0.5.x Release Notes

The 0.5 line was the first complete shape of "writing mode": story structure, world building, the AI assistant and the AI-tell detector all came together and settled down here.

0.5.7-canary - 2026-07-06

A broad release — the writing workbench, story structure, AI observability and the colour system all changed.

• You can finally see what the AI actually sent. A new request trace viewer shows, for every request, which model was used, how many tokens it cost, how long it took, the time to first token, and the exact content that went out. There's an entry point in the top bar. Traces stay out of shareable log bundles by default, so prompts and manuscript text don't leak.
• Story structure became two trees: one for how chapters are arranged (story → volume → chapter → prose), one for how cause and effect run (story → phase → thread → scene). Scenes attach to chapters by chapter reference. The chapter brief became a first-class thing, holding that chapter's goal, point of view, information control, pacing, opening and closing, and what must not be written.
• Theme system v2.1: colours collapsed to 36 variables across 8 built-in themes, plus a custom theme editor with live preview, regenerate-from-primary, JSON import/export and an on-screen colour picker.
• Groundwork for the AI's three modes (normal, read-only discussion, read-only planning), so that "read-only" genuinely is, and write operations go through your approval.

0.5.6-canary - 2026-07-03

• Fixed container deployments pulling an outdated image: a test-build installer no longer defaults to "latest stable".
• Admin creation commands are now listed per install method, so container users get the in-container script instead of accidentally running against host source.
• Clearer message when the database client is missing: running from source auto-generates it, while packaged builds tell you to pull a matching image rather than generating it on your machine.

0.5.3-canary - 2026-07-01

• Fixed packaged builds failing to load files generated at runtime. Those paths are only known while running, and the bundler couldn't resolve them; they now go through a dedicated loader.
• llmlint moved to its own development repository. The copy inside NeuroBook is now purely a runtime snapshot. The sync script strips repository metadata, dependency folders and evaluation corpora, so none of that ends up in your workspace.
• Corrected an llmlint evaluation metric: the document-burden score now counts deduplicated hits, so one sentence matching several rules no longer inflates the number.

0.5.2-canary - 2026-07-01

• World-engine config files load more reliably. You still write TypeScript, but it is transpiled before loading, so it no longer depends on your environment being able to run TypeScript directly.
• Temporary files left behind by an interrupted run are cleaned up and won't affect the next load.
• The AI's structured question panel is taller, so it no longer feels cramped when there are many options.
• llmlint docs corrected the install and run instructions: install via the skills command, and run under Bun or Node + tsx. Running the TypeScript sources under plain Node is not a supported path.

0.5.1-canary - 2026-06-30

• The project list opens faster. Added a short cache and a layered stats cache, and the main entry point no longer issues a query that bypasses them. After startup the caches warm up gradually in the background — this doesn't block startup, and no single unlucky request has to carry a full warm-up for everyone.
• The llmlint CLI feels like a real tool now: it accepts globs such as manuscript/**/*.md and !drafts/**, prints in colour in a terminal, and switches to plain text when piped or read by an agent.
• llmlint now declares and installs its own runtime dependencies instead of hoping the root project happens to have them.

0.5.0-canary - 2026-06-30

The close of writing mode v1 — story structure, world building, the AI assistant and AI-tell detection all stabilised.

• Story planning no longer lives in a separate screen. You do it inside the normal writing view. The structure got simpler too: only "scene" remains, and the old storyline/beat machinery is gone. Each scene hooks into your world building through when it happens, where, and who's in it.
• Scenes connect to world building. While writing a scene you can look up the relevant settings and each character's state at that moment. The workbench lets you edit those links, pick characters and locations, and view the context. The AI also gets a brief describing how this chapter should go.
• AI configuration is more reliable. Compiled configuration changed how it's stored, guaranteeing a stale or failed build can never be used. The settings page shows build status, and none of this slows the editor down any more.
• Smoother AI tool interaction. "The AI asks you a question" became its own mechanism, and reading files, editing files, approvals and planning mode all recover better after an interruption.
• llmlint got stronger: rules load automatically from a directory, whole-manuscript checks were added, and the CLI can scan an entire folder and auto-fix obvious mechanical tells like zero-width characters and duplicated punctuation.