Crash when using local model

[Vitalii953]

Xiaomi 15, 16GB RAM, Snapdragon 8s gen4, used the bigger local model. crashed without a backtrace. I suggest implementing logs to see the root of this problem and fixing the underlying issue unless it's not strong enough hardware

[KTheMan]

Similar issue on Samsung Galaxy Z Flip7. I haven't tried cloud, but get the same issue seemingly when either of the models attempt to load. Switching between them didn't resolve the issue. Using 0.6

[ithiria894]

I pushed the same local startup hardening into v0.6.1:\n\nhttps://github.com/agents-io/PokeClaw/releases/tag/v0.6.1\n\nThe point of this pass was to handle the "premium phone still crashes on local model startup" case better. If the GPU/OpenCL path is unstable, PokeClaw should recover into a CPU-safe mode instead of repeatedly trying the same bad backend.\n\nCould you update to 0.6.1 and try the local model again on your Xiaomi 15? I want to keep this open until we get a real-device confirmation.

[ithiria894]

I pushed another local-startup hardening pass in v0.6.2:

https://github.com/agents-io/PokeClaw/releases/tag/v0.6.2

This is the same crash family as #27, but I also added a release-build fix for the LiteRT-LM JNI wrapper classes. If the Xiaomi 15 crash was coming from that path, 0.6.2 is the one I want you to try.

If you can, update to 0.6.2 and test the local model again.

[Vitalii953]

okay, so it worked, i had to manually enable restricted settings for certain permissions. Model loaded properly. It was quite slow, but working. What i saw though: every task refreshes the context of conversation, multiple chats aren't possible due to "agent already running" - need to delete manually, which is probably a bug, and oftentimes the model unselects itself for some reason, in chat it says "Error: Attempt to invoke a virtual method..." yada yada yada.

[Vitalii953]

it also has a persistent overlay for some reason until i remove every conversation and close it from recents. If I don't delete every conversation and delete it from recents i still get the overlay

[ithiria894]

I shipped another round of fixes in v0.6.3 aimed at the follow-up problems you saw after the crash stopped:

• task/session cleanup is tighter
• overlapping tasks are rejected cleanly
• floating overlay cleanup is safer
• accessibility state is reported more truthfully

Please update to v0.6.3 and retry the exact Xiaomi flow:
https://github.com/agents-io/PokeClaw/releases/tag/v0.6.3

What I most want to know is whether these are still happening after the update:

• conversation gets wiped after a task
• "agent already running" when it shouldn't
• model unselects itself
• overlay gets stuck

If any of those are still there on 0.6.3, reply here and I'll keep chasing it.

[ithiria894]

We shipped another hardening pass in v0.6.4: https://github.com/agents-io/PokeClaw/releases/tag/v0.6.4\n\nThis one adds a safer local startup path plus GPU re-arm logic for devices that were getting stuck on CPU after an older local init failure. It also expands the debug report so we can see the backend state more clearly.\n\nPlease update to v0.6.4 and try the same local-model flow again. If it still breaks, send a debug report from Settings → About → Share Debug Report so we can compare the backend state on your device.

[ithiria894]

I dug a bit deeper on the local GPU path before shipping another patch.\n\nWhat I found is that our old check was too optimistic: on at least one real device, the LiteRT engine can initialize on the GPU path, but the first real inference still fails and drops back to CPU. So we were treating "engine started" as "GPU works", which is not actually true.\n\nI just tightened that logic on a branch. Going forward we only treat GPU as verified after a real inference succeeds, not just engine init.\n\nThat probably will not magically enable GPU on every device, but it should stop PokeClaw from claiming a broken GPU path is healthy.\n\nFor Xiaomi / other Snapdragon devices I still want a fresh debug report if the local model keeps crashing or silently falls back. Settings -> About -> Share Debug Report is the fastest way for me to compare backend state across devices.