vision-extract

原文涵盖的主题

• Scanned-Page Vision Extraction
• How to run
• Adapters (provider-neutral)
• Contract preserved from the source script
• PDF → page images
• Triggers
• Notes

Transcribe a directory of scanned page PNGs into Markdown source text — the acquisition step for inducting scanned books / pages. Provider-neutral: the vision call is a pluggable adapter, not hardcoded to one CLI. Save this wrapper as vision.sh; it receives the image, prompt file, and output paths as separate arguments: Outputs to .aiwg/research/sources/text/ -vision/: per-page pages/page-NNN.md, combined -vision.vision.md + stripped .txt, run.log. | --provider | Vision call | |---|---| | codex (default) | codex exec --sandbox read-only --image … --output-last-message (set AIWGCODEXBIN/--model as needed) | | command | a shell template (--command or AIWGVISIONCOMMAND) with {image} / {promptfile} / {out} placeholders — wire any vision CLI (Claude, a custom script). The command writes the transcription to {out} (or stdout). | Both require a vision-capable model to be reachable; needs-infrastructure. Command templates must use unquoted, unescaped placeholders: {image} → "$1", {promptfile} → "$2", {out} → "$3". Paths are passed as separate shell arguments; AIWGVISIONIMAGE, AIWGVISIONPROMPTFILE, and AIWGVISIONOUT are also exported. For example: vision-cli --image {image} --prompt-file {p

这里只展示节选，请以完整原文和当前要求为准。