[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6WYFBP1KYNGXK2QKA1B9PR5":3,"skill-same:01M6WYFBP1KYNGXK2QKA1B9PR5":79},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":22,"publishedAt":43,"linkOnly":44,"markdownContent":45,"markdownBody":46,"frontMatter":47,"refName":59,"commitSha":60,"skillDirectory":61,"forksCount":13,"repositoryHtmlUrl":62,"repositoryHomepageUrl":6,"ownerHtmlUrl":63,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":64,"install":65,"resources":78},"01M6WYFBP1KYNGXK2QKA1B9PR5","ollama-remote-verification",null,"Verify a local or remote Ollama instance, enumerate models, and run a smoke test safely over SSH. Includes a reliable pattern for remote checks on Hetzner\u002FVPS hosts and avoids brittle nested stdin parsing.","cryptopafi\u002Fnexusos-skills","cryptopafi","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F177822724?v=4","01M6RBPGFV10VSR55V7A819TGZ","01M6RBPGKCQAZH5QFGX99QW128",0,"MIT","https:\u002F\u002Fgithub.com\u002Fcryptopafi\u002Fnexusos-skills\u002Fblob\u002Fmain\u002Fmlops\u002Follama-remote-verification\u002FSKILL.md",[17],{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000008","测试与质量","SCENARIO","使用场景",[23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42],"ollama remote verification","verification","ollama","remote","models","for","hosts","vps","local","pattern","hetzner","nested","test","smoke test","smoke","hermes","verify","parsing","ssh","run","2026-11-10 10:12:30",false,"---\nname: ollama-remote-verification\ndescription: Verify a local or remote Ollama instance, enumerate models, and run a smoke test safely over SSH. Includes a reliable pattern for remote checks on Hetzner\u002FVPS hosts and avoids brittle nested stdin parsing.\nversion: 1.0.0\nauthor: Hermes Agent\nlicense: MIT\nmetadata:\n  hermes:\n    tags: [Ollama, SSH, Verification, Hetzner, VPS, Local-LLM]\n---\n\n# Ollama Remote Verification\n\nUse when you need to confirm that an Ollama instance is actually live on a local machine or remote VPS and that at least one model can answer a real prompt.\n\n## When to use\n- User says Ollama is installed on a server and wants live verification\n- You need fallback-model readiness before switching away from hosted models\n- You need a quick inventory of remote models and service health\n\n## Procedure\n\n### 1) Verify connectivity to the host\nUse SSH first. Do not assume the Ollama port is exposed publicly.\n\n```bash\nssh -o BatchMode=yes -o StrictHostKeyChecking=accept-new -o ConnectTimeout=5 user@host 'echo CONNECTED && hostname && command -v ollama || true'\n```\n\nIf this works, prefer querying Ollama through localhost on the remote machine.\n\n### 2) Query the remote Ollama API through SSH\nUse localhost on the remote host:\n\n```bash\nssh -o BatchMode=yes user@host 'curl -i --max-time 5 http:\u002F\u002F127.0.0.1:11434\u002Fapi\u002Ftags'\n```\n\nExpected:\n- HTTP\u002F1.1 200 OK\n- JSON body with `models`\n\n### 3) If parsing is needed, inspect raw output first\nIf a JSON parser fails, do not assume Ollama is broken. First print raw output and check for:\n- empty body\n- shell quoting issues\n- mixed stdout\u002Fstderr\n- broken pipeline input\n\nAlso collect process\u002Fservice state:\n\n```bash\nssh user@host '\nps aux | grep -i \"[o]llama\" || true\n(systemctl --no-pager --full status ollama 2>&1 | sed -n \"1,40p\") || true\n'\n```\n\n### 4) Smoke test a real model\nPrefer a minimal deterministic prompt:\n\n```bash\nssh -o BatchMode=yes user@host \"curl -s --max-time 180 http:\u002F\u002F127.0.0.1:11434\u002Fapi\u002Fgenerate -H 'Content-Type: application\u002Fjson' -d '{\\\"model\\\":\\\"qwen3-coder:30b\\\",\\\"prompt\\\":\\\"Reply with exactly: OLLAMA_OK\\\",\\\"stream\\\":false}'\"\n```\n\nSuccess criteria:\n- JSON response returns\n- `response` matches expected text\n- `done: true`\n\n### 5) Inventory useful model tiers\nWhen reporting, group models by role:\n- tiny\u002Ffast gate model\n- general workhorse\n- coding model\n- heavy strategic\u002Freasoning model\n- embeddings model\n\n## Important experiential findings\n\n### Avoid brittle nested stdin parsing over SSH\nA failed pattern encountered in practice:\n- generating JSON with a here-doc Python block\n- piping into curl\n- then piping into another here-doc Python block\n\nThis can lead to empty stdin \u002F JSON parse failures even though Ollama itself is healthy.