Neuralos

neuralOS — Headless Install & Operations Skill

neuralOS (neuralos on npm, CLI neuralos or gybackend) is the full RTerm runtime as a standalone Node daemon — no desktop UI. neuralOS is the headless AI-native backend for RTerm: the RTerm desktop app stays RTerm; neuralOS is the standalone backend daemon (also published as rterm-backend for backward compat). It boots the AI agent, SSH/WinRM/Serial/local terminals, fleet orchestration, scheduled automation, and change management, and serves them over a WebSocket JSON-RPC gateway (default ws://<host>:17888).

rterm-cli (rterm-cli on npm, bins rterm / rterm-cli) is the official command-line client for that gateway — a tiny zero-dependency script with 14 one-shot commands (ping, run, fleet, metrics, …) plus an interactive persistent chat (rterm chat) that gives the desktop chat experience in any terminal: streaming replies, session resume with history replay, command approvals, and slash commands. The engine (neuralos) does the work; the CLI is how scripts, CI, and humans talk to it.

Use this skill to install, configure, run, and administer the stack completely headlessly on macOS, Linux, or Windows — the bundled lifecycle script manages backend and CLI together or individually — then drive it with the rterm-gateway skill for the RPC reference.

1. The 60-second path (any OS)

# one command: installs BOTH packages (if missing), starts the daemon, verifies
node scripts/neuralos.mjs setup
# -> installs neuralos@latest + rterm-cli@latest
# -> starts the gateway (daemon by default)
# -> verify: gateway ping + CLI→gateway round-trip + terminals listing

# then use it from any shell:
rterm ping
rterm chat          # interactive persistent chat (desktop experience)

Prefer the manual path:

# 1. install (Node >= 18 required)
npm install -g neuralos rterm-cli

# 2. run it
neuralos
# -> [neuralos] WebSocket RPC endpoint: ws://0.0.0.0:17888

# 3. verify (in another shell)
rterm ping
# -> {"pong": true, ...}

The bundled scripts/neuralos.mjs CLI wraps every lifecycle step (doctor, install, update, setup, verify, start/stop/restart, status, logs, config, service, uninstall) into one cross-platform command — managing neuralos and rterm-cli together or individually (--backend / --cli flags). Run any of these with node scripts/neuralos.mjs <cmd>.

2. What runs inside (mental model)

┌──────────────┐  WebSocket JSON-RPC   ┌────────────────────────────────┐
│ your agent / │ ◄──────────────────► │ neuralos (neuralOS backend)      │
│ program / CI │                       │  AgentService  (LLM + tools)   │
└──────────────┘                       │  TerminalService SSH/WinRM/    │
                                       │    Serial/local PTY          │
                                       │  AutomationManager + cron    │
                                       │  ChangeManagement (MOP)      │
                                       │  Ledgers (SQLite)            │
                                       └────────────────────────────────┘
                  Data dir: settings.json + *.sqlite + session-logs/

• Requests: { "id": "1", "method": "<name>", "params": {...} }
• Responses: { "type": "gateway:response", "id": "1", "ok": true|false, "result"|"error" }
• Events (progress): { "type": "gateway:event" | "gateway:raw" | "gateway:ui-update", ... }

3. Install

3.1 Requirements

3.2 Install from npm (recommended)

# both packages in one shot (backend + CLI)
npm install -g neuralos rterm-cli

# or individually
npm install -g neuralos      # the daemon (bins: neuralos, gybackend)
npm install -g rterm-cli     # the CLI client (bins: rterm, rterm-cli)

neuralos --version 2>/dev/null || which neuralos || where neuralos
rterm version   # asks the running gateway; also proves CLI↔gateway connectivity

Or without a global install: npx -y neuralos / npx -y rterm-cli ping.

