Eve library

Phase 180: Remote Creative Control Plane - MacBook Remote Host, Blender + Unreal Deep Control, Browser Automation, Computer Use Fallback, Live Control Room, Cloud Gateway, MCP Agent Integration, and Multi-Network Agent Operations

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

Build the Oshun Remote Creative Control Plane: a secure, observable, agent-native system that allows Oshun agents running on a primary MacBook, Hetzner, AWS, or any other trusted compute environment to control Blender, Unreal Engine, Chrome, and selected desktop workflows on a second MacBook Air over the same local network or across different networks.

The full vision is not a generic remote desktop product. It is an agent control fabric for creative production. Agents should prefer deep structured integrations with Blender, Unreal, and browser APIs; use Computer Use only when the target application has no reliable API path; stream the work in real time so the human can see and intervene; and preserve audit-grade traces of every instruction, command, screenshot, render, file change, and approval decision.

Primary owning domains: Bellona, Psyche, Yemaya, Iris, Nous, Contracts, Proto, OpenAPI, Oshun, Shared. Deep integration targets: Blender, Unreal Engine, Chrome/Chromium, Playwright/CDP, Browser Use, Codex Computer Use, Claude Computer Use, local macOS Accessibility/Screen Recording APIs, WebRTC, Tailscale / WireGuard optional networking, and cloud-hosted Oshun agents on Hetzner/AWS.

Existing Oshun anchor points to extend:

  • apps/bellona/bridge-blender/
  • apps/bellona/bridge-unreal/
  • libs/bellona/bridge-core/
  • libs/bellona/blender/
  • libs/bellona/unreal/
  • libs/bellona/blender-agent/
  • services/psyche/computer-use/
  • services/psyche/orchestrator/
  • services/psyche/api-gateway/
  • apps/yemaya/studio-web/
  • apps/yemaya/studio-desktop/

Target new layout:

  • apps/bellona/remote-host/ - installable macOS host app for the controlled MacBook.
  • apps/bellona/remote-gateway/ - cloud or LAN gateway for device sessions, routing, tunnels, MCP, and audit.
  • apps/bellona/control-room/ - web/Electron operator UI for live viewing, command timelines, approvals, takeover, and replay.
  • libs/bellona/remote-protocol/ - typed envelopes, state models, command schemas, stream metadata, and transport contracts.
  • libs/bellona/remote-adapters/ - adapter interfaces for Blender, Unreal, browser, desktop, files, and process control.
  • libs/bellona/mcp-gateway/ - MCP server and tool registry for remote devices.
  • libs/bellona/mac-host-runtime/ - macOS permissions, process launching, window focus, screen capture, signing/update helpers.
  • libs/psyche/desktop-fallback/ - reusable screenshot/click/type/window fallback adapter shared with Computer Use.
  • testing/bellona/remote-control/ - integration, e2e, fixture, and network impairment tests.

Strategic principle: one control plane, many execution surfaces. The agent should be able to say "open the Obsidian project in Unreal, import the latest GLB from Isis, place it in the scene, adjust lighting in Blender, preview the browser-based production board, and show me progress live" while Oshun routes each action through the safest and most deterministic tool available.


180.C Canonical Sequential Implementation Checklist#

This is the checklist coding agents must work through and check off from top to bottom. Do not skip ahead. Do not mark an item complete until the item is implemented, verified, documented where relevant, and committed. Expanded architecture and backlog detail live in TODOS/phase-180-reference.md; that reference file is not the executable checklist. This 180.C sequence is the operating checklist.

180.C.01 Phase Control, Tracker, and Ground Rules#

  • 180.C.01.01 Create docs/domains/bellona/extras/remote-control/implementation-tracker.md with a table for every 180.C item, owner, status, changed files, verification command, hardware evidence, and follow-up blockers.
  • 180.C.01.02 Add a "do not skip stages" rule to the tracker: downstream stages may start only when the prior stage is checked off or explicitly marked blocked with a documented reason.
  • 180.C.01.03 Add a "definition of checked" rule: an item can be marked checked only after code/docs are merged, verification passes, and the item itself is updated from [ ] to [x].
  • 180.C.01.04 Add a "hardware blocker" rule for macOS, Blender, Unreal, and WebRTC tasks: if CI cannot verify the behavior, the task must include a manual/hardware evidence note before being checked.
  • 180.C.01.05 Add a "no hidden privileged path" rule: no arbitrary shell, arbitrary Blender Python, Unreal console, desktop control, or real browser profile access before policy, approval, and audit primitives are complete.
  • 180.C.01.06 Add a "deep integration first" rule: Blender, Unreal, and browser tasks must use typed adapters before desktop fallback is allowed.
  • 180.C.01.07 Add a "localhost adapter" rule: Blender, Unreal, and browser debug/control ports remain localhost-only and are never exposed directly to LAN or internet.
  • 180.C.01.08 Add a "single source of truth" rule: command schemas live in libs/bellona/remote-protocol and are imported by gateway, host, MCP, adapters, tests, and docs.
  • 180.C.01.09 Add the initial verification command list for this phase: markdown format check, affected lint/typecheck/test, protocol tests, gateway tests, host tests, adapter tests, and Control Room Playwright tests once those packages exist.
  • 180.C.01.10 Check off this stage only after the tracker exists, ground rules are documented, and future agents can see exactly how to update progress.

180.C.02 Repository Audit and Architecture Decisions#

  • 180.C.02.01 Audit libs/bellona/bridge-core and document reusable protocol, heartbeat, request/response, reconnection, and validation behavior in the tracker.
  • 180.C.02.02 Audit apps/bellona/bridge-blender and libs/bellona/blender and document which Blender commands are real, stubbed, missing, or unsafe for remote operation.
  • 180.C.02.03 Audit libs/bellona/blender/python/bellona_addon and document addon startup, threading, handler registration, dependency, and Blender main-thread risks.
  • 180.C.02.04 Audit apps/bellona/bridge-unreal and libs/bellona/unreal and document Unreal-side plugin gaps, command coverage, version support, and test constraints.
  • 180.C.02.05 Audit services/psyche/computer-use and document which screenshot, OCR, accessibility, click, key, file, browser, and screen sharing primitives can be reused for macOS fallback.
  • 180.C.02.06 Audit services/psyche/orchestrator and services/psyche/api-gateway and document which session, WebSocket, lifecycle, and event patterns should be reused or avoided.
  • 180.C.02.07 Audit apps/yemaya/studio-web and apps/yemaya/studio-desktop and document where Control Room should live for MVP and for the final product.
  • 180.C.02.08 Create the first architecture decision record: standalone Bellona Remote Gateway plus Remote Host, with reuse of Psyche fallback primitives and Bellona bridge adapters.
  • 180.C.02.09 Create a compatibility matrix for macOS, Blender, Unreal, Chrome/Chromium, Node, Python, WebRTC, same-LAN, and cloud gateway modes.
  • 180.C.02.10 Create a dated external implementation research note covering MCP transports/auth, Chrome DevTools Protocol, Playwright, Browser Use, Codex/Claude Computer Use, WebRTC/TURN, macOS TCC permissions, Blender addon APIs, Unreal Editor plugin APIs, remote desktop security patterns, and cloud tunnel options.
  • 180.C.02.11 Create a threat-model ADR covering STRIDE-style risks, prompt injection, browser credential exposure, malicious web pages, compromised gateway, compromised host, compromised agent, local network attacker, replay attacks, and supply-chain tampering.
  • 180.C.02.12 Check off this stage only after all audit notes and ADRs are committed and no implementation package has been created prematurely.
  • 180.C.02.13 Add an ADR deciding trusted-gateway versus end-to-end encryption trust model between agent and host, and if E2E is adopted, enumerate key-exchange, audit-integrity, and operator-consent implications that cascade into gateway, host, audit, and replay design.
  • 180.C.02.14 Add an ADR deciding cross-platform host scope: macOS only, macOS+Windows, macOS+Windows+Linux, or phased. The decision gates whether 180.C.36 and 180.C.37 are in-scope and drives the mac-host-runtime vs host-runtime abstraction split.
  • 180.C.02.15 Produce a hardware rig inventory matrix listing required physical and virtual test machines, Blender versions, Unreal versions, Chrome channels, and OS versions needed for signoff on every hardware-gated item across the checklist.
  • 180.C.02.16 Define the MVP release artifact in an ADR: which stages must be complete, which demo scenarios must pass, and what documentation ships at MVP. Resolve the ambiguity between 180.C.16.09 (alpha gate) and 180.C.29.08-.10 (final gate) before implementation begins.

180.C.03 Remote Protocol Package#

  • 180.C.03.01 Generate or scaffold libs/bellona/remote-protocol using the repo's established Nx/library pattern and tags for Bellona/shared protocol code.
  • 180.C.03.02 Implement RemoteDevice, RemoteCapability, RemoteAdapter, RemotePermission, and RemotePermissionScope schemas with fixtures and schema tests.
  • 180.C.03.03 Implement RemoteSession, RemoteSessionStatus, RemoteParticipant, RemoteSurface, and RemoteSessionLifecycleEvent schemas with fixtures and schema tests.
  • 180.C.03.04 Implement RemoteCommandEnvelope, RemoteCommandPriority, RemoteCommandTimeoutClass, RemoteCommandRiskClass, and RemoteCommandTarget schemas with fixtures and schema tests.
  • 180.C.03.05 Implement RemoteCommandResult, RemoteCommandProgress, RemoteErrorCode, RemoteWarning, and retry/idempotency metadata schemas with fixtures and schema tests.
  • 180.C.03.06 Implement RemoteAuditEntry, RemoteAuditEvidence, RemoteArtifact, RemoteApprovalRequest, and RemoteApprovalDecision schemas with fixtures and schema tests.
  • 180.C.03.07 Implement RemoteStreamDescriptor, RemoteStreamSource, RemoteStreamState, and WebRTC signaling metadata schemas with fixtures and schema tests.
  • 180.C.03.08 Implement command namespace constants for device.*, session.*, stream.*, blender.*, unreal.*, browser.*, desktop.*, file.*, process.*, approval.*, and diagnostic.*.
  • 180.C.03.09 Export JSON Schema/OpenAPI/Proto-ready artifacts or add a clear generation task that fails if protocol schemas drift from generated contracts.
  • 180.C.03.10 Add protocol version negotiation, backward-compatibility rules, deprecation policy, minimum host/gateway version checks, and fixture tests for old host/new gateway and new host/old gateway behavior.
  • 180.C.03.11 Check off this stage only after protocol build, typecheck, unit/schema tests, and docs for schema ownership pass.
  • 180.C.03.12 Extend RemoteErrorCode with a structured remediation field: human text, machine-parseable suggested-next-command template, and a boolean marking whether auto-retry is safe. Every error path must populate it so agents stop burning turns rediscovering what to do next.
  • 180.C.03.13 Add a RemoteActor schema modeling the full delegation chain (human operator, orchestrator agent, specialist agent, adapter) with stable identifiers. Thread it through RemoteCommandEnvelope, RemoteAuditEntry, RemoteApprovalRequest, and timeline events.
  • 180.C.03.14 Add a RollbackStrategy enum to command metadata with values undo_stack, snapshot_restore, compensating_action, and manual_remediation_only. Require every mutating command schema to declare its strategy; reject envelopes without one at validation.
  • 180.C.03.15 Add an agent.readiness and state.snapshot command schema returning current dynamic state (modal dialogs open, shader compile count, project-root read-only state, running jobs, disk/memory pressure) alongside the existing static RemoteCapability reports.
  • 180.C.03.16 Add a DryRunPlan schema with predicted changes, cost estimate (tokens, compute, bandwidth, render minutes, build minutes), time estimate, reversibility assessment, and impact summary. Require dry-run support for every command in the high-risk classes defined in 180.C.07.06.

