Devotd Github
ReleaseAI agents

wangzan101/mika v0.1.29 — Context compression calibration and Claude refusal fix

wangzan101/mika v0.1.29 calibrates the context compression decision for long tasks with actual usage, and fixes the mid-task refusal problem of the latest Claude models as well as the delivery of the server message when a model cannot be found.

Repositorywangzan101/mika ↗

wangzan101/mika v0.1.29 calibrates the context compression decision for long tasks with actual usage, and fixes the mid-task refusal problem of the latest Claude models as well as the delivery of the server message when a model cannot be found.

GitHub description: Mika — 你手机上的私人 AI 助理,替你操作 App、上网、把事办成。An on-device AI agent for Android that operates apps and the web to get things done for you.

stars
★ 2.7K
Forks
431
Language
Kotlin
License
Apache-2.0
Latest release
v0.1.29
Read original source ↗View on GitHub ↗

wangzan101/mika v0.1.29 is an on-device AI agent update for Android, released on November 1, 2026. The project is a personal AI assistant that operates apps on a phone and handles the user's tasks over the web. This release adjusts the point at which context is compressed during long tasks: it is first judged by token count and then calibrated by actual usage. It fixes the problem where the latest Claude models refuse partway through a task, and changes the behavior so that the server's original message is surfaced as-is when a model cannot be found. The source of these changes is the repository's v0.1.29 release.

Context compression refers to reducing the content when a conversation grows long, so as to adjust the context the model handles. The key point confirmed in this change is that actual usage is now reflected in the compression decision, which previously relied on token count alone. The criterion for setting the compression point in long tasks is therefore calibrated, but the release information does not explain how the actual usage is computed. The change that surfaces the server's original message verbatim when a model cannot be found is included as well.

The provided release information contains no accuracy comparison for context compression, no per-model refusal conditions, and no performance figures before and after the change. Nor is any detailed explanation given about the situations in which each fix applies. Therefore it cannot be confirmed how much this release changes the success rate or processing speed of specific tasks. The confirmable scope is the three fixes listed in v0.1.29, and it is difficult to assert any effect beyond that.