[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77JD7AFGM3KY9FNKD02NTH7":3,"skill-same:01M77JD7AFGM3KY9FNKD02NTH7":46},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":6,"updatedAt":31,"install":32,"resources":45},"01M77JD7AFGM3KY9FNKD02NTH7","docvqa-trace-skill",null,"Reusable document-image inspection, OCR, grounding, and answer-verification guidance for DocVQA tasks.","zz1358m\u002FAgentic-ESOpt","zz1358m","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F71759735?v=4","01M77JCPM94NH12HCQF1N7WJBN","Zhi Zheng","01M77JCPNDB4Y218PRPB0ZKH56",86,"MIT","https:\u002F\u002Fgithub.com\u002Fzz1358m\u002FAgentic-ESOpt\u002Fblob\u002Fmain\u002Fdocvqa-train-time\u002Fresults\u002Fskill\u002Fqwen35_4b_trace2skill\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: docvqa-trace-skill\ndescription: Reusable document-image inspection, OCR, grounding, and answer-verification guidance for DocVQA tasks.\n---\n\n# DocVQA Skill\n\nInspect the document image before answering. Read labels, table headers, nearby\ntext, and layout relationships carefully. Return only the requested answer unless\nthe task explicitly asks for explanation.\n\n## Tool \u002F Protocol Hard-Gates (fail fast; no guessing)\n- Use the exact accepted action envelope on every tool turn (match the environment’s schema exactly): `Action: {\"name\":\"bash\",\"arguments\":{\"command\":\"\u003Cshell command>\"}}`.\n- If the environment reports an action parsing\u002Fexecution rejection (e.g., action parsing error\u002Funknown action) treat it as a hard stop:\n  - immediately correct the next tool call format,\n  - retry only the minimum necessary step,\n  - do not continue OCR\u002Flayout reasoning or speculative answering based on the failed step.\n- After any tool\u002FOCR call, verify observable success:\n  - check tool stdout contains usable text when text is expected, and\u002For\n  - confirm any expected artifacts\u002Fcrops were created and are readable.\n- If OCR\u002Ftool execution fails or produces no usable OCR evidence, stop content-level reasoning for that tactic and either re-route to the smallest fixable inspection step or abstain if you still cannot obtain label\u002Fvalue OCR evidence.\n\n## OCR Readiness: smoke-test first; evidence-only gates\n- Before relying on OCR, confirm the OCR pipeline is runnable (engine\u002Ftool present; input image\u002Fobject type is valid).\n- Run a minimal OCR smoke test on a small obvious crop (or the full page) and ensure it returns non-empty, readable text.\n- After each OCR run, apply an OCR usefulness gate:\n  - if output is empty\u002Fgarbled\u002Funhelpful (or missing the intermediate evidence you intend to use), stop that tactic and retry with improved localization\u002Fcropping (avoid random transformations).\n\n## Inspection & Region Selection (anti-drift)\n- Briefly inspect the full image to infer layout type (table vs form vs list) and where evidence likely lives.\n- Use whole-page OCR only to locate\u002Fconfirm structure and approximate regions; do not derive the final answer from noisy whole-page OCR alone.\n- Perform deterministic two-stage OCR:\n  1) Whole-page recovery: OCR the full image, then use the question to localize the referenced instance\u002Farea (e.g., the relevant panel\u002Fsection\u002Ftable instance).\n  2) Tight ROI grounding: crop tightly around the localized label\u002Fvalue neighborhood (or the target row\u002Fcolumn intersection within the table) and re-OCR the crop for final extraction.\n\n## Grounded Extraction Rules (no drift; local co-occurrence)\n- Do not use whole-page OCR to infer instance-specific relationships; localize the exact referenced instance first, then OCR within that instance.\n- Label→adjacent value extraction:\n  - anchor by the exact label\u002Frole token as it appears in OCR,\n  - take the value from the same OCR line\u002Fitem where the label occurs,\n  - require label+value co-occurrence in the same local OCR context; reject candidates where the label does not co-occur.\n- If the required label\u002Fcue cannot be found in the tool-visible OCR text from the relevant region, return `cannot determine`.\n- If OCR output is fragmented\u002Ftruncated or multiple near-matches appear, re-crop tighter around the exact label\u002Fvalue area and re-OCR; do not “nearest match” across unrelated lines\u002Fcolumns.\n\n## Table Reading (header→row intersection)\n- Identify table header(s) in OCR and lock the requested column by header text (do not assume column order).