Projects
The Workbench — FPI-X and an LLM Wiki
A self-built, self-run workbench for developing with AI agents: a five-stage multi-agent workflow, a cross-project knowledge engine, and a library of on-demand skills. This website is one of the things it produced.
Claude Code / Codex / Trae / MCP / 多 Agent 编排 / Spec-Driven / 知识工程 / Python
Positioning
For whom: one person, or a very small team, using several AI coding tools on real projects. It isn't a product; it's a way of working plus the tools that hold it up — and the methodology Conflux grew out of. The way of working came first; productising it came second.
What it replaces: the default state where each tool is used on its own, the background is re-explained every time, and "it ran" passes for "it's right". Several AI tools together fragment three ways: context (re-explain to each), state (nobody knows who got how far), quality (running ≠ correct). The workbench is the answer to those three.
Why it earns its place: it's in daily use — this site, Conflux and the writing system all came out of it. The rules weren't designed; they were written after specific failures, and each can be traced back to the one that caused it.
Main modules
- FPI-X workflow: an orchestrator plus specialised roles (architecture, frontend, backend, test, red team, docs, research) over five stages — intake, contract freeze, parallel work, integration, validation. Workers isolated in git worktrees behind three gates: review, QA, run. One rule carries the rest: no parallel work before the contract is frozen.
- LLM Wiki: a cross-project knowledge base with three-tier progressive retrieval (index → summary → full text), Hot/Warm/Cold decay, provenance, and a daily crawl of its feeds. It replaces re-feeding background to a model every time — a linear cost that becomes a lookup once compiled.
- Capability layer: on-demand skills wiring Claude Code, Codex, Trae, Pencil MCP and rust-analyzer LSP into a design → build → review → distill loop.
- Verification discipline: the writer is not the checker; quality assurance is not outsourced; validate the measuring instrument before trusting its reading.
Stack
- Claude Code / Codex / Trae, each routed by task tier rather than "the strongest does everything".
- MCP for the design tool (Pencil) and language services (rust-analyzer).
- Python + Bash for gates, wiki compilation and feed crawling.
- Spec-driven: contracts and file ownership first, so parallel work doesn't collide.
Status
In continuous use. The principles it left behind — "the writer is not the checker", "quality assurance is not outsourced", "comprehension debt" — are unpacked in the writing section.