Repository navigation
v3.15.0
Summary
A work-order release. The AI work order (--format ai-md) groups its items
into families, binds each file to its scan-time SHA-256 once, and reads itself
back before it is written. The JSON Lines report, findings, scores and the
analysis are unchanged. Every change is listed in the Changelog.
What moves when you upgrade
- The work order has a new layout and names it. The header reads
format: ai-slop-work-order/2 markdown work order | records ai-slop-report/1.
Items are#### Wnnblocks under### Gnnfamily headers that state the
shared fix and channel once; a## FILEStable holds each file's SHA-256 and
binding, and items point to their row (B01). A tool that parses the
Markdown work order needs to follow this layout; the item fields are the same.
The JSON Lines report (--format ai) does not change. - A lossy work order is never written. The renderer reads the document
back: every item exactly once, every field it shows equal to its record,
every item bound to its file's hash, family counts that add up. Otherwise it
stops withthe work order is not losslessand writes nothing. - More evidence within the same token budget. Short functions whose whole
body costs less than an excerpt are drawn whole, and family headers are paid
for in the allocation. On a snapshot of this repository (227 items), the
auditor work order at 32,000 estimated tokens shows the evidence of 144 items
instead of 113 with 3.14.2. - VS Code extension: "Copy work order" carries the item's family header and
its FILES row; it still reads a work order from a 3.14 backend.
Verification
- Test suite: 2,010 passed, 2 skipped (Python 3.14); 1,969 passed, 34 skipped
(Python 3.8). - Mutation testing: 24 of 24 deliberately broken copies of the family key,
file table, read-back check and budget allocation were caught.
Known limitations
- Severity stands in for a fix-risk tier in the family key; the detector does
not compute a fix-risk tier yet. - The limitations listed for v3.14.0 still apply.