3.2b Install & update with the lifecycle script (handles both, individually or together)

node scripts/neuralos.mjs install              # BOTH neuralos + rterm-cli
node scripts/neuralos.mjs install --backend    # daemon only
node scripts/neuralos.mjs install --cli        # CLI only

node scripts/neuralos.mjs update               # both → latest (registry-checked)
node scripts/neuralos.mjs update --backend     # daemon only
node scripts/neuralos.mjs update --cli         # CLI only
# update compares installed vs registry version first; skips when already
# latest, and reminds you to `restart` the daemon when it did update.

node scripts/neuralos.mjs uninstall            # stop + remove BOTH
node scripts/neuralos.mjs uninstall --keep-cli # stop daemon, keep the CLI

The script auto-recovers from the classic EACCES … cache folder contains root-owned files npm failure by retrying once with a fresh temp cache — no sudo needed.

3.2c One-shot server setup (new machine → working stack)

# fresh server: install + start + verify in one command
node scripts/neuralos.mjs setup
# idempotent: re-running on a healthy machine changes nothing and just verifies

# with options
node scripts/neuralos.mjs setup --port 17889 --data /var/lib/neuralos --daemon

# end-to-end health check any time (also exits non-zero on failure — CI-friendly)
node scripts/neuralos.mjs verify
# -> ✔ gateway ping (ws://127.0.0.1:17888)
# -> ✔ rterm-cli 3.7.3 present
# -> ✔ CLI→gateway round-trip: backend 3.0.0, 132 RPC methods
# -> ✔ CLI terminals listing: 3 tab(s)
# -> ✔ verify PASSED

verify exercises the full path: raw WebSocket ping → CLI installed → CLI version round-trip through the gateway → CLI terminals (terminal subsystem). Use it after install, after update, and in CI health gates.

3.3 Install from a repo checkout (development)

git clone https://github.com/DrOlu/RTerm.git && cd RTerm
npm install
npm run build:backend-standalone      # dist-standalone/gybackend.js
npm run start:backend                 # or: node apps/gybackend/dist-standalone/gybackend.js

3.4 OS-specific service install (run as a daemon/service)

Use the bundled helper, or the unit files in service/:

node scripts/neuralos.mjs install-service     # prints the right unit + enable cmd for this OS

• Linux (systemd): service/neuralos.service → /etc/systemd/system/, then systemctl enable --now neuralos.
• macOS (launchd): service/ng.hyperspace.neuralos.plist → ~/Library/LaunchAgents/, then launchctl load <plist>.
• Windows (Task Scheduler): service/install-windows-service.ps1 → registers an at-logon task (schtasks). Native deps install via npm i -g first.

4. Configure

4.1 Environment variables

4.2 The data directory

4.3 Reuse desktop-app settings

# macOS
GYBACKEND_DATA_DIR="$HOME/Library/Application Support/rterm" gybackend
# Linux
GYBACKEND_DATA_DIR="$HOME/.config/rterm" gybackend
# Windows (cmd)
set GYBACKEND_DATA_DIR=%APPDATA%\rterm && gybackend
:: Windows (PowerShell)
$env:GYBACKEND_DATA_DIR="$env:APPDATA\rterm"; gybackend

Warn: two instances sharing one data dir should not run the same scheduled tasks at once (duplicate execution). For a dedicated automation server, give it its own data dir and recreate only what it needs.

4.4 Command policy (autonomy)

Pre-allowlist what a headless job needs, then run smart:

settings:addCommandPolicyRule {list:"allowlist", rule:"Update-MpSignature*"}
settings:addCommandPolicyRule {list:"allowlist", rule:"systemctl *"}

4.5 Securing the gateway

• Local-only: GYBACKEND_WS_HOST=127.0.0.1 when callers are on the same host.
• Token: non-localhost clients need Authorization: Bearer <token> (manage in access-tokens.json).
• CIDR allow-list: settings → gateway.allowedCidrs.
• Localhost bypass: 127.0.0.1/::1 skip the token by default.

