Tlive Agent Skill

tlive usage guide

tlive is a self-hosted monitoring/approval layer. Claude Code sessions report

through global hooks; Codex sessions are watched through an app-server

companion process (no hooks, no trust step). Completions and failures land in

IM (Telegram/Feishu) and the web dashboard, where you can reply-to-continue.

The posture (tlive mode, default notify) decides whether approvals are

held for a remote answer, in escalation order: off makes every hook a no-op;

notify only watches + notifies (the shim never holds an approval — prompts

stay 100% native) and reports to the machine (desktop toast + dashboard)

— IM stays quiet about a dialog only the terminal can answer, except once

ever per chat, when a card explains why and offers the switch to full;

full is the posture that puts approvals on your phone: it turns on

remote approval for the main session (Allow/Deny from IM/desktop/dashboard),

in parallel with the terminal dialog — first answer wins; all holds

sub-agent approvals too, with no terminal dialog for a held one until the

window ends, so use it only when nobody is at the keyboard (tlive mode full

goes back). Also settable from IM: /mode (a bare /mode replies with the

ladder). The daemon auto-starts with new sessions (disable via

daemon.autoStart: false).

Commands

• tlive setup — configure IM credentials + register the Claude/Codex plugins

(hooks ride the Claude plugin; Codex needs none). --hooks-only re-registers

plugins only; add --claude / --codex to pick a vendor.

• tlive status — daemon health, effective mode, channels, and the Codex

companion state (running / degraded / off; degraded or off = Codex

approvals local-only).

• tlive mode off|notify|full|all — set posture (see intro). Persisted to

config, takes effect on the next hook; notify is the default, full =

remote approval on, all = also holds sub-agent approvals.

• tlive run <cmd> — wrap a process: local terminal + live web terminal (QR to open).
• tlive url — print the dashboard link + QR code.
• tlive logs -f — follow the daemon log.
• tlive start / tlive stop — explicit lifecycle (start is rarely needed;

sessions lazy-start the daemon unless autoStart is off).

Diagnostics

• No IM messages: tlive status for channel config; tlive logs -f for send

errors; confirm the daemon is up after starting a session.

• Codex has no remote cards: check tlive status — the companion line must say

running. off means codex isn't on PATH (or Windows); degraded means the

app-server child keeps dying — see ~/.tlive/codex-appserver.log. Either way

Codex still prompts locally; nothing is ever auto-run.

• No approval card ever arrives: check tlive status — the mode: line must

say full or all. The default notify never sends approval cards (tool

prompts stay local); enable remote approval with tlive mode full.

• Claude approval card unanswered (in full or all): the local dialog stays

live the whole time (parallel channels, first answer wins); answering

locally resolves the remote card as "answered in terminal". The remote

window defaults to ~24h (approvals.windowSec, shared by both vendors).

• Web page unreachable: tlive url for the current link (token is in the URL);

phones need the same LAN (or your own reverse proxy/VPN — tlive has no

publicUrl config, and cards never carry the link).

Security model in one breath

• Never auto-allow: unanswered → Claude's local dialog governs / Codex's native

prompt governs. Deny always carries a reason.

• Read-only tools (Read/Glob/Grep) pass by default. /safe on also auto-allows

routine ops (non-dangerous Bash, non-sensitive edits) — the danger floor

(rm -rf, sudo, .env/.ssh writes…) still asks and no config can lower it.

/trust on pauses approvals entirely (high risk — pair with allowedSenders).

• Runtime switches flip the same state from either entrance: IM commands

(/mute /trust /safe on|off, /mode for the posture ladder) and the CLI

(tlive mute|trust|safe on|off, tlive mode off|notify|full|all). /mute on

= go quiet; it silences IM notifications ONLY. The desktop toast is

independent of /mute and has no on/off switch of its own: it appears

whenever something is blocking on you (a pending approval, or the idle

"waiting for your input" nudge) and disappears once it is answered (macOS

is the exception — Notification Center has no scriptable close or replace,

so there it accumulates one entry per change instead of recycling). Silence

it with your OS's Do Not Disturb, or tlive mode off to stop tlive entirely.

A finished turn stays on IM (a per-turn toast would flood the screen).

• Vendor-side permissions.deny always wins; tlive never overrides it.

First-time onboarding

When the user says "help me set up tlive" (or runs /tlive:setup), walk them

through:

• tlive status to check the engine; missing → npm i -g tlive.
• No channels → collect Telegram (bot token + chat id) or Feishu

(appId + appSecret) credentials and merge into ~/.tlive/config.json:

{ "allowedSenders": [], "adapters": { "telegram": { "token": "…", "chatIdAllowList": ["…"] }, "feishu": { "appId": "…", "appSecret": "…" } } }

• tlive start → tlive status to verify channels; tlive url for the dashboard.
• Offer remote approval: tlive defaults to notify (watch + notify only). If

the user wants to Allow/Deny tool calls from their phone, run tlive mode full

(holds each tool call for a remote answer; reversible with tlive mode notify).

Leave it in notify if they only want monitoring. If they say they're

stepping away and want sub-agent approvals on their phone too, that's

tlive mode all — flag the trade plainly: a held sub-agent has no terminal

dialog until the window ends, so it only pays off when nobody is at the

keyboard (tlive mode full to come back).

• Codex needs no extra step — the companion starts with the daemon. If status

says off/degraded, that's diagnostic info, not a setup task.