Repository navigation
0.8.1: Local AI recovery, catalogue and offline checks moved from 0.8.0 #379
Description
Moved from 0.8.0 on 14 Oct after the 0.8.0-alpha.16 Linux runs: these Local AI checks have no named test in #348 and no installed run yet. Each line links to the item it came from.
Done means
- LA-06: Activate and switch without restarting Home #100-2: model replacement renews file grants and rechecks memory (confirmed engine exit is already covered by
local_llama_stop_confirms_closure_and_unblocks_offer_refresh). - LA-06: Activate and switch without restarting Home #100-3: failed shutdown, orphan children, cancellation and stale activation recover without duplicate work or lost run ownership.
- LA-07: Make model state and recovery clear #101-1: the UI distinguishes download, preparation, ready, busy, running, stopped and actionable failure (alpha.16 Linux showed downloading, Finishing setup, ready and Stopped; busy was not checked).
- LA-07: Make model state and recovery clear #101-2: model selection, Keep and Remove have clear effects and preserve other users' protections.
- LA-07: Make model state and recovery clear #101-4: preparation Cancel/retry releases reservations; reload and expired-session recovery keep chats and drafts and reconnect to the same run.
- LA-07: Make model state and recovery clear #101 contract 1, 2, 4, 5: retained content shown apart from readiness, more than 64 attempts, unknown-settlement owner recovery, reload/expired-session reconnect.
- LA-08: Preserve model state through catalogue and Runtime updates #102-1, LA-08: Preserve model state through catalogue and Runtime updates #102-4, LA-08: Preserve model state through catalogue and Runtime updates #102-5: signed catalogue successor and bad signatures, removed/changed/incompatible entries, missing/corrupt catalogue, interrupted update and Undo.
- LA-09: Verify the local model journey on all targets #103-2: with external routes denied, an established Home signs in, replies and cold restarts with zero model-fetch or hosted traffic.
- LA-09: Verify the local model journey on all targets #103-3: receipts bind source, built, installed and running artifacts, model identity and observed UI behaviour.
Activity
andersalm commentedon Oct 14, 2026
RC finding (alpha.16, 24 GB Mac under memory pressure): Qwen reply refused "Not enough free memory" while Get said "Meets this model's minimum memory requirement". Refusal was correct (pressure warning, swap ~96% used); 0.8.0-alpha.17 only rewords the Get text.
0.8.1: show Runtime's needed vs free numbers in the refusal, an explicit "Wait and retry" with Cancel, "Answer with " only when it fits, and an advisory picker mark. Thresholds unchanged.
andersalm commentedon Oct 14, 2026
Also for 0.8.1 (RC finding, #84 comment 6169389408): inside a container with its own cgroup namespace (Docker default), model Get is refused because Runtime cannot see the host memory limits (host_memory.rs cgroup ancestry check). Refusal is safe; 0.8.1 should explain it or read the container limit. LXC not tested.
andersalm commentedon Oct 14, 2026
Anders (14 Oct): Marketplace Models looks like its own app: own sections and cards (.shared-model-card), a detail modal that repeats publisher/DID three times and shows DID, CID, byte count, Refresh and Manage storage by default. 0.8.1 plan: models use the standard store-row cards and the normal capsule modal (About, one "Verified publisher · size" line, one memory-fit line, Get/Retry/Open), technical facts behind Details; keep the CI journey selectors (model-detail, use, Preparation progress) and keep Refresh reachable for ci-installed-home-journey.mjs:340-345.
andersalm commentedon Oct 19, 2026
This issue is the single place for the moved lines from #100, #101, #102 and #103 (comments there point here). Tick them here.