Changelog

Release notes for LockStack. Updates within a major version are free for existing license holders.

v3.4.27 — 7 July 2026

Four new outputs per post + in-app cards for all of them

• 21 file formats per post (was 17) — 634 files per campaign. New per-post outputs: a Google Business Profile post (the listing local customers see in Google Maps and local search), an Instagram caption with 1–3 smart hashtags inline, YouTube title / description / tags to pair with the video script, and a featured-image recipe — the exact image prompt, filename, and upload path each blog page expects, so you can produce and place the right image for every post.
• Newsletter output now begins with a ready-to-send Subject: line.
• All four new outputs appear as copy/edit/download cards inside the app (SOCIAL CHANNELS), with real platform character limits (Google Business 1,500 · Instagram 2,200) — on both Solo and Pro.
• Quality guards: the article-grammar fixer now covers any word without ever touching real names ("An Nguyen" stays "An Nguyen"); https:// is added automatically when the site URL is typed as a bare domain; hashtags can no longer contain broken characters.

v3.4.26 — 6 July 2026

Major CPU speed fix + GPU watchdog

• Up to 5× faster generation on CPU-only machines. The engine previously launched with every logical core, which on modern hybrid Intel CPUs (P-cores + E-cores) forces the fast cores to wait on the slow ones. A 30-post campaign that could take ~13 hours on a no-GPU PC now finishes in ~2–3 hours. GPU machines are unaffected (~10–15 minutes).
• New GPU health watchdog: if the Vulkan engine process dies right after launch (broken driver, missing runtime), the app now detects it within seconds and automatically restarts on the CPU engine — no more indefinite spinner on machines with faulty GPU drivers.
• Generalised the "an/a" article-grammar guard from a fixed word list to a rule that covers any consonant word, with correct exceptions (an hour, an honest review, an SEO audit).

v3.4.25 — 13 May 2026

Eleven content-quality fixes layered on the GPU-capable line

• All 11 quality fixes from the production test cycle: hook varies per post, body sections deduplicated, headlines cleaner, friction vocabulary matches sector, hashtags multi-tag, plus six additional content guards (Bug A v2, Bug C extension, Bug D, Bug E, Bug F Local, Bug G Local).
• GPU acceleration for customers with a 4 GB+ discrete GPU (NVIDIA / AMD / Intel — all via Vulkan, one binary, no per-vendor build).
• CPU-only customers get behaviour identical to the proven v3.4.12.2 14-microarchitecture shard bundle. If the Vulkan binary fails to spawn (driver issue, missing vulkan-1.dll), the app automatically falls back to CPU. No customer is ever worse off than v3.4.12.2.
• All four SKUs (Local Solo, Local Pro, Global Solo, Global Pro) re-shipped as ~3.50 GB single-file Inno installers.

v3.4.22 — 8 May 2026

GPU dual-binary architecture + critical detect_gpu() fix

• New dual-binary architecture: at app launch, detect_gpu() queries Win32_VideoController via PowerShell Get-CimInstance. If a 4 GB+ discrete GPU is present (NVIDIA / AMD / Intel) → spawns the Vulkan llama-server.exe (5–20× faster generation). Otherwise → spawns the proven CPU binary.
• Fixed a 1-byte off-by-one in the earlier v3.4.14 detection: Win32 AdapterRAM is a uint32 that caps at exactly 4,294,967,295 for any GPU, so a "≥ 4 GB" threshold filtered out every discrete card on the planet. Threshold lowered to 3 GB; integrated GPUs (~128 MB) and Microsoft Basic Display Adapter (~0) still correctly fail through to CPU.
• Vulkan binary is llama.cpp release b9049 bundled alongside the existing CPU binary; total installer size grew ~100 MB.

v3.4.17 — 8 May 2026

Brand-protection layer (PERSONA + Twitter Short guard)

• Added an explicit Brand Positivity Rule to every prompt's PERSONA preamble: the customer’s own brand is always the trusted expert solution, never the antagonist. Industry frictions describe generic problems that other providers cause — the customer’s brand resolves them.
• Fixed a sub-bug in the Twitter Short generator that could produce brand-attack tweets like “Jindo’s pricing is NOT transparent” — rewrote P_X_TWEET to enforce positive-promotional voice.
• Per-platform anchor pools expanded from 5 → 30, with newsletter cycling independently so 30-post campaigns produce zero verbatim repeats.
• Two new QA guards: headline doubling detection (no “in X in X”) and brand-attack detection (no {Business} as antagonist subject).

v3.4.16 — 9 May 2026

Three blocker fixes from the production test cycle

• Slug uniqueness: 30-post campaigns no longer write three identical posts — new pre-write check rotates through alternate headline frames if a slug would collide within a campaign.
• Sector-aware tone: coaching / fitness / therapy / consulting / wellness sectors no longer get enterprise-procurement vocabulary like "procurement", "vendor", "RFP". The B2B SaaS path is unchanged.
• QA gate retry: posts whose headlines fail validation now retry once before writing, then fall back to a safe template — never ship a status: fail headline.

v3.4.12.2 — 7 May 2026

Five content-quality fixes (the CPU safe-baseline)

• Hook generation switched from JSON-mode to plain prose — eliminates the templated-fallback hook on every post.
• Body sections deduplicated in extract_body_section() — no more duplicate ## blocks within a single post.
• Headlines stronger: rejects [Adverb-ly] [Verb] tails (catches "Inadequately Address") and 4+ consecutive lowercase words (catches sector-token-pasted-twice headlines).
• Friction prompt vocabulary lock — no more SaaS / manufacturing example words leaking into coaching / services copy.
• Hashtag stop-word filter + multi-tag output (LSI + query + sector) — no more #OnlineCoachingAnd artefacts.
• This release is preserved as the CPU-only rollback baseline; v3.4.22+ ships layer the GPU dual-binary on top.

v3.4.5 — 5 May 2026

Path X subprocess refactor

• Bundled llama-server.exe with every installer. Customers no longer need to install Ollama separately — the app launches its own inference subprocess on first run.
• Reduces install friction from a 3-step setup to a 1-step setup.
• All four SKUs (Local Solo, Local Pro, Global Solo, Global Pro) rebuilt with the new bundled runtime.

v3.4.0 — 1 May 2026

4-SKU launch — Local & Global, Solo & Pro

• Two product axes: Local (postcode / area / landmarks for brick-and-mortar) and Global (B2B / SaaS / industry pick).
• Two tiers: Solo (full content engine) and Pro (adds three workflow tools: styled XLSX calendar, bulk social CSV for Buffer / Hootsuite / Later, and the in-app SEO DATA tab).
• One-time pricing locked: $147 / $247 / $347 / $797.
• 30 differentiated posts per campaign, 17 file formats per post, full schema graph (FAQPage, Speakable, BlogPosting, LocalBusiness or Organization+Product depending on tier).

v3.3.x — April 2026

Pre-launch hardening

• Two-pass generation pipeline with structural wipe pass for higher per-post quality.
• WordPress XML export and 1-click WordPress copy.
• Save / Open / Resume campaigns with full snapshot.
• PDF SUMMARY campaign report.
• Calendar exports (CSV + ICS).

All dates in UTC+9 (KST). Older release notes available on request — hello@lockstack.net.