Add agent-browser remote interop and browse-fleet-subagents skill#38
Open
Add agent-browser remote interop and browse-fleet-subagents skill#38
Conversation
No agent (Claude Code, OpenCode, OpenClaw, Cursor, Windsurf) reads setup.json — only SKILL.md is loaded into context. The setup.json pattern was invisible to all agents and therefore broken everywhere. Replace with standard frontmatter fields: - requires.bins: [browser] — agents skip/suppress the skill if the binary isn't installed, rather than always injecting it and hoping the agent notices setupComplete: false - install.kind/pkg — agents that support auto-install can invoke npm install @browserbasehq/stagehand-cli automatically Also remove the ".env file" reference for credentials — the env vars are read from the environment, not specifically from a .env file. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ed frontmatter requires.bins and install.kind/pkg are not part of any agent skills spec. Replace with: - compatibility field (valid per agentskills.io spec) to surface requirements - inline `which browser || npm install` check in the skill body so the agent can self-heal without relying on non-standard frontmatter fields Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The agentskills.io spec requires the directory name to match the name field in SKILL.md frontmatter. Since the skill is named "browser" (invoked as /browser), the directory should be browser/ not browser-automation/. Also fix inconsistent heading levels (### → ##). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Validated both skills against the skills-ref reference library and fixed all issues found: - Quote compatibility YAML value to fix strictyaml parse error - Rewrite functions description with trigger keywords (schedule, webhook, cloud, cron, Browserbase Functions) -- the spec requires triggers in the description, not in the body - Split functions SKILL.md into SKILL.md + REFERENCE.md for progressive disclosure (invocation examples, common patterns, troubleshooting) - Remove "When to Use" body section from functions (redundant with description, invisible during skill discovery) - Add license: Apache-2.0 and LICENSE.txt to both skills - Add table of contents to browser REFERENCE.md (535 lines) - Condense browser EXAMPLES.md from 8 repetitive examples to 4 diverse ones Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Update copyright year from 2025 to 2026 in both LICENSE.txt files - Fix package name from @browserbasehq/stagehand-cli (doesn't exist) to @browserbasehq/browse-cli (actual npm package) - Update CLI command from `browser` to `browse` to match the npm package binary name across SKILL.md, REFERENCE.md, and EXAMPLES.md Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add anti-bot stealth, CAPTCHA solving, residential proxy, and session persistence details to the browser skill description and mode comparison table. These trigger phrases help AI agents discover and select this skill when users need to interact with protected or JavaScript-heavy websites. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add requires.bins (browse CLI) and install spec so OpenClaw can gate the skill properly and auto-install the CLI - Add homepage for ClawHub trust score - Do NOT gate on env vars — local mode works without Browserbase keys - Rewrite "Environment Selection" section with clear guidance on when to use local mode (simple pages) vs remote mode (protected sites, CAPTCHAs, bot detection, Cloudflare, geo-restricted content) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Session logs show the agent screenshots after every action (expensive, slow) and ignores browse act/observe in favor of manual snapshot → click ref loops. Update the skill to: - Document browse snapshot and recommend it as default over screenshot - Add guidance on when to use snapshot vs screenshot - Steer toward browse act/observe over low-level ref-based commands - Rewrite best practices to reflect snapshot-first, act-first workflow Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The SKILL.md documented commands (navigate, act, extract, observe) that don't exist in the CLI. The actual commands are open, click, type, fill, snapshot, etc. This caused agents to run nonexistent commands and fall back to guessing. - Replace all command docs with actual CLI syntax from browse --help - Document snapshot-first workflow with element refs - Add session management commands (stop, status, pages, tab_switch) - Add "No active page" to troubleshooting - Fix quick example to use real commands Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
browse stop doesn't always kill the daemon process. Add pkill fallback for when the daemon is stuck with wsUrl: "unknown" after a SIGTERM. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
… CLI commands - Rename browser-automation → browser in README.md and marketplace.json - Rewrite EXAMPLES.md: replace nonexistent commands (navigate, act, extract, observe, close) with real browse CLI commands (open, snapshot, click, type, fill, get, stop). 4 concrete examples including remote mode escalation. - Rewrite REFERENCE.md: replace Stagehand/Playwright architecture with actual daemon-based CLI docs, all 20 real commands, env var config. - SKILL.md: add "Activating Remote Mode" progressive disclosure section, fix get text to require selector argument. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add browse mode to SKILL.md commands list and REFERENCE.md. Rewrite "Activating Remote Mode" as "Switching Between Local and Remote Mode" using browse mode as the primary mechanism. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…rowse-cli; gate on 'browse' bin only\n\n- Drop OpenClaw install block to avoid failed installs\n- Make compatibility + setup check generic (no npm package name)\n- Generalize REFERENCE note (avoid browse-cli v0.1.4 mention) Co-authored-by: Kyle Jeong <Kylejeong2@users.noreply.github.com>
…basehq/browse-cli; gate on 'browse' bin only\n\n- Drop OpenClaw install block to avoid failed installs\n- Make compatibility + setup check generic (no npm package name)\n- Generalize REFERENCE note (avoid browse-cli v0.1.4 mention)" This reverts commit 5b74333.
