Collection basis: today’s figures come from the GitHub Trending bot issue happydog-intj/ai-xiaohongshu-daily #566 (morning edition, 2026-10-16, 01:00 slot); the previous publication day’s figures come from #563 (morning edition, 10-15), with the later 10-15 snapshot #564 used as a cross-check. The primary metric is stars today. Repos with 40K+ total stars (morluto/rea 264.7K, K-Dense-AI/scientific-agent-skills 50.2K, caveman, superpowers, ruflo, ECC and others) were excluded. anthropics/claude-plugins-official (+525, 38.7K total, 1.4% of its total) was excluded as a long-standing large official directory. Yesterday’s pick ifixai-ai/iFixAi is not on today’s list and was not republished. Only one repo passed today’s criteria.
mksglu/context-mode — +902 ⭐ today (3 days in a row)
URL: https://github.com/mksglu/context-mode
Total stars: 29,689 · Daily gain: +902 (−4.6% vs. +946 on the previous publication day) · Language: TypeScript · License: Elastic License 2.0 (ELv2)
context-mode is an MCP server plus a set of hooks and plugins that keeps AI coding agents’ context windows from filling up with tool output. It holds raw output such as Playwright snapshots, GitHub issue lists and log files in a sandbox and returns only the results the model needs; the README claims a 98% reduction, for example 315 KB down to 5.4 KB. Its architecture has four main parts:
- Six sandbox tools (ctx_execute, ctx_batch_execute, ctx_execute_file, ctx_index, ctx_search, ctx_fetch_and_index) that run code in isolated subprocesses across 12 language runtimes and pass only stdout into context
- A session store that logs file edits, git operations, errors and user decisions in SQLite and retrieves them with FTS5 and BM25, so agents recover their working state after compaction or –continue
- A knowledge base that chunks Markdown by heading into a per-project SQLite database and fuses Porter-stemmed and trigram search with Reciprocal Rank Fusion
- Hooks (PreToolUse, PostToolUse, PreCompact, SessionStart and others) for routing and event capture
It supports 17 platforms, including Claude Code, Gemini CLI, Cursor, Codex CLI, VS Code and JetBrains Copilot, Kiro and Qwen Code (Zed and Antigravity IDE are MCP-only), and adds authenticated CLI passthrough for gh, aws and kubectl plus per-tool savings statistics. A steady gain of around 900 stars a day for three days reflects demand from teams for whom token costs and lost context have become the main bottleneck as agent sessions grow longer.
Practical takeaway: Adding it to in-house coding agents that handle large logs or browser snapshots can cut token spend and prevent context loss in long sessions (check the ELv2 terms before reselling it as a hosted service).
Tags: #ContextWindow #MCP #CodingAgents #ClaudeCode #TokenOptimization #AgentMemory
Leave a Reply