Skip to content

Agent Runtime

Describe the Borealis agent runtime, its roles, service modes, and how it communicates with the Engine.

Runtime Summary

  • Main entry: Data/Agent/cmd/agent builds one Go runtime binary named Agent.exe on Windows and Agent on Linux.
  • Service modes: SYSTEM/root plus same-binary helper mode. Windows CURRENTUSER tracks long-lived helper sentinels per active desktop session and executes signed quick jobs through SYSTEM-brokered CreateProcessAsUser. Linux CURRENTUSER reports unsupported until ported.
  • Role system: compiled Go role registry under Data/Agent/internal/roles.
  • Networking: SYSTEM/root runtime owns REST to Engine APIs plus the single Socket.IO connection.
  • Security: Ed25519 identity keys, public CA or Borealis local CA + hostname validation for the Engine FQDN, signed script payloads, and agent.json token/key storage.

Role Catalog (Go v1)

  • internal/roles/system_context - SYSTEM/root quick-job router and script execution for signed quick_job_run payloads.
  • internal/roles/current_user - Windows CURRENTUSER helper sentinel broker, active-session health, direct signed quick-job execution for active user sessions, and Windows tray/status UI. Linux CURRENTUSER reports unsupported in first PR.
  • internal/roles/device_audit - core CPU, memory, storage media type, removable media, network link speed, OS/build, hardware model/serial with motherboard serial fallback, last reboot, internal IP, device type, uptime, and last-user inventory published through heartbeat payloads.
  • internal/roles/file_management - SYSTEM/root file-management browse, upload-conflict preflight, lightweight text editing, copy/cut/paste mutations, delete, mkdir, rename, move, upload pull, and download artifact transfer.
  • internal/roles/patch_management - SYSTEM Windows pending Windows Update Agent inventory, installed KB history, Get-HotFix inventory, scheduler-driven WUA install requests, policy enforcement requests, reboot requests, refresh requests, and unsupported status on non-Windows agents.
  • internal/roles/process_management - SYSTEM/root live process snapshots, parent/child metadata, cache reuse, and operator-triggered process termination for the Device Summary Processes tab.
  • internal/roles/registry_management - SYSTEM Windows registry browse, key create/rename/delete, value create/update/delete, and unsupported status on non-Windows agents.
  • internal/roles/service_management - SYSTEM/root service inventory publishing plus operator-triggered start, stop, and restart through service_control_action.
  • internal/roles/software_management - SYSTEM/root Windows installed-app inventory with cached icon payloads, Linux dpkg/rpm inventory, refresh requests, and post-uninstall inventory refresh through the SYSTEM quick-job lane.
  • internal/roles/wireguard_tunnel - SYSTEM/root persistent WireGuard reverse tunnel lifecycle, Engine /api/agent/vpn/ensure polling, vpn_tunnel_start handling, Windows tunnel-service apply, Linux wg-quick apply, and /api/agent/vpn/ready reporting.
  • internal/roles/remote_shell - SYSTEM/root WireGuard-scoped TCP shell listener for Engine vpn_shell_* bridge traffic, using PowerShell on Windows and Bash/sh on Linux.
  • internal/roles/vnc - Windows UltraVNC always-on lifecycle, runtime credential broker, Engine /api/agent/vnc/ensure bootstrap, Socket.IO credential/start events, firewall scope, and listener readiness reporting. Linux VNC reports unsupported.
  • Pending ports are tracked in Data/Agent/Golang_Agent_Migration.md; Linux current-user/tray UI remains pending.

