Skip to content

Releases

v0.3.1 - docker run flowenginecloud/agentqs

Choose a tag to compare

@Ami3466 Ami3466 released this 10 Sep 15:14
docker run -d --name agentqs -p 3000:3000 -v agentqs-data:/data \
  -e SESSION_SECRET=$(openssl rand -hex 32) flowenginecloud/agentqs

Keep SESSION_SECRET stable across restarts or you get logged out. Open http://localhost:3000.

The image has a short name now. Every published tag is mirrored to Docker Hub as flowenginecloud/agentqs, so the quick start is a name you can remember instead of a path you have to copy. ghcr.io/ami3466/agentqs is the same image, digest for digest — anything already pinned there keeps working.

The mirror copies, it never rebuilds. It takes the exact index GHCR published, after that index passed the arch check and booted under the smoke test, so Hub cannot drift into shipping something the tests never ran. It re-checks both architectures on the Hub side anyway, because an amd64-only tag is the bug this project shipped for four versions.

Docs no longer point at a broken pull. The version-tag example said :0.2.2, which is amd64-only and dies on an Apple Silicon Mac.

Full log: v0.3.0...v0.3.1

v0.3.0 - a version tag you can pull on any machine

Choose a tag to compare

@Ami3466 Ami3466 released this 10 Sep 03:53
docker run -d --name agentqs -p 3000:3000 -v agentqs-data:/data \
  -e SESSION_SECRET=$(openssl rand -hex 32) ghcr.io/ami3466/agentqs:0.3.0

Pin 0.3.0, or track latest. Keep SESSION_SECRET stable across restarts or you get logged out.

The image

Version tags are multi-arch now. Every tag up to 0.2.2 was amd64-only, so pinning a version on an Apple Silicon Mac died with no matching manifest for linux/arm64/v8 — only latest had been rebuilt after the fix. 0.3.0, 0.3 and latest all carry linux/amd64 and linux/arm64, and CI fails the release if either is missing.

The container reports its own health. A HEALTHCHECK probes /api, the discovery manifest — unauthenticated and it reads nothing from the record or SQLite, so a probe every 30s costs nothing. The 40s start period is deliberate: a cold Next boot answers nothing before that, and without it a healthy container reports unhealthy on every start.

The published index carries its own metadata. Title, description, source and revision were on the per-arch children but never on the index that registries and UIs actually read, so a tag looked metadata-less. They shipped double-scoped on the first attempt (manifest:org.opencontainers.image.title — a key nothing reads), so the release now asserts a plain key survived on the index rather than trusting the scoping.

Since 0.2.2

Reads cost what is on screen, not what the record weighs. Coverage stopped recomputing per tab, the log ships a list instead of 200 captures, read views are cached by stamp and ETag, and no request can rebuild the cache. Idle models release their memory instead of holding the floor.

Dropped files behave. A file dropped twice is recognised by content and returned to the queue rather than failing; a photo is recognised by name as well as type, so an iPhone .HEIC or an UPPERCASE.JPG arrives as a picture instead of being read as text; PDFs are data, not a dead end.

Slack captures where the messages actually went, probing each conversation type separately so one missing scope no longer hides the rest, and a backlogged message keeps the timestamp Slack gave it.

Full log: v0.2.2...v0.3.0