5. Run & administer

The bundled scripts/neuralos.mjs handles the lifecycle cross-platform (uses only Node built-ins — no dependencies). It manages backend and CLI together or individually:

node scripts/neuralos.mjs doctor          # check Node, npm pkgs (both), data dir, port, gateway + CLI round-trip
node scripts/neuralos.mjs install         # npm i -g neuralos + rterm-cli (--backend / --cli for one)
node scripts/neuralos.mjs update          # registry-checked update to latest (both; --backend / --cli)
node scripts/neuralos.mjs setup           # install-if-missing + start + verify (idempotent one-shot)
node scripts/neuralos.mjs verify          # gateway ping + CLI round-trip + terminals (exit 1 on fail)
node scripts/neuralos.mjs start [--port N] [--host H] [--data DIR] [--daemon]
node scripts/neuralos.mjs stop
node scripts/neuralos.mjs restart
node scripts/neuralos.mjs status
node scripts/neuralos.mjs logs [--lines N]
node scripts/neuralos.mjs ping [--url ws://...]
node scripts/neuralos.mjs config-show     # effective env + data dir + installed versions of both pkgs
node scripts/neuralos.mjs install-service # print service unit + enable cmd for this OS
node scripts/neuralos.mjs uninstall [--keep-cli]  # stop + npm uninstall (both; keep CLI)

Boot output (healthy)

[WebSocketGatewayAdapter] Listening on ws://0.0.0.0:17888
[neuralos] Started.
[neuralos] WebSocket RPC endpoint: ws://0.0.0.0:17888
[neuralos] Data directory: /var/lib/neuralos

Foreground vs background

• Foreground: gybackend (Ctrl+C to stop) — good for first-run debugging.
• Background/service: systemd / launchd / Task Scheduler, or ... start --daemon (uses nohup/Start-Process and writes a pidfile + log).

5b. rterm-cli — the official command-line client

rterm-cli (npm) installs two bins: rterm and rterm-cli (identical). It speaks the gateway natively — zero dependencies, ~41 kB, works from any machine that can reach the gateway (point it at a remote daemon with --url or RTERM_URL).

One-shot commands (scripting / CI)

rterm ping                                 # liveness
rterm version                              # backend version + method count
rterm methods [--category agent]           # self-describing RPC surface
rterm call <method> [json]                 # raw JSON-RPC call (ANY method)
rterm terminals                            # list terminal tabs
rterm connections                          # saved SSH/WinRM/Serial connections
rterm open <connection-name>               # open a tab for a saved connection
rterm close <tabIdOrName>                  # close a tab
rterm run <tabIdOrName> <command>          # run a command in a tab (waits for output)
rterm fleet <tab1,tab2,...> <command>      # run on many tabs at once
rterm sessions                             # list chat sessions
rterm chat <sessionId> <message>           # one-shot agent message (blocking)
rterm dashboard                            # live dashboard state (JSON)
rterm metrics [--format prometheus]        # host metrics

Fleet example (the classic ops loop):

rterm open web-01 && rterm open web-02
rterm fleet web-01,web-02 "df -h / && systemctl is-active nginx"
rterm metrics

Remote gateway (CLI on your laptop, daemon on a server):

RTERM_URL=ws://10.0.0.5:17888 RTERM_TOKEN=<token> rterm ping
rterm --url ws://10.0.0.5:17888 --token <token> run web-01 "uptime"

Interactive persistent chat (rterm chat)

rterm chat with no arguments is the desktop chat experience in a terminal — same sessions, same events, same approvals the desktop app uses:

$ rterm chat
Connected to ws://127.0.0.1:17888 — session 96c153ac…
── resuming (4 messages) ──
you> Reply with exactly one word: PONG
· Reasoning... The user wants exactly one word: PONG
assistant> PONG
you> /exit
session 96c153ac… kept server-side — rerun "rterm chat" to resume.

• Streaming — replies, reasoning, and tool output render live (say, sub_tool_*, command_* gateway events).
• Persistent — sessions live server-side (SQLite); the last session id is saved in ~/.rterm-cli/chat-state.json and auto-resumed with history replay. Exit, kill the terminal, reboot — the conversation survives.
• Command approvals — when the agent asks to run a command, the CLI pauses for allow? [y/N] and replies via agent:replyCommandApproval (any non-y answer denies).
• Slash commands: /new /sessions /rename <t> /branch /export [--simple] /search <q> /stop /verbose /exit (or Ctrl-D).
• Flags: --session <id> resume a specific session; --verbose start with raw events on.