\n\nIf that happens, simplify immediately:\n- first print raw `curl` output\n- then use a direct one-line JSON payload for the smoke test\n- only add parsing after raw output is confirmed\n\n### Prefer localhost queries on the remote host\nEven if direct `http:\u002F\u002Fhost:11434` checks fail externally, Ollama may be healthy and intentionally bound to localhost. SSH + localhost curl is the reliable verification path.\n\n## Reporting template\n- SSH access: working \u002F failed\n- Hostname\n- Ollama binary path\n- Service state\n- Model count\n- Notable models by tier\n- Smoke test result\n- Recommended fallback routing\n\n## Benchmarking for fallback selection\nUse this when the user does not just want to know whether Ollama works, but which model is safe to rely on when premium credits are exhausted.\n\n### Goal\nChoose the best local\u002Fcloud Ollama fallback without sacrificing conversation quality or waiting unreasonably long for responses.\n\n### Recommended benchmark sequence\nTest from the strongest candidate down to smaller models. Example order:\n- `hermes3:70b`\n- other 70B \u002F premium local models\n- strong coding \u002F reasoning mid-tier models like `qwen3-coder:30b`\n- general workhorses like `gemma4:26b` \u002F `gemma3:27b`\n- small utility models like `llama3.1:8b`\n\n### Two-stage benchmark\n\n#### 1) Reachability + latency probe\nRun a minimal deterministic prompt first:\n\n```bash\nssh -o BatchMode=yes user@host \"curl -s --max-time 180 http:\u002F\u002F127.0.0.1:11434\u002Fapi\u002Fgenerate -H 'Content-Type: application\u002Fjson' -d '{\\\"model\\\":\\\"MODEL_NAME\\\",\\\"prompt\\\":\\\"Reply with exactly: READY\\\",\\\"stream\\\":false}'\"\n```\n\nCapture:\n- total duration\n- whether the instruction was followed exactly\n- whether the model timed out or stalled\n\n#### 2) Quality-under-constraint probe\nThen test a realistic operator prompt with strict formatting, multiple constraints, and a role-relevant task.\n\nGood pattern:\n- require a short answer\n- require exact sections or bullets\n- require tone and audience constraints\n- use a task similar to the actual production use case (chief-of-staff, planning, synthesis, crisis memo, etc.)\n\nExample criteria to evaluate:\n- follows required structure exactly\n- stays within length limits\n- does not invent nonexistent models \u002F entities \u002F assumptions\n- remains concrete, calm, and operational\n- gives usable output without excess fluff\n\n### Decision rule\nRank by:\n1. quality \u002F instruction discipline\n2. latency\n3. factual restraint\n\nReject models that:\n- hallucinate important details\n- miss format constraints repeatedly\n- exceed acceptable conversational latency even if output quality is decent\n\n### Practical finding from live testing\nA model can be “good” in absolute quality but still be a bad fallback if it takes 2-4+ minutes for a short structured response.\n\nIn one live Hetzner benchmark, this pattern appeared:\n- `qwen3-coder:30b` = strong practical fallback: usable structure + acceptable latency\n- `gemma4:26b` = decent quality but too slow for fluid conversation\n- `hermes3:70b` = coherent but too slow for chief-of-staff live dialogue\n- `llama3.1:8b` = fast but too weak \u002F too hallucination-prone for strategic continuity\n\nAdditional live finding for coding workloads on the same Hetzner host:\n- `qwen3-coder-next:latest` (79.7B Q4_K_M, ~51.7 GB on disk) performed surprisingly well as a local coding candidate on a 251 GiB RAM Hetzner machine.