Status: implementation and bounded native campaigns complete on Mac ARM64, Windows x64 and Linux x64; combined-source release qualification remains open. Owning topics: runtime and presentation, host control, and native distribution.
Make control independent of the settings UI, then offer personal, minimal and appliance presentation in one existing package/release stream. Preserve the same resident authority, providers and local/remote CLI contract. Operators can always reopen Settings and reset presentation through supported controls. Presentation changes never change trusted deployment policy.
First independently deliverable milestone: existing Mac personal presentation stays responsive during delayed, missing and interrupted helper status/session observations, with no unbounded poll queue and no late-result authority revival.
No new installer family, public profile release, provider rewrite, privilege expansion or protected-session redesign. Keep the current default and signed identity. Do not replace the controller user's installed app or alter a target's policy as part of planning. Dedicated target trials use doctor, exact target-use claims and cleanup; fresh provisioning also uses the private journal and credential handoff. Preserve existing appliance implementations during migration.
Create bounded child tacticals when implementation starts on later milestones; this document owns sequencing, not duplicate execution logs for every platform.
Implement milestones 1–4 and their native acceptance on Mac first. Run cheap shared frontend/contract checks and review other platform interfaces throughout; defer Windows/Linux integration campaigns to milestone 5. Prefer an available native Windows development host or CI runner for candidate builds and claimed targets for deployment/effects. Linux placement may use native/CI or VM facilities after checking availability. Do not assume a private host is available or force every platform build onto the Mac controller's VMs.
Inventory blocking work and thread-affinity requirements in ResidentServer,
ResidentService, locked-use permission/status, the Swift operator bridge and
Tauri command scheduling. Distinguish display-only reads from authoritative
pre-effect checks and lifecycle operations before changing call sites.
Introduce bounded asynchronous helper/session observation with injectable dependencies for deterministic tests. Use immutable display projections and coalesced refreshes for tray and settings. Separate physical-activity monitoring from slow helper discovery so a status delay cannot stop its publication. Treat unknown helper/session readiness conservatively without reusing a last good observation as new authority or clearing a pause based on missing data.
Replace blocking receives inside async Tauri commands with an awaited completion. Bound queued work as well as active work; caller timeout alone does not cancel a queued action. Use operation-specific deadlines, distinguish undispatched cancellation from unknown delivered effects, and reject superseded completion. Fence Stop locally before asynchronous cleanup. Keep independent watchdogs.
Acceptance: injected delayed/nonreplying status dependencies leave the main event loop, operator status and Stop responsive. Repeated polls stay bounded. Grant/session changes invalidate completions and prevent effects. Existing approval, physical takeover, pause, protected watchdog and update exclusion checks continue to pass. Investigate helper latency independently if still present; hiding it from the UI is not proof that the helper is healthy.
Centralize construction/lifecycle now duplicated by the AppKit resident and Tauri bridge. Define typed runtime state/events and operator intents. Retain one serialized authorization/session owner; isolate UI-affine methods and blocking providers individually. Keep the authenticated operator channel distinct from public agent operations.
Acceptance: existing native and Tauri shells use that boundary, client operations retain their contract, and starting a second presentation attaches to the same runtime. Verify outside-agent access still needs no in-target agent. Review the existing Windows/Linux service/session/operator interfaces when shaping the boundary; defer their native integration proof to milestone 5. Do not force identical process placement across platforms.
Create Settings on explicit demand instead of retaining a hidden polling WebView. Native notice/approval/Stop presentation remains usable independently of it. Closing/reopening Settings leaves runtime owners, policy and lifecycle unchanged. Minimal mode does not instantiate the full settings frontend until requested; hiding an already running window does not satisfy this milestone.
Acceptance: no settings WebView or its polling when dormant; required notices and approvals, Stop, open/close/reopen, single-instance forwarding and browser control work. Verify permission and updater identity/lifecycle on a signed candidate before accepting deployment.
Implement the versioned presentation setting, explicit personal default, Settings selector, supported local provisioning/reset commands and temporary open-Settings launch path. Reuse one atomic configuration implementation and define operator authority. Preserve active tasks or require an explicit bounded transition; never silently restart on a configuration edit.
Acceptance: all three profiles can open Settings and return to personal presentation without manual file removal. Invalid/missing configuration, reboot, second launch and interrupted writes have deterministic behavior. Changing presentation never changes grants, policy, protected installation or credential custody. Approval-required configurations cannot become unattended by selecting appliance presentation.
Adapt platform presentation adapters to the same semantics; reuse Windows and Linux residents and their platform permission requirements. Prove guest-local CLI and outside-agent parity on a claimed Windows appliance, with its VM console absent. Qualify Mac signed identity/TCC and Linux graphical-session and portal requirements separately. Preserve supported capability omissions.
Acceptance: each advertised profile passes startup, approval/policy, Pause/Stop, settings recovery, CLI/native effect, owner loss, restart and update checks for its claimed platform scope. Publish using the existing package stream only when explicitly authorized and qualified; defaults remain personal meanwhile.
Use exact-source Windows native contracts, format verification and x64/ARM64 publishes on the selected native build host/CI. Reuse Linux native checks and its GNOME/portal acceptance targets. Compilation alone is not installed behavior evidence, and successful Mac acceptance does not cover either OS.
Current: Mac milestones 1–4 are implemented and pass the bounded signed ARM64 campaign. Tactical 108 owns exact source, deterministic checks, both Mac bundle architectures, VM effects, browser-without-Settings, standing-policy preservation, host observations and restoration. It also records fixes found during live acceptance and remaining protected/physical limits. The host installation and public release are unchanged.
Tactical 109 adds bounded Mac hardening before the Windows handoff: asynchronous Resume, preflight deadlines and fencing, reply/handshake bounds, audit-free delegation discovery and signed host load. Protected lifecycle/update paths and broader physical qualification remain open.
Windows integration and its bounded development-candidate campaign are recorded in tactical 110. Linux integration and its bounded x64 GNOME Wayland campaign are recorded in tactical 111. Milestone 5 remains active for combined-source package, architecture and installed qualification, rather than further ordinary-profile integration. The acceptance review owns the cross-platform comparison and next campaign order. Each platform's earlier candidate retains its exact source and scope; it does not qualify later shared changes or a new public package.