\n- For a specific row\u002Fvalue request:\n  - anchor the row using its row key\u002Fidentifier as OCR shows it,\n  - extract the cell using row×column intersection where both the header and the row context appear in the same table-region OCR evidence.\n- If dense tables cause OCR confusion, re-crop the table region to include the target header and the target row context (header+row+cell) and re-OCR until the header+row+value are visible together.\n\n## Pagination \u002F Page-Number Extraction (grounded in a cropped band)\n- Treat pagination as field-specific: crop to the consistent header\u002Ffooter band first.\n- OCR the pagination crop and extract the page-indicator phrase from OCR lines that contain the pagination structure.\n- Apply tolerant normalization for common OCR confusions around pagination tokens only when the underlying pagination structure is present.\n- Evidence check: the final pagination answer must be derivable from the cropped pagination region’s OCR text; ignore incidental numbers elsewhere.\n- Output formatting:\n  - if only a single page index is present, return that index;\n  - if the composite pagination structure is present, return it in the required canonical pagination format.\n\n## Ordinal Title Selection (Nth heading without token drift)\n- For questions like “Nth\u002Fordinal title”:\n  - OCR the relevant heading\u002Ftitle block,\n  - split into distinct title-like line entries,\n  - order candidates by their vertical top-to-bottom appearance,\n  - select the Nth entry from that ordered list.\n- If OCR merges multiple headings into one line, refine with smaller crops around each heading and re-OCR until headings are separable.\n\n## Crop \u002F Preprocessing Discipline\n- Use one consistent coordinate system for all crops; validate image width\u002Fheight ordering once and reuse.\n- Prefer simple, consistent preprocessing; avoid heavy multi-step pipelines.\n- Mode-safe handling: if the image is grayscale, treat pixels as single-channel intensities (do not assume RGB tuples).\n- If disk-based crop-save\u002Fload is brittle, prefer in-memory cropping and ensure OCR receives an image object (not raw bytes).\n- If a crop OCR is garbled\u002Fnoisy, apply only light, deterministic enhancement (e.g., grayscale\u002Fcontrast\u002Fsharpen) plus at most one controlled upscale, then re-OCR.\n\n## Date\u002FToken Completeness & Exact Substring Copying\n- For date\u002Ftoken-like answers, require all needed components to be present in OCR evidence.\n- Preserve formatting fidelity by copying the cue-bearing OCR substring exactly as it appears in the relevant line (modifying only by removing surrounding label prefixes\u002Fsuffixes as needed).\n- If any component is missing\u002Fgarbled, re-crop\u002Fenlarge around the cue region and re-run OCR until the complete token sequence is visible.\n\n## Final Answer Contract (strict)\n- Return only the requested extracted value\u002Fcontent as a plain string (no OCR dumps, no reasoning, no extra surrounding label text).\n- Apply minimal normalization (trim\u002Fcollapse whitespace) while preserving punctuation\u002Fcasing as shown in OCR.\n- If you cannot deterministically link label→value (or row×column→cell) using verified OCR evidence, return `cannot determine`.\n","# DocVQA Skill\n\nInspect the document image before answering. Read labels, table headers, nearby\ntext, and layout relationships carefully. Return only the requested answer unless\nthe task explicitly asks for explanation.\n\n## Tool \u002F Protocol Hard-Gates (fail fast; no guessing)\n- Use the exact accepted action envelope on every tool turn (match the environment’s schema exactly): `Action: {\"name\":\"bash\",\"arguments\":{\"command\":\"\u003Cshell command>\"}}`.\n- If the environment reports an action parsing\u002Fexecution rejection (e.g., action parsing error\u002Funknown action) treat it as a hard stop:\n  - immediately correct the next tool call format,\n  - retry only the minimum necessary step,\n  - do not continue OCR\u002Flayout reasoning or speculative answering based on the failed step.\n- After any tool\u002FOCR call, verify observable success:\n  - check tool stdout contains usable text when text is expected, and\u002For\n  - confirm any expected artifacts\u002Fcrops were created and are readable.