Add Gemini native computer use with model-visible evidence
Adds Gemini native computer use through the same CUA bridge and canonical facade. System instructions and cancellation use the SDK configuration; normalized coordinates map to browser coordinates. Includes package, build, CI registration, and the shared harness guide.
Stores the actual model-visible tool response and screenshot against its call ID, with optional fields for older records and no duplicate per-action captures. Confirmed runner-owned browser loss or a typed core facade loss is terminal; arbitrary error text, including agent-marked errors, remains recoverable. Missing, blocked, truncated, and malformed provider responses fail as SDK errors, and partial tool calls from those responses are not executed. Every function-call batch is validated before any member executes. Arbitrary model/resource identifiers are passed to the provider, with only the optional google/ alias removed.
Drag uses the existing SDK dragAndDrop primitive. Hotkey arrays execute as one chord, wait honors the requested seconds, and the current scroll action uses its documented default while retaining the legacy action's default.
Budgets use EVAL_GEMINI_CUA_MAX_TURNS. Explicit empty answers remain empty, unfinished calls remain failed, and the last model-visible image is retained when terminal capture is unavailable. That fallback is the last captured frame, not a fresh terminal screenshot. Transcript clipping preserves Unicode pairs.
Validation: 72 Gemini SDK/eval tests across seven files pass, including scripted terminal-response cases and generated executor code exercised against the actual SDK Page and RPC schemas over an in-memory transport. The earlier combined Claude/Gemini/shared suite, including the compiled facade/MCP fixture, passed 101 tests across 14 files before this Gemini follow-up. Gemini SDK and full eval typechecks and the Gemini SDK build pass; parent propagation and final combined validation are separate checks. These are local contract checks; live provider performance and semantic judge compatibility remain separate gates.
Stack base: evals/consolidation-16-claude-cua. Review this PR relative to its immediate predecessor.
Reviewer entry points:
packages/integrations/gemini-cua-sdk/src/executor.tspackages/integrations/gemini-cua-sdk/src/session.tspackages/evals/framework/harnesses/geminiCuaAdapter.ts
Current open stack: #2893 → #2894 → #2902 → #2903 → #2889 → #2904 → #2905 → #2906 → #2907.