zhèngmíng 名 · Confucius 01 / 05
Name it right.
Your agent reads what the function does, then renames it to match, so the name stops lying to the next person who opens the file.
verstak is a set of skills plus an MCP server. As the agent works, it writes decisions, open questions, and module boundaries into a graph; the next session reads the graph and continues.
The agent reads and writes the graph while it codes. No step after the work, no summary pass.
A recorded run, uncut: the session ends mid-refactor; the next one orients on the graph and continues at the open question. In the same run, the agent hits a recorded decision and raises the conflict instead of coding through it. Then open the graph yourself: the reasoning reads as a map you can traverse, node by node.
Memory tools extract summaries after the work and inject them into context. In verstak, the graph is where the agent structures the work itself: decisions with their reasons, questions with a lifecycle. New work attaches to those nodes and gets checked against them.
Not a transcript kept after the fact. As the agent works, decisions and their reasons land in the graph as nodes, and new work attaches to them, so the structure of the problem is the thing you read back.
Open questions and unverified assumptions are nodes with a lifecycle the agent can't quietly drop. It raises them, parks them, brings them back, and flags where the model no longer matches the code.
The graph is a shared workspace, not a chat log. Several agents work the same map at once, one coding, one logging decisions, one triaging risks, coordinating through it while you set direction.
For one, a personal reasoning assistant; for a team, a shared reasoning substrate.
Each comes from a tradition that spent centuries on it, and each leaves one concrete mark on your code.
Your agent reads what the function does, then renames it to match, so the name stops lying to the next person who opens the file.
npx verstak setup. Skills land in the agent, the MCP server registers, and your first session seeds the graph from the repo.
npx verstak setup --claude, --cursor, or --codex to target one agent. The skills are public at github.com/verstak-ai/skills.The agent pulls the five skills from a public repo and installs them under the verstak prefix. Read them first if you want.
It wires the MCP server into your agent and signs you in over OAuth. Approve it once, or drop in an access token.
On the first session the agent reads your repo and seeds the graph in passes, keeping the structure the code already shows.
In your repo. The graph holds what the agent writes about it, not the source.
The nodes the agent writes: decisions, questions, boundaries, and their links.
One graph call from your agent: a read, a write, or a search.
Claude Code as a plugin; Cursor and Codex through the same skills and MCP server.
Same skills, same graph on both plans. The line is operations.
An operation is one call your agent makes to the graph: a read, a write, or a search.