Agent Settings and Storage

  • Installed configuration file: agent.json beside Agent.exe.
  • WireGuard runtime configuration file: wireguard.conf beside Agent.exe/Agent, generated from Engine tunnel material.
  • Startup cleanup removes Temp under the Agent install root so onboarding payload/state files do not persist after service start.
  • agent.json stores schema_version, server_url, optional server_ip_fallback, enrollment_code, agent.guid, agent.agent_id, agent.branch, agent.installed_build_id, agent.log_retention_days, agent.state, agent.liveness, agent.dependency_state, Ed25519 keys, access/refresh tokens, optional Internal-Only Engine CA PEM, and Engine script-signing trust material.
  • metadata-queue.json sits beside agent.json only while local CLI metadata updates are pending. It stores queued field updates as base64 value, modified_at, and source; Engine acknowledgement removes delivered entries.
  • Windows protection: ACL hardening is deferred in the current Go migration branch; files inherit permissions from C:\Borealis.
  • Linux protection: root-owned 0600 file with 0700 parent directory.
  • Writes are atomic temp-write + rename and serialized across processes through a sibling agent.json.lock file. Each write increments agent.state.revision and stamps agent.state.writer plus agent.state.last_write_at so stale writer/debug cases are visible.
  • Successful heartbeats clean stale atomic-write temp files such as .config-*.tmp from the Agent install root after they are older than one hour.

Agent CLI Flags

Use Agent CLI Flags for Windows Agent.exe and Linux Agent command-line arguments, including install and re-deploy inputs, metadata field get/set, update checks, service repair/removal, watchdog checks, validation, and internal helper modes.

Detailed Codex Breakdown

API endpoints (Engine-facing)

  • POST /api/agent/enroll/request (No Authentication) - start enrollment.
  • POST /api/agent/enroll/poll (No Authentication) - finalize enrollment after approval.
  • POST /api/agent/token/refresh (Refresh Token) - mint a new access token.
  • POST /api/agent/heartbeat (Device Authenticated) - heartbeat, metrics, Agent Metadata Field sync, and release-channel correction instructions when Engine detects a retired branch.
  • GET /api/agent/metadata/<field_number> (Device Authenticated) - read one decoded Engine metadata field for local Agent CLI.
  • POST /api/agent/status (Device Authenticated) - startup phase, boot ID, milestone timeline, and last-error telemetry for system:system_heartbeat.
  • POST /api/agent/details (Device Authenticated) - hardware, inventory, and cached service payloads.
  • POST /api/agent/script/request (Device Authenticated) - request work or receive idle signal.
  • POST /api/agent/vpn/ensure (Device Authenticated) - persistent WireGuard tunnel bootstrap.
  • POST /api/agent/vpn/ready (Device Authenticated) - active WireGuard tunnel readiness after service/config/firewall apply.
  • POST /api/agent/vnc/ensure (Device Authenticated) - advertise VNC readiness and reconcile always-on VNC state without returning the VNC password.
  • GET /api/agent/files/transfers/<transfer_id>/upload-item/<item_id> (Device Authenticated) - fetch one Engine-staged upload item for the File Management role.
  • GET /api/agent/files/transfers/<transfer_id>/status (Device Authenticated) - fetch one File Management transfer control snapshot so the agent can honor cancel requests mid-transfer.
  • POST /api/agent/files/transfers/<transfer_id>/progress (Device Authenticated) - update Engine-side File Management transfer progress.
  • POST /api/agent/files/transfers/<transfer_id>/content (Device Authenticated) - upload a completed File Management download artifact back to the Engine.
  • POST /api/agent/patches/install-progress (Device Authenticated) - update Engine-side scheduled patch install progress for the current device/run.

Source vs runtime

  • Edit only in Data/Agent/.
  • Windows installed runtime is C:\Borealis\Agent.exe plus C:\Borealis\agent.json, managed by the native BorealisAgent Windows service. Fresh bootstrap from an operator-downloaded Agent.exe stages the runtime copy. Re-deploy with install inputs stops Borealis-managed services, scheduled tasks, and processes, preserves existing agent.json identity/trust state, replaces Agent.exe, and restarts the Agent service. Providing a new enrollment code clears cached Agent auth tokens and remote-ops route state so cross-Engine or cross-site re-enrollment cannot keep using stale socket routing. If the old binary is still locked, bootstrap leaves a verified .redeploy binary and starts a deferred replacement loop that stops the old process, moves the new binary into place, and restarts the service.