180.C.04 Mock Harness and Contract Tests#

  • 180.C.04.01 Create testing/bellona/remote-control with mock device, mock gateway, mock host, mock adapter, command fixture, and audit fixture utilities.
  • 180.C.04.02 Add a fake host adapter that can register capabilities, receive commands, emit progress, return success/failure, and disconnect.
  • 180.C.04.03 Add a fake gateway transport that can accept outbound host connections and simulate command dispatch without network dependencies.
  • 180.C.04.04 Add a fake MCP/client transport that can call command tools against the gateway mock.
  • 180.C.04.05 Add tests for command validation failure, command timeout, command cancellation, duplicate idempotency key, and host disconnect.
  • 180.C.04.06 Add tests for policy allowed, policy denied, approval required, approval granted, approval denied, and approval timeout.
  • 180.C.04.07 Add tests for audit emission on success, failure, deny, approval, cancel, timeout, and disconnect.
  • 180.C.04.08 Add snapshot fixtures for device capability reports, session timelines, Blender scene query, browser page query, and stream descriptor.
  • 180.C.04.09 Add the mock harness to CI or documented local verification commands.
  • 180.C.04.10 Check off this stage only after future gateway, host, MCP, and adapter work can reuse the mock harness.
  • 180.C.04.11 Implement a record-and-replay sandbox that captures real adapter responses from live sessions and replays them deterministically for offline agent development and regression testing without needing a live Blender, Unreal, or Chrome.
  • 180.C.04.12 Add a protocol fuzzing harness covering malformed envelopes, oversized payloads, Unicode edge cases (RTL marks, normalization forms, surrogate pairs), schema-downgrade attacks, and deeply nested structures. Block release if the fuzzer finds a crash, panic, or policy bypass.
  • 180.C.04.13 Add property-based tests for envelope invariants (idempotency key determinism, priority ordering, timeout-class monotonicity, risk-class escalation, actor-chain well-formedness) using fast-check or equivalent.

180.C.05 Remote Gateway Walking Skeleton#

  • 180.C.05.01 Scaffold apps/bellona/remote-gateway with health, readiness, config validation, logging, graceful shutdown, and test target.
  • 180.C.05.02 Implement in-memory device registry using remote-protocol schemas before adding persistent storage.
  • 180.C.05.03 Implement host outbound WebSocket connection acceptance with handshake validation, heartbeat, disconnect detection, and tests.
  • 180.C.05.04 Implement command enqueue and dispatch to a connected fake host with idempotency key support and tests.
  • 180.C.05.05 Implement command result, progress event, and error event routing back to the requesting client with tests.
  • 180.C.05.06 Implement session lifecycle endpoints or internal service methods for start, pause, resume, stop, and takeover state.
  • 180.C.05.07 Implement basic audit emission for gateway-received command, gateway-dispatched command, host-result, and session-ended.
  • 180.C.05.08 Implement gateway metrics/log fields for correlation ID, session ID, device ID, command ID, actor ID, and adapter target.
  • 180.C.05.09 Add gateway docs showing local startup, fake host connection, fake command dispatch, and expected logs/tests.
  • 180.C.05.10 Check off this stage only after the gateway can complete a mocked command round trip through tests.

180.C.06 Remote Host Walking Skeleton#

  • 180.C.06.01 Scaffold apps/bellona/remote-host or the chosen host package with configuration loading, logging, health, and test target.
  • 180.C.06.02 Implement host identity loading from a development file store first, with a clear future replacement by macOS Keychain.
  • 180.C.06.03 Implement outbound connection from host to gateway with handshake, heartbeat, reconnect, and disconnect event tests.
  • 180.C.06.04 Implement local adapter registry with fake adapter registration, capability reporting, health reporting, and command routing.
  • 180.C.06.05 Implement host command execution lifecycle: accepted, started, progress, completed, failed, cancelled, and timed out.
  • 180.C.06.06 Implement host-side policy preflight stub that denies unknown commands before the real policy package is added.
  • 180.C.06.07 Implement host diagnostics command returning OS, host version, adapter list, missing permissions placeholder, and resource placeholder.
  • 180.C.06.08 Implement cleanup on shutdown: cancel running command, close gateway connection, stop fake adapters, flush logs, and emit final event.
  • 180.C.06.09 Add host docs showing local fake-host startup against local gateway.
  • 180.C.06.10 Check off this stage only after fake host and fake gateway pass a complete command round trip both from gateway tests and host tests.

180.C.07 Device Pairing, Auth, Policy, and Approvals#

  • 180.C.07.01 Add persistent device, session, command, approval, and audit storage for the gateway using the repo's established database pattern or an explicitly documented first-stage store.
  • 180.C.07.02 Implement one-time pairing code creation, expiry, single use, and tests.
  • 180.C.07.03 Implement host pairing flow that exchanges pairing code for device identity and stores credentials in the development credential store.
  • 180.C.07.04 Replace development host auth with signed tokens or mTLS scaffolding and document the final macOS Keychain storage task.
  • 180.C.07.05 Implement device revocation and prove revoked hosts cannot reconnect or receive commands.
  • 180.C.07.06 Implement shared permission scopes and policy evaluator for observe-only, project write, desktop control, browser profile, arbitrary code, shell, destructive, publishing, and financial actions.
  • 180.C.07.07 Implement approval request creation, approval response, timeout, deny-by-default, and command continuation after approval.
  • 180.C.07.08 Add policy tests for every high-risk command namespace: Blender Python, Unreal console, shell, desktop control, file delete, browser profile, form submit, and upload.
  • 180.C.07.09 Add audit tests for pairing, revocation, policy allow, policy deny, approval request, approval grant, and approval deny.
  • 180.C.07.10 Check off this stage only after mutating adapter work is impossible to execute without policy and audit coverage.

180.C.08 Audit, Observability, and Timeline Foundations#

  • 180.C.08.01 Implement append-only audit service with correlation IDs, command hashes, result hashes, actor, device, session, and policy decision fields.
  • 180.C.08.02 Implement timeline query API/service returning ordered session events, command events, approval events, stream events, and artifact events.
  • 180.C.08.03 Implement artifact metadata registration for screenshots, renders, logs, browser traces, exports, and diagnostic bundles.
  • 180.C.08.04 Implement redaction utilities for tokens, API keys, passwords, browser cookies, personal paths, and denied window titles.
  • 180.C.08.05 Add structured logging to gateway and host with consistent correlation fields.
  • 180.C.08.06 Add basic metrics counters/timers for command count, command duration, command failure, host online state, reconnect count, approval latency, and audit write failure.
  • 180.C.08.07 Add tests proving audit entries are written before and after command execution for allowed, denied, failed, and approved commands.
  • 180.C.08.08 Add docs for reading a session timeline and exporting an audit bundle.
  • 180.C.08.09 Add failure behavior for audit store unavailable: block privileged commands and allow only explicitly safe diagnostics.
  • 180.C.08.10 Check off this stage only after every existing command path emits timeline and audit events.
  • 180.C.08.11 Implement a stored-data schema migration framework for device, session, command, approval, and audit records with forward and backward migrations, version stamps, and tests that load fixtures from at least two prior protocol versions.
  • 180.C.08.12 Thread the RemoteActor identity and delegation chain (from 180.C.03.13) through every audit entry so attribution traces human operator, orchestrator, specialist, and adapter with no ambiguity when multiple agents share a session.

180.C.09 Blender Read-Only Adapter MVP#

  • 180.C.09.01 Create or adapt the Blender adapter package so the Remote Host can discover a local Blender addon or bridge without exposing a public port.
  • 180.C.09.02 Implement Blender adapter health command reporting Blender installed/missing, addon installed/missing, Blender version, addon version, connection state, and active file path.
  • 180.C.09.03 Implement read-only blender.scene.info schema in remote-protocol with scenes, collections, objects, materials, cameras, lights, frame range, selected objects, and render engine.
  • 180.C.09.04 Implement host-to-Blender read-only command route using the existing addon/bridge where possible and a mock adapter where Blender is unavailable.
  • 180.C.09.05 Add unit tests for scene-info schema and adapter mapping.
  • 180.C.09.06 Add integration tests using a mocked Blender addon response.
  • 180.C.09.07 Add optional hardware/headless Blender smoke test that opens a tiny fixture scene and returns scene info.
  • 180.C.09.08 Route blender.scene.info from gateway/client to host to adapter and back to timeline/audit.
  • 180.C.09.09 Document how to run the Blender read-only MVP on the second MacBook.
  • 180.C.09.10 Check off this stage only after an agent/CLI can query Blender scene info through the full remote path.

180.C.10 Blender Safe Mutation MVP#

  • 180.C.10.01 Implement blender.object.create_primitive schema with name, primitive type, transform, collection, material reference, and timeout/risk metadata.
  • 180.C.10.02 Implement Blender undo checkpoint creation before the primitive command mutates the scene.
  • 180.C.10.03 Implement host-side policy check for Blender mutation and block if the session is observe-only.
  • 180.C.10.04 Implement before/after scene evidence capture through scene-info snapshots and optional viewport screenshot artifact.
  • 180.C.10.05 Implement primitive creation through the Blender addon or bridge with structured result containing object ID, object name, and evidence references.
  • 180.C.10.06 Add mocked adapter tests for success, invalid primitive, missing Blender, timeout, policy deny, and undo failure.
  • 180.C.10.07 Add optional hardware/headless Blender smoke test creating a primitive in a fixture scene and verifying scene-info changed.
  • 180.C.10.08 Add Control Room or CLI timeline visibility for the Blender mutation result before broad UI work begins.
  • 180.C.10.09 Document rollback/undo behavior and limitations.
  • 180.C.10.10 Check off this stage only after a remote command can safely create one Blender object with audit, evidence, and rollback notes.

180.C.11 Browser CDP MVP#

  • 180.C.11.01 Implement browser adapter capability discovery for Chrome/Chromium availability, managed profile path, CDP launch support, and Playwright availability.
  • 180.C.11.02 Implement isolated managed browser profile creation with cleanup and no personal profile access.
  • 180.C.11.03 Implement browser.launch_managed, browser.navigate, browser.snapshot, browser.screenshot, and browser.close schemas.
  • 180.C.11.04 Implement adapter launch of managed Chrome/Chromium with CDP bound to localhost only.
  • 180.C.11.05 Implement read-only navigation and page snapshot through Playwright/CDP with blocked-domain enforcement.
  • 180.C.11.06 Implement screenshot artifact creation with redaction hook placeholders and audit references.
  • 180.C.11.07 Add browser tests for launch, navigate, snapshot, screenshot, blocked domain, timeout, and cleanup.
  • 180.C.11.08 Route browser commands through gateway-host-adapter with timeline and audit visibility.
  • 180.C.11.09 Document browser profile isolation and why personal profile access remains blocked at this stage.
  • 180.C.11.10 Check off this stage only after a remote agent/CLI can open an isolated browser, navigate, snapshot, screenshot, and clean up.
  • 180.C.11.11 Implement a browser compensating-action framework. Browsers have no native undo for form submissions, downloads, or account changes, so each mutating browser command declares a documented reversal path (back-navigation, cancel, delete download, revoke form) populated into the RollbackStrategy metadata from 180.C.03.14.
  • 180.C.11.12 Require mutating browser command envelopes to carry compensating_action_id references and enforce their presence before dispatch. Reject mutating browser commands without a declared compensating action at schema validation.

