采用 MIT 许可、本地优先的 Python 工具，将性能分析、基准测试与追踪转化为智能体可据以行动的边界化证据——通过 `uv` 运行，或作为 DeepSeek Harness 的 MCP 服务器使用。
DSH 适配
原生运行时
作者声明
安全审计
未审计
最后核验
2026-09-01
许可证
MIT
01它能帮你完成什么？
Trace and profile hotspots in application and native code, GPU kernels, and inference stacks from a coding agent
bounded, inline runtime evidence — profiler / benchmark / trace output — without a durable control plane
Developers and coding agents (including DeepSeek Harness) that need local-first performance evidence
Capture runtime evidence from a live command or an existing native artifact inside an agent loop
process-lifespan evidence with optional preservation to `<project>/.flameox`
Performance engineers who want agents to burn down hotspots with explicit, auditable evidence
02如何接入 DeepSeek Harness？
前置条件
- Python project managed with `uv` (the README installs extras via `uv sync` / `uv run`)
- A local target: an existing native artifact (with its exact path and format) or a live command to benchmark
安装步骤
- 01
Sync the optional extras you need: `uv sync --extra dev --extra memory --extra trace --extra cpu`
- 02
Discover capabilities for an intent: `uv run flameox capabilities discover --intent "CPU hotspots"`
- 03
Analyze an existing artifact: `uv run flameox analyze artifact.preview /absolute/path/to/artifact.json`
- 04
Or capture live evidence: `uv run flameox capture --provider direct -- python benchmark.py`
- 05
Run the MCP server for an agent: `uv run flameox mcp serve --project-root "$PWD"` (the MCP server fixes its project root at startup)
验证接入成功
- Print the stdio client configuration with `flameox setup` to confirm the MCP wiring
- Confirm capabilities are discovered for your intent before analyzing or capturing
回滚
- Stop the `flameox mcp serve` process and drop the `--project-root` MCP config printed by `flameox setup`
- Version 0.2 has no SQLite control plane, so removing `<project>/.flameox` leaves no stale state
03DSH 适配与能力边界
MCP server (`flameox mcp serve`) exposing profiler / benchmark / trace capabilities to DeepSeek Harness agents; also runnable as direct `uv` CLI commands
Capabilities discovery
an intent such as "CPU hotspots"→the matching profiler / benchmark / trace capabilities for an agent to use
Artifact analysis
an explicit native artifact path (e.g. artifact.preview with exact path and format)→bounded inline evidence for the supplied artifact
Live capture
a direct target command (e.g. `python benchmark.py`) via `--provider direct`→bounded runtime evidence from the live command
writes optional session scratch/cache and, on explicit preservation, to `<project>/.flameox`
04适合谁？何时不该用？
适合
- Developers and coding agents (including DeepSeek Harness) that need local-first performance evidence
- Performance engineers who want agents to burn down hotspots with explicit, auditable evidence
不适合
- Version 0.2 is a clean break: old `.diagnostics` state is not migrated, and there is no workspace to initialize, no `flameox.toml`, and no SQLite control plane.
05兼容性、维护与安全提示
- Version 0.2 is a clean break: old `.diagnostics` state is not migrated, and there is no workspace to initialize, no `flameox.toml`, and no SQLite control plane.
- Flameox is local-first: evidence stays in a bounded process-lifespan runtime, and preservation to `<project>/.flameox` is optional and only happens on explicit request.
MIT · actively maintained (latest release v0.1.15, 2026-08-30)
06常见问题
如何将 flameox 接入 DeepSeek Harness？
使用 `uv run flameox mcp serve --project-root "$PWD"` 启动 MCP 服务器（服务器在启动时固定其项目根目录）；`flameox setup` 会打印等价的 stdio 客户端配置供智能体使用。
flameox 具体做什么？
它协调性能分析器、基准测试工具、追踪处理器与直接本地目标，使智能体能够从明确的原生产物或实时命令快速获得有边界的证据，且结果保存是可选的。
需要初始化工作区或配置文件吗？
不需要。0.2 版本是一次彻底重构：无需初始化工作区、没有 `flameox.toml`，也没有 SQLite 控制面——你直接将精确的产物路径与格式传给 `analyze` 即可。
我的数据会被上传吗？
不会。flameox 是本地优先的；证据保留在生命周期受控的运行时中，仅当你显式请求时，才会将可选的会话缓存保存到 `<project>/.flameox`。
07相关的 DSH 工作流
ruflo
作者 ruvnet
原版智能体元框架（meta-harness）：支持多智能体集群部署、自主工作流编排与对话式 AI 系统构建，内置自适应记忆、自学习智能与 RAG 检索，原生集成 Claude Code / Codex / Hermes 等大量工具。
reactive-resume
作者 reactive-resume
开源简历制作插件，强调隐私保护，支持自定义、可移植和自托管。
everos
作者 evermind-ai
为所有 AI 智能体提供的便携统一记忆层，本地优先、Markdown 原生、数据归用户所有，跨应用与工作流自进化。
openpencil
作者 zseven-w
全球首个开源 AI 原生矢量设计工具，支持并发智能体团队，可在画布上直接通过提示词生成 UI。
08数据与来源
Bounded local runtime evidence for coding agents.
Flameox coordinates profilers, benchmark tools, trace processors, and direct
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-09-01。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。