    • Linux installed runtime is a single compiled Agent binary managed by systemd. Fresh execution from an operator-downloaded binary with both --server-url and --site-enrollment-code self-stages into /opt/Borealis/Agent/Agent, writes or updates agent.json, and installs the service plus updater and watchdog timers. Re-deploy with install inputs stops Borealis-managed systemd units and the WireGuard interface, preserves existing agent.json identity/trust state, replaces the runtime binary, and restarts the Agent service.
    • Go Agent updates use Engine release channels and the local --update-check path.

    Service modes and context

    • Service mode is used for elevated tasks (Windows service runtime, support scheduled tasks, VPN, system scripts).
    • The SYSTEM runtime is the only Borealis process that authenticates to the Engine, enrolls, refreshes tokens, or opens a Socket.IO connection.
    • Interactive user quick jobs now run through the SYSTEM broker by launching signed PowerShell/Batch payloads into active Windows user sessions. Same-binary helper sentinels run in active desktop sessions so role health can report ready helper sessions without exposing Engine tokens to user context.
    • Direct Session 0 UI is not supported; Borealis keeps the Engine-facing socket in SYSTEM and bridges into desktop sessions when current-user interaction is required.
    • Windows same-binary helpers host the Borealis tray icon. The right-click menu shows each role as Role Name: Healthy or Role Name: Unhealthy, plus Restart Agent and Check For Updates; it does not open a local webpage or expose diagnostics/Engine links. Remote access remains silent in this first UI version; the UI does not add remote-session banners, notifications, or consent prompts.
    • The agent still labels Engine traffic with X-Borealis-Agent-Context, but the supported Windows service path no longer relies on a standalone CURRENTUSER Engine identity.
    • Headless Linux agents without an active graphical desktop report desktop-only health surfaces as No Desktop Environment Active instead of unhealthy/recovering. This applies to Current User Context helper dispatch and UI-side UltraVNC presentation so server-class Linux hosts do not look broken for missing desktop roles.

    Role discovery and extension

    • Go roles are compiled under Data/Agent/internal/roles.
    • Add new role packages to the explicit registry in cmd/agent/runtime wiring instead of relying on dynamic Python module discovery.
    • Add regression coverage under package-local Go *_test.go files and run the lane through Data/Agent/Unit_Tests/Agent_Unit_Tests.sh or Data/Agent/Unit_Tests/Agent_Unit_Tests.ps1.