\n- Example timings from live coding probes on the same-host Ollama setup:\n  - codegen prompt: ~11.5s total\n  - Python debugging prompt: ~28.8s total\n- `qwen3-coder:30b` remained faster on some small tasks (~6s codegen) but should still be compared against `qwen3-coder-next:latest` on debugging and instruction-following before final promotion.\n\nDo not hardcode these conclusions globally. Re-test on the actual machine and current load, but use them as a strong prior.\n\n## Coding-specific benchmark extension\nUse this when the user wants the best local coding fallback, not just general conversation fallback.\n\n### Scope\nBenchmark only the models that are local to the execution domain you are testing. If Hermes and Ollama are on the same Hetzner host, do NOT mix MacM4 coding agents into the local-Ollama benchmark.\n\n### Capacity check before testing\nRun these first on the target host:\n\n```bash\nfree -h\nswapon --show || true\nollama ps || true\nps -o pid,ppid,%mem,rss,vsz,etime,cmd -C ollama || true\ncurl -s http:\u002F\u002F127.0.0.1:11434\u002Fapi\u002Ftags\n```\n\nInterpretation:\n- If available RAM is very high (e.g. >200 GiB available on a 251 GiB host), 70B-80B Q4 coding models are realistic to test.\n- Use `ollama ps` to see which model is currently resident and its memory footprint.\n- Same-host Hermes→Ollama testing removes most network noise, so measured latency is mostly inference latency.\n\n### Coding benchmark suites\nAt minimum run two suites per model:\n\n#### 1) Codegen probe\nPrompt pattern:\n- require code only\n- require a tiny but complete script\n- require stdin \u002F JSON \u002F fallback behavior\n- no prose allowed\n\nGood example:\n- “Return only Python code. Write a small script that reads JSON from stdin and prints the value of key `status`. If missing, print `UNKNOWN`. No explanations.”\n\n#### 2) Debugging probe\nPrompt pattern:\n- real traceback or bug pattern\n- strict length cap\n- exact response structure\n- require root cause, minimal fix, verification\n\nGood example:\n- “Give exactly 3 bullets: Root cause, Minimal fix, Verification.”\n\n### Coding benchmark decision rule\nRank by:\n1. debugging quality\n2. instruction-following discipline\n3. code validity \u002F minimalism\n4. latency\n\nReject models that:\n- hallucinate APIs, files, or nonexistent context\n- ignore code-only constraints\n- cannot identify the real root cause in simple debugging probes\n- are too slow for live iterative coding even if the answer is decent\n\n### Shell \u002F pipeline pitfalls found in practice\nWhen building remote benchmark scripts:\n- quote `-H 'Content-Type: application\u002Fjson'` correctly; dropping the quotes can break the curl invocation\n- if you pipe Python JSON generation into curl, the pipeline exit code may reflect the producer side and show non-zero RC values even when Ollama returned a valid JSON response\n- do not treat an `RC:6` style shell artifact as immediate model failure without inspecting the actual Ollama output\n- use `set -uo pipefail` with care; for benchmark logging scripts, prefer explicit response capture and explicit error markers over brittle shell pipeline assumptions\n- for multi-line Python benchmark scripts sent over SSH, nested here-doc quoting is fragile and can silently corrupt the script; a safer pattern is to base64-encode the script locally, decode it remotely into `\u002Ftmp\u002F\u003Cname>.py`, then launch it with `nohup python3 ... &`\n- immediately verify launch by checking all three: the script file exists, the log file exists and has a start marker, and `ps` shows the Python process still running\n\n### Practical capacity rule: \"fits\" is not the same as \"usable\"\nOn large-RAM Ollama hosts, separate two questions:\n1. Can the model be loaded?\n2. Is it fast enough for the actual workflow?