1.4.4
Fixed
- The assistant's Termix CLI is 1.3.2, so running commands works on Windows hosts
1.4.3
Fixed
- With the assistant on, the tabs and other plugins no longer disappear on Termix 26.11.2. The docked assistant now sits over the page instead of pushing it over
- The AI tab no longer closes right after it opens, such as from the mobile app, before the AI settings have loaded
1.4.2
Fixed
- The bundled Termix CLI is 1.3.1, which drops an old zod with a known security problem
1.4.1
Changed
- Runs in its own process with only the access it needs (plugin API 2, needs Termix 26.11.2)
- The assistant beside a terminal fills the box the terminal gives it, so it shows in the right place when it draws in its own frame (needs Termix 26.11.2)
- Form fields use the same background as the rest of Termix
Fixed
- Save as preset button now matches the Save button size
- Termix CLI tool calls now work in the desktop app, where they run in a worker thread
1.4.0
Added
- The assistant can find and run actions other plugins offer, such as restarting a container or service or waking a machine. Changes wait for your approval in Operations by default
- Read only commands for drive and array health: smartctl, nvme, mdadm, btrfs, lvs, vgs and pvs
- A Failing drive task preset: when Storage Health finds a failing or worn drive or a degraded pool or array, the assistant reads SMART and array state and reports what to do
- A homepage widget with recent auto-ops tasks and plans waiting for approval
Changed
- Needs Termix 26.11.1 or newer
- Toolbars, loading screens, back buttons and Save bars match the rest of Termix
Fixed
- Buttons next to fields in the assets and exposure settings are as tall as the fields
1.3.1
Fixed
- The assistant no longer offers to open the SSH Tools panel, which Plugin SSH Terminal 1.3.0 removes
1.3.0
Added
- The assistant starts on the model you used last
- Advanced settings for OpenAI and OpenAI compatible providers: endpoint (Auto, Responses API or Chat Completions), reasoning effort, reasoning summary, verbosity, max output tokens and temperature. Left on Default, nothing extra is sent
Changed
- The terminal assistant lives in the panel bar beside the SSH terminal on Termix 26.10.3 and later, so it can be resized, moved to a side or the bottom, and ordered with the other panels
- Panel beside the terminal host setting and panel settings for where it opens and how big it is
- OpenAI providers on api.openai.com use the Responses API, so reasoning models such as GPT-5.6 and GPT-6 keep their reasoning while calling tools. OpenAI compatible providers and custom base URLs still use Chat Completions
- The Thinking block shows markdown, and images in model text are never loaded
Fixed
- OpenAI models that refuse tools with their default reasoning now work, by asking again with reasoning off
1.2.1
Fixed
- The AI Agents button no longer shows on hosts where the AI assistant is turned off
1.2.0
Added
- Asset inventory: an Assets view in the AI tab lists hosts, Proxmox guests, containers and the external addresses they talk to, collected daily, on host changes or on request by one read-only script per host; zones come from built-in and your own CIDR rules, hosts in more than one zone are marked, and the new
ai.assets.view and ai.assets.manage permissions are admins only by default
- Asset annotations: after each collection a model writes each changed asset's role, importance (with who depends on it) and business system in the collector's language, a host and its containers in one call and several hosts per call; only assets whose summary changed and their neighbours are sent, never external addresses, environments or command lines
- The model can suggest
depends_on relations, kept only when they cite a connection, a container port or a runs_on the collector saw
- Edit role, importance and system yourself to lock them against later runs, and unlock them to hand them back; the table shows each with an AI or Locked mark
- A Graph view of the inventory, grouped by zone and sized by importance, with containers folded into their host and edge type filters; a node's panel shows its annotations, facts and edges
- Impact analysis: what fails with an asset, through
runs_on and depends_on and, weaker, connects_to, with the path to each and counts by importance, highlighted on the graph
- Annotation runs are recorded with their tokens, counted as Asset annotation usage, stop at their own token budget and when auto-ops is paused, and use a provider and model you can pick for assets, by default the collector's auto-ops choice
- Internet entries in the Assets view: a read-only list, for hosts you manage, of services that may be reachable from the internet, found from listening ports, Docker published ports, port forwards and reverse proxies or tunnels. Each entry shows its evidence path and a review priority with the reasons, nothing is probed from outside
1.1.0
Added
- One AI assistant with two views that share their conversations: a floating panel on every page, opened with Ctrl+J or from the AI tab, and the full-screen AI tab
- The assistant manages Termix through the Termix CLI, which ships inside the plugin and runs as the signed-in user, and can open terminals, tabs, settings and sidebar views
- With the new
ai.ops.use permission (admins only by default), the assistant runs shell commands on target hosts in parallel, with live results per host; pick targets by fleet, folder, tag or host, type @ to add them, or let the assistant choose
- Every command on a host and every Termix CLI call is recorded and written to the audit log, with secret arguments masked
- Admin settings for the assistant's parallel hosts, command timeout and output kept per host
- Pick up past conversations from Conversations, with the tokens each has used
- The assistant shows its replies in the order they happened, text and tool calls interleaved, with the model's thinking in a collapsed Thinking block of its own
- A context meter in the assistant's header, with the window read from the provider:
