Summary

There is no end-to-end coverage for the post-onboard messaging-channel lifecycle. Three bugs have shipped to main in this area in the last few weeks because no test exercises these flows:

• [Messaging] Cannot start a channel after a stop #3381 — channels start after channels stop reported success but the bot stayed offline (fixed via fix(onboard): preserve disabled channels through rebuild so channels start can recover #3395)
• bug(messaging) channels stop <name> does not stop the bridge through rebuild — openclaw.json still bakes the channel in #3453 — channels stop followed by rebuild still bakes the channel into openclaw.json (the build-time disable filter regression that surfaced after [Messaging] Cannot start a channel after a stop #3381 was closed)
• bug(messaging): channels add <mesaging> doesn't apply the network policy #3437 — channels add doesn't apply the channel's network policy preset, so the bridge boots with no egress

All three involve a full sandbox rebuild (>10 min each), so they belong as E2E scripts on Brev rather than as unit tests.

This issue tracks adding two E2E scripts under test/e2e/ that exercise the channel lifecycle end-to-end and assert against the baked image (/opt/nemoclaw/openclaw.json — the real source
of truth at runtime), not only the host-side registry. The #3381 → #3453 sequence is the exact reason: #3381's fix made the registry-side state correct, which masked the fact that the baked image
was still wrong until #3453 surfaced it.

Why two scripts instead of one

Each rebuild takes >10 min, so a single onboard → add → stop → start → remove script would run ~50 min and have a long mean-time-to-feedback when any one step regresses. Splitting into two
scripts:

• Bounds each run to ~30 min
• Lets us run them in parallel on separate Brev instances
• Makes regressions easier to attribute (channel-add bug vs. channel-stop bug)

Test 1 — onboard telegram → channels stop → channels start

Regression test for #3381 and #3453.

1. nemoclaw onboard --agent openclaw with telegram enabled
2. Verify the telegram bridge is up and openclaw.json contains the channel; bot responds to a DM
3. nemoclaw channels stop telegram; accept the rebuild prompt
4. Assert (baked image, bug(messaging) channels stop <name> does not stop the bridge through rebuild — openclaw.json still bakes the channel in #3453): docker exec <sandbox> jq .channels /opt/nemoclaw/openclaw.json does not contain telegram; gateway logs show no telegram-bridge provider attached
5. Assert (registry): disabledChannels: ["telegram"] and messagingChannels still includes telegram
6. nemoclaw channels start telegram; accept the rebuild prompt
7. Assert (baked image, [Messaging] Cannot start a channel after a stop #3381): openclaw.json contains telegram again
8. Assert (runtime, [Messaging] Cannot start a channel after a stop #3381): bridge is up and sending a DM to the bot gets a response — not just that the CLI printed success

Test 2 — onboard empty → channels add → channels remove

Regression test for #3437.

1. nemoclaw onboard --agent openclaw with no messaging channel selected
2. nemoclaw channels add telegram; provide token; accept the rebuild prompt
3. Assert (policy): nemoclaw policy list contains the telegram preset — no manual policy-add step required
4. Assert (network): bridge can reach api.telegram.org:443; sending a DM gets a response
5. Assert (baked image): openclaw.json contains telegram
6. nemoclaw channels remove telegram; accept the rebuild prompt
7. Assert: openclaw.json does not contain telegram; bridge process is gone; policy list no longer contains the telegram preset; registry shows telegram removed from messagingChannels