180.C.12 Live Stream and Minimal Control Room MVP#

  • 180.C.12.01 Decide and document whether the MVP Control Room lives in apps/bellona/control-room or an existing Yemaya Studio app.
  • 180.C.12.02 Scaffold the chosen Control Room app/view behind a feature flag with route, app shell, test target, and no production exposure by default.
  • 180.C.12.03 Implement device list UI showing paired devices, online state, active session state, adapter health, and missing permission placeholders.
  • 180.C.12.04 Implement session timeline UI showing command queued, started, progress, completed, failed, denied, and approval-required states.
  • 180.C.12.05 Implement approval UI for approve once, deny, timeout, and policy explanation using the approval service from Stage 07.
  • 180.C.12.06 Implement stream descriptor flow and screenshot-only placeholder mode before full WebRTC media is ready.
  • 180.C.12.07 Implement WebRTC MVP for one source with start, stop, reconnect, stream health, and TURN/direct status if the repo tooling is ready.
  • 180.C.12.08 Add Playwright tests for device list, timeline states, approval states, stream placeholder, pause, and stop controls.
  • 180.C.12.09 Document operator workflow for watching an agent run the Blender and browser MVP commands.
  • 180.C.12.10 Check off this stage only after an operator can observe command progress and approve/deny in a UI, even if stream is screenshot mode first.
  • 180.C.12.11 Implement operator first-run onboarding: Control Room empty states (no devices paired, no sessions), a guided pairing walkthrough from the operator side, permission-meaning tooltips explaining observe-only versus project write in plain language, and a "what do I do next?" step after first load.
  • 180.C.12.12 Implement keyboard shortcuts for approve, deny, pause, resume, stop, and takeover with a documented shortcut map, a conflict check against browser defaults, and a visible cheat sheet inside the Control Room.
  • 180.C.12.13 Implement operating-system desktop notifications (macOS Notification Center at MVP, cross-platform via 180.C.31.06) for approval-required, host-offline, long-running completion, and policy-denial events with per-event user preferences.
  • 180.C.12.14 Implement a mobile approval-only PWA (responsive UI, push-notification capable) so operators can approve or deny from a phone without waiting for the full mobile Control Room deferred to 180.C.30.09. Approvals are the highest-value mobile moment.
  • 180.C.12.15 Implement an operator command palette for ad-hoc remote commands with audit capture, approval-required-by-default posture, and a searchable command registry drawn from the protocol namespaces.
  • 180.C.12.16 Implement a first-session guided tour walking through Blender read-only, browser navigate, and screenshot MVP flows so a new operator sees a success path within five minutes of opening the app.

180.C.13 Desktop Fallback MVP#

  • 180.C.13.01 Implement macOS permission detection for Screen Recording and Accessibility and surface missing permission diagnostics to gateway and Control Room.
  • 180.C.13.02 Implement desktop.screenshot for selected display or selected app window with redaction hooks and artifact registration.
  • 180.C.13.03 Implement desktop.window.list and desktop.window.focus with app name, title, bounds, display, and focus result.
  • 180.C.13.04 Implement coordinate normalization for Retina scaling, window-relative coordinates, and stream/screenshot coordinates.
  • 180.C.13.05 Implement desktop.click only for a focused, allowlisted app/window and only after desktop-control permission is granted.
  • 180.C.13.06 Implement visual verification after click using before/after screenshot artifact references.
  • 180.C.13.07 Implement emergency stop that interrupts queued and running desktop fallback commands.
  • 180.C.13.08 Add tests for missing permission, denied window, coordinate normalization, screenshot artifact, click dispatch mock, and emergency stop.
  • 180.C.13.09 Document when desktop fallback is allowed and when Blender/Unreal/browser adapters must be used instead.
  • 180.C.13.10 Check off this stage only after a remote agent/CLI can focus an allowed window, screenshot it, click a known target, verify, and audit the result.
  • 180.C.13.11 Implement a desktop compensating-action framework. Desktop clicks, drags, and key events have no native undo, so each mutating desktop command declares a documented reversal recipe (click-back, key-undo, explicit remediation) populated into RollbackStrategy metadata from 180.C.03.14.
  • 180.C.13.12 Implement keyboard-layout-aware desktop.type: detect active layout (QWERTY, AZERTY, Dvorak, Colemak, ISO versus ANSI), handle IME/composition, and emit scancodes or Unicode matching the target system's active layout rather than assuming US QWERTY.
  • 180.C.13.13 Implement multi-display targeting with stable display identities (not just primary/secondary) so a command targeting "the Blender reference display" still works after the operator rearranges displays between sessions.
  • 180.C.13.14 Implement a desktop.clipboard.capture command with redaction hooks so operators can copy error text visible on screen into session notes without agents inventing fake OCR content.

180.C.14 MCP and Cloud-Agent MVP#

Planning split for future refactors:

  • 180.C.14.A / MCP Gateway MVP: .01-.10.

  • 180.C.14.B / Cloud-Agent Operations and Policy: .11-.25.

  • This phase keeps the existing numbering stable; do not renumber Stage 14 while executing this checklist.

  • 180.C.14.01 Scaffold libs/bellona/mcp-gateway with local stdio MCP server support and test target.

  • 180.C.14.02 Expose MCP tools for device.list, session.start, session.stop, blender.scene.info, blender.object.create_primitive, browser.navigate, browser.snapshot, and desktop.screenshot.

  • 180.C.14.03 Add MCP resources for active device capabilities, active session timeline, recent artifacts, and last command result.

  • 180.C.14.04 Add MCP auth/session binding for remote/cloud clients through the gateway.

  • 180.C.14.05 Add tests proving MCP tools cannot bypass policy, approval, device revocation, or command schemas.

  • 180.C.14.06 Deploy or document a local cloud-agent simulation where an agent process connects through the gateway rather than direct LAN.

  • 180.C.14.07 Add Hetzner reference deployment notes for gateway, TLS, TURN placeholder, environment variables, logs, and smoke test.

  • 180.C.14.08 Add AWS reference deployment notes for gateway, TLS, storage, Redis/Postgres, logs, and smoke test.

  • 180.C.14.09 Run the cloud-agent MVP smoke flow: list device, start session, query Blender, navigate browser, capture screenshot, stop session.

  • 180.C.14.10 Implement an agent tool-selection policy that prefers domain APIs, then Blender/Unreal/browser adapters, then high-level browser agents, then desktop fallback, then human handoff.

  • 180.C.14.11 Implement agent preflight checks for device online state, permissions, app availability, project root grants, stream availability, budget limits, and approval requirements before any remote task starts.

  • 180.C.14.12 Implement long-running job handling for agents: progress polling, cancellation, checkpointing, resume, and timeout escalation for renders, builds, file transfers, and imports.

  • 180.C.14.13 Implement visual verification policy after high-impact edits, requiring before/after evidence or a documented reason evidence is impossible.

  • 180.C.14.14 Implement dry-run plan and human-readable impact summary generation before risky Blender, Unreal, browser, desktop, file, or shell actions.

  • 180.C.14.15 Implement multi-agent locking so two agents cannot concurrently mutate the same Blender file, Unreal map, browser profile, project root, command queue, or approval request.

  • 180.C.14.16 Implement agent handoff paths between cloud model, local model, browser specialist, Blender specialist, Unreal specialist, desktop fallback specialist, and human operator.

  • 180.C.14.17 Implement capability memory so agents remember stable device/app capability results and stop repeatedly probing unchanged remote hosts.

  • 180.C.14.18 Implement autonomous stop criteria: task done, approval denied, policy blocked, repeated failure, host busy too long, stream lost, low confidence, or budget exceeded.

  • 180.C.14.19 Add evaluation tasks comparing deep-integration agents against Computer Use-only baselines for Blender, browser, desktop, and Unreal workflows.

  • 180.C.14.20 Check off this stage only after an MCP client can drive the remote Mac path through the gateway with no inbound MacBook ports.

  • 180.C.14.21 Implement the agent.readiness and state.snapshot commands defined in 180.C.03.15 across all adapters so agents can query current dynamic state rather than rediscovering blockers through failed commands.

  • 180.C.14.22 Implement structured error-remediation handling. Every RemoteErrorCode returned must populate the remediation field from 180.C.03.12, and the MCP layer surfaces suggested-next-command templates directly to the agent for reduced wasted LLM turns.

  • 180.C.14.23 Implement DryRunPlan per-namespace planners (180.C.03.16) for Blender, Unreal, browser, desktop, file, and shell so risky commands always produce a plan object before execution.

  • 180.C.14.24 Implement a cost and time estimator populating DryRunPlan with token, compute, TURN bandwidth, render-minute, and build-minute estimates drawn from historical execution telemetry.

  • 180.C.14.25 Implement MCP composition rules for Bellona MCP running alongside other MCP servers (Isis, Yemaya, third-party): namespacing, tool-name collision policy, discovery ordering, and transport multiplexing documented and tested.

  • 180.C.14.26 Treat this stage as two logical units for future work planning: items .01-.10 are MCP Gateway MVP, items .11-.25 are Cloud-Agent Operations and Policy. Future refactors may split them into 180.C.14.A and 180.C.14.B.

180.C.15 Unreal Read-Only and Safe Mutation MVP#

  • 180.C.15.01 Document the Unreal plugin architecture using lessons from the completed Blender/browser/gateway/host path.
  • 180.C.15.02 Implement Unreal adapter capability discovery for Unreal install path, engine version, project path, plugin state, editor running state, and supported command list.
  • 180.C.15.03 Implement read-only unreal.project.info schema and mocked adapter response.
  • 180.C.15.04 Implement read-only unreal.world.query schema and mocked adapter response.
  • 180.C.15.05 Build or complete the Unreal Editor plugin connection to the local host runtime with authenticated localhost transport.
  • 180.C.15.06 Route unreal.project.info and unreal.world.query through gateway-host-adapter with timeline and audit visibility.
  • 180.C.15.07 Implement unreal.actor.spawn as the first safe mutation with policy check, editor transaction/undo, result object reference, and screenshot/evidence placeholder.
  • 180.C.15.08 Add mocked tests for Unreal missing, plugin missing, project closed, read-only query success, spawn success, policy deny, and transaction failure.
  • 180.C.15.09 Add optional hardware Unreal fixture project smoke test for query and actor spawn.
  • 180.C.15.10 Check off this stage only after Unreal follows the same verified control path as Blender: schema, policy, adapter, audit, evidence, and docs.

180.C.16 Alpha Hardening and MVP Release Gate#

  • 180.C.16.01 Replace development credential storage with macOS Keychain storage for Remote Host device keys and tokens.
  • 180.C.16.02 Add signed/notarized macOS packaging plan or initial implementation for Remote Host, LaunchAgent, helper tools, Blender addon, and Unreal plugin.
  • 180.C.16.03 Add update channel, signed update manifest, rollback, and adapter compatibility plan.
  • 180.C.16.04 Add network impairment tests for gateway restart, host reconnect, host sleep/wake, packet loss, high latency, and TURN fallback placeholder.
  • 180.C.16.05 Add security tests for prompt injection, denied paths, denied app/window, revoked device, approval bypass attempt, and secret redaction.
  • 180.C.16.06 Add performance budgets and measurements for command latency, stream startup, screenshot capture, browser navigation, Blender command time, and gateway-host reconnect.
  • 180.C.16.07 Add runbooks for two-MacBook LAN setup, cloud gateway setup, permissions troubleshooting, plugin repair, stream failure, revocation, audit export, and uninstall.
  • 180.C.16.08 Add deployment smoke tests for local, LAN, Hetzner, and AWS reference modes as those modes become available.
  • 180.C.16.09 Add MVP alpha release gate requiring protocol, gateway, host, policy, audit, Blender, browser, desktop fallback, Control Room, MCP, and Unreal checks before MVP alpha release.
  • 180.C.16.10 Check off this stage only after the MVP is stable enough that deeper Blender, browser, Unreal, deployment, security, and fleet stages can extend the same architecture without replacing it.
  • 180.C.16.11 Define explicit MVP demo scenarios and acceptance criteria: solo creator pairs two MacBooks on a LAN, agent queries Blender scene, creates a primitive, navigates a browser, captures a screenshot, and operator approves plus replays the session. The demo scripts become release evidence.
  • 180.C.16.12 Define SLO targets and error-budget policy for command acknowledgement latency, stream start latency, approval round-trip time, and audit write durability. Burn-rate alerts gate release.
  • 180.C.16.13 Add a soak-test plan for multi-day host uptime: sleep/ wake cycles, LaunchAgent crash-loop recovery, gateway reconnect stability, audit store growth under sustained load, and memory-leak detection across 72+ hour runs.
  • 180.C.16.14 Align this Alpha/MVP gate with the final GA gate in 180.C.29.08-.10. Document which stages each gate requires so the two do not drift into divergent release criteria.
  • 180.C.16.15 Produce a bridge coexistence and deprecation plan for existing apps/bellona/bridge-blender and bridge-unreal consumers describing shim behavior, parallel operation, cutover date, and final removal gate. Execution is tracked in 180.C.38.

