ContextOS 2.7.1 — Focused code, reproducible results

ContextOS 2.7.1 — Focused code, compact receipts, reproducible results

ContextOS keeps AI coding conversations smaller by returning the code an agent needs and keeping command logs outside the conversation. This release fixes precise reads and installation verification, and replaces broad token-saving claims with a public benchmark.

What you can measure

Five local runs using o200k_base, median response-text tokens:

Controlled scenario	Native → ContextOS	Output reduction
Smaller repository whole-file read → symbol	1,459 → 336	76.97%
Large repository whole-file read → symbol	43,031 → 280	99.35%
Repeated unchanged symbol read	165 → 36	78.18%
Synthetic 1,000-line failed build	14,016 → 150	98.93%

Tradeoffs are published too: exact native slices were cheaper on first reads (165 vs 280 tokens in the large-file control); the compact tool, server instructions, and skill added about 2,504 tokens when loaded. These figures do not measure total-task tokens, model reasoning, cache discounts, or provider bills.

Read the benchmark and controls · Inspect every sample

Fixes that matter in daily use

• Parallel inspection preserves explicit symbols and line ranges.
• Non-contiguous code slices retain their original source line numbers.
• Installation refreshes Codex's registered version and verifies installed bytes.
• Read-only install checks reject stale versions and bundles; inactive historical caches stay intact.
• Plugin builds and installs include parser runtime assets; Python methods retain AST hashes and call information outside the repository.
• The benchmark checks content freshness, diagnostic preservation, and chain failure handling across five runs.

Try it

Requires Node.js 22+. For the tested source-install route:

git clone --branch v2.7.1 https://github.com/yubinbin32-ops/ContextOS.git
cd ContextOS
npm ci
npm run plugin:build
codex plugin marketplace add .
codex plugin add contextos@contextos-development
npm run plugin:install
npm run plugin:install:check

Start a new chat after updating an installed plugin. Other MCP hosts can launch the same local server; see the README.

Reproduce the measurements:

npm run benchmark:context -- --runs 5 --output .contextos/benchmarks/my-run.json

For a complete standalone MCP runtime, download contextos-plugin-2.7.1.zip, extract it, and configure your host to run node /absolute/path/contextos-plugin/server/contextos-mcp.mjs. It includes the parser WASM and grammars; Node.js 22+ is required. SHA256SUMS-plugin.txt covers the plugin archive and benchmark assets.

Release assets are listed below with checksums. Desktop availability is determined by the attached assets; the performance figures describe the MCP runtime on macOS arm64. Optional Cloud collaboration and Micro are outside this benchmark.

中文速览

本次更新修复了批量读取忽略区间、多区间行号错误，以及安装检查漏掉旧版本/旧文件的问题，同时补齐安装中的解析器资源，避免 Python 方法哈希与调用信息退化。公开测评包含五次运行、精确原生对照与负收益场景：两个源码整读改切片场景的返回 token 分别减少 76.97% 和 99.35%，但首次精确读取和极小任务可能更贵。数据不代表完整开发任务的模型账单节省。

中文 README · 完整测评

If this helps your workflow, star ContextOS and share a real task benchmark. Bug reports and measurements where it costs more are welcome.

Real development follow-up (2026-09-30)

A real installer-repair pilot now records actual model usage and context occupancy with gpt-6-luna / max. Native completed and passed six independent checks; both OS runs stopped before completing. Complete-task savings remain unproven.

Read the full pilot and sanitized accounting or review the candidate fixes. The candidate repairs search scope, precise inspection, installer integrity and edit/ownership recovery; all 408 tests and plugin smoke pass. Its final recovery fixes await another capped model evaluation. The runner defaults to request, token and time limits.