Run Brain synthesis on the helper model; make Memory a settings toggle

[mrubens]

Two coupled changes that remove the Brain's provider-key requirement and make Memory self-serve.

Synthesis via the deployment helper model

gbrain's gateway-mode chat now defaults to a roomote/helper sentinel model. The Brain inference gateway answers it before provider resolution: OpenAI-style messages are converted and run through generateTrackedNonTaskText with modelRole: 'small' (new brain_synthesis usage surface), so synthesis works with whatever providers the deployment already has — including Anthropic-only and trial inference — with usage tracked like every other non-task call. Details:

• stream: true → 400 (gbrain's gateway chat is non-streaming by design); response_format json_object/json_schema honored via a strict-JSON instruction (schema inlined).
• An operator's R_BRAIN_MODEL still wins: the sentinel body is rewritten to it and forwarded through the ordinary provider path.
• Auth unchanged and enforced first: closed 404 with no gateway token, 401 before any body parsing, path allowlist before the sentinel branch.
• Entrypoint: the sentinel default applies only in gateway mode; direct-provider defaults untouched. The chat model is re-derived from env every boot, so existing gateway brains converge on next boot without touching an explicit operator choice.

Memory becomes a settings toggle

New nullable deployment_settings.brain_enabled (migration 0063). isBrainEnabled() is the Brain gate now: an explicit choice wins; null falls back to the legacy provider-key signal, so existing Brains stay enabled with zero backfill. All former isBrainProviderConfigured() gate sites (brain connection resolution — which covers collectors/outbox/MCP proxy — MCP delivery, fast-agent saveMemory tool, saveTaskMemory, web commands) now flow through it; the provider-key check survives only where it genuinely means "which key serves this request" and in the run-completion outbox insert (deliberately env-only, no DB read in that transaction).

Settings → Memory gains an admin-gated enable toggle (with a note when the enabled state comes from the legacy key); disabled state collapses the page to the toggle + explanation. The Settings nav shows Memory when either the wiring or the flag is present, so a wired-but-disabled Brain can be turned on.

Known limitations (called out on purpose)

• The Configuration section still displays the static synthesis-model default; making it gateway-aware is a follow-up.
• /v1/responses does not answer the sentinel (gbrain's gateway chat uses chat-completions).
• Helper responses report zero token usage to gbrain; real usage is recorded deployment-side under brain_synthesis.

Testing

brain-inference 20/20, db model-runtime-config 56/56, sdk mcp-connections 36/36, cloud-agents fast-agent 79/79, web server 21/21 + client 23/23; typecheck clean across all touched packages; entrypoint bash -n clean.

[roomote-community]

No code issues found. See task

• Default gateway helper model is not recognized (apps/api/src/handlers/brain-inference/index.ts:328). — dismissed: gbrain strips the provider prefix before calling the OpenAI-compatible gateway.

_{Reviewed 49eae3b}

[roomote-community]

The entrypoint configures gateway chat with a provider-qualified sentinel (${BRAIN_PROVIDER}:roomote/helper), but this only recognizes the bare roomote/helper. When gbrain sends its configured model, the request falls through to provider resolution and forwards the nonexistent helper model instead of calling the deployment helper. Normalize or accept the qualified sentinel here, and cover the gateway-configured value in this test suite.

[roomote-community]

Dismissed: gbrain parses the configured provider:model value and sends the bare model ID to its OpenAI-compatible client, so this gateway receives roomote/helper as expected.