@fui-org/fui-cli 2.1.4 → 2.2.0

package/README.md

@@ -47,10 +47,10 @@ Installed fui with Bun? Install Playwright the same way, so fui finds it: `bun a
 
 ## Use with an AI agent
 
-`fui` installs two skills for each AI agent it finds, and refreshes them after every update:
-
-- `fui` teaches the agent the commands and safety rules. `fui skill` prints it.
-- `fui-skill` holds the FUI domain rules: module.json, components, tAPI and stored procedures, design systems.
+`fui` installs one skill, `fui`, for each AI agent it finds, and refreshes it after every update. It holds the commands, the
+safety rules and the FUI domain rules (module.json, components, tAPI and stored procedures, design systems); `fui skill`
+prints it. For Claude Code, `fui` also installs two agents, `fui-builder` and `fui-tester`, in `~/.claude/agents`, used
+by the larger tasks below.
 
 | Agent       | Skills folder                  |
 | ----------- | ------------------------------ |
@@ -60,12 +60,32 @@ Installed fui with Bun? Install Playwright the same way, so fui finds it: `bun a
 
 For Claude Code it also installs the `fui-target` mod into `~/.claude/skills/fui-target`. Once a session runs `fui`, a line above the prompt shows the session target, for example `🧩 monitor-manager › device-list · fui use`.
 
-Installed an agent later? The next `fui` run adds its skills. To install by hand:
+Installed an agent later? The next `fui` run adds its skill. To install by hand:
 
 ```bash
 fui skill install [--agent claude,codex,antigravity|all]
 ```
 
+## Larger tasks: a team of agents (Claude Code)
+
+You never pick a mode. For each request to build or change something, the agent decides: when you name the place or the
+exact change, it does exactly that and checks it; when you describe behavior and leave the places to it (a new screen, a
+feature across the UI and its stored procedures, a bug of unknown cause), it runs a team. A team run needs you twice: to answer the questions at the start, and to approve the release at the end. In between, a
+planner (your session) writes a plan with acceptance criteria, a builder writes the module and its stored procedures, a
+tester who has not seen the builder's work checks them, and the builder fixes what the tester found. The agents work one
+after another.
+
+There is nothing to set up: the two agents come with `fui`. A Claude Code session that started before they were
+installed does not see them, and then the agent does the same steps alone, with no independent tester. The builder and
+the tester cannot run anything that changes the server or real data (publish, deploy, `fui exec`, `--confirm-write`,
+deletes); the planner shows you the result and the one release command, and runs it only after you approve.
+
+Then ask as usual, for example `add a search box and an edit dialog to /my-project/students`. Say "do it directly" or
+"run it as a team" to overrule the choice.
+
+The run keeps its plan, test reports and the review packet in `<workspace>/_team/<run>/`. Stored procedures are checked
+by reading them, never by running them, so a SQL error can still appear at deploy.
+
 ## Quick tour
 
 ```bash
@@ -114,8 +134,8 @@ fui config set api-base https://<fui-api-host>
 printf %s "$FUI_TOKEN" | fui config set api-token -
 ```
 
-Connection files from before the multi-zone layout (`_connections.json` without `"version": 2`) stop DB commands with a
-hint; convert them once with `fui db migrate` (the old file is kept in `_history/`).
+Connection files from before the multi-zone layout (`_connections.json` without `"version": 2`) are upgraded the first
+time `fui` reads them (the old file is kept in `_history/`). `fui db migrate` does the same for every project at once.
 
 ## Uninstall