    Networking and authentication

    • All REST calls flow through the Go auth client in Data/Agent/internal/auth.
    • Internal-Only agents can persist server_ip_fallback as a bare IP route hint. Windows and Linux REST, update, file-transfer, software-override, and Socket.IO connections first try the Engine FQDN normally; if that TCP dial fails, the Agent dials the fallback IP while keeping the FQDN as HTTP host, TLS SNI, and certificate hostname. Linux WireGuard setup first applies the Engine FQDN endpoint, then rewrites only the local WireGuard endpoint to server_ip_fallback:<port> when wg-quick up fails because DNS cannot resolve the Engine FQDN.
    • EnsureAuthenticated handles identity generation, enrollment, approval polling, and token refresh.
    • During disaster-recovery restore, an already-enrolled Agent may temporarily reach a blank Engine before the backup import restores device trust. invalid_refresh_token and device_not_found refresh responses are treated as recoverable in that window, so the Agent preserves its existing refresh token and retries. Explicit revoke, purge, token expiry, fingerprint mismatch, and token-version mismatch still clear local token state.
    • Socket.IO is used by the SYSTEM runtime for:
    • quick_job_run dispatch (system jobs plus broker-backed current-user jobs).
    • file_management_request browse, upload-conflict preflight, lightweight text-edit, copy/cut/paste mutate, and transfer orchestration for the Device Summary File Management tab.
    • process_management_request live process snapshots and process termination for the Device Summary Processes tab.
    • registry_management_request Windows registry browse and mutation for the Device Summary Registry tab.
    • service_control_action start, stop, and restart requests for services discovered by the Service Management role.
    • patch_inventory_refresh_request operator-triggered Windows patch inventory refresh.
    • patch_install_request scheduled Windows Update Agent install for one matched pending update.
    • patch_policy_enforcement_request reversible Windows Update policy enforcement for managed, frozen, or unmanaged Patch Management modes.
    • patch_reboot_request scheduled Windows reboot after policy-approved patching, with logged-in-user skip by default.
    • software_inventory_refresh_request operator-triggered software inventory refresh after icon/override or software action changes.
    • vpn_tunnel_start (WireGuard lifecycle; tunnels are persistent and ignore stop events).
    • vnc_start, vnc_stop, vnc_refresh, and vnc_credential_request for Windows UltraVNC lifecycle and runtime password delivery.
    • agent_update_request to start the local platform updater path.
    • connect_agent registration (agent socket registry).
    • The SYSTEM socket advertises helper_contexts=["currentuser"] when the session broker is running so the Engine can route logical current-user work through the same socket.
    • The Socket.IO client waits for the site-worker connect_agent acknowledgement before recording the management socket as connected. Rejected or missing registration acknowledgements close the socket and force route refresh/reconnect.
    • The Socket.IO client keeps a 60-second read-idle deadline after namespace registration. If Engine.IO ping/activity stops, the connection loop returns, records the socket as disconnected, refreshes the remote-ops route, and reconnects instead of leaving the Agent in stale connected state. The local watchdog treats connected, connecting, or disconnected socket state older than 90 seconds as restart-worthy fallback recovery.
    • Helper processes never enroll, never refresh tokens, never open Socket.IO, and never talk to the Engine directly. Tray UI reads redacted role-health status from a local state file and writes file-based agent.restart / agent.update_check action requests that the SYSTEM runtime polls; no local web server or listener is created for the tray UI.
    • Windows CURRENTUSER helper processes take a per-session singleton mutex before starting the tray heartbeat. Duplicate helper launches exit cleanly and write helper-events.log under the local tray state directory so sleep/wake duplicate-prevention events are visible.
    • Current Go Windows CURRENTUSER support uses same-binary helper sentinels for session readiness, direct CreateProcessAsUser session launch from SYSTEM for signed quick jobs, and the Windows tray/status UI. Real-host PowerShell Desktop canary validation passed, including denial when the user context attempted to write to root C:\; tray UI real-host acceptance remains pending.
    • WireGuard tunnels are ensured via POST /api/agent/vpn/ensure on boot and refreshed periodically.
