F AWS work, with the right Python version and current boto3. Use this skill whenever Python code will be executed for a SageMaker deployment, training job, or any AWS automation — including when about to run `pip install`, when about to invoke `boto3`, when creating or activating a virtualenv, or when the user asks to \"set up the environment\". Never use system Python and never `pip install` into it. Always isolate. This skill prevents the most common failure modes: wrong Python version, dependency conflicts, and stale SDKs.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1269,1270,1271,1274],{"name":1223,"slug":1224,"type":14},{"name":1256,"slug":1257,"type":14},{"name":1272,"slug":1273,"type":14},"Engineering","engineering",{"name":1208,"slug":1209,"type":14},"2026-07-08T05:55:32.505017",{"slug":1277,"name":1277,"fn":1278,"description":1279,"org":1280,"tags":1281,"stars":1246,"repoUrl":1247,"updatedAt":1286},"hf-cloud-sagemaker-deployment-planner","plan model deployments to Amazon SageMaker","Plan and coordinate the deployment of a model to Amazon SageMaker AI. Use this skill whenever the user wants to deploy, host, serve, or expose a model on SageMaker or AWS — including phrases like \"deploy a model\", \"host this LLM on AWS\", \"serve this embedding model\", \"deploy a reranker\", \"deploy a text-to-image \u002F diffusion model\", \"host this for async inference\", \"create an endpoint\", \"serve my fine-tuned model\", or any request that involves making a model available for inference on AWS. Use this even when the user is vague (e.g. \"I just want to get this running on AWS, you figure it out\"). Works for text-generation LLMs, embedding models, rerankers, classifiers, text-to-image \u002F diffusion models — picks the right serving stack and chooses between real-time and async inference. This is the entry-point skill for SageMaker deployment work — it asks clarifying questions, picks a deployment pathway, and coordinates the other deployment skills.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1282,1283,1284,1285],{"name":1223,"slug":1224,"type":14},{"name":1256,"slug":1257,"type":14},{"name":22,"slug":23,"type":14},{"name":9,"slug":13,"type":14},"2026-08-06T05:39:28.520455",{"slug":1288,"name":1288,"fn":1289,"description":1290,"org":1291,"tags":1292,"stars":1246,"repoUrl":1247,"updatedAt":1301},"hf-cloud-sagemaker-iam-preflight","configure SageMaker IAM roles","Ensure a usable SageMaker execution role exists before deploying or training. Use this skill whenever about to create a SageMaker endpoint, model, training job, or any resource that requires an execution role. Use it especially when the user has not provided a role ARN explicitly, when scripts are about to call `iam:CreateRole`, or when an AccessDenied error mentions an IAM action. Never blindly call `iam:CreateRole` — always check for existing roles first. This skill prevents the most common SageMaker deployment failure: trying to create IAM resources from an SSO principal that has no IAM write permissions.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1293,1294,1295,1298],{"name":1223,"slug":1224,"type":14},{"name":1256,"slug":1257,"type":14},{"name":1296,"slug":1297,"type":14},"Permissions","permissions",{"name":1299,"slug":1300,"type":14},"Security","security","2026-07-08T05:55:34.948771",{"slug":1303,"name":1303,"fn":1304,"description":1305,"org":1306,"tags":1307,"stars":1246,"repoUrl":1247,"updatedAt":1313},"hf-cloud-sagemaker-production-defaults","create production-ready SageMaker endpoints","Create a SageMaker endpoint (real-time, real-time scale-to-zero, or async) with autoscaling, CloudWatch alarms, and tagging enabled by default. Use this skill whenever about to create a SageMaker endpoint, write deployment code that calls `create_endpoint`, or finalize a deployment after the image URI and IAM role are known. Provides deploy.py for real-time endpoints, deploy_ic.py for real-time endpoints that scale to zero instances via inference components, and deploy_async.py for async endpoints (also scale-to-zero). This is the last step in the SageMaker deployment workflow. Never generate a bare `create_endpoint` call without these defaults — endpoints without autoscaling or alarms are demos, not deployments.