CI / unattended usage of the same session machinery: rterm call agent:startTask '{"sessionId":"...","userInput":"..."}' (blocking) or agent:startTaskAsync + event watching (see the rterm-gateway skill).

6. Observability & SRE features (v2.0.0–v2.3.1)

The backend boots with a full observability stack wired in (createObservability), fed live by monitor snapshots. All of it is callable over the gateway (see the rterm-gateway skill).

SRE core

• MetricsLedger — time-series store for resource snapshots (cpu/mem/disk/load/net/gpu) per host, with trend slope + days-to-threshold forecasting ("disk full in N days").
• UptimeWatchdog — liveness probes (tcp/ssh/http/command) per host, up/degraded/down, with state transitions firing alerts.
• SloService — SLO/SLI definitions, error budget, burn rate, fast-burn alerting.
• AlertService — alertmanager-style routing: grouping, dedupe, silences, severity channels.
• IncidentLedger — auto-incidents with timelines, AI RCA, postmortems, runbook links.
• GoldenSignals — saturation/traffic/latency/errors per host + capacity forecast.
• SyntheticChecks — blackbox probes feeding the SLO SLI + golden latency/error.
• DriftDetector — template-vs-live config diff + MOP auto-remediation.

APM / DEM / Infra / ETW

• SpanLedger (APM) — OTLP distributed-trace store + analysis (per-service p50/95/99, error rate, slowest traces, bottleneck services).
• RumLedger (DEM) — Core Web Vitals (LCP/INP/CLS/TTFB) per page + error rate, slowest/poor pages.
• InfraMonitor (k8s/cloud) — cluster health (running/notReady/CrashLoop/restarts/nodes + cpu/mem % of limit), unhealthy instances.
• EtwService (Windows) — built-in ETW diagnostics (network/file/registry/process providers, logman sessions, Get-WinEvent/Get-Counter) — agentless, no install.

Predictive + behavioral + evals

• AnomalyDetector — z-score / robust z-score (median±MAD) anomaly detection over metric series.
• EarlyWarningService — predictive failure alerts (trend forecast + anomaly) + optional MOP auto-remediation.
• BehaviorLedger — UEBA-style baselines (runs/day, tokens/run, error rate, models) + deviations (run-spike, token-blowout, error-spike, unusual-model).
• EvalHarness — embedded evals measuring the agent's accuracy, tool selection, safety/policy, determinism/replay with an aggregate reliability report.

Notifications (Slack / Teams / SMTP / Telegram)

Wire alert channels into the AlertService with vaulted webhook URLs / SMTP creds (rich, severity-colored payloads):

# via the gateway (rterm-gateway skill) — add a Slack channel
# (see examples/notify-channels.mjs)

Unified live dashboard

A single dashboard:state object aggregates every ledger (fleet health, SLO board, uptime map, incident feed, APM bottleneck+slowest, DEM slowest/poor, k8s clusters, capacity forecast) — broadcast over the gateway for rich, live, cross-linked dashboards. A renderDashboardHtml renderer produces a browser-viewable HTML dashboard from that state (Aurora-themed, auto-refreshing) — serve it over HTTP to view the live dashboard in any browser.

dagu workflows (v2.4.0+)

Run declarative dagu YAML DAG workflows natively on RTerm's orchestrated playbook engine — no dagu server required. The daguParser compiles dagu YAML into a playbook:

• Steps — id/name, run/command/cmd/script/call (all forms) → step commands.
• Dependencies — depends (string or array) → dependsOn (fan-out/fan-in DAG waves).
• Failure handling — continue_on → onError: continue; retry_policy noted.
• Guards — preconditions → a desiredState skip-when-satisfied guard.
• Runbook params — params (string or object with default) → playbook params with defaults.

Paste a dagu YAML workflow to the agent ("run this dagu workflow") or compile it via parseDaguYaml and run the resulting playbook with run_playbook — it executes on RTerm's orchestrated DAG runner across your hosts with validation and rollback.

