remote-bridge

name	remote-bridge
description	Cross-target remote control via Zulip (default control) and optional Telegram mobile notify, with mailbox approvals/instructions for autonomous research loops. Not for OpenClaw.
user-invocable	true
disable-model-invocation	false
metadata	{"short-description":"Zulip/Telegram remote control mailbox for AAS agents","requires":{"bins":["python3"]}}

Remote Bridge

Cross-platform remote control plane for claude, codex, grok, kimi, deepseek, opencode, copilot, antigravity (not OpenClaw).

Channel	Role
Zulip	Default control + primary notify (Research / job/<job_id>)
Telegram	Mobile notify fallback only when Zulip send fails; inbound only with a dedicated bot

Notify policy (default)

Zulip first. Telegram only if Zulip fails. Sends never dual-spam both channels on success (stop_on_first_success).

Token	Behavior
auto / default	Zulip if configured, else Telegram
zulip	Try Zulip; fall back to Telegram on failure
both	Same as Zulip-primary + Telegram-fallback (alias, not dual fan-out)
telegram	Telegram only (explicit)
off	Silence

Does not inject messages into live TUI chats. Continuations use the on-disk mailbox and headless drive, or a local PreToolUse gate (Grok; Claude evidence-gated).

Windows Runtime Commands

On native Windows, use the managed Windows runner and the native runtime command target. Set $runtime to the installed runtime root. Multi-agent installs usually use %LOCALAPPDATA%\ai-agents-skills\runtime. Then run:

$runtime = if ($env:AAS_RUNTIME_ROOT) { $env:AAS_RUNTIME_ROOT } else { "$env:LOCALAPPDATA\ai-agents-skills\runtime" }
& "$runtime\run_skill.ps1" "skills/remote-bridge/run_remote_bridge.ps1" <args>
& "$runtime\run_skill.ps1" "skills/remote-bridge/run_remote_bridge.ps1" <args>

POSIX examples below use run_skill.sh and .sh command targets; use the Windows command target above on native Windows.

# Credential-bearing lanes must launch from a root-owned AAS component
# generation; the per-user runtime copy is refused by the credential gate.
# The generation directory is named for a git commit, so its name carries no
# ordering -- pick the newest publish, which the store records as its mtime.
launcher="${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh"
newest=0
for gen in /usr/local/libexec/coding-system/components/ai-agents-skills/*/; do
  gen="${gen%/}"
  [ -f "$gen/manifest/credential-runtime.json" ] || continue
  [ -x "$gen/canonical/runtime/runners/run_skill.sh" ] || continue
  stamp="$(stat -c %Y "$gen" 2>/dev/null || stat -f %m "$gen" 2>/dev/null)" || continue
  [ "${stamp:-0}" -gt "$newest" ] || continue
  newest="$stamp"
  launcher="$gen/canonical/runtime/runners/run_skill.sh"
done
bash "$launcher" \
  skills/remote-bridge/run_remote_bridge.sh <command> [args...]

Secrets

Never commit real tokens. Copy the example and fill placeholders:

• Linux/WSL: ~/.config/remote-bridge/secrets.json
• macOS: ~/Library/Application Support/remote-bridge/secrets.json or XDG
• Windows: %APPDATA%\remote-bridge\secrets.json

The selected secrets file must be an owner-private, single-link regular file inside an owner-controlled, non-writable-by-others directory; symlinked, hard-linked, permissive, oversized, or changing files are rejected. On POSIX, use mode 0600 for the file.

Env overrides: REMOTE_BRIDGE_SECRETS_FILE, ZULIP_*, TELEGRAM_BOT_TOKEN, TELEGRAM_ALLOWED_CHAT_IDS, AAS_REMOTE_JOB_ID.

Do not auto-read ~/.openclaw. Prefer a dedicated Zulip bot user and a dedicated Telegram bot (not OpenClaw’s).

OpenClaw boundary

The host CLI does not import from, export to, or synchronize with ~/.openclaw before or after commands. Host secrets and mailbox state remain host-owned. sync_remote_bridge_paths.py is shipped temporarily only as an inert revocation stub. Replacing a previously managed copy requires an explicitly reviewed backup-and-replace upgrade that preserves recovery data; the default installer does not overwrite divergent copies. The stub never inspects paths and is not part of normal remote_bridge.py execution. dispatch_aas.py is also an inert revocation stub: OpenClaw may not dispatch control commands or supply an authorization principal. Remove any previously published aas-remote-bridge workspace adapter from service. Any future compatibility transfer requires an explicit, separately reviewed one-way export that excludes secrets and treats workspace content as untrusted.

Source of truth

Reusable logic lives in ~/ai-agents-skills (this skill + canonical/runtime/skills/remote-bridge/). Agent homes are install products. OpenClaw control adapters are retired and are not install products. The publisher and dispatcher retained in the runtime are fail-closed revocation stubs only; neither reads or writes an OpenClaw workspace.

