[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77JD2GCDAX16ZPCGMTRHE45":3,"skill-same:01M77JD2GCDAX16ZPCGMTRHE45":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},"01M77JD2GCDAX16ZPCGMTRHE45","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_grpo_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## Protocol-safe OCR + Evidence-Grounded Extraction\n\n### 1) Tool\u002Faction protocol (hard stop before OCR)\n- Always emit a controller-parseable Action wrapper for every environment\u002Ftool step (do not output free-form commands).\n- If the runner reports an action parsing issue (e.g., nothing valid was parsed), stop and immediately retry with correct Action syntax.\n- Do not proceed to OCR\u002Ftable parsing after tool\u002Fprotocol parsing failures.\n- After each tool call, verify it actually produced usable evidence (expected text\u002Fartifact exists and is readable). If evidence is missing, do a minimal targeted re-run of the same step; otherwise switch to a simpler fallback approach.\n\n### 2) Inspect + decide the plan (no answering yet)\n- First visually inspect the page to determine whether the target is: a labeled field, a header\u002Ffooter token (including page-number style queries), or a table.\n- Plan to anchor extraction locally (fields\u002Ftables) rather than searching globally.\n\n### 3) OCR baseline, validation, and retry escalations\n- Prefer a baseline whole-page OCR pass.\n- Gate progress on evidence: only continue to reasoning\u002Fextraction if OCR produced non-empty, usable text for the relevant cues.\n- If baseline OCR is weak\u002Fnoisy\u002Fempty:\n  - Preprocess escalation: apply basic readability improvements (e.g., grayscale\u002Fcontrast\u002Flight binarization) and rerun the same simple OCR call.\n  - Crop escalation: crop a tight region around the visually identified (or OCR-located) area and rerun OCR on that crop.\n- Dependency safety: use only OCR entrypoints that are available in this runtime.\n  - If the intended OCR path fails due to missing functions\u002Fmodules or unsupported arguments, immediately fall back to a known-good OCR route.\n- Known-good fallback OCR: load the image with PIL and run `pytesseract.image_to_string(...)` on the full image (or on the tight crop if full-page OCR is too noisy), then extract by evidence from the OCR text.\n- Treat empty\u002Firrelevant OCR outputs as a hard extraction failure (do not answer without grounded evidence).\n\n### 4) Evidence-only grounding rules (no mismatched nearby values)\n- Anchor-and-adjacent extraction for non-table fields:\n  - Locate the label\u002Fdescriptor in OCR.\n  - Extract the value from OCR text that is layout-proximate to that label (same line first; otherwise the immediately neighboring line\u002Fblock).\n  - Require OCR co-location: accept a value only if it appears in the same local region\u002Fcrop that contains the matched label relationship.\n\n- Table grounding:\n  - Identify the header(s) first.\n  - Locate the correct row using the question constraint and local OCR matches.\n  - Read the value from the cell at the intersection of the matched row and the requested column\u002Fheader.\n  - If row\u002Fcolumn alignment is ambiguous, expand the crop to include the full relevant header + the full matched row, then rerun OCR.\n  - Never mix values across different rows\u002Fcolumns.\n\n- Category\u002Fdescriptor specificity:\n  - When the question distinguishes categories\u002Ftiers\u002Fgroups, extract only from the exact category-specific cell\u002Fline tied to the requested context—do not substitute adjacent summaries\u002Faverages from other categories.\n\n### 5) Header\u002Ffooter + page-number style recovery (high-impact)\n- For page-number requests, search within the whole-page OCR for the closest header\u002Ffooter phrase that contains the page-index pattern.\n- If the token is garbled or missed, retry with crops focused on header\u002Ffooter regions.\n- Recover the final page-index output only when the required components can be extracted from the matching header\u002Ffooter evidence; otherwise treat it as not found.\n\n### 6) OCR matching hygiene\n- Normalize OCR text for matching (case-insensitive key comparison; remove harmless punctuation\u002Fspacing differences).\n- Normalize units\u002Fformatting enough to match the expected pattern without fabricating missing characters.\n\n### 7) Final answer discipline (strict)\n- Output exactly one line: `Final answer: \u003Crequested_value>`.\n- If the value cannot be verified from the relevant OCR region(s), output: `Final answer: Not found`.\n- Do not include explanations or multiple candidates.