180.C.17 Blender Production Command Coverage#

  • 180.C.17.01 Implement blender.file.open, blender.file.save, blender.file.save_as, blender.file.pack_resources, and blender.file.recover_last with project-root policy checks, audit, and file snapshot behavior.
  • 180.C.17.02 Implement blender.object.delete, blender.object.rename, blender.object.duplicate, blender.object.parent, blender.object.set_transform, and blender.object.set_visibility with undo checkpoints and mocked tests.
  • 180.C.17.03 Implement blender.collection.create, blender.collection.move_object, collection visibility, collection renderability, and collection hierarchy queries with scene fixture tests.
  • 180.C.17.04 Implement blender.material.create_principled, blender.material.assign, blender.material.set_parameter, texture binding, and material semantic search with before/after evidence.
  • 180.C.17.05 Implement blender.asset.import for GLB/GLTF, FBX, OBJ, USD/USDZ where supported, Alembic, image textures, and audio references with file allowlists and artifact provenance.
  • 180.C.17.06 Implement blender.asset.export for GLB/GLTF, FBX, USD, OBJ, Alembic, image sequences, and .blend packaging with transfer registration and hash verification.
  • 180.C.17.07 Implement semantic selectors for "all emissive objects", "selected character rig", "large static props", "objects in collection", and "materials missing textures" with deterministic query results.
  • 180.C.17.08 Add golden-scene tests for object edits, material edits, collection edits, import/export, undo, file save, and scene-info round trips.
  • 180.C.17.09 Add hardware/headless Blender verification notes for every command that cannot run in normal CI.
  • 180.C.17.10 Check off this stage only after Blender has production scene/object/material/file/import/export command coverage through the same gateway-host-adapter-audit path.

180.C.18 Blender Advanced Automation and Render Coverage#

  • 180.C.18.01 Implement safe modifier stack commands: add, configure, reorder, disable viewport, disable render, apply with snapshot, and rollback note generation.
  • 180.C.18.02 Implement structured Geometry Nodes graph macros for scatter, cable runs, kitbash arrays, modular facades, terrain dressing, and procedural prop layout.
  • 180.C.18.03 Implement safe mesh edit macros for bevel, extrude, inset, merge by distance, normal recalculation, UV unwrap, decimate, and triangulate with selection preconditions.
  • 180.C.18.04 Implement camera and lighting presets for product shot, three-point lighting, HDRI lookdev, cinematic blocking, orthographic asset preview, and turntable render.
  • 180.C.18.05 Implement render commands for viewport screenshot, OpenGL preview, Cycles/Eevee still render, playblast, thumbnail capture, progress events, cancel, and artifact registration.
  • 180.C.18.06 Implement rigging setup macros for armature creation, IK/FK setup, retarget preparation, control shapes, and pose-library operations behind approval gates.
  • 180.C.18.07 Implement simulation setup macros for cloth, rigid body, soft body, particle, fluid, and smoke as guided setup only, with bake commands blocked until explicit approval exists.
  • 180.C.18.08 Implement arbitrary Blender Python as a privileged command with static preflight, project-root constraints, approval, redaction, and full command/result audit.
  • 180.C.18.09 Integrate libs/bellona/blender-agent validation rules before applying high-level automation plans and record validation findings in the timeline.
  • 180.C.18.10 Check off this stage only after advanced Blender automation is policy-gated, test-covered, reversible where possible, and visible in the Control Room timeline.

180.C.19 Browser Production Automation Coverage#

  • 180.C.19.01 Implement browser.back, browser.forward, browser.reload, browser.new_tab, browser.close_tab, browser.switch_tab, and browser target inventory with CDP tests.
  • 180.C.19.02 Implement DOM-aware browser.click_element, browser.type_into, browser.select, browser.submit, and browser.wait_for with accessible selector fallback and failure evidence.
  • 180.C.19.03 Implement download handling with project-root destination, file hash, content-type recording, blocked extension policy, and artifact registration.
  • 180.C.19.04 Implement upload handling with approved source paths, file-size limits, origin policy, operator approval for sensitive uploads, and browser trace evidence.
  • 180.C.19.05 Implement optional persistent automation profiles with explicit grants, per-origin restrictions, session expiry, and revocation.
  • 180.C.19.06 Implement real personal Chrome profile access only as a privileged mode with explicit human consent, visible stream, restricted origins, and credential-entry privacy controls.
  • 180.C.19.07 Implement browser trace recording for failed tasks and attach trace artifacts to the command timeline.
  • 180.C.19.08 Implement prompt-injection test pages that try to exfiltrate cookies, bypass policy, submit forms, upload files, or instruct the agent to ignore Oshun rules.
  • 180.C.19.09 Add task templates for production-board updates, asset marketplace search, documentation lookup, account-free research, and authenticated admin workflows.
  • 180.C.19.10 Check off this stage only after browser automation covers deterministic web tasks with profile isolation, origin policy, artifacts, prompt-injection tests, and audit.

180.C.20 Browser Use and Optional Extension Integration#

  • 180.C.20.01 Evaluate Browser Use local mode, Browser Use cloud mode, and direct CDP against the same browser task fixtures and document reliability, privacy, latency, and cost tradeoffs.
  • 180.C.20.02 Implement a browser-agent provider interface that can route tasks to direct CDP, Browser Use local, Browser Use cloud, or another MCP browser provider without changing planner code.
  • 180.C.20.03 Normalize Browser Use or external browser-agent step events into Oshun command progress, timeline events, screenshots, traces, and audit entries.
  • 180.C.20.04 Add policy gates that prevent third-party browser providers from receiving authenticated profiles, sensitive origins, file uploads, or form submissions without explicit approval.
  • 180.C.20.05 Add tests proving Browser Use integration cannot bypass Oshun command schemas, device policy, approvals, redaction, or audit.
  • 180.C.20.06 Evaluate whether a Chrome extension adds concrete value beyond CDP for overlays, profile grants, page annotations, and consent UI.
  • 180.C.20.07 If extension mode is approved, implement the minimal manifest, domain-scoped permissions, extension-origin authentication, and visible automation indicator.
  • 180.C.20.08 Add extension tests for malicious iframes, clipboard access, file upload prompts, form submission, blocked origins, and extension disable/uninstall.
  • 180.C.20.09 Document when direct CDP, Browser Use, extension mode, and desktop fallback are each allowed.
  • 180.C.20.10 Check off this stage only after browser-agent delegation is optional, policy-controlled, observable, and never required for the deterministic browser MVP.

180.C.21 Control Room Production Experience#

  • 180.C.21.01 Expand the Control Room device list with adapter versions, missing permissions, active app state, resource pressure, last seen, and diagnostic actions.

  • 180.C.21.02 Expand the session page with active agent task, plan, current tool call, command queue, command timeline, approval queue, artifact panel, and session notes.

  • 180.C.21.03 Implement live stream source switching between display, app window, browser tab, Blender viewport, Unreal viewport, and screenshot-only low-bandwidth mode.

  • 180.C.21.04 Implement stream overlays for agent cursor, clicks, drags, key events, selected window, selected Blender object, active Unreal actor, and browser target element where available.

  • 180.C.21.05 Implement privacy controls for excluded windows, blurred regions, credential-entry pause, recording disable, and screenshot redaction preview.

  • 180.C.21.06 Implement operator controls for pause, resume, stop, takeover, release takeover, approve once, approve for session, deny, edit constraints, and ask agent to revise plan.

  • 180.C.21.07 Implement audit replay view synchronized across timeline, screenshots, optional stream segments, commands, approvals, and artifacts.

  • 180.C.21.08 Implement search and filters across commands, artifacts, file paths, Blender object names, Unreal actor names, URLs, approvals, status, and risk class.

  • 180.C.21.09 Add Playwright and accessibility tests for all Control Room states: loading, empty, offline, online, streaming, approval, takeover, failed command, replay, and redacted content.

  • 180.C.21.10 Check off this stage only after a human can supervise, interrupt, approve, take over, replay, and understand a full remote session from the UI.

  • 180.C.21.11 Implement a live cost visibility panel showing real-time per-session token, gateway compute, TURN bandwidth, render-minute, and build-minute spend with a running total and per-command attribution so operators can decide before approving the next step. Uses the abstraction from 180.C.31.07.

  • 180.C.21.12 Implement a multi-session grid/mosaic view with mini-streams, a consolidated cross-session approval queue, and a cross-session risk ranking so an operator supervising several agents in parallel knows where to look next.

  • 180.C.21.13 Implement frame-accurate stream scrubbing backed by a keyframe index (pairs with 180.C.25.15) so operators can jump to any moment in a replay without replaying from the start.

  • 180.C.21.14 Implement a timeline annotation layer: operator notes tied to specific timestamps, stream moments, or commands, persisted in audit and visible in replay.

  • 180.C.21.15 Implement named bookmarks with persistence, export, and per-user bookmark lists across sessions.

  • 180.C.21.16 Implement secure expiring share links for specific timeline moments: signed URLs, audit-linked access, configurable expiry, and automatic revocation when the underlying session is redacted or tombstoned via 180.C.27.17.

  • 180.C.21.17 Implement a diff view for file and scene changes: visual before/after for text files, image diffs for screenshots, structural diff for Blender scene-info snapshots, and actor diff for Unreal world queries.

  • 180.C.21.18 Implement i18n/l10n in Control Room: string extraction, per-locale date/number/currency formatting, RTL layout, and translator-friendly resource files. Consumes scaffolding from 180.C.31.01.

  • 180.C.21.19 Implement screen-reader-friendly timeline semantics, high-contrast and reduced-motion themes, and WCAG 2.1 AA conformance tested at every Control Room milestone per 180.C.31.02.

  • 180.C.21.20 Run a structured usability test with new operators: time-to-approve under pressure, clarity of policy explanations, approve/deny error rate, takeover discoverability. Report findings and fix the top three friction points before release.

    _Blocked:_ Requires recruiting and observing real new human operators
    under time pressure; cannot be run in this sandbox. The UI capabilities it
    would evaluate (approve/deny flow, policy explanations, takeover
    discoverability) are implemented and tested in 21.06/21.09/21.19. A local
    `operator-usability-study` plan/evaluator now fail-closes the evidence
    contract and rejects malformed or placeholder participant
    identities/roles, duplicate participant identities including case variants,
    duplicate per-participant scenario observations, metric-only observations
    without required evidence-kind artifacts, duplicated observation evidence
    including case variants, observation artifacts that do not
    identify their participant/scenario/evidence-kind context with concrete
    evidence non-placeholder suffixes, stale, missing, or non-canonical
    timestamp observation metadata,
    published reports that omit recruited participants or observation
    evidence, published reports that cite invalid or incomplete observations,
    reports outside or escaping the canonical evidence path or published
    before observations, stale, non-canonical, or malformed report metadata,
    zero-duration human timing metrics, and weak top-three fix evidence with
    duplicate IDs including case variants, without concrete non-placeholder
    suffixes, or verified before the findings report or without canonical UTC
    millisecond verification timestamps, but real recruited operators and
    top-three friction fixes remain required.
    