AWS APerf deep-dive (v2.6.0+)

Deploy the AWS APerf CLI to any Linux host via SSH, record deep system performance metrics (CPU, memory, disk, network, PMU counters, processes, hotspot data), generate the aperf analysis report, and parse the findings into structured results that feed the metrics ledger + agent RCA. Combines aperf's deep profiling with RTerm's agent reasoning.

Ask the agent: "Run an APerf deep-dive on web-01 and report the top performance issues" — the agent installs aperf on the host (if needed), records for the sampling period, parses the report, and returns findings with severity thresholds (critical ≥90%, warning ≥75%, process ≥50% CPU).

Plugin system (v2.5.0+)

Anyone can develop a custom plugin and have it auto-integrate. A plugin is a folder with a plugin.json manifest (name, version, entry, tools, triggers, panels, permissions) and an index.mjs entry module exporting register(ctx). The PluginRegistry discovers plugins in:

1. ~/.gybackend-data/plugins (user-installed)
2. ./plugins (repo/dev)
3. {bundle}/../plugins (npm package)
4. {resourcesPath}/plugins (desktop app)

The backend ships with 14 official plugins out of the box (61 tools, 12 triggers, 14 panels):

Audit trail + evidence sealing (v2.7.1)

Hash-chained, tamper-evident audit ledger — every audit-relevant event (agent runs, command evaluations, approvals, MOP changes, playbook steps, trigger firings, alert ingestions) is appended with the SHA-256 hash of the previous record. Any tampering breaks the chain and is detectable via verify(). The evidence sealer computes a Merkle-tree root over records → sealed, independently-verifiable evidence bundles (KLA audit framework domain 11). 18 event kinds recorded.

Monitor diagnostics (v2.7.6)

monitorStatus diagnostic — reports exactly why monitor stats aren't displaying per terminal: publisher wired? session exists? collection stuck in-flight? terminal connected? platform detected? last-collect time? Diagnoses: terminal_not_connected, no_monitor_session, collection_stuck_in_flight, never_collected, stale_collection (>30s), publisher_not_wired.

Ask the agent: "Run monitor status diagnostics" — instantly shows which terminals aren't collecting and why.

AGT policy engine (v2.7.7)

Microsoft AGT-style policy engine — evaluates every consequential action against a YAML policy before execution. Decisions: allow / deny / escalate (route to approval). Features: glob-style action patterns ("read" matches "read /etc/passwd"), target wildcards (prod-*), first-match-wins, case-insensitive matching, agent identity + sponsoring principal for zero-trust. Built-in default policy: allow read/status/list; deny delete/drop/format; escalate restart/patch/deploy on prod-*. Drop a custom policy.yaml in the data dir to override.

Review model / maker-checker (v2.7.8)

The review model (a second LLM, the "checker") independently verifies the action model's (the "maker's") output on 5 dimensions: correctness, completeness, safety, compliance, and accuracy.

• Verdicts: approved / needs_revision / escalate.
• Modes: strict (block on any issue), advisory (flag but allow), auto-approve (skip review for low-risk actions).
• Fast output mode: if no reviewModelId is set in the model profile, reviews are skipped entirely (zero added latency).

Configure in settings.json → models.profiles[].reviewModelId + reviewMode — or in the desktop Settings UI (v2.7.9+).

v2.9.x platform capabilities

v2.9.0 added 9 backend modules; v2.9.2 exposed them as 41 observability:* gateway RPC methods + 9 agent tools (see the rterm-gateway skill §4b); v2.9.3 made the tools visible in the Tools section. All are wired into createObservability and live on a stock install.

Agent tools (visible in the Tools section since v2.9.3): get_metrics, manage_secret, manage_oncall, get_cost, manage_recording, manage_gitops, manage_playbook_version, get_cloud_inventory, get_live_dashboard. Ask the agent: "add this API key to the vault", "show my AI spend today", "page the on-call", "lint this playbook", "list my AWS instances".

v2.9.5 — APM/DEM/Infra/ETW ingestion. The observability ledgers are now genuinely fed out of the box via observability:apmIngestSpans (OTLP spans → trace store), observability:demIngestBeacon (Core Web Vitals RUM beacons → per-page p75), observability:infraCollect (k8s cluster health from kubectl), and observability:etwStartTrace/etwStopTrace/etwParse (Windows ETW diagnostics) — plus the matching agent tools ingest_apm_spans/get_apm_summary, ingest_dem_beacon/get_dem_summary, collect_infra, manage_etw.