\n\nLive finding on a ~251 GiB Hetzner host:\n- 70B–80B Q4 models fit comfortably in RAM\n- but `llama3.3:70b`, `gemma4:26b`, and `hermes3:70b` were still too slow for live coding iteration despite fitting\n- therefore the practical max coding size should be chosen by latency-under-debugging-tests, not by memory headroom alone\n\nUse this decision rule when reporting capacity:\n- `maximum possible size` = what the host can load without memory pressure\n- `maximum practical coding size` = largest model that still gives acceptable iterative latency on codegen + debugging probes\n\nDo not report only RAM fit; report both thresholds explicitly.\n\n## Example summary\n- SSH access works to `user@host`\n- Ollama active on `127.0.0.1:11434`\n- `\u002Fapi\u002Ftags` returns model inventory\n- Smoke test passed on `qwen3-coder:30b`\n- Use workhorse + coding + heavy model triad for fallback planning\n- If credits are nearly exhausted, benchmark live before promoting any Ollama model to primary fallback\n\n## Do not\n- Do not rely on public port exposure as the first verification method\n- Do not assume JSON parse failure means Ollama is down\n- Do not close the task without a real model smoke test\n- Do not promote a model to conversation fallback based only on \"it answered once\"; benchmark structure + latency + hallucination risk","# Ollama Remote Verification\n\nUse when you need to confirm that an Ollama instance is actually live on a local machine or remote VPS and that at least one model can answer a real prompt.\n\n## When to use\n- User says Ollama is installed on a server and wants live verification\n- You need fallback-model readiness before switching away from hosted models\n- You need a quick inventory of remote models and service health\n\n## Procedure\n\n### 1) Verify connectivity to the host\nUse SSH first. Do not assume the Ollama port is exposed publicly.\n\n```bash\nssh -o BatchMode=yes -o StrictHostKeyChecking=accept-new -o ConnectTimeout=5 user@host 'echo CONNECTED && hostname && command -v ollama || true'\n```\n\nIf this works, prefer querying Ollama through localhost on the remote machine.\n\n### 2) Query the remote Ollama API through SSH\nUse localhost on the remote host:\n\n```bash\nssh -o BatchMode=yes user@host 'curl -i --max-time 5 http:\u002F\u002F127.0.0.1:11434\u002Fapi\u002Ftags'\n```\n\nExpected:\n- HTTP\u002F1.1 200 OK\n- JSON body with `models`\n\n### 3) If parsing is needed, inspect raw output first\nIf a JSON parser fails, do not assume Ollama is broken. First print raw output and check for:\n- empty body\n- shell quoting issues\n- mixed stdout\u002Fstderr\n- broken pipeline input\n\nAlso collect process\u002Fservice state:\n\n```bash\nssh user@host '\nps aux | grep -i \"[o]llama\" || true\n(systemctl --no-pager --full status ollama 2>&1 | sed -n \"1,40p\") || true\n'\n```\n\n### 4) Smoke test a real model\nPrefer a minimal deterministic prompt:\n\n```bash\nssh -o BatchMode=yes user@host \"curl -s --max-time 180 http:\u002F\u002F127.0.0.1:11434\u002Fapi\u002Fgenerate -H 'Content-Type: application\u002Fjson' -d '{\\\"model\\\":\\\"qwen3-coder:30b\\\",\\\"prompt\\\":\\\"Reply with exactly: OLLAMA_OK\\\",\\\"stream\\\":false}'\"\n```\n\nSuccess criteria:\n- JSON response returns\n- `response` matches expected text\n- `done: true`\n\n### 5) Inventory useful model tiers\nWhen reporting, group models by role:\n- tiny\u002Ffast gate model\n- general workhorse\n- coding model\n- heavy strategic\u002Freasoning model\n- embeddings model\n\n## Important experiential findings\n\n### Avoid brittle nested stdin parsing over SSH\nA failed pattern encountered in practice:\n- generating JSON with a here-doc Python block\n- piping into curl\n- then piping into another here-doc Python block\n\nThis can lead to empty stdin \u002F JSON parse failures even though Ollama itself is healthy.