\n- If OCR\u002Ftool execution fails or produces no usable OCR evidence, stop content-level reasoning for that tactic and either re-route to the smallest fixable inspection step or abstain if you still cannot obtain label\u002Fvalue OCR evidence.\n\n## OCR Readiness: smoke-test first; evidence-only gates\n- Before relying on OCR, confirm the OCR pipeline is runnable (engine\u002Ftool present; input image\u002Fobject type is valid).\n- Run a minimal OCR smoke test on a small obvious crop (or the full page) and ensure it returns non-empty, readable text.\n- After each OCR run, apply an OCR usefulness gate:\n  - if output is empty\u002Fgarbled\u002Funhelpful (or missing the intermediate evidence you intend to use), stop that tactic and retry with improved localization\u002Fcropping (avoid random transformations).\n\n## Inspection & Region Selection (anti-drift)\n- Briefly inspect the full image to infer layout type (table vs form vs list) and where evidence likely lives.\n- Use whole-page OCR only to locate\u002Fconfirm structure and approximate regions; do not derive the final answer from noisy whole-page OCR alone.\n- Perform deterministic two-stage OCR:\n  1) Whole-page recovery: OCR the full image, then use the question to localize the referenced instance\u002Farea (e.g., the relevant panel\u002Fsection\u002Ftable instance).\n  2) Tight ROI grounding: crop tightly around the localized label\u002Fvalue neighborhood (or the target row\u002Fcolumn intersection within the table) and re-OCR the crop for final extraction.\n\n## Grounded Extraction Rules (no drift; local co-occurrence)\n- Do not use whole-page OCR to infer instance-specific relationships; localize the exact referenced instance first, then OCR within that instance.\n- Label→adjacent value extraction:\n  - anchor by the exact label\u002Frole token as it appears in OCR,\n  - take the value from the same OCR line\u002Fitem where the label occurs,\n  - require label+value co-occurrence in the same local OCR context; reject candidates where the label does not co-occur.\n- If the required label\u002Fcue cannot be found in the tool-visible OCR text from the relevant region, return `cannot determine`.\n- If OCR output is fragmented\u002Ftruncated or multiple near-matches appear, re-crop tighter around the exact label\u002Fvalue area and re-OCR; do not “nearest match” across unrelated lines\u002Fcolumns.\n\n## Table Reading (header→row intersection)\n- Identify table header(s) in OCR and lock the requested column by header text (do not assume column order).\n- For a specific row\u002Fvalue request:\n  - anchor the row using its row key\u002Fidentifier as OCR shows it,\n  - extract the cell using row×column intersection where both the header and the row context appear in the same table-region OCR evidence.\n- If dense tables cause OCR confusion, re-crop the table region to include the target header and the target row context (header+row+cell) and re-OCR until the header+row+value are visible together.\n\n## Pagination \u002F Page-Number Extraction (grounded in a cropped band)\n- Treat pagination as field-specific: crop to the consistent header\u002Ffooter band first.\n- OCR the pagination crop and extract the page-indicator phrase from OCR lines that contain the pagination structure.\n- Apply tolerant normalization for common OCR confusions around pagination tokens only when the underlying pagination structure is present.\n- Evidence check: the final pagination answer must be derivable from the cropped pagination region’s OCR text; ignore incidental numbers elsewhere.\n- Output formatting:\n  - if only a single page index is present, return that index;\n  - if the composite pagination structure is present, return it in the required canonical pagination format.\n\n## Ordinal Title Selection (Nth heading without token drift)\n- For questions like “Nth\u002Fordinal title”:\n  - OCR the relevant heading\u002Ftitle block,\n  - split into distinct title-like line entries,\n  - order candidates by their vertical top-to-bottom appearance,\n  - select the Nth entry from that ordered list.\n- If OCR merges multiple headings into one line, refine with smaller crops around each heading and re-OCR until headings are separable.\n\n## Crop \u002F Preprocessing Discipline\n- Use one consistent coordinate system for all crops; validate image width\u002Fheight ordering once and reuse.\n- Prefer simple, consistent preprocessing; avoid heavy multi-step pipelines.