New env vars: OTEL_EXPORTER_OTLP_ENDPOINT / RTERM_OTLP_METRICS_ENDPOINT (OTel push), RTERM_SECRETS_MASTER_KEY (unlock the secrets vault).

v2.9.6 — settings-driven cost, alerts, on-call & cloud (no placeholders)

Four capabilities that were constructor-injected but never wired to settings/UI are now persisted settings blocks (schema v4→v5), editable in Settings UI, and live-reloaded without a restart (via SettingsService.onDidChange → refreshCost / refreshAlertChannels / refreshOncallChannels / refreshCloudAccounts). Secrets are never inline — secretRef into the AES-256-GCM vault, resolved at send/sync.

Set them via settings:set ({cost:{modelPrices:{…},budgets:[…]}}, {alerts:{channels:[…]}}, {oncall:{pagingChannels:[…]}}, {cloud:{accounts:[…]}}) or the desktop UI — both persist and apply live. Seed the actual secret values into the vault separately (manage_secret / RTERM_SECRETS_MASTER_KEY).

v2.9.7 — security republish + 3 bug fixes

• SECURITY (npm only): neuralos@2.9.6/rterm-backend@2.9.6 accidentally bundled a local .gybackend-data/settings.json (created during a boot-verify) containing a live API key. Both 2.9.6 packages were deprecated then unpublished; clean 2.9.7 adds a hardened .npmignore. Git history was never affected. If you installed 2.9.6 from npm, rotate the affected OpenRouter/provider key and upgrade.
• Cost attribution: new normalizeModelId() collapses self-doubled provider model ids (e.g. moonshotai/kimi-k3moonshotai/kimi-k3) at the run-ledger boundary so pricing matches the configured model (was falling to default). Fixes forward.
• GitOps gateway: assertManifest() guard — calling observability:gitopsDrift/inSync/reconcile with no manifest returns a clear actionable error instead of an opaque Cannot read properties of undefined.

v2.9.8 — backend typecheck fully green + CHANGELOG-driven release notes

• Backend typecheck is now exit 0 across the whole backend. ESM-safe native loaders via createRequire(import.meta.url): commandParser.ts (web-tree-sitter Language type-vs-value), NodePtyBackend.ts (lazy node-pty), and SSHBackend.ts (the let ssh2 variable shadowed the ssh2 type namespace → renamed ssh2Lib + import type * as ssh2, ~50 errors cleared).
• build-release.yml builds the GitHub release body from CHANGELOG.md (extracts the current version's section) instead of a stale hardcoded template — v2.9.8's notes were generated by this new path.

v2.9.9 (in progress) — AgentSpan/Conductor durable-agent bridge

The agentspan-bridge plugin connects neuralOS/RTerm to an AgentSpan (Netflix Conductor) server, adding what RTerm didn't already have: true durable agent execution (a crashed run resumes from the last completed step, not just a ledger entry), plan-execute determinism (LLM plans once → immutable sub-workflow), enterprise event triggers (Kafka/SQS/AMQP/DB), and the server's visual execution UI.

• Configure: Settings → AgentSpan (agentspan.serverUrl, default http://localhost:6767, + optional agentspan.authSecretRef → a vault key holding AGENTSPAN_AUTH_KEY/AGENTSPAN_AUTH_SECRET). Run the server with agentspan server start.
• 6 agent tools: agentspan_health, agentspan_run (AgentConfig or named workflow → executionId), agentspan_status, agentspan_approve (HITL respond), agentspan_list, agentspan_stop.
• 1 trigger: agentspan_execution_failed (fires on FAILED/TERMINATED/TIMED_OUT). 1 panel: agentspan-executions (live execution feed).
• Resilient: if the server is down, tools return a clear "server unreachable" hint instead of throwing. See the agentspan skill for the standalone AgentSpan SDK/CLI.

