unkempt merlin
Two days in. Chat replies work fine; tool dispatch blocked at gateway level. Fresh sessions reproduce.
Versions: Host CLI 2026.4.24 (npm, upgraded yesterday from 2026.2.19-2). Container 2026.4.6 (locally built, commit 6e1f18efe7). macOS → Docker.
Symptom: Agent replies "I can't execute commands in this session right now." sandbox explain shows Elevated.channel: (unknown) + allowedByConfig: false — but same values appear for agent:main:main (which IS working for chat), so those look like CLI display artifacts.
Sandbox state:
sessionKey: telegram:slash:<user-id>
runtime: direct
mode: off scope: agent
workspaceAccess: rw
Elevated: enabled:true channel:(unknown) allowedByConfig:false
Tried, no fix:
Onboarded openai-codex (started here)
Re-paired device — has all 6 scopes (operator.admin/approvals/pairing/read/write/talk.secrets)
Disabled channels.telegram.execApprovals.enabled
Added * wildcard + channel-prefix variants to tools.elevated.allowFrom
agents.defaults.sandbox = { mode:"off", workspaceAccess:"rw" } (confirmed flipped from none)
Telegram session reset
Ruled out plugin-runtime-deps/openclaw-unknown-* (it's the optional-channel-plugin dep pool)
Bug refs considered: #51245, #65690, #68634, #17607 — none quite fit.
Questions:
Where does telegram:slash:* get resolved into a channel string for the dispatch path?
Could host/container version split (4.24 ↔ 4.6) cause this — host emits a key shape container doesn't recognize, falls to default-deny?
Anyone hit this after upgrading host from 2026.2.x to 2026.4.x?
Thanks.
fervent token