Browser

What the source does

• FilesystemNone observedNo reads or writes in the source
• NetworkNone observedNo outbound calls in the source
• ShellNone observedNo commands run in the source
• DependenciesPinnedNothing floats
• SecretsNone foundNo keys in source

Five audit layers

Gates applied: no behavioural pass.

backend-rs/config/example.env

CLAUDE.md

backend-rs/CLAUDE.md

Audit 0.4.1 · source sha 52a5031418c3 · 16 Oct 2026Report an issue or request a re-scan

• Publisherwecode-ai GitHub organizationOwner of the repository this skill is read from
• Source folderwecode-ai/Wegent/backend/init_data/skills/browserThe exact folder we audited
• LicenceApache-2.0As the repository states it
• Commit pinned for this verdict52a5031418c3A new audit runs when the source changes
• Repository statusNot archivedPushed today
• Sandboxed runNot runSkills are read, not executed; layer L4 is skipped

Its description: Complete real user web tasks end-to-end via browser-tool, navigate, interact, wait for page state, extract results, and provide evidence when needed.

---
name: browser
description: Complete real user web tasks end-to-end via browser-tool, navigate, interact, wait for page state, extract results, and provide evidence when needed.
---

# Browser Control Skill

## Goal

Finish the user’s real task reliably.  
Prioritize successful completion and correct results over aggressive call minimization.

## Operating Rules

1. Start with the intended action directly (`navigate`/`open`/`act`/`evaluate`). Do not run `status` as a pre-check.
2. Use `snapshot` only when refs are required for interaction (click/type/select/drag/scrollIntoView).
3. Prefer `evaluate` for extraction. Return structured data in one comprehensive call when possible.
4. Use condition waits by default (`loadState`/`url` → `selector`/`text`/`textGone` → `fn`). Avoid `timeMs` unless explicitly needed.
5. Before clicking potentially off-screen elements, run `act.scrollIntoView` on the ref first.
6. Keep context stable: once `targetId` is known, pass it in follow-up calls when supported.
7. Avoid blind loops: every extra call must have a clear purpose.

## Reliability and Recovery

1. If `Ref not found`, do not reuse stale refs. Take one fresh `snapshot`, retry once, then stop if still failing.
2. For repeated failures with the same cause, stop and explain the blocker clearly instead of retrying endlessly.
3. Connection recovery is built into the tool. Allow auto-recovery once; if still disconnected, instruct user to install/connect extension.

## Screenshot Policy

1. Default: no screenshot.
2. Use screenshots only when user asks, or when visual proof is required.
3. Prefer element screenshots (`ref` or `element`) over full-page screenshots.
4. Use full-page screenshots only for page-level evidence.

## Recommended Flow

1. Direct action first (`navigate`/`open` or immediate `act`/`evaluate`).
2. If interaction needs refs, run `snapshot` (`interactive: true` preferred).
3. Wait for readiness using `act.wait` with explicit conditions.
4. Interact (`scrollIntoView` → `click/type/select/drag` as needed).
5. Extract/verify with `evaluate` (preferred) or `snapshot`.
6. Provide screenshot evidence only when necessary.

## Connection Handling

Connection recovery is built into the tool. On connection failure, let the tool auto-attach/launch/retry once. If still disconnected, stop and instruct the user to install/connect the extension.

## Minimal CLI Usage

Use `<BROWSER_TOOL_CMD>` for commands:

- macOS/Linux: `~/.wegent-executor/bin/browser-tool`
- Windows: `~/.wegent-executor/bin/browser-tool.cmd`

```bash
<BROWSER_TOOL_CMD> '<json>'
```

## Quick Examples

```bash
# Navigate directly
<BROWSER_TOOL_CMD> '{"action":"navigate","url":"https://example.com"}'

# Snapshot only when refs are needed
<BROWSER_TOOL_CMD> '{"action":"snapshot","interactive":true}'

# Act on ref
<BROWSER_TOOL_CMD> '{"action":"act","request":{"kind":"click","ref":"e1"}}'

# Ensure element is visible before click (recommended on long pages)
<BROWSER_TOOL_CMD> '{"action":"act","request":{"kind":"scrollIntoView","ref":"e1"}}'

# Condition wait (preferred over fixed sleep)
<BROWSER_TOOL_CMD> '{"action":"act","request":{"kind":"wait","loadState":"domcontentloaded","timeoutMs":15000}}'

# URL-based wait
<BROWSER_TOOL_CMD> '{"action":"act","request":{"kind":"wait","url":"checkout","timeoutMs":10000}}'

# Run JS in page context via act.evaluate (function or expression)
<BROWSER_TOOL_CMD> '{"action":"act","request":{"kind":"evaluate","fn":"() => ({title: document.title, href: location.href})"}}'

# Run JS against a target element ref via act.evaluate
<BROWSER_TOOL_CMD> '{"action":"act","request":{"kind":"evaluate","ref":"e1","fn":"(el) => ({text: el.textContent?.trim() || \"\"})"}}'

# Close current tab (or pass targetId)
<BROWSER_TOOL_CMD> '{"action":"act","request":{"kind":"close"}}'

# Element screenshot (prefer over full-page when only target proof is needed)
<BROWSER_TOOL_CMD> '{"action":"screenshot","ref":"e1","type":"jpeg"}'

# Comprehensive extraction in one evaluate
<BROWSER_TOOL_CMD> '{"action":"evaluate","expression":"(() => ({title:document.title,url:location.href}))()"}'
```

Excerpt, first 4,130 characters of SKILL.md, verbatim from the audited commit. It is text from someone else’s repository, quoted so you can judge it before your agent reads it.

git clone https://github.com/wecode-ai/Wegent
git -C Wegent checkout 52a5031418c3

Works in

1. 2026-10-16
   ✓SAFE B 89 / 100

   first audit · source 52a5031418c3

Questions about Browser

• github.com/wecode-ai/Wegent/tree/main/backend/init_data/skills/b — the audited source at commit 52a5031418c3 and every finding aboveObserved · 16 Oct 2026

Spotted something wrong? Report a correction · Methodology