agent-browser-relay（Agent 浏览器中继）

Agent Browser Relay

Use this skill to operate through the local relay gateway on an explicitly attached browser tab, verify readiness before reads, and keep concurrent tab workflows lease-scoped instead of falling back to direct browser-control tools.

Chrome is the documented target. Chromium-based browsers that can load the extension may also work, and relay status now surfaces detected host-browser identity plus a persistent profile id, but the skill does not yet provide first-class multi-browser orchestration.

Quick start

Fresh-machine rule:

• skills add guarantees the installed skill at ~/.agents/skills/agent-browser-relay
• the guaranteed Chrome load path after skills add is ~/.agents/skills/agent-browser-relay/extension
• ~/agent-browser-relay/extension is only an optional visible convenience copy created by npm run extension:install
• npm run extension:path, npm run relay:start, and npm run relay:global:install print the primary folder to load
• npm run extension:path also prints stable absolute paths for read-active-tab.js and preflight.sh so agents can avoid cwd-dependent invocation
• opening the extension popup once after relay startup now wakes the extension and lets npm run extension:status confirm that Chrome actually loaded it
• on a new machine, the human must load the primary path in chrome://extensions before attach/read steps. Do not treat a missing visible convenience copy as a sandbox or socket-permission issue.

Defaults are set in code:

• Host: 127.0.0.1
• Port: 18793
• Attach timeout: 120000 ms Override per command with --host, --port, and --attach-timeout-ms when needed.

1. Install dependencies and start relay This prints the primary Chrome extension path to load in Chrome and refreshes the optional visible convenience copy when possible. npm run relay:start -- --status-timeout-ms 3000 Or pin host/port explicitly: npm run relay:start -- --host "127.0.0.1" --port "18793" --status-timeout-ms 3000 relay:start keeps running until you stop it or the process is restarted. If you want a bounded session, pass --auto-stop-ms explicitly: node scripts/relay-manager.js start --auto-stop-ms 10800000
2. Load extension in Chrome

- chrome://extensions - Enable developer mode - Load unpacked from ~/.agents/skills/agent-browser-relay/extension after skills add - If the command printed the extension path, treat that printed path as the source of truth - Run npm run extension:path from the installed skill directory any time you want the exact path printed again - ~/agent-browser-relay/extension is optional; create it with npm run extension:install if you want a visible shortcut

1. Confirm the extension is loaded in Chrome Open the toolbar popup once after relay is running. The popup should show Relay connected on <port>. Agent requirement: before any attach step on a fresh machine, ask the human to open the popup once, then confirm: npm run extension:status -- --port "18793" --wait-for-connected --connected-timeout-ms 120000 This now also reports the detected browser host and persistent profile id for the connected extension instance.
2. Attach the extension to the target tab (open toolbar popup and click attach) Optional per-tab relay: in the popup, set Tab port before clicking attach if this tab should use a non-default relay port. If you want the agent to create its own first background tab instead, enable Allow agent to create new background tabs in the popup. Agent requirement: after extension:status confirms Chrome loaded the extension, pause and ask the human to do this attach step, then wait for confirmation before continuing. If your .agents skill folder drops extension/ after a git fetch or pull, repair it from the repo: cd ~/.agents/skills/agent-browser-relay 2>/dev/null || cd ~/.agents/skills/private/agent-browser-relay git sparse-checkout disable git config --unset-all core.sparseCheckout || true git config --unset-all core.sparseCheckoutCone || true git checkout --.
3. Check readiness and attach state npm run relay:doctor -- --port "18793" --tab-id "<TAB_ID>" --json Resolve <TAB_ID> from status first (npm run relay:status -- --all --status-timeout-ms 3000). For all agent runs, use the assigned tab id: npm run relay:doctor -- --host "127.0.0.1" --port "18793" --tab-id "<TAB_ID>" --json Continue only if this command returns success.

