 Matthew Meszaros
|
d51de7db3a
|
feat: address the CodeRabbit review by quoting a fragment in the copy's own casing rather than the model's retyping of it, extracting the case-fold offset map into internal/pkg/casefold so the AI half gets the same Unicode safety the rules half has, giving a trigger term a span in each half it appears in instead of losing the second one to deduplication, scanning subject links before body anchors so the display cap cannot drop the subject's own, requiring WRITE_TEMPLATES on the credit-spending analyze route so a read-only key cannot spend the workspace balance, refusing to tell a customer their credits came back when the refund is what failed, and no longer letting a stale analysis retire the newer rules request that was about to replace it
|
2026-09-10 09:50:28 -07:00 |
|
|
Matthew Meszaros
|
e7b9491975
|
feat: drop an AI finding's category when the model returns one outside the closed set the API documents, so a client validating the response against that enum is never handed something outside it, and correct the parse comment that still described an unreadable response as falling back to the rules pass when Analyze now refuses it and refunds
|
2026-09-10 09:28:00 -07:00 |
|
|
Matthew Meszaros
|
ade18d1650
|
feat: leave an AI finding's field empty when the model labelled neither half and nothing in the finding could be anchored in the copy, instead of defaulting it to the body and rendering a badge that sends the writer to the wrong box on the one panel whose whole purpose is saying which box to open
|
2026-09-10 09:28:00 -07:00 |
|
|
Matthew Meszaros
|
3a9b1747dc
|
feat: map a trigger term's offset back through the case fold before slicing the copy, so a rune that changes byte length when lowercased (U+0130, U+212A) no longer shifts every following span into bytes the writer never typed or cuts one in half into invalid UTF-8, read an issue's field from the whole span list before the display cap trims it so a subject-first list of more than eight can no longer relabel a both-halves issue, and refuse an AI response carrying a verdict with no score and no findings instead of pricing its empty finding list into a confident 100 out of 100 above a verdict saying the opposite
|
2026-09-10 09:28:00 -07:00 |
|
|
Matthew Meszaros
|
331db196d8
|
feat: locate every content-check issue in the subject or the body with the exact fragments that caused it and a one-line fix, add POST /templates/analyze running the configured LLM over a campaign template for located spam findings quoted verbatim from the copy plus a rewritten subject and an overall score, verify every model quote against the draft so an invented sentence is dropped rather than shown, pin the analysis temperature so re-checking unchanged copy returns the same number, and give the editor panel a Re-check button that re-runs both passes and reports the movement since the last check
|
2026-09-10 09:28:00 -07:00 |
|