max_model_len and the other fields of OpenAI-compatible model lists (vLLM, OpenRouter, Groq and more), Ollama's num_ctx and trained length, Gemini's and Anthropic's model limits; a guess from the model name only when the provider says nothing, marked with ~
- A context window override per model in the provider settings, which list each model's window and where it came from, with Refresh to ask the provider again
- The assistant loads Ollama models with the conversation's window as
num_ctx, and uses prompt caching on Anthropic
- Automatic compaction: as a conversation nears what the context window leaves after the reply's tokens, older tool results are trimmed and then the earlier part is summarised, marked with "Context compacted"; a header button compacts on request
- Token usage for every AI call, with a usage view in the assistant and in settings, a dashboard card, and every user's usage for those with the new
ai.usage.view_all permission (admins only by default)
- Auto-ops: the
ai.tasks service lets an automation start the assistant unattended, to investigate read-only and report, or to propose a plan per host that runs verbatim only after the owner or someone with ai.ops.use approves it; plans expire after 24 hours
- An Auto-ops view in the AI tab with tasks and their reports, plans waiting for approval with live progress, and presets for a daily patrol, disk space cleanup and host down triage that become automations in one step; save your own presets, privately or shared
- More auto-ops presets: service failure and container anomaly (propose a single restart or start for approval), CPU and memory overload, weekly security updates and weekly certificate expiry (report only); the enable form lets you pick the container event
- Auto-ops guardrails under Settings, AI Assistant: Pause AI auto-ops, AI tasks per automation per hour, a token budget per task and a daily token budget per user; duplicate triggers for a host with an open task are skipped
- Alerts when an auto-ops report is ready, a plan waits for approval, a plan has run or a task failed, linking to the AI tab
- The read-only command list also allows
docker system df, find without writing actions, ping -c, nc -z, qm and pct status and list, timedatectl, openssl x509 -noout, update checks and reboot checks
- The read-only command list also allows
docker inspect with a --format limited to a container's state, memory limits and restart policy, dnf and yum updateinfo summary and list, needs-restarting -r, and reading /var/run/reboot-required.pkgs
- The read-only command list also allows
kubectl and k3s kubectl get, describe, top, logs --tail, events, version, cluster-info, api-resources and auth can-i, without secrets and with full objects only for kinds that hold no pod template or data; crictl ps, pods, images, stats, version and logs --tail; nvidia-smi queries and topo -m; zpool status, list and iostat with a count, zfs list and zfs get of native properties; pvesm status and list, pvecm status and nodes, pveversion, pvesh get of listed status, task, storage and log paths; and systemctl list-timers
- The daily patrol and overload presets (version 2) no longer take an LXC container's load for its own when lxcfs passes through the hypervisor's, and
systemd-detect-virt is on the read-only list
- With Termix 26.10.2+ auto-ops tasks can read Termix through the Termix CLI as the task's owner, limited to read-only commands, with a five-minute session per call that is never stored; plans still hold only shell commands
- With Termix 26.10.2+ auto-ops reports are written in the owner's Termix language, falling back to the language the AI tab saved, then English; the plugin now asks for
users:read for this, which an admin approves on update
- Assistant runs go on on the server: closing the panel, switching pages, reloading, closing the browser or losing the connection no longer stops them, and coming back shows the conversation with the run continuing live; only Stop ends a run early
- Manage models at the bottom of the assistant's provider list opens the provider settings
- With Termix 26.10.2+ the assistant docked on the right moves the page over instead of covering it
Changed
- The plugin name and description show in your language
- AI in the sidebar now opens the assistant full screen, replacing the old chat; the terminal assistant still proposes changes for you to approve
- Without
ai.ops.use, the assistant has no target hosts and cannot run commands on hosts through the CLI, including exec, fleets exec, snippets run and execute routes
- The assistant has no step limit any more and runs until it is done or stopped; it only stops itself when it repeats the same call with the same result five times
- A tool result sent back to the model is capped at 12,000 characters, keeping its start and end
Fixed
- Change proposals showed a host id instead of the host name
- Anthropic providers failed every chat with "signal: Extra inputs are not permitted"
- Thinking from reasoning models, sent as
reasoning_content, <think> tags, Gemini thought parts or Ollama's thinking field, no longer runs into the answer
- Read-only commands no longer allow
journalctl --vacuum-* and --rotate, ip changes such as ip link set, date -s or setting the hostname
- Read-only
docker logs needs --tail (up to 1000 lines) and cannot follow, and docker stats needs --no-stream, so neither runs until the command times out
1.0.1
Changed
- Translated into every supported language
1.0.0
Added
- First release
- Ask about your hosts, snippets, fleets, alerts, automations and more, or about one host right from its terminal
- Proposes changes instead of making them, and nothing runs until you approve it
- Can run safe read-only commands like
df and uptime if you allow it
- Works with Ollama, Anthropic, OpenAI, Google Gemini and any OpenAI compatible endpoint
- AI Agents: run Pi, OpenCode, Claude Code or Codex on an SSH host using one of your Termix providers, with tool approvals, queued messages, attachments, Git diffs and worktrees
- One-click install of the selected agent and Node.js, and a button to set up SSH loopback forwarding