MCP stdio: a data call landing between the transport attach and `kernel:bootstrapped` still memoizes a pre-bind localization for the life of the transport

[os-zhuang]

Found while implementing #11580 (PR on claude/issue-11580-mcp-stdio-settings-bind-window). Not fixed there — out of that card's scope, and closing it is a design choice with a cost, not a mechanical follow-on.

What #11580 fixed, and what it left

#11580 moved the stdio transport's localization read out of MCPServerPlugin.start(). It now resolves from a kernel:bootstrapped hook — strictly after SettingsServicePlugin binds its data engine (kernel:ready) — memoized so it stays one resolution for the life of the transport (#7279).

The memo has a second, lazy entry point: resolvePrincipal() awaits the same memo, so a read that arrives before the hook fires resolves it then instead of deadlocking on a hook a bare host may never fire. That lazy entry is deliberate, and it is also the residual:

MCPServerPlugin.start() ← runtime.start() attaches StdioServerTransport to stdin/stdout
 … other plugins' start() …
kernel:ready handlers ← SettingsServicePlugin binds the engine HERE
kernel:bootstrapped handlers ← the localization read HERE

The transport is live from the first line. A client that sends initialize + a data call while the boot is still working through the remaining start() bodies and every kernel:ready handler (schema sync, backfills — seconds on a real deployment) reaches resolvePrincipal() first. That call resolves localization pre-bind, gets UTC / en-US from the manifest defaults, and memoizes it — so the narrow race turns into the permanent wrong value #11580 was about, for that process.

Why it was left

Both cheap closures have a real cost, and picking one is a judgement, not a repair:

1. Make the early call WAIT for the bind (await a deferred resolved by the hook). Correct value always; introduces a hang on any host that never fires kernel:bootstrapped — worse than a wrong locale.
2. Memoize only post-bind answers (cache only once the hook has run; resolve fresh before that). No hang, correct steady state; makes the read per-call inside the window, and makes finding: a THIRD hand-written ExecutionContext assembly survives in the stdio MCP plugin, and it drops tabPermissions / accessToken #7279's "one resolution for the life of the transport" conditional on a hook firing, which several bare/lite hosts do not.
3. Do not attach the transport until kernel:bootstrapped. Closes the window at the source and is arguably where the transport belongs, but it changes when the stdio server starts answering at all — a visible behaviour change on the MCP surface.

Reproduction sketch

No test in the tree exercises it: packages/mcp/src/__tests__/plugin-settings-bind-window.test.ts drives every read after bootstrap() resolves. Reproducing it needs a read issued between MCPServerRuntime.start() and the kernel:bootstrapped dispatch — e.g. a kernel:ready handler registered after the settings provider's that calls the record reader.

Scope note

Narrow: needs a client fast enough to land inside the boot, and only bites deployments whose persisted localization differs from the platform defaults. Filed rather than fixed so the choice above is made deliberately.

[os-zhuang]

Triage: lands in packages/mcp (stdio transport's pre-bind localization memoization) → domain:cli, type Bug, pm:blocked.

Blocked-by: #11580

Rationale: #11580 is in flight on this exact seam — it moves the stdio transport's settings read out of the pre-bind window. Whether a data call landing between transport attach and kernel:bootstrapped can still memoize a pre-bind localization is decided by the shape #11580 lands; this card's remaining gap (if any) can only be measured on its merged ref. Unlock action: when #11580 closes, re-run this card's repro against the merged ref — if the memoization window is gone, close as absorbed; if a narrower window survives, re-queue with the measured residue.

---

Generated by Claude Code

[os-zhuang]

Triage unlock scan (session session_01LrqnyLhY3kYBBRird4peBT, 2026-08-24): upstream blocker #11580 closed 12:52Z (its PR merged). Surface re-verified: this card was filed from that same PR with the residual explicitly declared out of its scope ("Not fixed there"), so the merge that closed the upstream cannot have absorbed it — premise intact. pm:blocked → pm:queue, same stroke.

Note for the claiming cli seat: closure options 1 and 2 (wait-for-bind vs post-bind-only memoization) are within seat discretion; option 3 (deferring transport attach to kernel:bootstrapped) changes when the stdio server starts answering at all — a visible MCP-surface behaviour change. If the implementation converges on option 3, escalate for a maintainer ruling before dispatch rather than after.

---

Generated by Claude Code