Commands

Command	Purpose
selftest	Offline smoke (no network)
show-config	Redacted config
doctor	State root, jobs, channels (--live optional)
arm --job ID --provider P --cwd DIR [--loop DIR]	Create mailbox job
status	List jobs + pending requests
send --text "…" [--channel zulip|telegram|both|auto] [--dry-run]	Notify (Zulip-primary; Telegram fallback)
request-approval --job ID --tool T [--wait --timeout N]	Create approval + optional wait
instruct --job ID --text "…"	Push inbox item
handle-command --text "/aas …" --allow-local-cli or --text-stdin --allow-local-cli	Process one explicitly authorized local-operator command
format-inbox --job ID [--consume]	Claim/format inbox for prompts
check-approval --job ID --digest HEX	Consume matching allow reply

Authenticated host transports may translate sender identity to a trusted local control event. The public CLI deliberately rejects --principal; command-line text cannot assert a remote identity. Chat control commands are /aas help|status|approve|deny|say|instruct|stop|pause|resume|focus|doctor.

ARL / drive integration

arm and drive use --notify auto (default): if Zulip and/or Telegram credentials are present in ~/.config/remote-bridge/secrets.json (or env), progress events are sent without an extra channel flag for legacy/off/monitor loops. Enforce mode also requires AAS_AUTOLOOP_EXTERNAL_NOTIFY_EGRESS=allow; without that explicit egress consent it produces local notification audit only. Prefer:

Notification and panel egress consent is independent of provider execution trust. Selecting trusted-local never implies permission to send notification, panel, PII, or secret-bearing payloads externally; the existing exact consent and admission gates still apply.

# one-time arm (persists notify_channel on loop_state + registry)
… run_autonomous_research_loop.sh arm --dir <loop> --root <proj> --notify auto

# drive inherits arm/env/secrets (auto by default)
… run_autonomous_research_loop.sh drive --dir <loop> --root <proj> \
  --provider codex
# equivalent explicit: --notify auto
# enforce-mode network consent: AAS_AUTOLOOP_EXTERNAL_NOTIFY_EGRESS=allow
# silence: --notify off   or   AAS_AUTOLOOP_NOTIFY=off

Optional job id for topic routing / channel override:

export AAS_REMOTE_JOB_ID=example-job
# default when secrets present: Zulip primary, Telegram only if Zulip fails
# export AAS_AUTOLOOP_NOTIFY=zulip     # same as default primary
# export AAS_REMOTE_STRICT_NOTIFY_CHANNEL=zulip  # forbid every Telegram fallback
# export AAS_AUTOLOOP_NOTIFY=telegram  # Telegram only
# export AAS_AUTOLOOP_NOTIFY=both      # alias for primary+fallback (not dual)
# export AAS_AUTOLOOP_NOTIFY=off       # silence

AAS_REMOTE_STRICT_NOTIFY_CHANNEL=zulip is a send-boundary restriction, not just a launch preference: explicit Telegram requests and Telegram credentials added later are ignored while it is set. An invalid strict-channel value fails closed.

Events notified (best-effort, never abort the loop): drive_start, drive_stop, iteration_ok / iteration_failed, quota_wait, paused, terminal, driver_dead.

Not remote-notified (local progress only):

• iteration_start — pairs with iteration_ok ~1s later
• watch ledger tick iteration — drive already owns completion notifies; sending both produced duplicate Zulip posts when drive and watch ran together

Identical bodies are deduped for 15s (in-process + per-loop disk). Materially equivalent retries whose event ids or timing fields were rebuilt are deduped for 120s. Remote Bridge locks the semantic retry identity across the complete check -> send -> remember sequence, so concurrent processes deliver it once; material changes to status, content, agents, or compute remain deliverable.

Headless iterations inject a labeled data-only inbox block when AAS_REMOTE_JOB_ID is set. Approvals for auto-approve providers (--yolo, full-auto) are advisory unless a live PreToolUse gate is installed.

Grok live gate (optional)

Example hook: hooks/grok-remote-bridge-gate.json.example
Script: hooks/pretooluse_deny_until_approved.py (local FS only; deny-until-approved). Grok hooks fail-open on crash/timeout — not hard OS security.

Security notes

• Allowlist users on Zulip/Telegram.
• Compromised allowlisted chat ≈ operator authority for headless soft path.
• Zulip/Telegram endpoints require HTTPS, reject embedded URL credentials and redirects, and allow localhost HTTP only with AAS_REMOTE_BRIDGE_ALLOW_HTTP_LOCALHOST=1.
• Structured event redaction covers every nested string value before transport output or delivery-registry persistence; secret-bearing event ids are replaced by opaque digest-derived ids. Common explicit email, phone, government-id, address/DOB, and participant/patient/subject data is also redacted; the heuristic is not a completeness guarantee.
• Notification sends fail closed when the v2 secret/PII redactor cannot be loaded or raises; exact configured-secret replacement is not an acceptable fallback for unconfigured credentials or personal data.
• Delivery locks and the dedupe registry require private host-owned regular files and no-follow directory traversal.
• Prefer structured --notify over raw --notify-cmd (set AAS_ALLOW_RAW_NOTIFY_CMD=1 only if needed).
• Platform “supported” claims need dated native smoke evidence per OS.