# Computer Use Model-Loop Foundation ## Product Boundary The primary model path is the provider-neutral `maka_computer` function tool through `AiSdkBackend`. The model observes an Accessibility tree, then uses `click_element`, `set_value`, or another semantic action with IDs from that observation. This matches the recovered Codex layering: ```text model function/tool call -> Computer Use facade -> observed AX element identity -> stale-element refetch and uniqueness checks -> element executor -> AXPick / AXPress / AXValue when supported -> synthetic event fallback only inside the native executor ``` Codex does not currently expose a top-level native `computer` tool in the captured production request. Its deferred Computer Use wrapper exposes the AX facade. Accessibility-first dispatch happens inside element execution, not by implicitly turning every model coordinate into an AX element. ## Current Safety Policy Maka keeps the same separation: - `maka_computer` is the primary model-facing path; - semantic element actions and verified AX/CDP value updates are retained; - coordinate click, scroll, drag, key input, and pixel fallback are described as disabled and fail closed; - provider adapters use the same `maka_computer` contract rather than a separate native Computer Use loop. No provider adapter may infer a missing observation ID or silently bind an action to the current frame. ## Real Provider Results The local Azure Responses bridge at `127.0.0.1:8538` and coproxy Anthropic endpoint at `127.0.0.1:8537` were used without persisting credentials or raw provider responses. ### OpenAI Responses `gpt-5.6-sol` completed the product path: ```text list_apps -> observe -> set_value -> verified finish ``` The full product path passed: ```text getAIModel -> OpenAI Responses model -> AiSdkBackend / streamText -> ToolRuntime -> maka_computer -> synthetic AX semantic backend ``` The product path persisted tool calls and results, emitted permission-safe telemetry, and reached the verified final value. ### Anthropic `claude-sonnet-4-6` completed the same semantic task through coproxy. One run omitted `observation_id` from `set_value`. The harness returned a typed tool error requiring another observation; the model recovered instead of the executor guessing a frame. This behavior is a required regression scenario for future provider adapters. ### Kimi and MiniMax No live Kimi or MiniMax credential is configured on this machine. Their product paths are covered as hermetic protocol evidence, not real-provider evidence. Both `kimi-coding-plan` and `minimax-coding-plan` complete the same multi-step semantic loop through their exact Anthropic-compatible URL/auth contracts: ```text getAIModel -> streaming tool_use -> AiSdkBackend -> ToolRuntime -> maka_computer -> list_apps -> observe -> set_value -> final response ``` ## Provider Schema Findings OpenAI strict function schemas require every property to be listed in `required`. Optional fields must be represented as nullable. Because one function schema serves several action variants, the model can populate known fields that are irrelevant to the selected action. The OpenAI adapter therefore: 1. emits all properties as required and nullable; 2. rejects unknown keys and accessor properties; 3. projects only the known keys allowed for the selected action; 4. records discarded non-null keys; 5. passes the projected value to the existing strict action parser. The core `maka_computer` parser remains strict and provider neutral. ## Non-Claims This foundation does not: - reconnect a compatibility CGEvent path; - add the real AppKit AX provider runner (that is the next evidence-layer PR); - resolve PID reuse, stale driver nodes, or executor lifecycle hardening; - replace the executor-hardening and stacked-PR restack work.