\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## Protocol-safe OCR + Evidence-Grounded Extraction\n\n### 1) Tool\u002Faction protocol (hard stop before OCR)\n- Always emit a controller-parseable Action wrapper for every environment\u002Ftool step (do not output free-form commands).\n- If the runner reports an action parsing issue (e.g., nothing valid was parsed), stop and immediately retry with correct Action syntax.\n- Do not proceed to OCR\u002Ftable parsing after tool\u002Fprotocol parsing failures.\n- After each tool call, verify it actually produced usable evidence (expected text\u002Fartifact exists and is readable). If evidence is missing, do a minimal targeted re-run of the same step; otherwise switch to a simpler fallback approach.\n\n### 2) Inspect + decide the plan (no answering yet)\n- First visually inspect the page to determine whether the target is: a labeled field, a header\u002Ffooter token (including page-number style queries), or a table.\n- Plan to anchor extraction locally (fields\u002Ftables) rather than searching globally.\n\n### 3) OCR baseline, validation, and retry escalations\n- Prefer a baseline whole-page OCR pass.\n- Gate progress on evidence: only continue to reasoning\u002Fextraction if OCR produced non-empty, usable text for the relevant cues.\n- If baseline OCR is weak\u002Fnoisy\u002Fempty:\n  - Preprocess escalation: apply basic readability improvements (e.g., grayscale\u002Fcontrast\u002Flight binarization) and rerun the same simple OCR call.\n  - Crop escalation: crop a tight region around the visually identified (or OCR-located) area and rerun OCR on that crop.\n- Dependency safety: use only OCR entrypoints that are available in this runtime.\n  - If the intended OCR path fails due to missing functions\u002Fmodules or unsupported arguments, immediately fall back to a known-good OCR route.\n- Known-good fallback OCR: load the image with PIL and run `pytesseract.image_to_string(...)` on the full image (or on the tight crop if full-page OCR is too noisy), then extract by evidence from the OCR text.\n- Treat empty\u002Firrelevant OCR outputs as a hard extraction failure (do not answer without grounded evidence).\n\n### 4) Evidence-only grounding rules (no mismatched nearby values)\n- Anchor-and-adjacent extraction for non-table fields:\n  - Locate the label\u002Fdescriptor in OCR.\n  - Extract the value from OCR text that is layout-proximate to that label (same line first; otherwise the immediately neighboring line\u002Fblock).\n  - Require OCR co-location: accept a value only if it appears in the same local region\u002Fcrop that contains the matched label relationship.\n\n- Table grounding:\n  - Identify the header(s) first.\n  - Locate the correct row using the question constraint and local OCR matches.\n  - Read the value from the cell at the intersection of the matched row and the requested column\u002Fheader.\n  - If row\u002Fcolumn alignment is ambiguous, expand the crop to include the full relevant header + the full matched row, then rerun OCR.\n  - Never mix values across different rows\u002Fcolumns.\n\n- Category\u002Fdescriptor specificity:\n  - When the question distinguishes categories\u002Ftiers\u002Fgroups, extract only from the exact category-specific cell\u002Fline tied to the requested context—do not substitute adjacent summaries\u002Faverages from other categories.\n\n### 5) Header\u002Ffooter + page-number style recovery (high-impact)\n- For page-number requests, search within the whole-page OCR for the closest header\u002Ffooter phrase that contains the page-index pattern.\n- If the token is garbled or missed, retry with crops focused on header\u002Ffooter regions.\n- Recover the final page-index output only when the required components can be extracted from the matching header\u002Ffooter evidence; otherwise treat it as not found.\n\n### 6) OCR matching hygiene\n- Normalize OCR text for matching (case-insensitive key comparison; remove harmless punctuation\u002Fspacing differences).\n- Normalize units\u002Fformatting enough to match the expected pattern without fabricating missing characters.\n\n### 7) Final answer discipline (strict)\n- Output exactly one line: `Final answer: \u003Crequested_value>`.\n- If the value cannot be verified from the relevant OCR region(s), output: `Final answer: Not found`.\n- Do not include explanations or multiple candidates.\n",{"name":5,"description":7},"main","3c6cd6992df3c353299cb9c6d33f733ede7854c8","docvqa-train-time\u002Fresults\u002Fskill\u002Fqwen35_4b_grpo_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_grpo_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",[],[]]