REMOTE-2925 Add --computer-use-model to warp agent run

[warp-agent-staging]

Description

Adds --computer-use-model <MODEL> to warp agent run (hidden, alongside the existing --computer-use / --no-computer-use flags) and plumbs it through agent-run setup so it overrides the AI execution profile's computer_use_model when the client builds model_config.computer_use_agent for MAA requests.

Today cloud runs use the default CLI profile, where computer_use_model is None, so the computer use (CU) subagent always falls back to auto. This flag is what lets a run's configuration choose the CU model.

This is the client half of a two-repo change and must land first. warpdotdev/warp-server only emits --computer-use-model from AugmentArgsForTask once this flag exists in the agent build used by Warp-hosted workers; otherwise a run configured with a CU model would dispatch an unknown flag and fail at sandbox startup. Server-side work (run-config snapshot field, cloud run API field, named-agent defaults, factory-file computerUseModel, dispatch emission) is a separate PR in warpdotdev/warp-server.

Behavior notes:

• Override only. With the flag absent, behavior is byte-for-byte what it is today; no GUI change and no change to profile behavior.
• Oz-harness only: ignored when --harness selects a third-party harness, matching the design decision that the CU model is an Oz-harness sub-setting.
• A CU model together with --no-computer-use is accepted and simply unused, matching the server-side configuration semantics.
• The override beats the active profile, not just the CLI default profile, because the server dispatches --profile and --computer-use-model from the same run-config snapshot; an override that only touched the default CLI profile would be silently dropped whenever a profile is configured.
• The value is forwarded verbatim rather than resolved against the client's cached CU catalog. Unknown/disallowed ids are rejected server-side (at run creation, and again by MAA's getComputerUseAgentModel), so resolving locally would silently swap a misconfigured — or newly added, not-yet-cached — model for the default.

Linked Issue

REMOTE-2925 — https://linear.app/warpdotdev/issue/REMOTE-2925/factory-configurable-computer-use-model-factory-config-cloud-run-api

Testing

Automated tests cover the full chain from flag to request settings:

• crates/warp_cli/src/lib_tests.rs: flag parses, trims blank values, requires a value, and coexists with --no-computer-use.
• app/src/ai/agent_sdk/mod_tests.rs: the parsed flag becomes the driver task's computer_use_model for Oz and is dropped for a third-party harness.
• app/src/ai/llms_tests.rs: the run override wins over the profile's computer_use_model, is forwarded verbatim even when it isn't in the cached CU choices, and leaves other surfaces and the UI-facing resolution on the profile value.
• app/src/ai/agent/api/impl_tests.rs: RequestParams::computer_use_model lands in model_config.computer_use_agent (the request-settings construction was extracted into build_model_config to make this assertable).

Commands run in the agent sandbox:

• cargo test -p warp_cli --lib computer_use — 11 passed, 0 failed.
• cargo clippy -p warp_cli --all-targets --all-features --tests — clean.
• cargo fmt -p warp -p warp_cli — clean.

Not run: anything that compiles the warp app crate. cargo check -p warp --lib is OOM-killed (SIGKILL) in this 4 GB sandbox even at -j 1, so the app-crate edits and their tests (llms_tests, mod_tests, impl_tests) are unverified locally and need CI. No manual ./script/run testing for the same reason.

• I have manually tested my changes locally with ./script/run

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

[warp-agent-staging]

This PR was generated with Warp.

Comment @warp-factory on this PR to send it follow-up work.