The adapter adds T3's MCP server and prompts at once; OpenCode still
reports the server pending, so its tools arrive mid-turn.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

OpenCode 2 answers mcp.add for a fresh Location while the server is still
pending, so T3's tools joined the Code Mode catalog mid-turn and restarted
Claude-backed turns. After adding, read mcp.list and wait for
mcp.status.changed until the server settles, bounded by INVENTORY_TIMEOUT.
A server that does not connect keeps the instructions without T3's tools.

Fixes #30240

Co-Authored-By: Claude Opus 5.5 <[email protected]>

OpenCode reports T3's MCP server connected before its tool registry picks
up the tools on a 100 ms debounced reload, and nothing reports that reload.
A prompt sent in that gap still gets the tools mid-turn, which restarts
Claude-backed turns.

Watch the server's mcp.status.changed from before mcp.add. When this turn's
add (re)connected it, wait 300 ms before prompting. An unchanged re-add of a
connected server emits no status change on OpenCode 2.2.5, so steady-state
turns pay nothing.

Fixes #30240

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…rails

OpenCode's status event can reach T3 after the mcp.list reply that already
reads connected, which skipped the settle on a thread's first registration.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…rompt time

A server still pending when the connected wait times out could finish its
handshake after the prompt and add T3's tools mid-turn.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…atus before the add

A reconnect's status event can trail the list reply, so an add after idle
eviction could prompt before OpenCode loaded T3's tools. List the server's
status before re-adding instead of racing the event.

AI-assisted; verified with the tests and tools listed above.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

OpenCode keeps connecting a server whose add request timed out, so its tools
could still join mid-turn. A failed add now takes the same withdrawal as a
server that is still connecting, under one timeout for the add and the wait.

AI-assisted; verified with the tests and tools listed above.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…e add

A server that read connected before the add could still be reconnected by
it, so the turn skipped the tool settle. Seeing it pending after the add, or
a reconnect that forgot it, now counts as a join too.

AI-assisted; verified with the tests and tools listed above.

Co-Authored-By: Claude Opus 5.5 <[email protected]>