Clean release scope

This NorthCUA release contains the NorthCUA Harness and is based on main 315508f3c08a7dec7f753691f38a609c9b8d0601 in a separate worktree. The original working checkout and historical research artifacts are not modified by the cleanup.

Disposition	Scope
Keep and validate	Desktop/Web agents, typed tools, observations, visual history, semantic compaction, ordinary skills, working memory, artifact/live evaluators, controller/device, generic external-task APIs, batch and artifact/reporting code
Remove	Private driver/PiP and companion control UI; training/data conversion/generation pipelines; environment-builder actor; CUA-Gym/WebHarbor/NorthJetty workflows; reusable warm pools; dated experiment wrappers, inventory acceptance suites and private/public NexAU migration release administration
Replace	Private model defaults → public env-based configs; dated template catalog → public Desktop/Web substrates; source-only install paths → packaged resources/wheelhouse; OSWorld hosted-dataset runner → local task JSON adapter; historical docs → user workflows and paper map
Test-only	Deterministic model/provider streams, fixture pages, fake provider/task scenarios and regression schemas; these are not live experiment claims

Decisions applied

• Breaking CLI/config/Python changes are allowed. Old outputs may be inspected; old runs are not promised to resume in place.
• Desktop targets Linux X11, macOS and Windows. Controller/batch's first-release host is Linux.
• Desktop actor defaults to screenshot/input; accessibility is optional. Web includes DOM/ref actions.
• Setup and evaluators may use constrained shell/file tools; they are not default Desktop actor tools.
• Public NexAU is pinned. OpenAI-compatible Chat, Responses, Anthropic and Gemini adapters remain.
• Every fresh/prepared environment belongs to one attempt; no task state is reused through a warm pool.
• Ordinary skills remain; recursive skill/checkpoint-search research systems are not shipped.
• License: Apache-2.0, as selected by the project owner.

Retained because removal is not yet justified

Area	Why it remains	Boundary
NexAU creation/history/provider adapters	Focused regressions depend on their semantics	Pin and contract tests; no private dependency
AgentKit and shared runtime abstractions	Used by both agents, evaluators and device transport	Public extension mechanism, not a third bundled agent
External/managed task custody and policy-route interfaces	Generic task/model-routing integrations	No compatibility promise to private RL clients
Case recipes and script evaluation	Needed for task setup and externally supplied oracles	Users supply app catalogs and task data
Legacy browser/shell MCP contracts	Still used by evaluator and generic tool integration paths	Not default Desktop actor tools
E2B SDK compatibility and reliable cleanup	Runtime cancellation/late-create tests depend on them	Cloud canary remains a separate release gate
Protocol extension values	Used by generic plugin contracts	No bundled mobile-agent support is advertised
Fake provider and deterministic fixture traces	Exercise failure/recovery without cloud credentials	Always identified as simulated, never proof of live task success

Further removal needs dependency and behaviour evidence. An uncertain component is kept with a stated boundary rather than silently deleted or presented as supported in every environment. The detailed original file inventory and cleanup audit are maintained alongside the implementation review outside the distributable packages.

Follow-up audit, 2026-10-03

A second review traced surviving entry points and failure paths. It removed the duplicate CLI registry/adapter chain and export-only visual planner, replaced historical test-template identifiers with synthetic fixture names, and deleted the unused root configuration stub and environment-builder contract kind. Active CLI handlers and provider adapters remain.

The review also found behavior that a deletion audit alone cannot establish: WebSocket admission/shutdown races, a missing Desktop failure terminal, logical session-ID mismatch, and remote cleanup reporting/cache retention. These are repaired and covered by targeted regressions; see verification for current run and review status.

The subsequent Desktop simplification keeps one tool-returned-image path. It deletes the pre-round probe policy, user-observation projection, their strategy factory and configuration types, custom observation-chain configuration, the empty window/control compatibility chain, and the duplicate legacy collector. Unused window/control/text probe adapters and export-only prompt interfaces are removed with their callers. Unused Desktop wait/clipboard configuration fields and orphaned driver test helpers are also removed; their presence did not correspond to active runtime behavior. Both agents use the paper's sliding overlay/clean allocator; the generic-budget branch and its selector are removed. Initial capture, action screenshots, history budgets, optional accessibility records and evaluator screenshots remain active functionality.

“Clean” describes a maintained, bounded implementation, not proof that every retained line is necessary or every environment has been exercised. The retained-uncertainty table above remains deliberate. New deletion proposals should trace actual callers and preserve the behavioral coverage of surviving paths.