diff --git a/.changeset/capability-loop-host-copy-first.md b/.changeset/capability-loop-host-copy-first.md
new file mode 100644
index 0000000000..5619243047
--- /dev/null
+++ b/.changeset/capability-loop-host-copy-first.md
@@ -0,0 +1,25 @@
+---
+"@objectstack/cli": patch
+---
+
+Stop claiming the `os serve` capability loop loads a "host copy first" (#10909).
+Two module-header comments in `packages/cli/src/commands/serve.ts` — above the
+`@objectstack/plugin-email` and `@objectstack/service-sms` imports — described
+the capability loop (`Serve.CAPABILITY_PROVIDERS`, the `for (const cap of
+requires)` block) as resolving `EmailServicePlugin`/`SmsServicePlugin` "host
+copy first". Measured at head, the loop does a bare `await import(spec.pkg)` /
+`await import(ex.pkg)` — no `importFromHost` in either path — which Node ESM
+resolves against **this CLI's own** realpath, so the CLI's bundled copy always
+wins; the host app's copy is never consulted. The comments described a
+behaviour the code does not have.
+
+The corrected comments also name the contrast the file now actually contains:
+`Serve.importConfigPlugin` (the served app's own `plugins: [...]` entries) IS
+host-anchored — an app-declared package wins there — while the capability
+loop is not. Making that split legible is the point of the fix, so the next
+reader does not assume one resolution rule governs the whole file.
+
+Comment-only: no runtime path, resolution order, or accepted specifier changes.
+All 21 `CAPABILITY_PROVIDERS` packages remain CLI-declared, so bare resolution
+still finds every one of them today — this only corrects what the comment
+claims about *how* that resolution happens.
diff --git a/packages/cli/src/commands/serve.ts b/packages/cli/src/commands/serve.ts
index 424a822d1f..15add70083 100644
--- a/packages/cli/src/commands/serve.ts
+++ b/packages/cli/src/commands/serve.ts
@@ -33,15 +33,19 @@ import { missingProviderMessage } from '../utils/capability-preflight.js';
 // transports rather than restated here (#5132) — `resolveEmailCapabilityArg`
 // has to refuse exactly the configurations `makeTransport` cannot build, and
 // two literal lists for one vocabulary is the drift #5094 was filed for. Values
-// only (no plugin class): `os serve` loads `EmailServicePlugin` itself through
-// the capability loop's dynamic import, host copy first.
+// only (no plugin class): the capability loop loads `EmailServicePlugin` itself
+// with a bare `import()`, resolved against THIS CLI's own realpath — its
+// bundled copy always wins, never the host app's. Contrast `importConfigPlugin`
+// below, which IS host-anchored: an app-declared package wins there (#10909).
 import { isEmailTransportProvider, emailProviderRequiresApiKey, unsupportedProviderFix } from '@objectstack/plugin-email';
 // The SMS provider vocabulary, read from the package that materialises the
 // transports, for the same reason and by the same rule as the mail one above
 // (#5713). `resolveSmsCapabilityArg` has to refuse exactly the tags
 // `makeSmsTransport` cannot build — restating `log`/`aliyun`/`twilio` here would
 // be the second literal #5094 was filed for. Values only (no plugin class): the
-// capability loop dynamic-imports `SmsServicePlugin` itself, host copy first.
+// capability loop dynamic-imports `SmsServicePlugin` itself the same way — a
+// bare `import()` resolved against this CLI's own realpath, so its bundled
+// copy wins, never the host's (#10909).
 import { isSmsTransportProvider, SMS_TRANSPORT_PROVIDERS } from '@objectstack/service-sms';
 import { resolveObjectStackHome } from '@objectstack/runtime';
 import { LOG_LEVELS, resolveLogLevel, readLogLevelEnv } from '../utils/log-level.js';