Codex from Telegram

Keep Codex moving when you leave the computer.

Codex UI is still where the serious coding happens. OpenClawBrain gives you the missing mobile layer: observe progress, follow one thread briefly, attach a passive note, send an explicit action, steer active work, and detach when Telegram should go quiet.

The point is not more notifications. The point is knowing when to look, when to wait, and when to give Codex one clear instruction at the right moment.

Install OpenClawBrain See the examples

What is Codex doing? What did Codex just say? Tell me when it replies Attach a passive note Send an explicit action Steer active work Detach cleanly Brief me before I sit down

Why this is useful

Sometimes Codex is running tests, fixing a bug, or waiting on a decision while you are away from the machine. Without a bridge, you either stay at the computer or lose the thread.

OpenClawBrain keeps the loop small and useful. Telegram becomes a calm operator channel, not a second IDE.

Check progress

See what Codex appears to be working on, how many threads are visible, and whether anything needs attention.

Read the latest reply

Copy the actual Codex message text into Telegram. No extra LLM summary, no waiting for another model.

Follow one thread

Tell OpenClawBrain which Codex thread matters and receive only completed assistant replies or attention-worthy events for a short window.

Note, act, steer

Attach passive context without starting Codex, send an explicit action when you want work to happen, or steer an active turn.

Clear examples

Example 1

Check what Codex is doing

You are on your phone and want a quick pulse without opening the computer.

/brain codex status

You get a short operator summary: visible threads, active work, watched items, app-server readiness, and the latest thread.

Example 2

Read the last thing Codex said

You missed the final answer and want the exact text in Telegram.

/brain codex last --latest

OpenClawBrain copies the completed Codex message directly from local rollout records. It does not summarize it through another LLM.

Example 3

Follow one Codex thread while away

You know which thread matters and want only useful updates.

/brain codex threads openclawbrain
/brain codex bind <thread-id>
/brain codex tail --bound

New completed assistant replies are forwarded for a short window. Routine tool chatter stays quiet. Use /brain codex detach to stop.

Example 4

Attach a passive note

You want to add context for the handoff, but you do not want Codex to start editing or running tools.

/brain codex note The failure looks like an auth mock issue, not a UI issue.

Notes stay in OpenClawBrain's local bridge state. They do not call Codex app-server and do not start a turn.

Example 5

Ask Codex to act

Codex is in the right thread and you want a real instruction to run.

/brain codex act --with-notes Please fix only the focused test.

Actions may edit files, run tools, or request approvals under Codex's normal sandbox. OpenClawBrain refuses fuzzy or latest-thread writes.

Example 6

Steer active work

Codex is currently working, and you want to redirect the active turn without starting a new one.

/brain codex steer --bound Please stop after this test run and report the exact failing assertion.

Steer only works for an exact or bound thread with an in-progress Codex turn. If the thread is idle, use /brain codex act instead.

Example 7

Get a return-to-computer brief

You are about to sit back down and want the current state without rereading the whole thread.

/brain codex handoff --bound

The brief separates observed facts from Codex-reported claims, so you know what was actually seen versus what still needs verification.

Example 8

Detach when Telegram should go quiet

You are done operating from your phone, or a tail got too noisy.

/brain codex detach

Detach removes the chat binding and pauses matching active watches in one move.

Why OpenClawBrain belongs in the middle

This is more than a notification pipe. The hard part is deciding what matters enough to send, what should stay quiet, and which thread is safe to write into.

Quiet by default

Plain watches send blockers, failures, approvals, and completions. Assistant-message tailing is explicit.

Exact when it should be

Recent-message copy is direct transport from Codex records, not an AI summary of an AI answer.

Intent is explicit

note is passive. act starts real work. steer redirects live work. detach stops the bridge.

Stock OpenClaw

The bridge lives in OpenClawBrain, so OpenClaw core can stay clean and upgradeable.

What it will not do

The bridge is useful because it is restrained.

No noisy mirror

Telegram is not a full Codex UI and does not stream every tool call or hidden internal event.

No database writes

OpenClawBrain reads Codex state and sends replies through Codex app-server. It never edits Codex SQLite or rollout files.

No wrong-thread shortcuts

Read commands can use latest. Write commands must be exact or bound.

No durable raw telemetry

OpenClawBrain stores bridge-local cursors and redacted audit events, not raw Codex transcripts as permanent memory.

Install or upgrade

openclaw plugins install clawhub:openclawbrain@0.2.33 --force
openclaw plugins enable openclawbrain
openclaw gateway restart

After install, start with /brain codex status. That single command tells you whether the bridge is seeing Codex and gives you the next useful thing to ask.