keys-MiniMax-Code-CLI-Browser-Scroll-Context-Enhancement-Pack-with-Jev-Ultrafast-Integration
Enhancement pack for MiniMax Code CLI on local DGX Spark hardware, including a Jev-style local decision browser agent and context patches.
Overview
This pack collects validated enhancements for the MiniMax Code (MCode) CLI, built on a four-node NVIDIA DGX Spark (arm64) running a local GLM-5.3 model through vLLM with no cloud API. Most components are not about decisions: patches add a transcript scrollbar and a context-meter status item, a continuous context-renewal patch triggers recurring compaction after 25% utilisation, and a compaction fix repairs checkpoint failures on local vLLM with enforced limits and a watchdog.
The decision-model part is jev-ultrafast, a structured browser agent on a verified arm64 Chrome DevTools lane. A local djev-spark policy serves a /v1/systemone-compatible endpoint to pick browser actions while the local GLM model handles text, with the Syntra client for adaptive routing and feedback; cloud TypeSafe is opt-in. Playwright MCP integration and notes on running Chromium headless under AppArmor are included, along with a usage ledger. It is aimed at people running coding agents on local high-end hardware.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which browser action should the agent take next on this page?(inferred) | — |
How it works
- Run MCode with the local GLM model via vLLM
- Apply TUI and compaction patches
- Browser agent observes the page over CDP
- Local djev-spark policy answers action choices via /v1/systemone
- GLM handles text; usage is recorded in a ledger
Strengths
- Fully local stack with no token cost
- Decision endpoint compatible with Jev's API
- Documented arm64 browser setup
Caveats
- Specific to DGX Spark arm64 and MiniMax Code
- Some fixes still pending verification after restart
- Mixed collection; decision part is one component
At a glance
- Type
- MCP server
- Status
- prototype
- Runtime
- MiniMax Code CLI patches, MCP, local vLLM
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Validated enhancement pack for MiniMax Code CLI on arm64 DGX Spark + local GLM-5.3-EXL3: TUI scrollbar + context-meter patches, compaction repair for local vLLM, Jev Ultrafast + Playwright MCP browser integration, real usage ledger.
hellogumbo/awesome-jev
Frequently asked questions
Where does a decision model come in?
In jev-ultrafast, a local djev-spark policy picks browser actions through a /v1/systemone-compatible endpoint.
Does it need TypeSafe?
No, cloud TypeSafe is opt-in; the default is local.
What hardware is it for?
NVIDIA DGX Spark (arm64) running a local GLM model.
Related projects
RAM-efficient Rust coding-agent harness, listed as using Jev decisions for memory pruning, browser navigation and voice routing.
ServiceproductionRustLocal browser agent that picks each step's operation and element with a Jev-shaped model such as a fine-tuned Laya, no cloud needed.
MCP serverbetaPythonOpenCode V2 port of fast-jev-compaction that asks Jev or a local Jev-style model which old tool outputs still matter.
ProjectprototypeTypeScriptSelector-free pytest UI tests for saucedemo using a jev-ultrafast-style agent run fully locally with laya-browser and a small Ollama model.
ProjectprototypePythonOpen 2B browser-agent decision model fine-tuned from MiniCPM5 that serves the System One protocol for jev-ultrafast.
ProjectresearchPythonZero-build local playground for Jev with browser-agent, intent routing and context-pruning presets, including a Playwright agent.
AppbetaJavaScript