v2.9.10 — AgentSpan Phase 2 (playbooks as workflows + delegate)

Deepened the bridge both directions: agentspan_export_playbook (dry-run a playbook as a Conductor WorkflowDef), agentspan_register_playbook (register an RTerm playbook as a reusable Conductor workflow other agents call via SUB_WORKFLOW), and agentspan_delegate (hand a prompt to a durable AgentConfig agent that runs start-to-finish → executionId survives restart). The plugin now has 9 tools.

v2.9.11 — agent-tool session recording fix

Recordings started via the agent's manage_recording tool captured 0 events (the tool called SessionRecorder.start() directly, which never registered the terminal in TerminalService.activeRecordings — the live-output feed checks that map). start now routes through TerminalService.startRecording() (registers the terminal) and stop deregisters. Agent-started recordings now capture, replay, and export .cast. No asciinema needed.

v2.9.12 — agent-created triggers fire live

Triggers created via manage_trigger were persisted but never fired (the TriggerEngine loaded persisted triggers once at startup; new ones weren't synced into the live engine). manage_trigger create/update/delete/enable/disable now upserts/removes them in the live engine, so they fire without a backend restart.

v2.9.13 — version check 403 fix + silent background updates + no GitHub in UI

The updater fetched version.json from the GitHub API contents endpoint (rate-limited → red "Check Failed: HTTP 403" every hour). Now it fetches from the raw GitHub URL (raw.githubusercontent.com, no rate limit), checks silently in the background (transient network failures keep last-good and stay quiet instead of a red error), and the UI shows the app website rterm.app as the source and rterm.app/#download as the download URL — no GitHub URL visible.

v3.0.0 — API self-discovery (gateway:describe + method registry + list_gateway_methods)

The gateway now describes itself. A single-source methodRegistry.ts holds the whole RPC surface (name, category, description, since, params) that the adapter's dispatch, the gateway:describe endpoint, the list_gateway_methods agent tool, and the reference docs all derive from — so they can never drift. 123 methods across 12 categories. gateway:describe returns {version, count, categories, methods} with optional category/prefix filters; the agent tool does the same. Ask the gateway what it can do instead of reading WebSocketGatewayAdapter.ts or a static doc.

v3.0.2 — live browser dashboard at /dashboard (same port as the WS gateway)

The unified dashboard is now visible in any browser. A new httpRoutes option on WebSocketGatewayAdapter lets the default server factory create ONE node http.Server — plain HTTP requests hit a route table, WS upgrades hit the WSS on the same socket/port (ESM-safe createRequire for node:http; no routes = old behavior). startGyBackend registers /dashboard (live HTML) + /dashboard/json (state): renderLiveDashboardHtml() renders initial state server-side, then an embedded client subscribes via observability:liveDashboardSubscribe and updates each section in place on every monitor-snapshot push (falls back to polling /dashboard/json 5s). Auth mirrors the WS gateway (loopback open, remote needs an access token via Bearer/header/query). Startup logs the dashboard URL. open http://localhost:17888/dashboard.

v3.0.5 — terminal/session core + chat navigation + memory improvements

• SSH auto-reconnect with exponential backoff + jitter (1s→2s→5s→15s→60s cap, 10 max attempts); tab.reconnectState surfaces "reconnecting (attempt N)…" in the UI; manual kills cancel the schedule.
• WinRM persistent runspaces — runCommandOnShell() reuses one shell across commands (was 4 WS-Man round trips per command); streaming output via onChunk; persistent cwd so cd sticks; auto-recovery on dead shell.
• Serial break/DTR/RTS — sendBreak() (Cisco password recovery / ROMMON) + setControlLines().
• Chunked ring buffer — ChunkedRingBuffer replaces the single re-sliced string (O(1)-ish appends, no O(n) copy per chunk on busy tabs).
• Chat user-message navigation — Prev/Next/Latest user buttons + Top/Bottom scroll; programmatic-scroll guard prevents the "can't scroll back to bottom" bug.
• Memory manager — memoryManager.ts: search, dedupe, append-with-cap, relevance-ranked recallForPrompt() (caps injected memory at 12k chars instead of the whole file).

