[SPEC] Unify per-turn sandbox execution (streamSandboxPrompt/driveSandboxTurn) onto shared seam

[vutuanlinh2k2]

Status

Under consideration — NOT approved for implementation. Decide whether to proceed only after #75 ships and is confirmed connected + working well across all apps (creative + gtm + legal + tax). This is the "validate the provisioning unification holds across all four apps first, then decide" gate.

Blocked by #75.

Summary

After provisioning is unified onto the shared SandboxRuntimeConfig seam (#75), the next-largest duplicated surface across the apps is per-turn execution: each app hand-rolls streamSandboxPrompt / driveSandboxTurn / runSandboxPrompt. agent-app already ships a shared per-turn path (streamSandboxPrompt(shell, box, …) with severed-stream + interactive-question detection), but it is not yet rich enough for creative-agent to drop its own version.

Why deferred / why bigger

• 6 hot, user-facing call sites in creative alone (chat, design-canvas agent, sequences agent, mission dispatch, api-chat-action, designs/$documentId agent) — each must be re-verified.
• creative's per-turn path assembles 8 MCP servers per turn (delegation / proposal / mission / task / event / asset / media / media-decisions) plus surface-overlay merge, spend-key guard, and thread-id logic.
• creative's version also has model-call-failure detection that the shell's streamSandboxPrompt currently lacks.
• The shell's per-turn path must therefore be extended additively (MCP assembly via seams + model-call-failure detection) before creative can drop its version — otherwise features are lost.

Definition of done (if approved later)

• Shell per-turn path extended (additively) to cover creative's MCP assembly + model-call-failure detection — no behavior loss.
• creative-agent's bespoke streamSandboxPrompt / driveSandboxTurn / runSandboxPrompt deleted; all 6 call sites delegate to the shell.
• Provisioning unification ([TASK] Unify workspace sandbox provisioning onto shared SandboxRuntimeConfig seam #75) confirmed working across creative + gtm + legal + tax before this starts.

Context

Companion to #75 (provisioning unification). See the architecture analysis: the shell owns mechanism + control flow; per-app domain (which MCP servers, which model policy) is supplied through typed seams, never baked into the shell.