180.C.22 File, Artifact, and Cross-App Workflow Coverage#

  • 180.C.22.01 Implement project root grants with allowed paths, denied paths, read/write scopes, conflict policy, and audit-visible grant state.
  • 180.C.22.02 Implement resumable file transfer to host with hash verification, destination validation, disk-space preflight, overwrite approval, and transfer progress events.
  • 180.C.22.03 Implement file transfer from host for screenshots, renders, exports, logs, traces, and diagnostic bundles with artifact metadata and retention policy.
  • 180.C.22.04 Implement file snapshots before destructive operations and restore for project-local files.
  • 180.C.22.05 Integrate artifact storage with S3/MinIO or the existing Oshun artifact store and record hashes, sizes, content types, and provenance links.
  • 180.C.22.06 Implement "Isis to Blender": generated asset selected, transferred, imported, materialized, preview rendered, and provenance attached.
  • 180.C.22.07 Implement "Blender to Unreal": selected asset exported, transferred, imported into Unreal, placed in map, screenshot verified, and audit-linked.
  • 180.C.22.08 Implement "Unreal render to Yemaya": Movie Render Queue or screenshot output collected, uploaded, attached to review, and timeline linked.
  • 180.C.22.09 Implement "browser research to asset action": browser references captured with citations/screenshots, then applied through Blender or Unreal after approval.
  • 180.C.22.10 Check off this stage only after core cross-app creative flows work through transfers, artifacts, provenance, approvals, and audit.
  • 180.C.22.11 Implement an asset dependency graph indexing which Blender materials, Unreal material instances, Substance packages, and downstream assets reference each texture, mesh, or source asset. Update on import, export, and mutation events.
  • 180.C.22.12 Implement reference integrity queries: find broken references, detect missing source files, report stale assets, and flag upstream changes that invalidate downstream artifacts.
  • 180.C.22.13 Implement an impact-preview command used by the dry-run plans in 180.C.14.23 that lists downstream artifacts affected by a proposed edit, including which collaborators' working files may need refresh.
  • 180.C.22.14 Implement an optional production-tracking integration (Autodesk Flow/Shotgun, ftrack, Kitsu) so agent actions reflect as task status updates, note attachments, and version publishes where configured. Cross-references 180.C.34.06.

180.C.23 Unreal Production Command Coverage#

  • 180.C.23.01 Implement unreal.map.open, unreal.map.save, unreal.map.create, and unreal.map.list with source-control and dirty state reporting.
  • 180.C.23.02 Implement unreal.actor.destroy, unreal.actor.set_transform, unreal.actor.set_property, unreal.actor.attach, and unreal.actor.select with editor transactions and policy gates.
  • 180.C.23.03 Implement unreal.blueprint.call_function with schema validation, allowlist support, structured result, and approval for non-allowlisted functions.
  • 180.C.23.04 Implement unreal.material.set_parameter for scalar, vector, texture, and material instance targets with screenshot evidence.
  • 180.C.23.05 Implement unreal.asset.import for GLB/FBX/USD/textures and unreal.asset.reimport for source asset updates with provenance.
  • 180.C.23.06 Implement unreal.asset.find, unreal.asset.rename, unreal.asset.move, unreal.asset.delete, and redirector fixup with approval for destructive operations.
  • 180.C.23.07 Implement unreal.console.execute with allowlist, approval for arbitrary commands, command output capture, and high-risk audit.
  • 180.C.23.08 Add Unreal fixture tests for map open/save, actor mutation, material update, asset import, Blueprint call, console allowlist, undo/transaction, and policy denial.
  • 180.C.23.09 Add hardware Unreal verification notes for every command that cannot run in normal CI.
  • 180.C.23.10 Check off this stage only after Unreal editor control is production-capable for maps, actors, materials, assets, Blueprint, console allowlist, evidence, and audit.

180.C.24 Unreal Sequencer, Render, Build, and Diagnostics#

  • 180.C.24.01 Implement unreal.sequencer.list, unreal.sequencer.open, unreal.sequencer.set_frame, unreal.sequencer.play, and unreal.sequencer.render_preview.
  • 180.C.24.02 Implement Unreal viewport, active camera, and named camera screenshot capture with artifact registration.
  • 180.C.24.03 Implement Movie Render Queue/Movie Render Graph job creation, launch, progress, cancel, output collection, and artifact linking.
  • 180.C.24.04 Implement Play-in-Editor controls for start, stop, eject, state query, and restricted input injection where policy allows.
  • 180.C.24.05 Implement build/cook/package commands with platform target, configuration, map list, output path, logs, progress, cancel, and artifact collection.
  • 180.C.24.06 Implement shader compile, derived data cache, asset registry, and source-control status reporting so agents can wait and recover intelligently.
  • 180.C.24.07 Implement Unreal diagnostic bundle collection for editor logs, crash reports, plugin logs, project settings, adapter telemetry, and recent command timeline.
  • 180.C.24.08 Add virtual-production observation hooks for Live Link status, nDisplay/Switchboard status, camera tracking state, and stage monitor screenshots as read-only capabilities.
  • 180.C.24.09 Add performance benchmarks for world query, actor mutation, screenshot capture, asset import, render job, and build command overhead.
  • 180.C.24.10 Check off this stage only after Unreal long-running render/build/editor workflows are observable, cancellable, artifact-linked, and recoverable.

180.C.25 Networking, Cloud Deployment, and Transport Hardening#

  • 180.C.25.01 Implement mDNS/Bonjour discovery for same-LAN hosts while keeping identity verification certificate/token based.
  • 180.C.25.02 Implement direct LAN TLS/WebSocket command routing when policy allows it, with fallback to gateway tunnel.
  • 180.C.25.03 Implement host-initiated outbound gateway tunnel over port 443 with reconnect, session resume tokens, duplicate suppression, and backoff.
  • 180.C.25.04 Implement WebRTC TURN relay fallback with stream metrics showing direct versus relay mode.
  • 180.C.25.05 Separate transport channels for control, events, logs, file transfer, artifact metadata, stream signaling, and emergency stop.
  • 180.C.25.06 Implement backpressure and priority handling for emergency stop, approval, observe, interactive control, engine command, file transfer, batch job, and telemetry.
  • 180.C.25.07 Add network impairment tests for high latency, packet loss, gateway restart, host sleep/wake, NAT change, Wi-Fi disconnect, and laptop lid close.
  • 180.C.25.08 Create production-ready Hetzner deployment with TLS, gateway service, TURN, database, object storage, logs, monitoring, backups, and smoke test.
  • 180.C.25.09 Create production-ready AWS deployment with load balancer, gateway service, RDS/Postgres, Redis, S3/object storage, logs, monitoring, TURN, backups, and smoke test.
  • 180.C.25.10 Add optional Tailscale/WireGuard private-network mode with setup docs, trust boundaries, diagnostics, and explicit guidance that it does not replace gateway auth or policy.
  • 180.C.25.11 Evaluate QUIC/WebTransport for future multiplexed control and media signaling, document whether to adopt, defer, or reject it, and capture migration requirements if adopted later.
  • 180.C.25.12 Check off this stage only after same-LAN and cloud-agent modes both work, recover from network failures, and expose clear diagnostics.
  • 180.C.25.13 Implement HDR and wide-gamut stream support (PQ/HLG transfer functions) with color-space metadata carried in RemoteStreamDescriptor and preserved end-to-end from capture to Control Room preview.
  • 180.C.25.14 Implement multi-track streaming: separate tracks for cursor, audio, and region-of-interest video, enabling bandwidth- adaptive encoding and selective recording.
  • 180.C.25.15 Implement stream recording with a keyframe index for fast seek in replay (pairs with scrubbing in 180.C.21.13).
  • 180.C.25.16 Implement hardware codec selection with a documented fallback chain: VideoToolbox on macOS, NVENC on Windows, VA-API on Linux, software fallback otherwise. Record the selected codec in session metadata.
  • 180.C.25.17 Implement bandwidth-adaptive bitrate with sender-side bandwidth estimation and automatic quality switch under congestion, exposing direct-versus-relay mode and current bitrate in operator UI.

180.C.26 macOS Host Packaging, Permissions, and Updates#

<!-- STAGE 26 STATUS (2026-07-07): 26.01 and 26.02 are now CLOSED with live hardware evidence from the physical macOS rig MAC-HOST-03 (macOS 26.5.2, Apple Silicon): the native SwiftPM AppKit menu-bar shell in apps/bellona/remote-host/macos (NSStatusItem UI rendered from the host's loopback token-guarded /v1/local-control/menu-bar-model, which derives the model through deriveMacOSMenuBarModel — single source of truth), the host-side local-control surface (pause/resume/stop-session/unpair/diagnostics) in src/local-control.ts with focused tests, and a full launchd lifecycle run (bootstrap/enable/kickstart in the gui domain, KeepAlive crash restart with new pid, shell-supervised host-child restart, launchctl kill TERM graceful stop with last exit code 0 and no restart, bootout + plist + bundle removal). Evidence: docs/domains/bellona/extras/remote-control/evidence/180.C.26.01/ and .../180.C.26.02/. The locally testable contracts remain in src/macos-host-support.ts + macos-host-support.test.ts, and the local gate still rejects malformed signing env, missing per-scenario QA artifacts, incomplete, stale, or non-canonical timestamp evidence metadata, QA artifacts that do not identify their scenario/evidence-kind context, QA records from non-macOS host runners, generic install-QA operator identities, duplicated artifact IDs including case variants, stage artifacts that do not identify both the checklist item and exact evidence slot, placeholder suffixes, and placeholder Stage 26 booleans. The remaining open items (26.06, 26.09, 26.14) still require an Apple Developer ID signing identity (0 codesigning identities on MAC-HOST-03), signed-package install QA (chronologically after signing evidence), and non-developer acceptance evidence. -->

  • 180.C.26.01 Build the macOS Remote Host app shell with menu bar status, local pause/stop/unpair controls, diagnostics, and session-active indicator.

  • 180.C.26.02 Implement LaunchAgent registration, restart after crash, graceful shutdown, and clean uninstall.

  • 180.C.26.03 Implement first-run onboarding for Screen Recording, Accessibility, Automation, Files and Folders, local network access, and exact missing-permission diagnostics.

  • 180.C.26.04 Implement app and adapter discovery for Blender versions, Unreal versions, Chrome/Chromium, Playwright browsers, Bellona bridge services, plugins, and addon install state.

  • 180.C.26.05 Implement install/repair/uninstall actions for Blender addon, Unreal plugin, browser helper, and stream/desktop helper processes.

  • 180.C.26.06 Code-sign and notarize Remote Host, helper tools, LaunchAgent, Blender addon package, Unreal plugin package, and update installer.

    _Blocked:_ Needs an Apple Developer ID certificate plus macOS
    codesign/notarytool; no signing identity or macOS toolchain is available
    in this sandbox. The readiness contract now checks required environment,
    Developer ID/notary/Team ID shape, artifacts, and commands in remote-host
    macos-host-support.test.ts.
    
  • 180.C.26.07 Implement signed update manifests, channels, canary rollout, compatibility checks, rollback, and emergency update revocation.

  • 180.C.26.08 Implement diagnostic bundle creation with redaction and user-approved upload.

  • 180.C.26.09 Add clean install, upgrade, permission revoked, adapter repair, update rollback, crash-loop, and uninstall tests/manual QA.

    _Blocked:_ The full scenario matrix must run against SIGNED, notarized
    packages on a real macOS host: clean-install/upgrade/update-rollback
    require signed package receipts and the evidence chronology gate requires
    QA runs captured after the 26.06 signing evidence, which is still blocked
    (no Apple Developer ID identity on this rig). The crash-loop and
    uninstall lifecycle mechanics were already exercised live on macOS rig
    MAC-HOST-03 during the 26.02 launchd run (see
    evidence/180.C.26.02/), and the local scenario matrix and evidence
    evaluator remain covered by remote-host macos-host-support.test.ts.
    
  • 180.C.26.10 Add SBOM generation, dependency vulnerability scanning, license checks, malware scanning expectations, and release evidence for host app, gateway, Control Room, Blender addon, Unreal plugin, browser helper, and native helpers.

  • 180.C.26.11 Add macOS hardened runtime, entitlement, sandbox/not sandboxed rationale, Accessibility/Screen Recording permission rationale, and notarization failure recovery documentation.

  • 180.C.26.12 Add tamper detection for modified host binaries, unsigned helper tools, unexpected plugin files, mismatched update manifests, and local adapter binaries outside expected hashes.

  • 180.C.26.13 Add MDM-friendly and offline/private-studio install/update paths with configuration profile guidance and no-cloud update option.