    • The ensure loop re-establishes the tunnel automatically after network hiccups.
    • Go startup posts full timeline milestones before entering the Socket.IO connect loop and keeps heartbeat/status telemetry on the SYSTEM/root runtime.
    • Runtime heartbeat cadence is 20 seconds plus bounded jitter. Heavy device-audit data is cached so rapid role-health updates do not force full hardware/software collection every cycle.
    • After a successful /api/agent/heartbeat, the Agent prunes stale .config-*.tmp and .metadata-queue-*.tmp files beside agent.json when they are older than one hour. Fresh temp files and lock files are left alone so active config writes are not interrupted.
    • Patch inventory publishes through /api/agent/details as details.patches and uses state=pending|installed, source=wua_pending|wua_history|quick_fix_engineering, and KB-first de-duplication. Scheduled patch install work reaches the Agent through patch_install_request with a selected patch identity. The Agent serializes patch installs per device so overlapping same-device patch jobs wait instead of failing immediately. The Agent matches WUA updates by update identity/revision first, then falls back to KB/title when the WUA identity is stale. If the update was already installed by the time the job runs, the Agent returns a completed result with already_installed=true. The install path waits for Windows Update Agent to become idle, uses WUA async BeginDownload/EndDownload and BeginInstall/EndInstall when available, falls back to synchronous WUA Download() or Install() if async callback startup fails, posts live download/install progress to /api/agent/patches/install-progress, captures WUA result-code names and per-update result/HRESULT details when available, does not reboot unless a separate patch_reboot_request arrives, and refreshes inventory after the attempt.
    • Patch policy enforcement reaches the Agent through patch_policy_enforcement_request. Managed/frozen modes back up Borealis-owned Windows Update policy values under HKLM:\SOFTWARE\Borealis\PatchManagement, set reversible Windows Update policy values to prevent automatic installs, and keep Borealis SYSTEM WUA install paths available. Unmanaged mode restores values Borealis backed up.
    • Patch reboot requests reach the Agent through patch_reboot_request. The default script checks Win32_ComputerSystem.UserName and returns skipped_logged_in_user when a user is signed in. Policies must explicitly set force behavior before the Agent schedules shutdown for logged-in-user sessions.
    • Heartbeats include queued metadata_fields from metadata-queue.json when present. Engine responses return metadata_field_acks for accepted or superseded queued fields; acked entries are removed from the queue. Engine remains source of truth.
    • Heartbeats/details also carry centralized role-supervisor health snapshots so the Device Details Agent Health tab can show current role/service status with desired state, observed state, last-checked timestamps, last healthy time, last error, and recovery-attempt counts. Startup status uses POST /api/agent/status under the separate startup context so later SYSTEM role-health heartbeats do not erase the timeline row.
    • The VNC role generates one shared UltraVNC password when the role starts, rotates it again every 24 hours by default (BOREALIS_VNC_CREDENTIAL_ROTATION_SECONDS), keeps it in memory only, and returns it to the Engine only through live Agent Socket.IO vnc_credential_request calls. The Agent does not probe UltraVNC auth locally by default because each loopback auth probe consumes an UltraVNC login attempt and can trip lockout before Guacamole connects. Set BOREALIS_VNC_LOCAL_AUTH_VERIFY=1 only for focused diagnostics. The role keeps UltraVNC continuously running once it has the Engine /32 firewall scope, writes UltraVNC config under %ProgramData%\UltraVNC\ with loopback allowed for local diagnostics, and reports ready, service_state, listener_state, last_ready_at, Windows display_topology, and Windows display_virtual_bounds through VNC ensure, credential, and role-health payloads even when no operator is currently connected.
    • VNC role trace logs (vnc_trace ...) are disabled by default because the always-on health loop can otherwise produce high-volume logs during normal operation. Set BOREALIS_VNC_TRACE=1 only for short diagnostic captures.
    • The UltraVNC config writer enables capture performance flags (TurboMode, full-screen polling defaults, EnableDriver, and EnableHook) when the official UltraVNC helper DLLs are present beside winvnc.exe.
    • Software inventory suppresses UltraVNC's known EXE/MSI-wrapper ARP noise row (UNREGISTERED - Wrapped using MSI Wrapper) and keeps the real UltraVNC installed-software row.