\n\nIf that happens, simplify immediately:\n- first print raw `curl` output\n- then use a direct one-line JSON payload for the smoke test\n- only add parsing after raw output is confirmed\n\n### Prefer localhost queries on the remote host\nEven if direct `http:\u002F\u002Fhost:11434` checks fail externally, Ollama may be healthy and intentionally bound to localhost. SSH + localhost curl is the reliable verification path.\n\n## Reporting template\n- SSH access: working \u002F failed\n- Hostname\n- Ollama binary path\n- Service state\n- Model count\n- Notable models by tier\n- Smoke test result\n- Recommended fallback routing\n\n## Benchmarking for fallback selection\nUse this when the user does not just want to know whether Ollama works, but which model is safe to rely on when premium credits are exhausted.\n\n### Goal\nChoose the best local\u002Fcloud Ollama fallback without sacrificing conversation quality or waiting unreasonably long for responses.\n\n### Recommended benchmark sequence\nTest from the strongest candidate down to smaller models. Example order:\n- `hermes3:70b`\n- other 70B \u002F premium local models\n- strong coding \u002F reasoning mid-tier models like `qwen3-coder:30b`\n- general workhorses like `gemma4:26b` \u002F `gemma3:27b`\n- small utility models like `llama3.1:8b`\n\n### Two-stage benchmark\n\n#### 1) Reachability + latency probe\nRun a minimal deterministic prompt first:\n\n```bash\nssh -o BatchMode=yes user@host \"curl -s --max-time 180 http:\u002F\u002F127.0.0.1:11434\u002Fapi\u002Fgenerate -H 'Content-Type: application\u002Fjson' -d '{\\\"model\\\":\\\"MODEL_NAME\\\",\\\"prompt\\\":\\\"Reply with exactly: READY\\\",\\\"stream\\\":false}'\"\n```\n\nCapture:\n- total duration\n- whether the instruction was followed exactly\n- whether the model timed out or stalled\n\n#### 2) Quality-under-constraint probe\nThen test a realistic operator prompt with strict formatting, multiple constraints, and a role-relevant task.\n\nGood pattern:\n- require a short answer\n- require exact sections or bullets\n- require tone and audience constraints\n- use a task similar to the actual production use case (chief-of-staff, planning, synthesis, crisis memo, etc.)\n\nExample criteria to evaluate:\n- follows required structure exactly\n- stays within length limits\n- does not invent nonexistent models \u002F entities \u002F assumptions\n- remains concrete, calm, and operational\n- gives usable output without excess fluff\n\n### Decision rule\nRank by:\n1. quality \u002F instruction discipline\n2. latency\n3. factual restraint\n\nReject models that:\n- hallucinate important details\n- miss format constraints repeatedly\n- exceed acceptable conversational latency even if output quality is decent\n\n### Practical finding from live testing\nA model can be “good” in absolute quality but still be a bad fallback if it takes 2-4+ minutes for a short structured response.\n\nIn one live Hetzner benchmark, this pattern appeared:\n- `qwen3-coder:30b` = strong practical fallback: usable structure + acceptable latency\n- `gemma4:26b` = decent quality but too slow for fluid conversation\n- `hermes3:70b` = coherent but too slow for chief-of-staff live dialogue\n- `llama3.1:8b` = fast but too weak \u002F too hallucination-prone for strategic continuity\n\nAdditional live finding for coding workloads on the same Hetzner host:\n- `qwen3-coder-next:latest` (79.7B Q4_K_M, ~51.7 GB on disk) performed surprisingly well as a local coding candidate on a 251 GiB RAM Hetzner machine.\n- Example timings from live coding probes on the same-host Ollama setup:\n  - codegen prompt: ~11.5s total\n  - Python debugging prompt: ~28.8s total\n- `qwen3-coder:30b` remained faster on some small tasks (~6s codegen) but should still be compared against `qwen3-coder-next:latest` on debugging and instruction-following before final promotion.\n\nDo not hardcode these conclusions globally. Re-test on the actual machine and current load, but use them as a strong prior.