plat 275

← Pulse platform issue index

PLAT-275 — Every agent start spent ~19 s per pass (38 s per Builder turn) retrying an MCP connector that can never connect without a login

Coordination	Value
Assigned agent	Claude Code
Ticket state	fixed — deployed to RTS 2026-09-03 and verified live; see Verification
Last synchronized	2026-09-03

• Priority: P1 — a 52 s hi on RTS, 38 s of it before the model launched; every turn and every scheduled run of a workflow with an unauthenticated connector paid it.
• Owner: mcpagent/mcpclient/client.go (Connect, ConnectWithRetry, isNonRetryableConnectError), mcpagent/agent/connection_session.go (lazy path only after a first successful connect).
• Origin: live, RTS rtslatency Builder chat 2026-09-03 ([LATENCY_DEBUG] T+38375ms | Starting StreamWithEvents for hi); reported by the operator.

What happened

rtslatency selects the Notion connector. The admin user has no Notion OAuth login on RTS, so its tool list was never cached and the lazy path (cache hit → defer) never applied; every agent build attempted the connection. The chat agent and the Workshop session it creates each ran their own pass: 2 × 19 s.

Root causes

• oauth.ErrNoValidToken ("no valid token available - run authentication flow": the token file <tokens root>/<user>/Notion.json does not exist; no network is involved) was retried like a transient failure.
• The retries nested by accident: ConnectWithRetry (4 attempts, 1/2/4 s backoff) called Connect, which has its own 3-attempt loop (1 s + 2 s). 4 × 3 + 7 = 19 s of pure sleep around twelve stats of a missing file.

Fix

• Missing OAuth token is non-retryable in both loops (mcpagent 13dabbb, TestConnectFailsFastWithoutOAuthToken).
• ConnectWithRetry makes single attempts (connectOnce), so transient failures get four tries, not twelve (b39c978, TestConnectWithRetryDoesNotNestConnectRetries).

Verification

Live on RTS after release 9d48c370c-20260903135124: the same hi reached Starting StreamWithEvents at T+399ms; both connector passes completed in under 0.5 ms. Remaining time-to-first-token (~14 s) is cursor-agent's own startup and is not platform time.

Follow-ups (not done)

• Skip OAuth connectors that have no token at agent build and say "Notion needs sign-in" in the chat instead of attempting and logging an error.
• Reuse the chat agent's connection results for the Workshop session instead of a second pass.
• Classify retries by cause (connection refused, timeout, 5xx) instead of by count; a 4xx or a refused registration should fail once.
• The hosted "Cloudflare Docs" catalogue entry returns HTTP 410 and costs 19 s at server boot; update or drop it.
• Cross-reference: PLAT-102 measured perceived send latency but never saw this pre-launch cost because its traces had authenticated connectors.