v3.0.6 — chat scroll fix + top/bottom buttons

Fixed the scroll bug where "Prev user" latched auto-scroll off permanently. programmaticScrollRef guard distinguishes programmatic jumps from user scrolls. Added ⇤ Top / ⇥ Bottom one-click buttons. Nav bar always present.

v3.0.8 — SSH legacy/cisco algorithm preset hotfix

Removed ssh2-unsupported algorithms from the legacy/cisco presets (added in v3.0.6 but ssh2 1.17 throws on any offered algo it can't load). filterToSupported() defensively intersects presets with ssh2's SUPPORTED_* constants.

v3.0.9 — web-intel plugin: local-first web intelligence (via wigolo)

The agent now has first-class web tools it didn't have: multi-engine search, clean-page fetch, site crawl, research, and page-watch → RTerm trigger automation. Built as a first-class plugin (plugins/web-intel/) following the agentspan-bridge pattern.

9 tools / 1 trigger / 1 panel:

• webintel_health — daemon status, lean-vs-full warmup, auto-start state.
• web_search — multi-engine ranked search with citations (keyless, $0).
• web_fetch — clean markdown + metadata + links (tiered router escalates to browser engine for JS/SPA/anti-bot).
• web_crawl — multi-page crawl (BFS/DFS/sitemap/map-only).
• web_research — decompose a question → ranked evidence + citations. Synthesis uses RTerm's own agent — no LLM key needed or stored.
• web_find_similar — pages similar to a URL/concept (keyword + semantic + live web fusion).
• web_watch_add / web_watch_list / web_watch_remove — watch a vendor/CVE/status page; the webintel_page_changed trigger fires so a playbook/MOP can react.
• Panel web-intel — watched pages + daemon status.

Lean by default (stock RTerm stays lean):

• The wigolo daemon starts lazily on first use (npx -y wigolo serve) — nothing downloaded at install time.
• Default is WIGOLO_NO_WARMUP=1 — the ~1.5 GB browser engine + on-device models are not downloaded until a tool that needs them runs, or until webIntel.warmupOnInit: true (which kicks off a background wigolo init).
• Search/fetch/crawl work keyless without the heavy models.

Settings block webIntel (schema v5 + normalizeWebIntelSettings): {enabled, restUrl, token, autoStart, warmupOnInit} — defaults keep everything lean and local. Token is optional (only if the daemon uses WIGOLO_API_TOKEN).

Plugin infrastructure upgrades (shared):

• PluginContext.spawnProcess (optional) — plugins can spawn local sidecar daemons; wired in observability.ts via createRequire('node:child_process').
• PluginContext.settings / getSettings — live settings snapshots for plugins that read config blocks.
• registerPanel accepts both (name, render) and {name, title?, render} (pre-existing signature drift fixed).

Resilient: if the daemon is down and can't auto-start, every tool returns {error, hint} instead of throwing — the agent stays usable.

v3.1.x — NATS event mesh, Synapse + Numbat bridges, serial fix

v3.1.1 — serial transport fix + standalone transports. SerialBackend.loadSerial() returned require('serialport') (the module namespace) but spawn() called it as a constructor → SerialPort is not a constructor. Fixed by resolving the class (mod.SerialPort ?? mod) and tolerating both call signatures (v9 positional (path, opts) vs v10+ object {path, ...opts}) — works across serialport v9–v13+. Also declared serialport in optionalDependencies so the standalone npm packages ship it (SSH/serial/local transports now install automatically via optionalDependencies {serialport, ssh2, node-pty}).

v3.1.2 — comprehensive NATS event mesh. NatsEventBus rewritten from a thin core-pub/sub adapter (no auth) into a full NATS client. Auth (token, username/password, NKey seed, JWT jwt+seed, .creds, TLS mutual-auth) — the key gap; before, only open/localhost servers worked. Plus core pub/sub (queue groups + hea

…(truncated)