\n- Mode-safe handling: if the image is grayscale, treat pixels as single-channel intensities (do not assume RGB tuples).\n- If disk-based crop-save\u002Fload is brittle, prefer in-memory cropping and ensure OCR receives an image object (not raw bytes).\n- If a crop OCR is garbled\u002Fnoisy, apply only light, deterministic enhancement (e.g., grayscale\u002Fcontrast\u002Fsharpen) plus at most one controlled upscale, then re-OCR.\n\n## Date\u002FToken Completeness & Exact Substring Copying\n- For date\u002Ftoken-like answers, require all needed components to be present in OCR evidence.\n- Preserve formatting fidelity by copying the cue-bearing OCR substring exactly as it appears in the relevant line (modifying only by removing surrounding label prefixes\u002Fsuffixes as needed).\n- If any component is missing\u002Fgarbled, re-crop\u002Fenlarge around the cue region and re-run OCR until the complete token sequence is visible.\n\n## Final Answer Contract (strict)\n- Return only the requested extracted value\u002Fcontent as a plain string (no OCR dumps, no reasoning, no extra surrounding label text).\n- Apply minimal normalization (trim\u002Fcollapse whitespace) while preserving punctuation\u002Fcasing as shown in OCR.\n- If you cannot deterministically link label→value (or row×column→cell) using verified OCR evidence, return `cannot determine`.\n",{"name":5,"description":7},"main","3c6cd6992df3c353299cb9c6d33f733ede7854c8","docvqa-train-time\u002Fresults\u002Fskill\u002Fqwen35_4b_trace2skill",9,"https:\u002F\u002Fgithub.com\u002Fzz1358m\u002FAgentic-ESOpt","https:\u002F\u002Fgithub.com\u002Fzz1358m","https:\u002F\u002Fzz1358m.github.io\u002Fzhizheng.github.io\u002F\u002F","2026-09-10 15:49:28",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add zz1358m\u002FAgentic-ESOpt --skill docvqa-trace-skill",{"tool":38,"command":39},"bunx","bunx skills add zz1358m\u002FAgentic-ESOpt --skill docvqa-trace-skill",{"tool":41,"command":42},"pnpm","pnpm dlx skills add zz1358m\u002FAgentic-ESOpt --skill docvqa-trace-skill","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fzz1358m\u002FAgentic-ESOpt\u002Fblob\u002Fmain\u002Fdocvqa-train-time\u002Fresults\u002Fskill\u002Fqwen35_4b_trace2skill\u002FSKILL.md",true,[],[47,54,59,64,71,77],{"skillNo":48,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":51,"categories":52,"keywords":53,"publishedAt":19},"01M77JDS53Z4D602YM692PZBPX","webarena-sft-trace-skill","Skill instructions for WebArena agents using WebRL id actions.","https:\u002F\u002Fgithub.com\u002Fzz1358m\u002FAgentic-ESOpt\u002Fblob\u002Fmain\u002Fwebarena-train-time\u002Fresults\u002Fskills\u002Ftrace2skill_no-finetune\u002FSKILL.md",[],[],{"skillNo":55,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":56,"categories":57,"keywords":58,"publishedAt":19},"01M77JDPJ0XK35CG8ZHDD5RGEQ","https:\u002F\u002Fgithub.com\u002Fzz1358m\u002FAgentic-ESOpt\u002Fblob\u002Fmain\u002Fwebarena-train-time\u002Fresults\u002Fskills\u002Ftrace2skill_agentic_esopt\u002FSKILL.md",[],[],{"skillNo":60,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":19},"01M77JDMZXNQW6C2B02P6DVD5K","https:\u002F\u002Fgithub.com\u002Fzz1358m\u002FAgentic-ESOpt\u002Fblob\u002Fmain\u002Fwebarena-train-time\u002Fmethods\u002Ftrace2skill\u002Fskills\u002Fwebagent\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":19},"01M77JDHAQQ8N7PCZYZ3QWGD3M","math-reasoning-trace-skill","Follow strict tool\u002Fprotocol and exact mathematical proof-and-verification workflow for math problems; always end with the required single-line final answer format.","https:\u002F\u002Fgithub.com\u002Fzz1358m\u002FAgentic-ESOpt\u002Fblob\u002Fmain\u002Fmath-train-time\u002Fresults\u002Fskill\u002Fqwen35_4b_trace2skill\u002FSKILL.md",[],[],{"skillNo":72,"name":66,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M77JDFF9B4FZ3CR1STFRVJ92","Reusable reasoning, computation, and verification guidance for mathematical problem solving.","https:\u002F\u002Fgithub.com\u002Fzz1358m\u002FAgentic-ESOpt\u002Fblob\u002Fmain\u002Fmath-train-time\u002Fresults\u002Fskill\u002Fqwen35_4b_grpo_trace2skill\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"publishedAt":19},"01M77JDCBSHY0PWZDH2Z0T4H5W","qwen35_4b_agentic_esopt_trace2skill","https:\u002F\u002Fgithub.com\u002Fzz1358m\u002FAgentic-ESOpt\u002Fblob\u002Fmain\u002Fmath-train-time\u002Fresults\u002Fskill\u002Fqwen35_4b_agentic_esopt_trace2skill\u002FSKILL.md",[],[]]