Releases

Changelog

What changed in each published version of ScissorCV, newest first.

Every ScissorCV release, newest first. Version numbers match the manifest.json version in the published extension.

v2.89.51 — 30 July 2026

Current release. Mostly a correctness and cost release: one scoring engine, no silent truncation, smaller prompts.

NVIDIA NIM is now the recommended free default provider. Groq is still supported and still free, and existing setups are untouched.
Raw AI call log. Every model call is recorded locally and can be exported to scissorcv_raw_ai_calls.json (the last 40 calls) so you can see exactly what was sent and returned on your own machine.
An automatic retry when a provider stops generation at its output cap, with the salvaged result clearly flagged rather than presented as complete.

One scoring calculation feeds both numbers. The Real score and the Potential ceiling are now derived from a single source instead of separate calculators that could disagree.
Output ceiling raised from 16,384 to 32,768 tokens, so long resumes finish generating.
Smaller prompts. Evidence is de-duplicated when it is saved — roughly 55% of lines were repeated across sources — and the achievement bank sent to the model is capped at 14,000 characters, with a 60,000-character ceiling on rewrite prompts. Your full evidence bank is still used for local scoring; only what is sent to the provider is trimmed.
No provider switching mid-generation. Fallback is disabled for generation, so a single resume is never written half by one model and half by another.

An improved score could snap back to a lower number when moving between the checklist and the dashboard. Both screens now report the same figure.
Long resumes could stop mid-sentence when a provider hit its output limit, without saying so. That case is now retried, and flagged if it still cannot complete.

Full manual acceptance testing pass on a clean browser profile.

Groq as the recommended free default provider, with free-first provider labelling during onboarding so no card is needed to start.
Skills Map graph view for inspecting evidenced, implied, confirmed and manual capabilities.
Regional CV renderers for US, UK, EU Europass-style, India, Canada and Australia/New Zealand.

Truth Guard now unions all evidence sources — resume, profile, confirmed claims and achievement bank — when validating skills, credentials, institutions, languages, certifications and numeric metrics.
Semantic hints are explicitly advisory and can no longer influence the match score.
Friendlier provider error copy with per-provider timeout and output caps.

Checklist cards no longer jump position while a generation is running.
Clear, non-crashing message on browser-restricted pages such as chrome:// URLs and the built-in PDF viewer.
IndexedDB warnings removed in both normal and incognito windows.
Long two-page source resumes no longer truncate mid-generation.

Conversion-layer work: store listing assets, onboarding copy and the founder pricing flow (5 free runs, then $19 lifetime).
Provider contract tests locked provider defaults, labels, caps and free-provider ordering.

Stabilised weighted score mathematics and aligned Truth Guard output with the reported score, with regression tests to prevent drift.
Rejected malformed generation output across all providers before it could reach a saved result.

Releases before v2.89.20 were internal builds and were not published to the Chrome Web Store. Public release notes start with the entries above.