Remove: get html, drag, highlight, is, execute references. Add docs for: hover, newpage, eval, viewport, network capture, snapshot --compact, screenshot --full-page, type --delay/--mistakes, fill --no-press-enter, stop --force. Replace get html with get value in SKILL.md. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ked commands These commands are functional again after PR browserbase/stagent-cli#11 fixed the daemon startup (EPIPE crash) and restored selector command surfaces. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…NCE.md These CLI features were verified working in browse-cli v0.1.5 but were missing from the reference documentation: - `refs` command for cached ref map lookup - `open --wait` flag (networkidle/domcontentloaded) for SPAs - `--json` global flag for structured output - `--session` global flag for concurrent browser sessions Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Match the #### heading + description + code block pattern used throughout the rest of REFERENCE.md. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…hing - Add Table of Contents to REFERENCE.md (matches functions/REFERENCE.md and top skills like terraform-skill, mcp-builder) - Remove Typical Workflow and Local vs Remote Mode sections from REFERENCE.md — these were near-identical copies of SKILL.md content - Condense "Switching Between Local and Remote Environment" in SKILL.md from 38 lines to 16 — keeps the signal detection list and credential setup, drops redundant env commands already shown in Commands section Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
shrey150
changed the title
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 2 potential issues.
Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
| "strict": false, | ||
| "skills": [ | ||
| "./skills/browser-automation" | ||
| "./skills/browser" |
There was a problem hiding this comment.
New skills not registered in marketplace discovery catalog
High Severity
The two new skills (agent-browser-remote and browse-fleet-subagents) are added with full SKILL.md files and listed in the README.md skills table, but neither is registered in marketplace.json. Since marketplace.json is the discovery catalog that Claude Code uses to find and load skills, these skills won't be discoverable or installable through the plugin system. The browse plugin's skills array only contains ./skills/browser.
Additional Locations (1)
| When done: | ||
|
|
||
| ```bash | ||
| node scripts/browserbase-session.mjs close --session-id "$BROWSERBASE_SESSION_ID" |
There was a problem hiding this comment.
Script path unresolvable from user working directory
High Severity
The SKILL.md instructs Claude to run node scripts/browserbase-session.mjs as a relative path, but the script lives at skills/agent-browser-remote/scripts/browserbase-session.mjs relative to the repo root. When Claude Code follows these bash instructions, commands execute in the user's project directory, where no scripts/browserbase-session.mjs exists. Every other skill in this repo uses globally-installed CLI tools (e.g., browse, pnpm bb) rather than relative file paths.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
agent-browser-remoteskill to teach explicit Browserbase CDP handoff foragent-browserbrowse-fleet-subagentsskill as a sub-agent orchestration pattern (not a CLI primitive)browserbase-session.mjs) to create/close Browserbase sessions and export CDP env varsbrowserskill docs to emphasize focused snapshots and parallel work via sub-agentsNaming decisions
agent-browser-remotefor discoverability on explicit interop intent (agent-browser+ remote/CDP)browse-fleet-subagentsto make the sub-agent orchestration model explicit in the slugNotes
Note
Low Risk
Mostly documentation and skill catalog updates, plus a small standalone Node script that creates/closes Browserbase sessions via API; no runtime/CLI behavior in this repo is modified. Risk is limited to potential user confusion or session leaks if the helper script is misused.
Overview
Adds two new skills:
agent-browser-remote(teaches explicit Browserbase CDP handoff foragent-browser, including abrowserbase-session.mjshelper to create/close sessions and export env vars) andbrowse-fleet-subagents(documents a sub-agent fanout pattern with retries and cleanup).Replaces the prior
browser-automationskill/docs with a newbrowserskill centered on the daemon-basedbrowseCLI (focused snapshots, remote-mode guidance, and examples), and updates the marketplace/README skill catalog accordingly. Updates thefunctionsskill docs (adds MIT license, streamlinesSKILL.md, and moves invocation/patterns/troubleshooting intoskills/functions/REFERENCE.md).Written by Cursor Bugbot for commit 9e5c1a3. This will update automatically on new commits. Configure here.