fix(server): wait for Codex MCP startup before the turn snapshot

[macodev00]

Fixes pingdotgg#13437

Problem

Codex snapshots the model-facing tool catalog at turn/start. Optional stdio MCP servers that are still starting are omitted after a short grace, so the first turn only binds the servers that finished first. Later servers can be ready before the model answers and still be missing from that snapshot.

What Changed

CodexAdapterV2 records mcpServer/startupStatus/updated and holds turn/start until every server that has reported in for the thread leaves starting.

• A later ready replaces a spurious cancelled (app-server: thread/start emits duplicate MCP startup notifications, including a spurious cancelled, for servers that were only started once openai/codex#36682). cancelled does not replace ready.
• A server that has already reached ready or failed does not delay the turn. A thread with no startup notifications returns immediately.
• The wait is bounded at 30 seconds. Servers still starting or cancelled at the deadline are marked unavailable for that thread, so a hung server does not add 30 seconds to every later turn. A new starting after that blocks again.
• Notifications without a threadId apply to every thread. A thread-specific phase wins.

This does not call config/mcpServer/reload. That RPC only queues a refresh and an explicit refresh reconnects every configured server.

Scope and approval

Accepted bug, triaged in pingdotgg#13437. Julius confirmed the race on main and prescribed a bounded wait before turn/start until each expected server leaves starting, keyed by server name, with a later ready winning over a spurious cancelled.

Verification

./node_modules/.bin/vp test run apps/server/src/provider/CodexMcpCatalog.test.ts

Test Files 1 passed (1). Tests 8 passed (8). Duration 242ms. Started 2026-10-04 06:43:16 UTC.

The tests cover a partial catalog (one server ready, one still starting), a spurious cancelled that later becomes ready, the idle fast path, a timeout that does not block the next turn, threadless notifications, thread isolation, and failed not blocking.

./node_modules/.bin/vp test run apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.test.ts -t "waits for native start before interrupting an acknowledged queued turn"

Tests 1 passed, 132 skipped (133). Duration 4.98s. Started 2026-10-04 06:43:35 UTC. startTurn still completes when no MCP startup status has been reported.

./node_modules/.bin/vp run --filter t3 typecheck

Exit 0. Pre-existing Effect language-service suggestions only.

Limitations

The wait only includes servers that have already emitted mcpServer/startupStatus/updated when the turn samples the catalog. A first starting event that arrives after that sample can still miss that turn.

A server that is still starting, or stuck on cancelled without a later ready, is omitted after 30 seconds. Later turns do not wait for it again unless it reports starting again.

Not run against a live Codex app-server or the 10–20s stdio servers from the report.

  

[macodev00]

Opened upstream.