feat(runtime): expose a runtime-native model access lifecycle to hosts

[Yevanchen]

Context

Mosoo Computer currently implements runtime-native subscription access outside mosoo-agent-driver for Codex, Hermes, and OpenClaw. It works, but every host that wants subscription-backed runtimes would need to recreate the same runtime-specific adapters.

This issue records that implementation as a reference and asks whether the Driver should expose the smallest runtime-neutral lifecycle needed by hosts. It is not a request to move product billing, account storage, or OAuth token custody into the Driver.

Reference implementation

Mosoo Computer currently has one product-facing compatibility map:

Runtime	ChatGPT	GitHub Copilot	xAI
Codex	openai	—	—
Hermes	openai-codex	copilot	xai-oauth
OpenClaw	openai	github-copilot	xai

The implementation separates these concerns:

1. Compatibility and native provider names — one runtime/provider map plus binding validation (reference).
2. Normalized runtime actions — status, connect, disconnect, catalog, and select, translated to each runtime native CLI/protocol (contract, runtime adapters).
3. Cancellable streaming login — device URL/code output is streamed to the host, each attempt has an ID, and host cancellation terminates the native process (client lifecycle, sandbox execution).
4. Runtime-owned credentials — OAuth/device-login state stays in the runtime native HOME; the host persists/restores that HOME without reading or returning tokens (reference, persistence).
5. Run-time validation — before dispatch, the host distinguishes disconnected, reconnect-required, entitlement, transient, and invalid-model failures; only a ready binding reaches the turn driver (reference).

Current Driver gap

The Driver already states that hosts own credentials and persistence while it owns provider lifecycle (boundary). ACP initialization also publishes available auth methods and an authenticated/failed event (methods, session result).

However:

• auth discovery/result is ACP-specific and tied to Driver startup;
• Codex app-server and runtime-native CLI login are not represented by the common contract;
• the closed capability set has no model-access lifecycle capability (current capabilities);
• hosts cannot uniformly query connection state/catalog or start, cancel, reconnect, and disconnect a native account through the Driver.

Proposed minimum contract to explore

Expose a capability-gated runtime-native model access surface, implemented by provider backends where supported:

• discover supported account/provider IDs and native labels;
• query normalized connection state: disconnected | connecting | ready | reconnect_required;
• start/reconnect a login attempt and stream only user-actionable output;
• cancel an identified in-flight attempt;
• disconnect the native account;
• return the authenticated account model catalog;
• validate or apply the selected provider/model before a run.

The exact shape can reuse Driver commands/events rather than adding a separate local API. Existing ACP auth events are the closest starting point. Unsupported backends should advertise the capability as unsupported instead of emulating it.

Security and ownership invariants

• The host remains responsible for user/account authorization, durable HOME storage, and access policy.
• The runtime owns native OAuth/device credentials. The Driver must not return access tokens, refresh tokens, or raw auth files.
• Login output is bounded and sanitized; secrets and diagnostic stderr are not emitted as ordinary user-visible transcript.
• Cancellation must stop the native login process and leave a retryable state.
• A transient catalog/provider failure must not erase a valid sign-in.

Non-goals

• Billing, credits, entitlements accounting, or pricing.
• BYOK proxy/grant storage.
• A generic OAuth framework.
• Product UI or provider branding policy.
• Supporting speculative runtimes/providers before a backend can prove the lifecycle.

Suggested first slice

Define the capability and normalized status/event vocabulary, then prove one native backend (Codex + ChatGPT device login) and one ACP backend. Add live coverage for connect/cancel/status/catalog/reconnect without exposing credential material. Hermes/OpenClaw mappings above can remain reference data until their backends are in scope.