跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
视频与音频

doc-report

Summarize TeXada CLI checks, Studio repair candidates or local image/PDF transcription evidence. Use when the user needs an auditable account of changes, review decisions and remaining uncertainty, keeping OCR fidelity separate from syntax and compilation.

CacinieP · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add CacinieP/TeXada-Studio-with-agentskills --skill doc-report

技能说明


name: doc-report description: Summarize TeXada CLI checks, Studio repair candidates or local image/PDF transcription evidence. Use when the user needs an auditable account of changes, review decisions and remaining uncertainty, keeping OCR fidelity separate from syntax and compilation.

Quality Check Report

Select the evidence source

WorkflowExisting recordsReport route
CLI node checksRun directory state.jsonl, available crops/, generated report.mdUse the harness report generated for the relevant run_id; report-generation functions are in harness/docforensics/__main__.py
Studio .tex repairStudio data directory jobs/<job_id>/job.json and events.jsonl; authenticated job/status APIsExport the completed task's .audit.md from Studio, supplementing from persisted evidence when requested
Studio image/PDF transcriptionimports/<import_id>/source.<ext> and candidate.jsonSummarize the import record and human comparison with the source; OCR import has no separate built-in audit export

Use the configured Studio data directory, not an assumed repository-relative location: the local launcher defaults to the user's application data directory. Report generation must not rerun OCR, repair, compilation or the CLI pipeline merely to refresh a summary.

Report what actually happened

  • Identify the workflow, provider and model from recorded evidence. CLI and Studio formula-repair requests send formula text; Studio OCR sends page images to the local vision endpoint. Fixture repairs are offline test doubles, never measured model performance. Deterministic table checks have no model request.
  • For checks and repairs, include original and candidate text, source line or node, checker outcomes, attempted/rejected candidates and unresolved items. Preserve recorded reasons; missing responses after interruption are missing evidence. Bounded response excerpts may be truncated.
  • For Studio repair tasks, use result.original_sha256, candidate_sha256, edits, problems_after, trace, elapsed_seconds and compilation records when present. compile_ok and preview_ok are separate: a PDF may compile while image conversion fails. A NOT_CHECKED compilation was not a failed compilation.
  • Separate checker acceptance from human adoption. CANDIDATE_ACCEPTED records checker acceptance; USER_ACCEPTED/USER_REJECTED and top-level job review record the user's decision. Adoption into the editor is separate from saving a document. The current exported .audit.md includes result trace, but does not automatically include later review events from events.jsonl; inspect those records for a requested review audit.
  • For OCR, record source name/hash, import ID, configured model, page reported_model, content_sha256, imported page range and has_more_pages. Summarize unreadable spans and only the discrepancies actually compared against the original. requires_review: true marks a candidate, not completed review. The import record does not establish that the candidate was adopted or saved.
  • OCR pages[].raw_response is the cleaned transcription, not a complete raw transport response. Per-page usage may be absent; the import record has no repair-style model_calls or per-request event journal. Do not invent HTTP-attempt counts, elapsed time, OCR scores, bbox accuracy or confidence from page count/token metadata. Import source and field details are in the input contract.
  • Include crop paths only when the files exist. Studio retains the original image/PDF; temporary PDF page images are not retained as CLI crops. A page image hash is provenance metadata, not a surviving image attachment. Mark missing evidence instead of fabricating it.
  • Keep OCR fidelity, syntax, compilation and mathematical correctness as separate conclusions. Successful checks or compilation do not establish faithful transcription or correct mathematics.

Calls and Skill evidence

Distinguish configured capabilities from operations recorded as executed. Report Skill mode and recorded skill_loaded/skill_sha256 evidence rather than assuming every named Skill ran. Studio OCR uses a fixed prompt and does not load doc-layout-parse; report generation is fixed code and does not load doc-report into a model.

For CLI request counts, use current-run PROVIDER_RESULT records and model_calls_known; unmatched starts or unknown counts mean a lower bound. An attempted request that times out may have reached the server, while completion remains unknown. For Studio, keep recorded repair model_calls scoped to that task; do not add earlier OCR requests to it or treat missing counts as zero.

Execution identity

CLI terminal reuse requires matching canonical input and execution-context hashes. A matching node identifier alone is insufficient. Reports must distinguish terminal results from skipped matching work and avoid counting resumed events as fresh quality measurements. Skill-content hashes identify the instructions loaded by the formula host; they do not prove better semantic accuracy.

Guardrails

Report generation must not trigger another model call or silently edit source files. Exclude credentials, allocated network details, and private document content from material intended for publication. Package existing evidence with relative paths when sharing a CLI run.