\n\n## Coding-specific benchmark extension\nUse this when the user wants the best local coding fallback, not just general conversation fallback.\n\n### Scope\nBenchmark only the models that are local to the execution domain you are testing. If Hermes and Ollama are on the same Hetzner host, do NOT mix MacM4 coding agents into the local-Ollama benchmark.\n\n### Capacity check before testing\nRun these first on the target host:\n\n```bash\nfree -h\nswapon --show || true\nollama ps || true\nps -o pid,ppid,%mem,rss,vsz,etime,cmd -C ollama || true\ncurl -s http:\u002F\u002F127.0.0.1:11434\u002Fapi\u002Ftags\n```\n\nInterpretation:\n- If available RAM is very high (e.g. >200 GiB available on a 251 GiB host), 70B-80B Q4 coding models are realistic to test.\n- Use `ollama ps` to see which model is currently resident and its memory footprint.\n- Same-host Hermes→Ollama testing removes most network noise, so measured latency is mostly inference latency.\n\n### Coding benchmark suites\nAt minimum run two suites per model:\n\n#### 1) Codegen probe\nPrompt pattern:\n- require code only\n- require a tiny but complete script\n- require stdin \u002F JSON \u002F fallback behavior\n- no prose allowed\n\nGood example:\n- “Return only Python code. Write a small script that reads JSON from stdin and prints the value of key `status`. If missing, print `UNKNOWN`. No explanations.”\n\n#### 2) Debugging probe\nPrompt pattern:\n- real traceback or bug pattern\n- strict length cap\n- exact response structure\n- require root cause, minimal fix, verification\n\nGood example:\n- “Give exactly 3 bullets: Root cause, Minimal fix, Verification.”\n\n### Coding benchmark decision rule\nRank by:\n1. debugging quality\n2. instruction-following discipline\n3. code validity \u002F minimalism\n4. latency\n\nReject models that:\n- hallucinate APIs, files, or nonexistent context\n- ignore code-only constraints\n- cannot identify the real root cause in simple debugging probes\n- are too slow for live iterative coding even if the answer is decent\n\n### Shell \u002F pipeline pitfalls found in practice\nWhen building remote benchmark scripts:\n- quote `-H 'Content-Type: application\u002Fjson'` correctly; dropping the quotes can break the curl invocation\n- if you pipe Python JSON generation into curl, the pipeline exit code may reflect the producer side and show non-zero RC values even when Ollama returned a valid JSON response\n- do not treat an `RC:6` style shell artifact as immediate model failure without inspecting the actual Ollama output\n- use `set -uo pipefail` with care; for benchmark logging scripts, prefer explicit response capture and explicit error markers over brittle shell pipeline assumptions\n- for multi-line Python benchmark scripts sent over SSH, nested here-doc quoting is fragile and can silently corrupt the script; a safer pattern is to base64-encode the script locally, decode it remotely into `\u002Ftmp\u002F\u003Cname>.py`, then launch it with `nohup python3 ... &`\n- immediately verify launch by checking all three: the script file exists, the log file exists and has a start marker, and `ps` shows the Python process still running\n\n### Practical capacity rule: \"fits\" is not the same as \"usable\"\nOn large-RAM Ollama hosts, separate two questions:\n1. Can the model be loaded?\n2. Is it fast enough for the actual workflow?\n\nLive finding on a ~251 GiB Hetzner host:\n- 70B–80B Q4 models fit comfortably in RAM\n- but `llama3.3:70b`, `gemma4:26b`, and `hermes3:70b` were still too slow for live coding iteration despite fitting\n- therefore the practical max coding size should be chosen by latency-under-debugging-tests, not by memory headroom alone\n\nUse this decision rule when reporting capacity:\n- `maximum possible size` = what the host can load without memory pressure\n- `maximum practical coding size` = largest model that still gives acceptable iterative latency on codegen + debugging probes\n\nDo not report only RAM fit; report both thresholds explicitly.