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1308,1309,1310,1311,1312],{"name":1223,"slug":1224,"type":14},{"name":1256,"slug":1257,"type":14},{"name":22,"slug":23,"type":14},{"name":9,"slug":13,"type":14},{"name":19,"slug":20,"type":14},"2026-08-06T05:39:26.526636",{"slug":1315,"name":1315,"fn":1316,"description":1317,"org":1318,"tags":1319,"stars":1246,"repoUrl":1247,"updatedAt":1323},"hf-cloud-serving-image-selection","select SageMaker serving containers","Pick the right serving container for a SageMaker model deployment and find its current image URI. Use this skill whenever about to deploy a model to a SageMaker endpoint and an image URI needs to be chosen — including when the user says \"deploy this LLM\", \"host this HuggingFace model\", \"serve this fine-tuned model\", \"deploy this embedding model\", \"host a reranker\", \"serve a sentence-transformers model\", or when about to hardcode any container URI in deployment code. HuggingFace-curated Deep Learning Containers are ALWAYS preferred: HuggingFace vLLM (LLMs and generative rerankers), HuggingFace vLLM-Omni (multimodal), TEI (embeddings\u002Fcross-encoder rerankers), HF Inference Toolkit (other transformers). Generic images (AWS vLLM, DJL-LMI, SGLang) are used only when no HuggingFace image is compatible — never merely because they carry a newer version. Never hardcode a container URI from memory and never default to TGI. Prevents stale-image failures and wrong-region URIs.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1320,1321,1322],{"name":1223,"slug":1224,"type":14},{"name":1256,"slug":1257,"type":14},{"name":22,"slug":23,"type":14},"2026-07-08T05:55:36.173465",{"slug":1325,"name":1325,"fn":1326,"description":1327,"org":1328,"tags":1329,"stars":1246,"repoUrl":1247,"updatedAt":1336},"hf-mcp","access Hugging Face Hub via MCP","Use Hugging Face Hub via MCP server tools. Search models, datasets, Spaces, papers. Get repo details, fetch documentation, run compute jobs, and use Gradio Spaces as AI tools. Available when connected to the HF MCP server.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1330,1331,1332,1333],{"name":1242,"slug":1243,"type":14},{"name":9,"slug":13,"type":14},{"name":1205,"slug":1206,"type":14},{"name":1334,"slug":1335,"type":14},"MCP","mcp","2026-04-06T18:25:50.364185",{"slug":1338,"name":1338,"fn":1339,"description":1340,"org":1341,"tags":1342,"stars":1246,"repoUrl":1247,"updatedAt":1350},"hf-mem","estimate memory for Hugging Face models","Hugging Face CLI to estimate the required memory to load Safetensors or GGUF model weights for inference from the Hugging Face Hub",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1343,1344,1345,1346,1347],{"name":1223,"slug":1224,"type":14},{"name":1239,"slug":1240,"type":14},{"name":9,"slug":13,"type":14},{"name":1205,"slug":1206,"type":14},{"name":1348,"slug":1349,"type":14},"Performance","performance","2026-06-13T07:23:57.101435",{"slug":1352,"name":1352,"fn":1353,"description":1354,"org":1355,"tags":1356,"stars":1246,"repoUrl":1247,"updatedAt":1365},"huggingface-best","find and compare Hugging Face models","Use when the user asks about finding the best, top, or recommended model for a task, wants to know what AI model to use, or wants to compare models by benchmark scores. Triggers on: \"best model for X\", \"what model should I use for\", \"top models for [task]\", \"which model runs on my laptop\u002Fmachine\u002Fdevice\", \"recommend a model for\", \"what LLM should I use for\", \"compare models for\", \"what's state of the art for\", or any question about choosing an AI model for a specific use case. Always use this skill when the user wants model recommendations or comparisons, even if they don't explicitly mention HuggingFace or benchmarks.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":8},[1357,1360,1361,1362],{"name":1358,"slug":1359,"type":14},"Analytics","analytics",{"name":9,"slug":13,"type":14},{"name":1205,"slug":1206,"type":14},{"name":1363,"slug":1364,"type":14},"Research","research","2026-08-06T05:39:25.529773",38]