Devotd Github
NewAI agents

wuyhong715/rikkahub-agent-pure — an Android agent built for long-running operation

wuyhong715/rikkahub-agent-pure is a derivative project that adds operational features for long-running operation to a RikkaHub-based Android agent.

Repositorywuyhong715/rikkahub-agent-pure ↗

wuyhong715/rikkahub-agent-pure is a derivative project that adds operational features for long-running operation to a RikkaHub-based Android agent.

GitHub description: RikkaHub Agent · Pure — an Android AI agent that survives long runs. Independent hardening pass on the rikkahub on-device agent: usage ledger & budgets, on-demand tools, expert sub-agents, cold memory, chained workflows. Kotlin/Compose, Shizuku/Termux/MCP. QQ group 1030362371.

stars
★ 72
Forks
11
Language
Kotlin
License
AGPL-3.0
Latest release
v0.8.4-moxw
Read original source ↗View on GitHub ↗

wuyhong715/rikkahub-agent-pure is an Android on-device AI agent based on RikkaHub. The latest release tag is v0.8.4-moxw. The provided materials do not list what changed in this release, so per-version changes cannot be confirmed.

According to the repository description and README, this project keeps the functionality of the upstream ExTV/rikkahub-agent while adding the operational features needed for long-running operation. It manages usage with a usage/cost ledger and budget, and provides tool result limits and model-driven context compression. It also includes run retries and keep-alive. It further offers on-demand tool loading, reusable specialist sub-agents, Markdown cold memory, workflows that pass data, and an encrypted secrets store, with these additional features off by default. It is written in Kotlin/Compose and uses Shizuku, Termux, and MCP together. The license is AGPL-3.0.

However, what the materials show is feature enhancement built on an existing project, and no benchmarks backing a novel independent algorithm or a performance advantage are presented. It can therefore be viewed as a project centered on usage management and tool/memory configuration for long-running operation, but execution performance or effectiveness is hard to judge from the provided evidence alone.