  • 180.C.26.14 Check off this stage only after a non-developer can install, pair, grant permissions, run, update, diagnose, and uninstall the Remote Host safely.

    _Blocked:_ Stage gate still depends on the gated items 26.06 (Apple
    Developer ID signing/notarization — 0 codesigning identities on rig
    MAC-HOST-03) and 26.09 (signed-package install QA), plus a real
    non-developer human performing the install/pair/update/uninstall
    acceptance run (`operator:non-developer:` evidence). 26.01/26.02 closed
    on macOS rig MAC-HOST-03 on 2026-07-07. The roll-up gate is encoded in
    remote-host macos-host-support.test.ts so it stays open until concrete
    macOS evidence records exist for every required run. The final-GA
    evidence gate requires this Stage 26 macOS packaging/acceptance artifact
    before `180.C.29.09` can pass.
    
  • 180.C.26.15 Implement continuous DCC plugin integrity checks: verify Blender addon and Unreal plugin hashes on every session start (not only at install time), block session start on mismatch, and emit a tamper-detection audit event. Extends 180.C.26.12.

  • 180.C.26.16 Implement the cross-platform permission abstraction interface in libs/bellona/host-runtime so Windows and Linux hosts (180.C.36, 180.C.37) can plug in without duplicating adapter and command code.

  • 180.C.26.17 Treat this stage as two logical units for future work planning: items .01-.08 are macOS Host App, items .09-.16 are Signing, Updates, SBOM, and Integrity. Future refactors may split them into 180.C.26.A and 180.C.26.B.

180.C.27 Security, Privacy, and Governance Completion#

  • 180.C.27.01 Implement device key rotation, device revocation, suspicious clone detection, local unpairing, and trusted gateway pinning.
  • 180.C.27.02 Implement host-side deny rules that cannot be overridden by a compromised gateway.
  • 180.C.27.03 Implement project-root enforcement blocking home directories, SSH keys, password stores, browser secrets, system settings, and unrelated repositories by default.
  • 180.C.27.04 Implement credential-entry mode where the human enters secrets locally while stream capture is paused or blurred and no keystroke data is logged.
  • 180.C.27.05 Implement redaction for logs, screenshots, traces, browser content, command payloads, file paths, API keys, tokens, and personal data.
  • 180.C.27.06 Implement organization controls for device pairing, session start, stream access, risky approval, audit export, MCP clients, and external browser providers.
  • 180.C.27.07 Implement spend and usage controls for cloud agent tokens, gateway compute, TURN bandwidth, artifact storage, render jobs, and build jobs.
  • 180.C.27.08 Implement audit retention policies, data residency options, access review, and administrator reports for devices/sessions.
  • 180.C.27.09 Add security review tests for prompt injection, denied paths, sensitive windows, approval bypass, revoked device, token redaction, third-party browser delegation, and local deny rules.
  • 180.C.27.10 Complete privacy/legal review for remote screen capture, browser session handling, employee monitoring risk, consent, recording retention, audit export, and cross-border artifact storage.
  • 180.C.27.11 Add organization policy templates for solo creator, small studio, enterprise studio, education lab, local-only private mode, and locked-down review-only mode.
  • 180.C.27.12 Check off this stage only after the system is safe enough for real authenticated browser, file, Blender, Unreal, and desktop sessions.
  • 180.C.27.13 Implement agent identity and delegation-chain audit end-to-end: operator, orchestrator, specialist, and adapter threaded through every audit entry, approval decision, command envelope, and timeline event using RemoteActor from 180.C.03.13.
  • 180.C.27.14 Implement consent capture for incidentally-recorded subjects in stream content: FaceTime windows, dictation overlays, bystander cameras, visible documents. Prompt-based consent with redaction fallback when consent is unavailable.
  • 180.C.27.15 Implement stream and recording watermarking (visible and cryptographic) for leak attribution back to session, operator, and agent chain.
  • 180.C.27.16 Implement jurisdiction-aware retention defaults (EU, California, UK, US federal) with operator override, documented defaults per organization template, and automated enforcement.
  • 180.C.27.17 Implement a right-to-erasure workflow reconciling user-deletion requests with append-only audit: tombstoning, cryptographic erasure of redaction targets, and a documented compliance pathway preserving audit integrity while honoring deletion requests.
  • 180.C.27.18 Extend tamper detection from install-time (180.C.26.12) to continuous: re-verify host binaries, helper tools, Blender addons, Unreal plugins, and native helpers at every session start and periodically during long sessions.

180.C.28 Observability, Reliability, and Replay Completion#

  • 180.C.28.01 Add dashboards for gateway health, host fleet health, command latency, command failure, stream quality, approval latency, adapter health, reconnects, and audit write failures.
  • 180.C.28.02 Add alerts for host offline, command failure spike, stream failure, policy violation attempt, crash loop, update failure, audit store unavailable, and relay cost spike.
  • 180.C.28.03 Implement tamper-evident audit hash chains for high-trust command sessions.
  • 180.C.28.04 Implement replay bundles containing timeline, screenshots, optional stream segments, command payloads, result summaries, approvals, artifacts, and redaction metadata.
  • 180.C.28.05 Implement replay export and machine-readable audit export for compliance, debugging, and incident review.
  • 180.C.28.06 Implement failure taxonomy and user-facing error codes for host offline, gateway offline, adapter offline, app missing, app crashed, permission missing, timeout, policy denial, approval timeout, network degraded, stream failed, and artifact transfer failed.
  • 180.C.28.07 Implement degraded modes: command-only without stream, observe-only without control, screenshot-only low bandwidth, and local-only without cloud.
  • 180.C.28.08 Implement recovery flows for adapter restart, app crash, failed mutation rollback, interrupted file transfer, browser crash, and Unreal/Blender restart.
  • 180.C.28.09 Add chaos tests for process crash, gateway restart, network loss, permission revocation, host sleep, stream freeze, and audit store outage.
  • 180.C.28.10 Check off this stage only after operations can diagnose, replay, recover, and audit real failures without guessing.
  • 180.C.28.11 Implement outbound event sinks: webhooks with retry and HMAC signing, Slack approval cards, Discord notifications, and email digests for approvals, failures, and completions. Operators choose channels per event class. Full subsystem in 180.C.34.
  • 180.C.28.12 Implement ticketing integration (Linear, Jira) for session outcomes and failure tickets, and on-call integration (PagerDuty, Opsgenie) for operational alerts. Bidirectional where APIs allow, for example marking a ticket resolved on session success.
  • 180.C.28.13 Implement SLO error-budget automation: track the budgets defined in 180.C.16.12, block release when burn rate exceeds policy, and surface remaining budget in operational dashboards.

180.C.29 Final Verification, Documentation, SDKs, and Release#

<!-- STAGE 29/30 STATUS (2026-07-02): the final GA verification/evidence/ acceptance gates are now encoded in testing/bellona/remote-control/final-ga-release-gate.ts with focused coverage in tests/final-ga-release-gate.test.ts. The evaluator blocks by default and only passes when verification logs, release evidence artifacts, final product acceptance scenarios, conditional platform completion evidence or accepted scope-waiver evidence, and one-or-many host fleet closure evidence are attached as structured dated records with canonical UTC millisecond timestamps captured on or after the current final-GA evidence window, with command, full 40-character commit SHA, and environment context for automated release logs. Bare requirement ID summaries are compatibility metadata only; artifact IDs must name the final-GA record kind, requirement ID, and concrete evidence suffix; final-GA environment, reviewer, operator, and fleet-run identities must use contextual release prefixes with non-empty suffixes; and malformed or duplicated structured evidence records keep the gate blocked. Stage 26 macOS host packaging/acceptance evidence must identify both the non-developer operator rig and macOS host rig. Acceptance and fleet artifact IDs must name the required evidence slots they prove, and any context appended after the slot must be concrete and non-placeholder, so generic bundle artifacts or placeholder slot sub-suffixes cannot satisfy a valid evidence-kind list. Duplicate artifact or rig IDs and reused artifact IDs across final-GA evidence groups, including case variants, also invalidate every dependent record. Operator acceptance signoff must be captured after every final product acceptance scenario record, the final test-log release evidence must be captured after every verification record it summarizes, and fleet closure records must be captured after the Stage 29 verification, release evidence, and acceptance records they close over. Conditional Windows/Linux scope waivers require canonical macOS GA release notes, current contextual waiver approval, residual-risk ownership, target-platform scope text, and remote-control evidence links. Bare conditional platform stage IDs no longer satisfy final GA; Windows/Linux completion must be proven by structured conditional-platform records plus the matching Stage 36/37 roll-up gate evidence from @bellona/remote-host, unless the platform is explicitly waived. The remaining open items still require real macOS/Windows hosts, cloud deployment smoke, stream/diagnostic evidence, and operator acceptance signoff. -->

  • 180.C.29.01 Add or update all architecture docs for topology, trust boundaries, protocols, adapters, policy, gateway, host, Control Room, and cloud deployments.

  • 180.C.29.02 Add user/operator docs for two-MacBook LAN setup, Hetzner/AWS setup, pairing, permissions, Blender, Unreal, browser, Control Room, approvals, takeover, replay, and uninstall.

  • 180.C.29.03 Add developer docs for adding commands, adapters, MCP tools, policy rules, approval gates, audit evidence, and tests.

  • 180.C.29.04 Add TypeScript SDK coverage for device, session, command, stream, approval, artifact, audit, Blender, Unreal, browser, desktop, and file operations.

  • 180.C.29.05 Add Python client generation or SDK coverage for automation scripts and notebooks.

  • 180.C.29.06 Add bellona remote CLI commands for pair, devices, sessions, stream, command, approve, diagnose, artifacts, audit export, and unpair.

  • 180.C.29.07 Add end-to-end examples: Blender create/render/export, Unreal import/place/capture, browser production-board update, cloud-agent control, and desktop fallback dialog handling.

  • 180.C.29.08 Run final verification: format, lint, typecheck, unit, integration, contract, Playwright, security, performance, deployment smoke, and hardware-in-the-loop checks.

    _Partially blocked:_ Focused local verification continues to pass for the
    currently touched scope, re-verified 2026-07-03: remote-protocol Unreal
    contract spec 110 tests, remote-host focused Unreal/Windows specs 3 files
    / 102 tests, remote-host focused Windows contracts 1 file / 39 tests,
    remote-host full suite 68 files / 911 tests, remote-host lint with 0
    errors and 5 pre-existing warnings, remote-host TypeScript app compile on
    TypeScript 6.0.3 without CLI or inherited tsconfig deprecation overrides,
    remote-host focused Stage 26 macOS contracts 1 file / 30 tests,
    changed-file Prettier,
    stub-indicator, and diff
    whitespace checks. The complete 29.08 matrix still cannot be
    truthfully closed here because deployment smoke needs deployed cloud
    infrastructure and hardware-in-the-loop checks need real macOS/Windows
    hosts. The final-GA gate evaluator now codifies the full matrix and keeps
    this item blocked until every required verification check is attached as a
    single structured passing record with one of the planned commands, full
    40-character commit SHA, environment, canonical UTC millisecond timestamp,
    and unique artifacts; bare checklist IDs and duplicated structured
    records, placeholder environment/artifact suffixes, or artifact IDs reused
    across evidence groups do not satisfy the gate. The deployment-smoke
    harness also now exposes `releaseSignoffEligible=false` for LAN, Hetzner,
    and AWS fixture results until live external smoke evidence is attached.
    
  • 180.C.29.09 Produce release evidence: test logs, screenshots, stream captures where allowed, diagnostic bundles, security review, known risks, rollback plan, and compatibility matrix.

    _Blocked:_ A consolidated release-evidence bundle needs stream captures
    and diagnostic bundles from real sessions across the gated macOS/Windows
    hosts (depends on 29.08/29.10). The final-GA gate evaluator now
    enumerates the required logs, captures, diagnostics, security review,
    known risks, rollback, compatibility, supply-chain, and operator signoff
    artifacts as single structured dated records with canonical UTC
    millisecond timestamps and unique artifact identifiers; operator signoff
    must be captured after the final acceptance scenario records; bare,
    duplicated, or cross-group reused artifact IDs and placeholder
    reviewer/operator/artifact suffixes or duplicate structured records for
    the same requirement do not satisfy the gate.
    