Per-tab relay port behavior

• If you run one relay process with multiple ports, the extension can manage different relay ports per attached tab.
• A tab with no saved relay-port mapping uses the global default relay port (18793).
• After a successful attach, the extension saves that tab’s mapped relay port and reuses it automatically.
• Closed tabs have their mapping removed automatically.
• This is per-tab port routing, not first-class multi-browser support.

Mandatory behavior for agents

• Use fixed commands from this repo. Do not try to "discover" alternate script names.
• Gateway-only rule: always communicate through the local relay gateway (/status and node scripts/read-active-tab.js).
• Never use direct browser-control tooling for this workflow (for example Playwright, Puppeteer, Selenium, agent-browser, or ad-hoc Chrome control scripts).
• Never take control of a random Chrome window/profile. Only operate on the explicitly attached target tab leased via --tab-id.
• On a fresh machine, explicitly tell the human to load the primary extension path from npm run extension:path before any attach/read attempt. After skills add, that is normally ~/.agents/skills/agent-browser-relay/extension.
• Canonical commands:

- npm run extension:path - npm run extension:status - npm run relay:start - npm run relay:status - npm run relay:doctor - npm run relay:stop - node scripts/read-active-tab.js

• For relay health checks, always use explicit timeouts to avoid hangs:

- npm run relay:status -- --status-timeout-ms 3000 - curl --max-time 3 -sS "http://127.0.0.1:18793/status" - npm run relay:status -- --all --status-timeout-ms 3000

• After relay:start, pause and ask the human to open the popup once so npm run extension:status -- --port "18793" --wait-for-connected --connected-timeout-ms 120000 can confirm Chrome actually loaded the extension.
• Only after extension:status succeeds, either ask the human to attach the target tab before reads, or confirm that Allow agent to create new background tabs is enabled before first-tab creation workflows.
• Run npm run relay:doctor -- --host "127.0.0.1" --port "18793" --tab-id "<TAB_ID>" --json before reads and proceed only when it succeeds.
• If the workflow will open tabs via Target.createTarget, run npm run relay:doctor -- --host "127.0.0.1" --port "18793" --require-target-create --json and proceed only when it succeeds.
• For all agent runs (single-agent and concurrent), always pass --tab-id <tabId> on check/read commands so every operation is lease-scoped.
• In concurrent runs, attached tabs are shared extension state, but an attached tab can only have one active lease at a time. Inspect npm run relay:status -- --all --status-timeout-ms 3000 and prefer leaseSummary.availableAttachedTabIds or attachedTabs[].leasedSessionId when choosing the next --tab-id.
• If doctor returns TAB_LEASED_BY_OTHER_SESSION, do not attempt takeover. Wait for that session to finish or retry with another attached tab that has no active lease.
• When Target.createTarget is enabled, the extension may create and auto-attach the first agent-controlled tab for the session without a human seed attach step.
• Do not stop/restart relay during the task unless the human requests it or recovery is explicitly required.
• Do not restart relay only because code was updated locally; updates are applied on next explicit human-approved restart.
• If the requested tabId is missing from relay status attachedTabs, stop and ask the human to re-attach the target tab in the popup before continuing.
• If relay:start times out, report the actual relay log/error. Do not guess about sandbox restrictions unless the command output shows a concrete permission error.
• If the page shows human-verification gates (for example "Are you human?" or CAPTCHA), stop immediately, alert the human with $attention-please, and wait for explicit human confirmation before continuing.

1. Read structured tab payload node scripts/read-active-tab.js --host "127.0.0.1" --port "18793" --tab-id "<TAB_ID>"
2. Optional one-command smoke test ./scripts/preflight.sh This is a local smoke wrapper. For agent workflows, the canonical preflight remains npm run relay:doctor -- --port "18793" --tab-id "<TAB_ID>" --json for attached tabs, or npm run relay:doctor -- --port "18793" --require-target-create --json for first-tab creation workflows.

