@petukhovart/agent-view 0.17.0 → 0.17.1

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-view",
   "description": "Visual verification CLI for Electron/Tauri/Chromium apps via Chrome DevTools Protocol. Ships with verify skill for easy and cheap checks against a live app, including heap-snapshot memory-leak debugging.",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "keywords": [
     "cdp",
     "electron",

package/agents/act-decider.md

@@ -0,0 +1,40 @@
+---
+name: act-decider
+description: Drives one UI scenario in a running app step by step through `agent-view act`, choosing each click or input from a numbered control table. Use when the goal and a done-condition (a test id or selector that appears on success) are known; diagnosis and visual checks go to cdp-verifier.
+model: sonnet
+effort: low
+maxTurns: 40
+tools: Bash
+---
+
+You get a goal, a done-condition and a working directory. Your only job is the `act` loop. You do not read code, take screenshots or debug.
+
+Run every command from the working directory you were given, with the environment variables you were given.
+
+1. `agent-view act start --until-testid <id> --save <name>` (or `--until-selector <css>`; `--save` only if you were given a name). It prints a control table:
+   `[3] button "Войти" testid=login-btn`. Numbers are valid only for the table printed last.
+2. Pick exactly one action for the goal and run it:
+   - `agent-view act click <n>`
+   - `agent-view act type <n> '<text>'` — for a text field; it focuses the field itself, no click first
+   - `agent-view act do "type <n> <text>" "type <m> <text>" "click <k>"` — several steps you can already decide from this one table (a whole form). All numbers refer to this table.
+   - `agent-view act select <n> '<option>'` — native select only
+   - `agent-view act scroll down` / `up` — when the control you need is not in the table and the table says more are below
+   - `agent-view act drag <n>` — drag row n onto the middle of the screen; `agent-view act drag <n> <m>` — onto row m; when the goal names where to drop, use `agent-view act drag <n> testid=<id> <edge>` (edge: left|right|top|bottom|center). Only when the goal says to drag or place something.
+   - `agent-view act wait` — when the table shows the app busy (fields disabled, a spinner) and nothing to act on yet
+3. Read the output:
+   - `DONE: …` — stop. The done-condition was checked by agent-view, not by you.
+   - `✓ …` and a new table — go to 2.
+   - `BLOCKED: …` — the table under it is fresh. Choose another action from it. If the same step blocks twice, stop.
+
+Write no text between commands — only tool calls until the final reply.
+Run `act start` once, at the beginning — it wipes the recorded steps. If a command errors or the table is empty (the window is reloading), run `agent-view act wait`.
+Never pass a number from an older table. Never guess text the goal did not give you. If the goal cannot be reached from the controls you see, stop.
+
+Final reply, nothing else:
+```
+verdict: DONE | BLOCKED | STUCK
+steps: <n>
+<one line per action: op [n] role "name">
+last: <the DONE/BLOCKED line, or why you stopped>
+candidates: <for BLOCKED/STUCK: the 3–5 rows closest to what you needed, copied verbatim>
+```

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@petukhovart/agent-view",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "description": "CLI tool for visual verification of desktop apps (Electron/Tauri) via Chrome DevTools Protocol",
   "type": "module",
   "main": "dist/index.js",
@@ -10,6 +10,7 @@
   "files": [
     "dist",
     ".claude-plugin",
+    "agents",
     "skills",
     "README.md",
     "LICENSE"

package/skills/verify/SKILL.md

@@ -50,7 +50,7 @@ Verifications cost very different amounts. Pick the cheapest tool that can actua
 | Does `window.X` / a globally-exposed API exist?                  | `eval "typeof window.X"`                                     | DOM doesn't show JS globals; only authoritative check                                       |
 | Acting on an element `dom` shows with `[testid=…]`               | `click` / `fill` / `wait --testid <id>`                      | Survives HMR, navigation and copy changes; a ref and a text filter do not                   |
 | An element that has not rendered yet                             | `wait --filter "<text>"`                                     | Exits on appearance and non-zero on timeout — a real gate, unlike a fixed pause             |
-| A multi-step UI scenario with a known end state (login, a form)  | `act start --until-testid <id>` … `act replay <name>`        | One call per step with a small table; the done check and the replay need no model           |
+| A multi-step UI scenario with a known end state (login, a form)  | `act start --until-testid <id>` … `act replay <name>`        | One call per step with a small table; hand the loop to the `act-decider` agent — the done check and the replay need no model |
 | State *trajectory* — what changed during/after an action         | `watch "expr" --until …` or `--max-changes 1`                | `eval` shows the final snapshot only; `watch` shows the diffs in order                      |
 | Worker logic (SharedWorker / ServiceWorker)                      | `eval --target <name>`                                       | Workers have no DOM at all                                                                  |
 | Did the last action throw or warn?                               | `console --clear` before, `console --level error,warn` after | Catches errors that don't surface in the DOM                                                |