\n\n## Example summary\n- SSH access works to `user@host`\n- Ollama active on `127.0.0.1:11434`\n- `\u002Fapi\u002Ftags` returns model inventory\n- Smoke test passed on `qwen3-coder:30b`\n- Use workhorse + coding + heavy model triad for fallback planning\n- If credits are nearly exhausted, benchmark live before promoting any Ollama model to primary fallback\n\n## Do not\n- Do not rely on public port exposure as the first verification method\n- Do not assume JSON parse failure means Ollama is down\n- Do not close the task without a real model smoke test\n- Do not promote a model to conversation fallback based only on \"it answered once\"; benchmark structure + latency + hallucination risk",{"name":5,"author":48,"license":14,"version":49,"metadata":50,"description":7},"Hermes Agent","1.0.0",{"hermes":51},{"tags":52},[53,54,55,56,57,58],"Ollama","SSH","Verification","Hetzner","VPS","Local-LLM","main","5c34213899a7e6ba7597a65f60c542d4b7cb9521","mlops\u002Follama-remote-verification","https:\u002F\u002Fgithub.com\u002Fcryptopafi\u002Fnexusos-skills","https:\u002F\u002Fgithub.com\u002Fcryptopafi","2026-08-24 14:21:06",{"source":8,"skillName":5,"commands":66,"prompt":76,"zipAvailable":77},[67,70,73],{"tool":68,"command":69},"npx","npx skills add cryptopafi\u002Fnexusos-skills --skill ollama-remote-verification",{"tool":71,"command":72},"bunx","bunx skills add cryptopafi\u002Fnexusos-skills --skill ollama-remote-verification",{"tool":74,"command":75},"pnpm","pnpm dlx skills add cryptopafi\u002Fnexusos-skills --skill ollama-remote-verification","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fcryptopafi\u002Fnexusos-skills\u002Fblob\u002Fmain\u002Fmlops\u002Follama-remote-verification\u002FSKILL.md",true,[],[80,120,145,161,183,203],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":100,"publishedAt":43},"01M6WYTX72QQN7J0ZKSES38NEN","writing-plans","Use when you have a spec or requirements for a multi-step task. Creates comprehensive implementation plans with bite-sized tasks, exact file paths, and complete code examples.","https:\u002F\u002Fgithub.com\u002Fcryptopafi\u002Fnexusos-skills\u002Fblob\u002Fmain\u002Fsoftware-development\u002Fwriting-plans\u002FSKILL.md",[86,89,92,97],{"categoryNo":87,"categoryName":88,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000015","视觉设计",{"categoryNo":90,"categoryName":91,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000018","写作与文案",{"categoryNo":93,"categoryName":94,"taxonomyType":95,"taxonomyTypeDesc":96},"01JBSEEDCAT000000000000030","开发工程师","PROFESSION","适用职业",{"categoryNo":98,"categoryName":99,"taxonomyType":95,"taxonomyTypeDesc":96},"01JBSEEDCAT000000000000036","内容创作者",[101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,28,118,119],"writing","writing plans","plans","design","review","workflow","documentation","tasks","development","code","planning","code-review","subagent","superpowers","paths","test-driven-development","you","test-driven","when",{"skillNo":121,"name":116,"nameZh":6,"description":122,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":123,"categories":124,"keywords":131,"publishedAt":43},"01M6WYTN5C9XC8EYWGPXPWKGGX","Use when implementing any feature or bugfix, before writing implementation code. Enforces RED-GREEN-REFACTOR cycle with test-first approach.","https:\u002F\u002Fgithub.com\u002Fcryptopafi\u002Fnexusos-skills\u002Fblob\u002Fmain\u002Fsoftware-development\u002Ftest-driven-development\u002FSKILL.md",[125,128,129,130],{"categoryNo":126,"categoryName":127,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000007","代码审查",{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},{"categoryNo":90,"categoryName":91,"taxonomyType":20,"taxonomyTypeDesc":21},{"categoryNo":98,"categoryName":99,"taxonomyType":95,"taxonomyTypeDesc":96},[35,132,109,133,134,101,135,110,136,113,114,137,138,139,140,141,142,143,144,119],"test driven development","driven","testing","debugging","refactor","red-green-refactor","cycle","green","systematic","bugfix","implementing","systematic-debugging","test-first",{"skillNo":146,"name":143,"nameZh":6,"description":147,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":148,"categories":149,"keywords":151,"publishedAt":43},"01M6WYTDP76WP7Y5QVSWT2303F","Use when encountering any bug, test failure, or unexpected behavior. 