    Token storage

    • Refresh/access tokens are stored in agent.json.
    • Device GUID and Engine agent ID are stored in agent.json.
    • When tokens are invalid or expired, the agent refreshes or re-enters enrollment.

    Logging

    • Agent runtime log: Logs/Agent/agent.log with daily rotation.
    • Agent error log: Logs/Agent/agent.error.log.
    • Agent remote shell log: Logs/Agent/remote_shell.log.
    • Agent role/watchdog recovery log: Logs/Agent/role_recovery.log.
    • Agent log retention defaults to agent.log_retention_days: 1; all Agent logs rotate daily and old rotated logs are pruned on the next write/start using that value.
    • Agent bootstrap/update diagnostics: <AgentInstallRoot>/Logs/Agent/bootstrap.log; Windows bootstrap truncates this file at each start and always writes verbose trace/command output there while keeping console/GUI output minimal. Deferred Windows self-replacement writes retry, hash verification, finalization, and task restart output to Logs/Agent/updater.log. Linux updater diagnostics use bootstrap.log.
    • WireGuard role log: Logs/WireGuard/wireguard.log.
    • WireGuard MSI install log: Logs/WireGuard/wireguard-msi-install.log.
    • UltraVNC role log: Logs/UltraVNC/vnc.log.
    • UltraVNC MSI install log: Logs/UltraVNC/ultravnc-msi-install.log.

    Troubleshooting flow

    • If enrollment fails, check:
    • Logs/Agent/agent.log for enrollment errors.
    • Engine/Services/api-backend/logs/engine.log for approval or auth failures.
    • If current-user execution fails, confirm the SYSTEM broker is advertising helper capability, inspect session inventory for helper_ready, and expect no_interactive_user_session when no eligible user session exists.
    • If CURRENTUSER execution fails, inspect the Go helper broker migration status in Data/Agent/Golang_Agent_Migration.md.
    • Operator-requested manual updates arrive over the SYSTEM Socket.IO channel as agent_update_request and start the local AutoUpdater path immediately so the same updater-owned path is used for both manual and hourly runs.
    • Engine-managed release channels cache a Go Agent binary bundle containing Data/Agent/dist/windows-amd64/Agent.exe and Data/Agent/dist/linux-amd64/Agent. Agents download that authenticated bundle, verify SHA-256 when provided, validate the candidate with --validate-config --config-path <agent.json>, stage the platform binary, and restart through the local service manager.
    • Branch installs persist the operator-selected branch in agent.json as agent.branch; local update checks use that branch when it is not main so feature-branch agents do not jump release channels accidentally. Repo-ref update checks resolve target commit through Engine /api/repo/current_hash first so Engine cache absorbs GitHub API usage; direct GitHub lookup is fallback only. If both lookups confirm the branch/ref no longer exists, the Agent writes stable/main to agent.json and runs the stable manifest update path instead of staying stranded on the retired branch.
    • Engine heartbeat recovery backs up Agent self-remediation. Devices without explicit release-channel override that still report unstable are immediately corrected to stable/main. When a device with an override keeps reporting a non-main unstable branch that no longer resolves, the Engine waits five minutes from first confirmed missing-ref observation, clears the stale override on the device row, and returns agent_release_channel_instruction in the heartbeat response. The Agent stores that instruction as a local update operation and starts the normal updater path.
    • Installed build tracking lives in agent.json as agent.installed_build_id; fresh unstable/source branch bootstraps stamp this from the resolved repository commit SHA so the Engine can prove the exact branch build before the first update cycle. The Go Agent does not create a standalone installed_build_id.txt.
    • Update checks do not persist update_status.json; transient state such as state, update_available, and last_checked_at is intentionally not stored by the Go Agent.
    • Metadata fields can be read or queued by scripts through Agent.exe --metadata get 1 / Agent.exe --metadata set 1 "text" on Windows or Agent --metadata get 1 / Agent --metadata set 1 "text" on Linux. set base64-encodes the supplied value into metadata-queue.json; the next heartbeat sends it to Engine. get returns a pending queued value first, otherwise reads Engine through device-auth API. Blank values queue clears. Decoded values are capped at 1024 characters.
    • Windows update archives and extracted repository payloads are staged under C:\Borealis\Temp\Updater; the updater removes update workspaces immediately, schedules full C:\Borealis\Temp cleanup after bootstrap exits so stdout/stderr handles are closed, and cleans old accidental C:\Borealis\Agent update workspaces.
    • Windows self-update stages Agent.exe.update beside Agent.exe only while replacing a running binary. The updater validates candidate config compatibility before stopping the service, and the detached replacement script repeats that validation before moving the pending binary. The detached updater retries the move, verifies the installed Agent.exe SHA-256, then runs Agent.exe --finalize-update so agent.json receives the new agent.installed_build_id only after verified replacement. Failed deferred replacements remove their staged files, and later bootstrap/update starts remove stale Agent.exe.tmp, Agent.exe.update.tmp, and abandoned Agent.exe.update artifacts.
    • The scheduled AutoUpdater cadence is hourly on Windows and Linux. Windows registers Borealis Agent (Watchdog) every minute; Linux registers borealis-agent-watchdog.timer every minute. Both watchdog paths repair missing/stopped services and restart stale service health when no update is active. Watchdog health validation checks service PID vs agent.liveness.pid, stale local ticks, stale heartbeat success, stale connected Socket.IO state, and prolonged socket disconnection, not only service process state.
    • If scripts do not run:
    • Confirm quick_job_run events and the correct role context.
    • Verify signatures with signature_utils logs.
    • If VPN fails:
    • Check agent WireGuard role logs and confirm /api/agent/vpn/ensure succeeds.
    • Ensure the Engine has an active tunnel session and the WireGuard service is running.
    • If VNC fails:
    • Check Logs/WireGuard/wireguard.log for tunnel lifecycle and WireGuard recovery events.
    • Call POST /api/agent/vnc/ensure and inspect ready, service_state, listener_state, detail, and last_ready_at.
    • Confirm the active collaboration session still exists from the Engine side with GET /api/vnc/sessions.

