[CMGS]

A handle kept the owner it was claimed on. After a live move (#445), an evacuation (#437) or the old node's retirement, every call answered 404 or failed to dial until the lease ran out; only Close looked the claim up. The MCP server and the LangChain and OpenAI adapters hold handles for a long time, so they broke until restart. This was left open by the 2026-11-06 audit (#481).

What changes

Both SDKs follow a moved sandbox:

• Trigger. The owner's refused dial, its agent-upgrade 404 or its control-call 404 is tagged where it arises, with the sandbox id and the owner address it came from. Only a tag for this handle's sandbox at its current owner runs Lookup; a different owner then replaces the old one and parked connections are drained. A tag whose address a concurrent call has already replaced retries at the current owner without a second Lookup.
• Replay only when nothing reached the old owner.
  • Go retries the control calls, the one-reply and streaming verbs, and Run without Stdin.
  • Python retries every call, since its inputs are bytes; find_iter is the one exception.
  • Calls that already streamed input or handed back a stream rebind and fail once, and the next call reaches the new owner: Go Run with Stdin, WriteFile, Push, OpenPty, Watch, DialPort, Lsp.Request, Logs and Attach.
• Caller code is never retried. An error from the caller's own writer (Go Run stdout/stderr, ReadFileTo, Pull) or output callback (Python) is returned as it is, whatever tag it carries.
• Derived handles from Fork, Checkpoint and Promote bind to the node that answered, not to a concurrently rebound owner.
• A released sandbox still answers 404. The MCP server and both adapters inherit the change without code.

Numbers

4 commits. Production +366 −145 (net +221: Go +~110, Python +~70, and the rest moved or re-indented code); tests +245; docs +11.

Gates

• make go-lint: 0 issues. ×10 (5 modules × linux/darwin).
• make go-test green; asl clean on both GOOS.
• sdk/python: ruff, mypy and 285 tests; langchain 15 tests.
• Every new test fails on the code before its fix (negative control): moved, retired node, writer/callback errors tagged for another sandbox and for this one.
• Outside review: Codex, four rounds on one thread.
  • Round 1: a caller writer/callback error could trigger a replay; derived handles bound to a rebound owner.
  • Round 2: a nested call's tag from another sandbox was accepted.
  • Round 3: a nested call on the same handle's find_iter was accepted.
  • All fixed with tests. Round 4: "No blocking findings."
• No hardware run: a live move or evacuation with the MCP server attached is not covered here.