  • 180.C.29.10 Run final product acceptance scenarios: same-LAN two-MacBook session, Hetzner/AWS cloud-agent session, Blender edit, browser workflow, desktop fallback, Unreal edit, artifact transfer, operator approval, takeover, replay export, revocation, and uninstall.

    _Blocked:_ Requires a same-LAN two-MacBook session and a Hetzner/AWS
    cloud-agent session running on real hardware/cloud; neither exists in this
    sandbox. The final-GA gate evaluator now enumerates every required
    acceptance scenario and keeps the roll-up gates blocked unless every
    scenario has a single structured passing record with operator, unique rig
    IDs, canonical UTC millisecond timestamp, unique artifact evidence, and
    required evidence-kind coverage where artifact IDs name the exact required
    evidence slots;
    duplicate structured records, placeholder operator/artifact suffixes,
    placeholder evidence-slot sub-suffixes, generic bundle artifact IDs, and
    artifact IDs reused across evidence groups do not satisfy the gate.
    
  • 180.C.29.11 Run final checklist completeness audit by comparing TODOS/phase-180.md, TODOS/phase-180-reference.md, implemented code, docs, tests, release evidence, and known gaps; add missing ordered tasks before release instead of relying on the reference file.

  • 180.C.29.12 Check off this stage only after the Remote Creative Control Plane is documented, SDK-accessible, operator-usable, tested, and releasable.

    _Blocked:_ Stage gate depends on 29.08/29.09/29.10 (full verification
    matrix, release evidence, product acceptance), which all need real
    hosts/cloud. Docs (29.01-03), SDKs (29.04-06), examples (29.07),
    quickstart (29.13), scaffolder (29.14), and evolution (29.15-16) are done.
    The final-GA gate evaluator now makes this dependency executable and
    rejects bare requirement ID summaries and duplicate structured records for
    the same requirement, plus artifact IDs reused across evidence groups.
    
  • 180.C.29.13 Produce a single top-level "5-minute from clone to round-trip" local dev quickstart combining gateway, host, mock adapter, MCP client, and Control Room so a new contributor sees the full spine without reassembling it from per-package docs.

  • 180.C.29.14 Build a tools/bellona/new-command scaffold generator that creates coordinated stubs across protocol, gateway, host, policy, audit, Control Room timeline, SDK, CLI, and docs for a new command namespace entry. Include a conformance test runner.

  • 180.C.29.15 Document and test protocol evolution: stored-data migrations (180.C.08.11), host/gateway version drift, deprecation policy, and forward-compatibility of stored audit across protocol upgrades.

  • 180.C.29.16 Produce cross-platform permission abstraction developer docs covering the interface defined in 180.C.26.16 and the Windows/Linux host implementations in 180.C.36 and 180.C.37.

180.C.30 Multi-Host Fleet and Advanced Production Expansion#

  • 180.C.30.01 Implement fleet scheduling by app version, project availability, resource load, network latency, permissions, current operator, and adapter capabilities.

  • 180.C.30.02 Implement host roles for Blender workstation, Unreal workstation, browser/admin workstation, render node, build node, QA viewer, and shared operator room.

  • 180.C.30.03 Implement multi-host session rooms where multiple agents operate separate devices under one production task with explicit locks.

  • 180.C.30.04 Implement distributed locks for shared project roots, source-control branches, Blender files, Unreal maps, browser profiles, and artifact output directories.

  • 180.C.30.05 Implement source-control-aware workflows for Git LFS, Perforce, Plastic/Unity Version Control, branch selection, checkout, changelist, revert, submit, and conflict reporting.

  • 180.C.30.06 Implement render/build queue integration so long-running jobs can continue after interactive sessions end.

  • 180.C.30.07 Implement fleet health dashboard, capacity planning, maintenance windows, adapter update cohorts, and fleet security posture scoring.

  • 180.C.30.08 Implement semantic scene diffing, visual regression, and style-guide checks across Blender and Unreal outputs.

  • 180.C.30.09 Implement mobile observer/approval mode and voice-driven Iris supervision after the desktop Control Room is stable.

  • 180.C.30.10 Check off this final stage only after the system supports one or many remote creative hosts with secure, observable, agent-driven, human-supervised production workflows end to end.

    _Blocked:_ Final fleet gate requires one-or-many real remote creative
    hosts running end-to-end production workflows; no real hosts are available
    in this sandbox. The fleet logic (scheduling/roles/rooms/locks/SCM/queue/
    health/scene-diff/red-team/mobile/voice) is implemented and tested in
    30.01-30.09, 30.11, and 30.12. The final-GA gate evaluator now requires
    explicit single fleet closure evidence records with fleet run, operator,
    rig, canonical UTC millisecond timestamp, and artifact context captured
    after the Stage 29 evidence before this item can pass; fleet artifact IDs
    must name the exact required evidence slots, and duplicate structured
    records for the same fleet
    requirement, placeholder evidence-slot sub-suffixes, generic bundle
    artifact IDs, and artifact IDs reused across evidence groups do not
    satisfy the gate.
    
  • 180.C.30.11 Implement voice and audio operator input (Iris supervision): push-to-talk approve/deny, spoken session commentary captured to audit, and voice-driven command palette. Expansion of the single bullet previously in 180.C.30.09.

  • 180.C.30.12 Implement an agent-in-the-loop red-team adversarial test suite: an agent actively attempting to bypass policy, escape the project-root sandbox, exfiltrate secrets, or manipulate approvals. Any successful bypass blocks release.

180.C.31 Cross-cutting UX Foundations#

This stage is staged alongside 180.C.12 and 180.C.21, not after them. Cross-cutting UX concerns must be codified before production UI work so they are not retrofitted.

  • 180.C.31.01 Extract all Control Room UI strings through an i18n resource system from the MVP scaffold, not after the fact.
  • 180.C.31.02 Adopt a WCAG 2.1 AA accessibility baseline tested at every UI milestone (180.C.12, 180.C.21), not retrofitted in 180.C.21.09.
  • 180.C.31.03 Implement reduced-motion, high-contrast, and light/ dark themes with automated visual regression.
  • 180.C.31.04 Stub RTL-language layout support from MVP and test at each milestone.
  • 180.C.31.05 Define a keyboard navigation baseline (tab order, focus rings, skip links, focus-trap rules for modals) applied at Control Room MVP.
  • 180.C.31.06 Define a notification-channel abstraction (in-app toast, desktop notification, mobile push, email, webhook) wired from MVP so 180.C.12.13 and 180.C.28.11 share one pipeline.
  • 180.C.31.07 Define a cost-display abstraction (tokens, compute, bandwidth, render minutes, build minutes) wired into every UI surface showing session state so 180.C.21.11 has a single source of truth.
  • 180.C.31.08 Produce an empty, loading, and error state library for consistent operator UX across all Control Room views.
  • 180.C.31.09 Document the icon, color, typography, and spacing system before production UI work in 180.C.21 begins.
  • 180.C.31.10 Check off this stage only after cross-cutting UX concerns are codified such that 180.C.21 does not need to retrofit them.

180.C.32 Adapter SDK and Conformance Suite#

Runs after Blender (180.C.10) and Unreal (180.C.15) MVPs have validated the adapter interface by example.

  • 180.C.32.01 Publish an Adapter SDK with a stable contract in libs/bellona/remote-adapters covering capability handshake, command dispatch, progress/result emission, audit hooks, policy preflight, rollback metadata, and health reporting.
  • 180.C.32.02 Publish an adapter conformance test suite that any adapter must pass: lifecycle, error codes, idempotency, cancellation, timeout, policy-deny propagation, evidence registration, and rollback-strategy declaration.
  • 180.C.32.03 Implement adapter hot-reload for dev: file-watcher based reload without host restart, isolated reload failure handling, and state preservation where safe.
  • 180.C.32.04 Define an adapter packaging format: signed bundle, version metadata, declared permission scope, declared command namespace, and runtime integrity check.
  • 180.C.32.05 Implement adapter telemetry hooks so third-party adapters report into the same dashboards as first-party adapters without extra plumbing.
  • 180.C.32.06 Implement an adapter policy sandbox: a third-party adapter cannot bypass device policy, approval gates, or audit even if it misbehaves.
  • 180.C.32.07 Ship a trivially-simple example adapter (a timer or text-file adapter) that demonstrates the full contract end to end.
  • 180.C.32.08 Document the Adapter SDK with versioning policy, deprecation policy, and migration guides.
  • 180.C.32.09 Define an adapter signing and distribution model (signed registry or documented ad-hoc distribution with trust implications).
  • 180.C.32.10 Check off this stage only after a third-party developer could ship a conforming adapter without modifying Bellona core.

180.C.33 Timecode, Color, and Media Pipeline#

  • 180.C.33.01 Implement frame/timecode sync primitives (SMPTE, TC-relative, offset) shared across Blender, Unreal, Sequencer, and DAW commands.
  • 180.C.33.02 Attach OCIO/ACES color-management metadata to every rendered or captured artifact and preserve it through transfer and storage.
  • 180.C.33.03 Implement audio capture alongside screen capture as a separately routed stream track with independent start/stop and redaction hooks that mute during credential entry.
  • 180.C.33.04 Implement HDR capture, transfer, storage, and Control Room preview including PQ and HLG tone-mapping for SDR fallback.
  • 180.C.33.05 Implement DAW transport control stubs (Logic, Pro Tools, Reason, Ableton) via OSC, ReWire, or AU hosts as read-only first, then gated write.
  • 180.C.33.06 Implement video editor project queries (DaVinci Resolve, Premiere) as read-only first with timeline/clip/bin access.
  • 180.C.33.07 Extend Live Link and virtual-production surfaces from read-only (180.C.24.08) to guarded write commands.
  • 180.C.33.08 Add color-pipeline verification tests: round-trip a scene through OCIO transforms and confirm no drift beyond documented tolerance.
  • 180.C.33.09 Implement audio-stream redaction hooks so credential entry, muting policies, and redacted windows also silence the audio track.
  • 180.C.33.10 Check off this stage only after media pipeline concerns are first-class across DCCs rather than improvised per command.

180.C.34 Notifications and External Event Sinks#

  • 180.C.34.01 Implement an outbound webhook subsystem with retry, dead-letter queue, HMAC signing, idempotent delivery, and per- subscription filters.
  • 180.C.34.02 Implement Slack integration with approval cards, session links, failure digests, and per-channel routing.
  • 180.C.34.03 Implement SMTP email integration for digests, critical alerts, and approval requests with operator preferences per event class.
  • 180.C.34.04 Implement mobile push notifications via APNs and FCM for approval and session events, coordinated with the mobile PWA from 180.C.12.14.
  • 180.C.34.05 Implement PagerDuty/Opsgenie on-call integration for host-offline, gateway-failure, and policy-bypass attempts.
  • 180.C.34.06 Implement ticketing integration (Linear, Jira) for failed-session ticket creation and session-outcome cross-references. Pairs with 180.C.22.14.
  • 180.C.34.07 Implement calendar integration (.ics export and Google/Outlook APIs) for scheduled render and build jobs.
  • 180.C.34.08 Implement per-user notification preferences, mute windows, escalation policies, and test-delivery helpers.
  • 180.C.34.09 Build a notification delivery test harness that verifies every channel end to end on each release.
  • 180.C.34.10 Check off this stage only after operators can choose reachability channels and configuration changes flow reliably without a restart.

180.C.35 Tier-2 DCC Adapter Onboarding#

Runs after the Adapter SDK (180.C.32). Each Tier-2 DCC participates in the same audit, approval, evidence, and rollback pipeline as Blender and Unreal.