    Borealis Agent Codex (Full)

    Use this section for agent-only work (Borealis agent runtime under Data/Agent -> /Agent). Shared guidance is consolidated in ui-and-notifications.md and the Engine runtime notes.

    Scope and runtime paths

    • Purpose: outbound-only connectivity, device telemetry, scripting, UI helpers.
    • Bootstrap: Agent.exe owns deploy, repair, update check, config write, native service registration, support-task registration, and runtime. Windows onboarding stages the Go binary from Data/Agent/dist/windows-amd64/Agent.exe; the installed copy runs from C:\Borealis\Agent.exe.
    • Windows support dependencies: Agent.exe can still install UltraVNC and WireGuard from official installers. Installed dependency versions and install state-machine phases live in agent.json under agent.dependency_state with phase/status/version/timestamp/error fields. When an existing WireGuard client executable is present, bootstrap records the detected file/registry/config version, marks the dependency healthy, and skips MSI reinstall. Transient installer payloads under C:\Borealis\Dependencies are removed after dependency reconciliation. It does not stage Python, create a venv, or call launch_service.ps1.
    • Existing Windows agents are repairable when C:\Borealis\Agent.exe, the BorealisAgent service, and an Engine-accepted token in agent.json are present.
    • Linux first install: download Data/Agent/dist/linux-amd64/Agent to the operator's current directory, mark that downloaded file executable, and run it as root with --server-url <url> --site-enrollment-code <code>. The binary self-stages into /opt/Borealis/Agent/Agent, writes agent.json, installs borealis-agent.service, and enables borealis-agent-updater.timer plus borealis-agent-watchdog.timer. Re-running the same command performs an in-place redeploy and preserves existing install identity. --update-check also preserves existing install state.
    • Edit in Data/Agent, not /Agent; runtime copies are ephemeral and wiped regularly.
    • Keep Linux Agent installation separate from deployed Engine runtime roots.

    Logging

    • Primary log: Logs/Agent/agent.log with daily rotation to agent.log.YYYY-MM-DD; retention defaults to one day through agent.log_retention_days.
    • Agent support logs: Logs/Agent/bootstrap.log, Logs/Agent/remote_shell.log, and Logs/Agent/role_recovery.log.
    • WireGuard logs: Logs/WireGuard/wireguard.log and Logs/WireGuard/wireguard-msi-install.log.
    • UltraVNC logs: Logs/UltraVNC/vnc.log and Logs/UltraVNC/ultravnc-msi-install.log.
    • Keep ad-hoc traces (for example, system_last.ps1) under Logs/ to keep runtime state self-contained.
    • Updater trace exception: Agent.exe writes bootstrap/update diagnostics to <AgentInstallRoot>/Logs/Agent/bootstrap.log; Windows bootstrap starts by truncating this file and keeps verbose output out of operator-facing stdout/stderr streams.
    • Troubleshooting: prefix lines with <timestamp>-<service-name>-<log-data>; ask operators whether verbose logging should stay after resolution.

