0.0.14-dev.0 advertises memwal_remember_status at cold start, but prod/staging relayers do not serve it

[harrymove-ctrl]

Found by driving the published bridges over stdio against all three relayers.

What happens

@mysten-incubation/memwal-mcp@0.0.14-dev.0 serves a 9-tool cold-start list that includes memwal_remember_status. The prod relayer does not implement that tool. Measured:

0.0.14-dev.0 -> PROD
  tools/list (cold start)  9 tools  ... includes memwal_remember_status
  tools/list (after swap)  8 tools  ... memwal_remember_status GONE
  memwal_remember_status   -> MCP error -32602: Tool memwal_remember_status not found

0.0.14-dev.0 -> DEV   (control, relayer has the tool)
  tools/list (after swap)  9 tools  ... includes memwal_remember_status
  memwal_remember_status   -> "remember job not found: 000...000"   (reaches the tool)

So during the cold-start window — before the relayer's tools/list_changed swap — an agent on 0.0.14-dev.0 against prod or staging can call a tool that does not exist there.

Why it matters more than a normal drift

The same run produced this:

0.0.14-dev.0 -> PROD
  memwal_remember  90.67s  ERROR
  Tool error: remember job timed out after 90000ms (job_id=f3430edf-66c8-4b40-b3a9-c9732438768f)

The agent is handed a job_id and told to settle it — and on prod there is no memwal_remember_status to settle it with. The recovery path the message points at does not exist in that channel.

For comparison, 0.0.13 against the same prod relayer returned 34.86s ok with a real blob_id.

Full matrix, same session

bridge	relayer	recall	remember
0.0.13	prod	2.87s	34.86s ok
0.0.13	staging	3.88s	25.92s ok
0.0.13	dev	7.08s	29.91s ok
0.0.14-dev.0	prod	2.08s	90.67s ERROR
0.0.14-dev.0	staging	2.73s	31.41s ok
0.0.14-dev.0	dev	6.48s	27.47s ok

The single prod timeout is one sample and could be queue variance rather than a deterministic regression — but the missing status tool is deterministic and verified in both directions.

Suggestions

• Gate the cold-start advertisement of memwal_remember_status on the relayer actually serving it, or drop it from the static list and let the post-swap list introduce it.
• Do not let a pending/timeout message name a tool that is absent from the current tools/list.
• More generally: the cold-start list is written against dev's sidecar, but a published bridge is routinely pointed at prod and staging, which lag. Worth asserting in the release check that the static list is a subset of what the oldest supported relayer serves.

Where

• packages/mcp/src/auth-required.ts (static cold-start tool definitions)
• packages/mcp/src/bridge.ts (the tools/list_changed swap)
• services/server/scripts/mcp/tools/remember.ts (the timeout message that names memwal_remember_status)

[github-actions]

Thank you for opening this issue, a team member will review it shortly.