feat(models): optional auto-alias creation for autosynced provider models

[kaspesi]

Problem

model_autosync keeps a provider's models map current, but those models are not routable: inference resolves the request model against aliases only (Router.resolveCandidates/resolve → findAlias, which reads config.models), and GET /v1/models enumerates aliases only. So a newly-discovered provider model never appears in /v1/models and can't be called by bare name until an operator hand-creates an alias for it. Autosync fills half the gap (the provider models map) but not the half clients actually use.

Change

Add an opt-in model_autosync.createAliases flag (default false). When enabled, after autosync inserts newly-discovered models into the provider's models map, the scheduler also creates a passthrough alias (slug === model id) for each new model that has no existing alias:

{
  "priority": "api_match",
  "sticky_session": true,
  "target_groups": [
    { "name": "<provider>", "selector": "random",
      "targets": [{ "provider": "<provider>", "model": "<id>", "enabled": true }] }
  ]
}

• Never overwrites an alias that already exists (a slug already taken by this or any other provider is left untouched).
• Fires the existing models-changed callback when aliases or provider models were added, so the change is picked up without a restart.
• preferred_api is intentionally omitted so the model family determines the default surface.

Details

• New model_autosync_create_aliases column (sqlite + postgres schema). Migrations not committed per repo workflow — CI regenerates on merge.
• ModelAutosyncScheduler.createMissingAliases() + wiring in runSyncNow().
• OpenAPI ProviderConfig.model_autosync.createAliases documented; asset regenerated.
• Admin UI: "Auto-create model aliases" checkbox in the provider Advanced editor (disabled unless autosync is enabled).
• Tests: scheduler create / skip-existing / off / callback-on-alias-only + DB round-trip of the flag.

Drive-by fix

Set an explicit Connection: keep-alive header on MCP SSE responses instead of relying on a runtime-implicit header. This fixes a pre-existing mcp-routes SSE test (expect(response.headers.connection).toBe('keep-alive')) that fails under Bun's inject/light-my-request harness, where the implicit header isn't synthesized. Needed to get a green bun run test locally.

Test plan

• bun run typecheck — all workspaces pass
• bun run test — full backend suite green (incl. new scheduler + DB tests)
• Manually verified on a live gateway: enabling the flag surfaces a newly-catalogued Codex model in /v1/models and routes an inference call to it end-to-end.

🤖 Generated with Claude Code

[mcowger]

Why not simply using the direct/ prefix?