    Security

    • Generates device-wide Ed25519 keys on first launch and stores PKCS8/SPKI base64 in agent.json.
    • Refresh/access tokens are stored in agent.json and bound to the device identity plus Engine-issued token state; mismatches force re-enrollment.
    • REST and Socket.IO traffic use the Engine FQDN with normal CA + hostname validation. Internal-Only installs persist the Borealis local CA PEM in agent.json and append it to the system trust pool. Internal-Only installs can also persist server_ip_fallback; this changes HTTPS TCP dial targets only after normal FQDN connection fails. On Linux, it also lets the WireGuard role rewrite the endpoint to the fallback IP only after FQDN resolution fails during wg-quick up.
    • Validates script payloads with backend-issued Ed25519 signatures before execution.
    • Outbound-only; API/WebSocket calls flow through the Go auth client for proactive refresh. The Agent never uses InsecureSkipVerify; local CA trust still requires FQDN SAN match. Logs bootstrap, enrollment, token refresh, and signature events under Logs/Agent/.
    • Helper processes inherit no Borealis token state and rely on the local SYSTEM broker for job delivery.

    Reverse VPN tunnels

    • WireGuard reverse VPN design and lifecycle are documented in remote-shell.md and remote-desktop.md.
    • The original references were REVERSE_TUNNELS.md and Reverse_VPN_Tunnel_Deployment.md (now consolidated into this knowledgebase).
    • Agent roles:
    • Data/Agent/internal/roles/wireguard_tunnel (Go tunnel lifecycle)
    • Data/Agent/internal/roles/remote_shell (Go VPN remote shell TCP server)
    • Data/Agent/internal/roles/vnc (Go Windows UltraVNC lifecycle and credential broker)

    Execution contexts and roles

    • Go roles are explicit packages under Data/Agent/internal/roles.
    • First PR supports SYSTEM/root quick-job script execution, Windows CURRENTUSER helper session health plus direct session PowerShell/Batch execution, core device audit inventory including Windows domain/workgroup telemetry, SYSTEM/root file management, SYSTEM/root process management, SYSTEM/root service management, SYSTEM/root software management, SYSTEM/root WireGuard tunnel lifecycle, SYSTEM/root Remote Shell over WireGuard, Windows VNC lifecycle/credential brokerage over WireGuard, and Go release-channel self-update.
    • Pending ports are tracked in Data/Agent/Golang_Agent_Migration.md.
    • Service tasks depend on Windows service and scheduled-task creation rights; failures should surface through Engine logging and Logs/Agent/role_recovery.log.

    Platform parity

    • Windows is the reference path and has the broadest tested feature surface.
    • Linux Go runtime builds as Agent, self-stages when server URL and enrollment code are provided, installs through systemd, enables an hourly updater timer, and supports root/SYSTEM Bash quick jobs in first PR.
    • Linux CURRENTUSER and tray UI are pending Go ports. Linux VNC is explicitly unsupported in the current Go role.

    Ansible support

    • The agent no longer hosts an Ansible playbook execution role.
    • Borealis Ansible control-node execution is Engine-side and should target devices over the Engine-managed WireGuard paths.
    • Agent responsibilities for the Ansible architecture are limited to:
    • maintaining device identity and inventory in the Engine
    • sustaining the reverse WireGuard tunnel and related remote-access services
    • exposing the device to Engine-driven automation over the VPN path