Skip to content

0.8.x Release Notes

The theme of the 0.8 line was making installation and deployment solid. 0.8.0 rewrote how long AI conversations load; nearly every patch release after it went into installing, updating and container shutdown — the parts nobody enjoys working on that ruin everything when they break.

Manager 0.1.0-canary.30 - 2026-07-25

Installer-only release; the app itself did not change.

  • start gained --no-health-check: it skips the startup probe and the automatic browser launch, for when the service is actually up but the check hangs. Migrations and data checks still run — it just doesn't wait for the probe.
  • Windows portable only; every other install method rejects the flag outright.
  • If you already unpacked the 0.8.19 portable build, run this from the portable root and then open http://127.0.0.1:3000 yourself:
cmd
.runtime\bin\bun.cmd x --bun @notnotype/neuro-book-manager@0.1.0-canary.30 --root "%CD%" start --no-health-check

This borrows the newer installer temporarily; it does not replace the one bundled in the zip. Double-clicking Start Neuro Book.cmd next time behaves exactly as before.

0.8.19-canary - 2026-07-20

The run of container fixes that started at 0.8.13 finally passed everything and shipped in full here.

  • Fixed Docker and Podman disagreeing about whether two images are the same image, which made a correct install report a version mismatch.
  • Verified across five install methods, both CPU architectures, upgrades from older versions, stop and restart, and interrupted-operation recovery.

0.8.18-canary - 2026-07-20

  • Unified how the two container engines look up a container ID, so Podman no longer fails on an incompatible command.
  • The container config NeuroBook generates now allows exactly one app service. If you need extra services, create a separate config of your own rather than editing NeuroBook's — otherwise diagnostics will reject the instance.

0.8.17-canary - 2026-07-20

  • Fixed the Podman health check using a command older podman-compose versions don't support. It now reads the image straight from the config file NeuroBook generates, instead of relying on vendor-specific extensions.

0.8.16-canary - 2026-07-20

  • Fixed the Podman health check reading a field only Docker provides.
  • Container images now build for x64 and ARM64 in parallel on native machines. ARM64 went from 30 minutes to 7, and the whole release pipeline dropped from nearly an hour to 27 minutes.

0.8.15-canary - 2026-07-20

  • Fixed containers being deleted when you stopped the app under Podman. In older podman-compose, "stop" also removed the container, so afterwards diagnostics, restart and recovery had nothing left to find. The container now survives a stop.

0.8.14-canary - 2026-07-20

  • Fixed containers refusing to stop. The shutdown signal never reached the process actually doing the work, so every stop sat through a 10-second wait and then got force-killed. Shutdown is clean now.

0.8.13-canary - 2026-07-20

  • Diagnostics after stopping a container no longer report a normal shutdown as a crash.
  • Podman now always uses podman-compose, instead of silently routing to Docker when both are installed on the same machine.

0.8.12-canary - 2026-07-20

  • Fixed a self-check script inside the release pipeline. Nothing user-visible changed.

0.8.11-canary - 2026-07-19

  • Fixed the installer occasionally reading command output incompletely on macOS, which made successful commands look like failures.

0.8.10-canary - 2026-07-19

  • Fixed the Windows portable build failing while packaging its bundled runtime (directory entries in the archive were treated as invalid paths).
  • The release pipeline gained a single preflight stage: all contract tests run before spending an hour building five platforms.
  • Publishing a new test build now cancels any still-running older test build instead of letting them pile up.

0.8.9-canary - 2026-07-19

The largest release in the 0.8 line — a batch of installation and permission fixes.

  • Windows portable login failure. The database path used to follow whichever directory you launched from, so it looked for the database in the wrong place and login failed immediately. The path is now resolved from the data directory, always.
  • Updates became ledger-based. Every step records what it intends to do before touching anything, then records the result. A power cut or a failure mid-update can be rolled back exactly, instead of leaving a half-old half-new instance.
  • Downloaded runtimes are verified before use. Bun, ripgrep and Git are checked for integrity, executability and real version in a staging area first; if verification fails, the copy you're currently using is left alone.
  • Stricter AI model identity. Changing the service URL or proxy now counts as a different connection, so old credentials are never silently reused. Deleting a model provider requires clearing references to it first.
  • Unified file permissions. The AI's read, write, edit, patch and memory tools all share one permission check now, and symlinks and cross-project paths are blocked consistently.

Upgrading from the 0.8.6 Windows portable build: back up the whole data/ folder, unpack the new version somewhere fresh, then copy data/ across. Do not copy the old .deploy, .runtime, .output or the launcher scripts in the root.

Fixing model references in old conversations

From this version on, an AI conversation records only "which provider, which model" rather than a full copy of the model configuration. If an older conversation can't prove which provider it originally used, the system will not guess for you — it blocks that one conversation instead of quietly falling back to a default model.

To fix them, write a mapping file pointing each affected conversation at the right provider and model:

json
{
    "mappings": [
        {
            "sessionId": 241,
            "entryId": "model-change-entry-id",
            "providerConfigId": "my-provider",
            "modelId": "my-model"
        }
    ]
}

Then do a dry run to see what would change before applying it for real:

text
bun scripts/maintenance/migrate-session-model-refs.ts --workspace-root <your workspace directory> --mapping <mapping.json> --dry-run
bun scripts/maintenance/migrate-session-model-refs.ts --workspace-root <your workspace directory> --mapping <mapping.json>

The command verifies that those providers and models actually exist in your configuration, and fails outright on a wrong or redundant entry. Back up the whole data directory before applying.

0.8.6-canary - 2026-07-17

  • Fixed one-off maintenance commands being treated as "start the service" during container installs, which left the installer waiting forever.
  • Known issue (fixed in 0.8.9): with login enabled, the Windows portable build could return a 500 when signing in.

0.8.5-canary - 2026-07-17

  • Fixed fresh installs aborting and rolling back because there were no AI conversations yet. Having no data is now a valid state rather than an error.

0.8.4-canary - 2026-07-17

  • Fixed the packaged app wrongly reporting that its built-in AI configuration was stale, which stopped both the Linux build and the Windows portable build from starting at all.

0.8.3-canary - 2026-07-17

  • Fixed containers failing to start under a normal user account, because the container tried to write to read-only files inside its own image.
  • Temporary files no longer land in the program directory; they go to the matching data directory, leaving the program directory read-only.

0.8.2-canary - 2026-07-17

  • Images in AI conversations are no longer stored inline. They used to sit in the conversation record as base64 and grew the file without bound. They now go into an attachment store, the conversation keeps only a reference, and the image is loaded only when it actually has to be sent to the model.
  • Fixed the AI looking in the wrong place for files under the Windows portable build and under custom data directories.

0.8.0-canary - 2026-07-15

This minor release rewrote how long AI conversations load and stream, so that hundreds of turns stay responsive.

  • Long conversations no longer load all at once. You get the most recent page on open, and older content loads as you scroll up. Your scroll position stays put, and a reply never gets split from the tool results that belong to it.
  • Live messages send only what's new. Previously every update resent the whole accumulated blob — tool arguments, diffs, images and all.
  • Tool cards show a summary only: path, short preview, original size. Whole file contents no longer get pushed to the browser.
  • AI images became attachment references, the same approach 0.8.2 describes.
  • Model configuration is validated strictly. Broken configuration can't reach the runtime; the settings page validates as you type and offers a one-click fix that edits your draft without auto-saving it.

A note on the gaps

0.8.1, 0.8.7 and 0.8.8 failed during the build and produced no installable files. Don't download whatever partial assets are attached to them.