Raw Upload Content Scale — Scope Note
Version. 0.2 · Date. 2026-07-02
Status. Open finding, not yet scheduled for build.
Supersedes. v0.1 (loomworks-raw-upload-content-scale-v0_1.md) — same finding, adds new evidence below. Prior content preserved, not deleted.
The problem (unchanged from v0.1)
Uploaded file content is stored raw — not interpreted, not chunked, not summarized. Some formats (HTML) fail to read at all.
New evidence (2026-07-02, real repo, 8 files)
Operator uploaded a real 8-file GitHub repo folder (partspilot-repo-init). Result:
- 3 files unreadable (
.gitignore, .env.example, CODEOWNERS) — same "couldn't recognize this file" failure as the HTML case.
- 5 readable files dumped as full raw text, each becoming its own held assertion in Memory — README, SECURITY.md, PR template, etc. Held panel reached 12 items, several are entire files pasted verbatim.
- New symptom: for each unreadable file, the interface stops and presents an empty textarea, waiting for the Operator to manually type replacement content before continuing. At 8 files with 3 unreadable, this is 3 forced manual stops mid-upload. This does not scale — at higher file counts this becomes an unworkable number of manual interruptions.
Combined finding (now three symptoms)
- Format support. Common file types (HTML,
.gitignore-style dotfiles, no-extension files like CODEOWNERS) fail to read.
- Interpretation. Readable files are dumped raw as individual held assertions, not summarized or chunked.
- Blocking behavior. Each unreadable file halts the batch and demands manual input before proceeding, rather than skipping and continuing or batching failures into one follow-up.
Status
Not yet scheduled. Should be resolved before any workflow involving real-world folders or repos, where unreadable/dotfile-style files are common.