  • 180.C.35.01 Maya adapter: read-only scene/project info, safe mutation (createNode, setAttr), playblast, arbitrary MEL/Python behind approval, reference/referenceload tracking.
  • 180.C.35.02 Houdini adapter: HIP read, node graph query, parameter set, cook/render, arbitrary Python behind approval.
  • 180.C.35.03 Cinema 4D adapter: read-only scene, safe object creation, render, Python Tag behind approval.
  • 180.C.35.04 Substance Painter/Designer adapter: project read, bake/export, JavaScript scripting behind approval.
  • 180.C.35.05 Nuke adapter: script read, node/parameter set, render, TCL/Python behind approval.
  • 180.C.35.06 DaVinci Resolve adapter: project timeline read, clip operations, render queue, LUA behind approval.
  • 180.C.35.07 Figma Desktop adapter: file read via plugin, comment/selection, export.
  • 180.C.35.08 Photoshop and After Effects adapters via UXP or ExtendScript: read-only MVPs before mutation.
  • 180.C.35.09 Run the adapter conformance suite (180.C.32.02) against every Tier-2 DCC adapter.
  • 180.C.35.10 Check off this stage only after Tier-2 DCCs participate in the same audit/approval/evidence pipeline as Blender and Unreal, end to end.

180.C.36 Windows Host Support#

Conditional on the scope decision in 180.C.02.14.

<!-- STAGE 36 STATUS (2026-07-02): the locally testable Windows contracts now exist in apps/bellona/remote-host/src/windows-host-support.ts with coverage in windows-host-support.test.ts: signed-MSI/Windows-Service packaging plan, DXGI + UIAutomation + SendInput desktop strategy, Windows permission projection through the shared host-runtime provider, isolated Chrome/Edge CDP launch policy, Blender/Unreal/Chrome/Edge discovery candidates, signed update-channel plan, Windows diagnostic bundle through the shared cross-platform bundle builder, render-node/browser-admin fleet roles, macOS<->Windows command parity comparison, Authenticode signing-readiness validation, and a Stage 36 roll-up evidence gate that rejects missing/placeholder Windows parity and operator acceptance records, generic parity summaries without a passing structured comparator result for browser, desktop, and Unreal command anchors, reused managed browser profile IDs that escape the Windows managed-profile root, managed browser CDP ports outside the valid TCP range, managed-browser launch plans that omit an explicit loopback CDP bind or carry launch-time initialUrl navigation instead of using browser.navigate policy checks, macOS/Windows parity runner identities, and artifact identifiers duplicated within or across stage evidence records, including case-variant duplicates, stale or non-canonical timestamp stage evidence, stage artifacts that do not identify the checklist item, exact evidence slot, and a concrete non-placeholder run/artifact suffix, stage records from the wrong Windows runner class or with placeholder/prefix-only runner suffixes, macOS parity records from non-macOS host runners or placeholder runner suffixes, and final acceptance without a concrete Windows acceptance operator identity or captured before every prerequisite Stage 36 evidence record exists. The gate also rejects signed-update rollback evidence captured before signed-MSI/service lifecycle evidence and macOS/Windows parity evidence captured before the Windows capability records it compares. The checklist items remain open because final completion still needs a real Windows host/build pipeline: actual MSI build and Windows Service install, live DXGI/UIAutomation/SendInput probes, UAC/Defender/Smart App Control verification, real Chrome/Edge/Blender/Unreal discovery, Authenticode signing, diagnostic bundle collection on Windows, cross-platform parity runs against a macOS host, and hardware evidence. -->

  • 180.C.36.01 Package the Windows Remote Host as a signed MSI running under Windows Services with graceful update and rollback.

    _Blocked:_ Requires a real Windows build/signing pipeline, Authenticode
    signing material, MSI build output, and live Windows Service install,
    update, and rollback evidence. Local contract coverage exists, but it
    cannot substitute for signed Windows host artifacts.
    
  • 180.C.36.02 Implement Windows screen capture (DXGI Desktop Duplication), window focus (UIAutomation), and input injection (SendInput) with parity to macOS primitives.

    _Blocked:_ Requires live Windows desktop capture, UIAutomation focus, and
    SendInput probes on real Windows hardware plus parity evidence against a
    macOS host. Linux sandbox tests can only validate the local contract.
    
  • 180.C.36.03 Implement Windows permission handling: UAC elevation prompts, Windows Defender Controlled Folder Access allowlisting, and Smart App Control compatibility.

    _Blocked:_ Requires real UAC, Defender Controlled Folder Access, and Smart
    App Control verification on Windows. The remaining evidence must be
    captured from Windows security surfaces, not fixture-only tests.
    
  • 180.C.36.04 Implement Windows Chrome/Chromium/Edge launch with CDP, managed-profile creation, and policy-isolated browser paths.

    _Blocked:_ Requires live Chrome/Chromium/Edge launch and managed-profile
    validation on Windows with policy-isolated paths and CDP connectivity.
    
  • 180.C.36.05 Implement Windows Blender and Unreal install discovery, version detection, and plugin installation paths.

    _Blocked:_ Requires real Windows Blender and Unreal installations,
    version-discovery output, and plugin installation evidence from the
    Windows filesystem/registry paths used by artists.
    
  • 180.C.36.06 Implement a Windows update channel with signed installer manifests, canary rollout, and rollback.

    _Blocked:_ Requires signed Windows installer/update manifests plus live
    canary rollout and rollback evidence from a Windows host or packaging
    pipeline.
    
  • 180.C.36.07 Implement a Windows diagnostic bundle with the same content and redaction policy as the macOS bundle.

    _Blocked:_ Requires a diagnostic bundle captured on Windows with real
    host, permission, adapter, browser, DCC, and redaction evidence comparable
    to the macOS bundle.
    
  • 180.C.36.08 Validate Windows as a render-node and browser/admin-workstation host role from 180.C.30.02.

    _Blocked:_ Requires Windows render-node and browser/admin-workstation
    role validation against real workloads and fleet role evidence.
    
  • 180.C.36.09 Add cross-platform tests verifying macOS and Windows agents see identical capabilities for shared command namespaces.

    _Blocked:_ Requires paired macOS and Windows host evidence from the same
    capability namespace matrix; local contract tests cannot prove live
    cross-platform parity.
    
  • 180.C.36.10 Check off this stage only after Windows hosts reach feature parity for Blender, Unreal, browser, and desktop MVPs.

    _Blocked:_ Stage roll-up depends on 36.01-.09 plus final operator
    acceptance evidence from a real Windows host. It remains open until signed
    packaging, live capability probes, diagnostics, parity, and acceptance are
    complete.
    

180.C.37 Linux Host Support#

Conditional on the scope decision in 180.C.02.14. Render-node focus. Local Stage 37 closure is now roll-up gated by evaluateLinuxHostStageGate, which requires concrete per-item Linux evidence plus build-node validation before 180.C.37.10 first-class-node acceptance can pass.

  • 180.C.37.01 Package the Linux Remote Host as deb, rpm, and AppImage with systemd user units and distribution-signed packages.
  • 180.C.37.02 Implement Wayland and X11 screen-capture and input- injection strategies with an explicit compositor compatibility matrix.
  • 180.C.37.03 Implement PipeWire audio capture alongside the media pipeline in 180.C.33.03.
  • 180.C.37.04 Implement Linux permission handling: PolicyKit, cgroups, AppArmor/SELinux profiles for the host and adapters.
  • 180.C.37.05 Implement headless Blender (bpy), Unreal commandlet, and Chromium-headless adapters for render-node use.
  • 180.C.37.06 Implement the Linux render-node host role without interactive DCC UI requirements, coordinated with 180.C.30.02.
  • 180.C.37.07 Implement a Linux diagnostic bundle with the same content and redaction policy as macOS and Windows bundles.
  • 180.C.37.08 Implement Linux package signing and repository distribution (signed deb repo, signed rpm repo, or signed AppImage updater).
  • 180.C.37.09 Add cross-platform tests verifying Linux render nodes participate fully in fleets from 180.C.30.
  • 180.C.37.10 Check off this stage only after Linux hosts function as first-class render, build, and headless-adapter nodes.

180.C.38 Bridge Coexistence and Deprecation#

Executes the plan produced in 180.C.16.15.

  • 180.C.38.01 Implement a compatibility shim so existing consumers of apps/bellona/bridge-blender and bridge-unreal continue to function during rollout without code changes.

  • 180.C.38.02 Document a parallel-operation guide: old bridges and new Remote Host running simultaneously with conflict detection and reconciliation rules.

  • 180.C.38.03 Produce a migration guide for consumers moving from old bridges to the new command envelopes with before/after examples.

  • 180.C.38.04 Publish a deprecation timeline with dates for warning emission, hard cutover, and final removal of old bridges.

  • 180.C.38.05 Implement a feature flag gating the gateway's acceptance of old-bridge protocols for forced-migration drills.

  • 180.C.38.06 Instrument telemetry on old-bridge usage to drive deprecation communication and identify holdouts.

  • 180.C.38.07 Submit removal PRs for old bridges with migration paths pinned in release notes.

    _Blocked:_ Requires Remote Host GA, zero-consumer telemetry, and actual
    submitted removal PRs with migration paths in release notes. Old bridges
    are still in coexistence mode.
    
  • 180.C.38.08 Run a post-removal verification that no Oshun service, test, or script depends on the old bridges.

    _Blocked:_ Requires the old bridge removal to have happened first, then a
    clean post-removal dependency scan across services, tests, scripts, docs,
    and workspace metadata.
    
  • 180.C.38.09 Clean up apps/bellona/bridge-blender, apps/bellona/bridge-unreal, and related packages once Remote Host is GA and no consumers remain.

    _Blocked:_ Requires Remote Host GA, zero-consumer telemetry, merged
    removal PRs, and package cleanup evidence for the legacy bridge apps.
    
  • 180.C.38.10 Check off this stage only after old bridges are fully retired with zero production consumers and documentation reflects the single supported path.

    _Blocked:_ 38.07-38.10 are sequenced removal tasks gated on Remote Host GA
    plus zero consumers. The old bridges are still in active coexistence
    (38.01-06 shim/feature-flag/telemetry) and are still consumed. Verified
    2026-06-22: apps/oshun/web has live studio routes
    (studio/bellona/bridge-blender and bridge-unreal pages), workspaces, route
    maps, navigation, and tests (100+ files reference the bridges). Therefore
    38.07 is premature, 38.08 depends on removal already having happened,
    38.09 is explicitly gated on Remote Host GA and no consumers, and 38.10 is
    the zero-production-consumer stage gate. A local
    `bridge-retirement-gate` plan/evaluator now fail-closes the PR,
    telemetry, cleanup, app/lib/service/script/test/workspace post-removal
    scan, docs, approval sequencing, merged-PR, and post-cutover
    zero-consumer-window evidence contract. Duplicate PR, bridge telemetry, or
    command telemetry records, unknown or nonzero command telemetry entries,
    reused per-bridge PR URLs or telemetry sources, non-canonical PR,
    migration, release-note, cleanup, telemetry, or approver identities, scan
    commands that do not use `rg`/`git grep`, placeholder telemetry source or
    bridge-retirement approver suffixes, scans that omit or only partially
    cover workspace metadata files, scans that only mention required paths as
    substrings instead of scanning the exact legacy bridge package/path
    references, stale or non-canonical timestamp removal PR submissions,
    premature or non-canonical timestamp PR merges, stale or non-canonical
    timestamp cleanup records, and final approvals that predate the
    retirement evidence or use non-canonical timestamp metadata do not satisfy the single authoritative evidence
    record expected for each bridge or compatibility command. Package
    cleanup removed-path and removed-package lists must be exact sets with no
    duplicate or unrelated removals, and allowed-reference paths must be
    unique including case variants, docs/release-note scoped residual
    references that cannot whitelist active app/lib/service/script/test/
    workspace dependencies directly or through path traversal, but the real
    GA/zero-consumer/removal evidence remains required.
    

Reference#

Detailed architecture notes, expanded requirements, and future backlog items were moved to TODOS/phase-180-reference.md. Do not execute that file as a checklist; use it only as supporting context while completing the 180.C items above from top to bottom.