[GH-ISSUE #470] Add Achbit: Computer Use and a shared agent loop for DeepSeek #136

New issue

Open

opened 2026-10-26 09:45:52 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 09:45:52 +00:00

Owner

Originally created by @Achbite on GitHub (Oct 25, 2026).
Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/470

Please add Achbit to the DeepSeek agent integration list. Achbit is an MIT-licensed coding agent with a desktop GUI, TUI and CLI, combining project tools and Computer Use in one persistent conversation.

What this integration adds

From code to observed interface behavior. DeepSeek can read and edit files, run commands, inspect a page, interact with its controls and capture the result in the same task. CLI/MCP handles programmatic work; the embedded browser and native desktop tools handle interface work. The user can review the implementation and visual result together.
One Session loop across interfaces. GUI, TUI and CLI use the same model turns, context, journal, waiting and cancellation state. Changing interface does not create a separate automation history. The Kernel owns tool execution and result facts; Session uses those facts to continue the task.
Tools selected for the run. Plugins are discovered and activated when needed, contributing their tools and guidance to the current run. This extends a small core toolset without adding a second loop. Activation retains the existing permission boundary.
Dependencies and outcomes stay explicit. Eligible file calls run with bounded concurrency according to Kernel-provided dependencies. Long-running commands keep their running state, and dependent steps wait for completion. Native actions use observations tied to the app, window and run; a failed post-action capture does not erase the dispatch result or authorize replaying the input.

A reproducible first task

The Launch checklist project contains a standalone HTML page and prompt. A recorded DeepSeek run selected two tasks, verified 2 / 4 and 50%, checked the Completed filter, reloaded and confirmed persistence. The run summary and capture are included. This demonstrates the embedded browser on macOS; external native apps and Windows were not exercised by that example.

The proposed contribution includes English and Simplified Chinese guides, an English GUI screenshot and entries in both README tables. Setup uses the Releases page, deepseek-flash, Responses API, 1M context, a 384K output ceiling, Thinking, max effort and image input, following the current DeepSeek documentation.

Implementation references: Session loop, tool catalog, Computer Use.

Originally created by @Achbite on GitHub (Oct 25, 2026). Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/470 Please add **[Achbit](https://github.com/Achbite/Achbit)** to the DeepSeek agent integration list. Achbit is an MIT-licensed coding agent with a desktop GUI, TUI and CLI, combining project tools and Computer Use in one persistent conversation. ## What this integration adds - **From code to observed interface behavior.** DeepSeek can read and edit files, run commands, inspect a page, interact with its controls and capture the result in the same task. CLI/MCP handles programmatic work; the embedded browser and native desktop tools handle interface work. The user can review the implementation and visual result together. - **One Session loop across interfaces.** GUI, TUI and CLI use the same model turns, context, journal, waiting and cancellation state. Changing interface does not create a separate automation history. The Kernel owns tool execution and result facts; Session uses those facts to continue the task. - **Tools selected for the run.** Plugins are discovered and activated when needed, contributing their tools and guidance to the current run. This extends a small core toolset without adding a second loop. Activation retains the existing permission boundary. - **Dependencies and outcomes stay explicit.** Eligible file calls run with bounded concurrency according to Kernel-provided dependencies. Long-running commands keep their running state, and dependent steps wait for completion. Native actions use observations tied to the app, window and run; a failed post-action capture does not erase the dispatch result or authorize replaying the input. ## A reproducible first task The [Launch checklist project](https://github.com/Achbite/Achbit/tree/0.6.60/examples/launch-checklist) contains a standalone HTML page and prompt. A recorded DeepSeek run selected two tasks, verified **2 / 4 and 50%**, checked the Completed filter, reloaded and confirmed persistence. The [run summary](https://github.com/Achbite/Achbit/blob/0.6.60/examples/launch-checklist/run-summary.json) and capture are included. This demonstrates the embedded browser on macOS; external native apps and Windows were not exercised by that example. The proposed contribution includes English and Simplified Chinese guides, an English GUI screenshot and entries in both README tables. Setup uses the [Releases page](https://github.com/Achbite/Achbit/releases), `deepseek-flash`, Responses API, 1M context, a 384K output ceiling, Thinking, `max` effort and image input, following the [current DeepSeek documentation](https://api-docs.deepseek.com/quick_start/pricing/). Implementation references: [Session loop](https://github.com/Achbite/Achbit/blob/0.6.60/userspace/session-core/src/local-agent/loop.ts), [tool catalog](https://github.com/Achbite/Achbit/blob/0.6.60/crates/achbit-kernel-daemon/src/local_agent_tool_catalog.rs), [Computer Use](https://github.com/Achbite/Achbit/blob/0.6.60/plugins/computer-use/README.md).

bsakaguchi referenced this issue

2026-10-26 09:47:00 +00:00

[PR #136] feat: add DeepLossless integration guide #222