Repository navigation
No safe way to get an agent token when the desktop socket is unavailable #1920
Description
Activity
- added a commit that references this issue
on Oct 8, 2026 The desktop socket didn't fail here. It was slow:
GET /agentstook about 26 s because each request ran session discovery, which waited on Codex IPC. The root cause is fixed in relay-desktop#333.The documented fallback is now the same socket with a longer (60 s) timeout. We're not adding a CLI token path: many users don't have the npm
agent-relayCLI, so it isn't the right fallback.#1921 now keeps only the rotation guard.
agent-relay agent registerfails for an existing name and leaves that identity's token unchanged unless you pass--rotateexplicitly.That PR also includes a separate fix for the MCP "Connection closed" startup exit. The published 13.1.5 still has that bug, and the PR has the evidence.
Found by a teammate's Claude Code session during a Teams sign-up and Sessions setup on 2026-10-07 (macOS arm64, Desktop 2026.10.10, Claude Code 2.1.153, agent-relay CLI 12.4.0).
When the desktop socket route fails, an agent's fallback for messaging a peer is
agent-relay message dm send, which needsRELAY_AGENT_TOKEN. The only documented way to get one isagent-relay agent register, which can rotate an existing identity's token and break the session currently using it. So there's no safe fallback.Ask: a read-only or non-rotating way to get a token for the current session's own identity (e.g.
agent-relay agent token --current, or reuse via the session's existing credentials). Or document a supported fallback that doesn't rotate.Related: the relaycast MCP from CLI 12.4.0 (
agent-relay-mcp.js) failed to connect at Claude Code startup with "Connection closed"; current setup skills expect CLI 13.x.