If the current working directory is not the installed skill root, use the absolute Stable read CLI path printed by npm run extension:path instead of a relative node scripts/read-active-tab.js path.

Capabilities

read-active-tab.js returns capability metadata in every successful payload under:

source.capabilities

Version compatibility checks are also included under source.extension:

• installedVersion
• sourceVersion
• relayVersion
• observedExtensionVersion
• versionMismatch

If a mismatch is detected, the command also prints a human-friendly update hint to stderr on every run.

• scripts/read-active-tab.js default extraction: url, title, text, links, metaDescription.
• Relay session leases (--tab-id) for concurrent agent isolation per tab on one relay port.
• Runtime.evaluate expression mode with --expression, --expression-file, or --expression-stdin.
• Screenshot capture mode via --screenshot (optional --screenshot-full-page, --screenshot-path).
• Preset extraction for WhatsApp and generic chat-auditing with regex filters.
• Attach-state polling with --check --wait-for-attach.

Concurrent workflow notes

• Inspect npm run relay:status -- --all --status-timeout-ms 3000 before selecting a tab for a concurrent run.
• Use attachedTabs plus leasedSessionId / tabLeases to understand which attached tabs are currently free.
• If doctor returns TAB_LEASED_BY_OTHER_SESSION, do not force takeover. Wait for that session to release the tab or pick another attached tabId.
• Lease isolation scopes relay command routing per attached tab. It does not imply a separate browser profile or stronger browser-process isolation.

Presets and filters

• --preset values: default, whatsapp, wa, whatsapp-messages, chat-audit, chat.
• Regex filters: --text-regex, --exclude-text-regex, --link-text-regex, --link-href-regex.
• WhatsApp/chat filters: --message-regex, --exclude-message-regex, --sender-regex, --exclude-sender-regex.

Common command examples

In agent workflows, use the --tab-id variants. Unscoped commands are for manual/local debugging only. Prefer --expression-file or --expression-stdin over inline --expression for non-trivial JavaScript.

node scripts/read-active-tab.js --host "127.0.0.1" --port "18793" --pretty false
node scripts/read-active-tab.js --host "127.0.0.1" --port "18793" --tab-id 123 --pretty false
node scripts/read-active-tab.js --host "127.0.0.1" --port "18793" --tab-id 123 --check --wait-for-attach --require-target-create --attach-timeout-ms 120000
node scripts/read-active-tab.js --host "127.0.0.1" --port "18793" --expression "document.documentElement.outerHTML"
node scripts/read-active-tab.js --host "127.0.0.1" --port "18793" --tab-id 123 --expression-file "./tmp/expression.js"
node scripts/read-active-tab.js --host "127.0.0.1" --port "18793" --tab-id 123 --expression-stdin < ./tmp/expression.js
node scripts/read-active-tab.js --host "127.0.0.1" --port "18793" --screenshot --screenshot-full-page --screenshot-path "./tmp/page.png"
node scripts/read-active-tab.js --host "127.0.0.1" --port "18793" --preset whatsapp-messages --max-messages 200 --selector "#main"
node scripts/read-active-tab.js --preset chat-audit --selector "body" --message-regex ".*"

All successful commands return a source object with relayHost, relayPort, relayStatusUrl, and relayWebSocketUrl.

Recommended flow with agents

Before fetching data in an automation flow, run a lightweight preflight once to ensure relay + attached tab state are ready.

For multiple agents on one relay:

• Resolve tab ids from relay status (npm run relay:status -- --all --status-timeout-ms 3000).
• Assign one tab id per agent.
• Use --tab-id in every read-active-tab.js call for that agent.
• Treat the extension instance as shared: leases prevent two controller sessions from driving the same attached tab, but they do not create separate Chrome profiles or separate extension processes.
• If relay:doctor reports TAB_LEASED_BY_OTHER_SESSION, inspect attachedTabs, tabLeases, and blocker detail, then pick another attached tab or wait for the owning session to release it.