4-phase root cause investigation — NO fixes without understanding the problem first.","https:\u002F\u002Fgithub.com\u002Fcryptopafi\u002Fnexusos-skills\u002Fblob\u002Fmain\u002Fsoftware-development\u002Fsystematic-debugging\u002FSKILL.md",[150],{"categoryNo":18,"categoryName":19,"taxonomyType":20,"taxonomyTypeDesc":21},[152,135,140,101,35,109,113,153,114,154,155,156,103,157,116,158,118,159,160,119],"systematic debugging","use","bug","root-cause","problem-solving","problem","fixes","phase","root cause",{"skillNo":162,"name":163,"nameZh":6,"description":164,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":165,"categories":166,"keywords":176,"publishedAt":43},"01M6WYT5P36B4M01BASQB9F0AY","subagent-driven-development","Use when executing implementation plans with independent tasks. Dispatches fresh delegate_task per task with two-stage review (spec compliance then code quality).","https:\u002F\u002Fgithub.com\u002Fcryptopafi\u002Fnexusos-skills\u002Fblob\u002Fmain\u002Fsoftware-development\u002Fsubagent-driven-development\u002FSKILL.md",[167,168,171,172,173],{"categoryNo":90,"categoryName":91,"taxonomyType":20,"taxonomyTypeDesc":21},{"categoryNo":169,"categoryName":170,"taxonomyType":20,"taxonomyTypeDesc":21},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":93,"categoryName":94,"taxonomyType":95,"taxonomyTypeDesc":96},{"categoryNo":98,"categoryName":99,"taxonomyType":95,"taxonomyTypeDesc":96},{"categoryNo":174,"categoryName":175,"taxonomyType":95,"taxonomyTypeDesc":96},"01JBSEEDCAT000000000000037","律师与法务",[109,177,113,133,101,105,106,178,108,110,179,112,180,181,114,103,182,116,118,119],"subagent driven development","code quality","delegation","parallel","delegate","fresh",{"skillNo":184,"name":185,"nameZh":6,"description":186,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":187,"categories":188,"keywords":191,"publishedAt":43},"01M6WYSXD1J366HVZRMRE7ENDA","requesting-code-review","Pre-commit verification pipeline — static security scan, baseline-aware quality gates, independent reviewer subagent, and auto-fix loop. Use after code changes and before committing, pushing, or opening a PR.","https:\u002F\u002Fgithub.com\u002Fcryptopafi\u002Fnexusos-skills\u002Fblob\u002Fmain\u002Fsoftware-development\u002Frequesting-code-review\u002FSKILL.md",[189,190],{"categoryNo":126,"categoryName":127,"taxonomyType":20,"taxonomyTypeDesc":21},{"categoryNo":93,"categoryName":94,"taxonomyType":95,"taxonomyTypeDesc":96},[192,105,110,193,101,194,112,24,195,196,197,113,198,199,200,114,116,201,202,118],"code review","requesting code review","commit","security scan","pipeline","loop","reviewer","fix","auto","static","auto-fix",{"skillNo":204,"name":205,"nameZh":6,"description":206,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":207,"categories":208,"keywords":209,"publishedAt":43},"01M6WYSNEBG74A5FCMCDGQ7QP4","plan","Plan mode for Hermes — inspect context, write a markdown plan into the active workspace's `.hermes\u002Fplans\u002F` directory, and do not execute the work.","https:\u002F\u002Fgithub.com\u002Fcryptopafi\u002Fnexusos-skills\u002Fblob\u002Fmain\u002Fsoftware-development\u002Fplan\u002FSKILL.md",[],[205,101,210,106,109,111,211,113,212,213,214,215,216,217,103,218,219,220,28,221],"directory","markdown","workspace","work","write","implementation","do","active","plan mode","inspect","mode","plan-mode"]