openLight v0.2.0

v0.2.0 widens the safe surface around the deterministic-first
runtime. The Telegram bot picks up an inline-button home menu and
multi-step skill forms. Voice notes, images, and screenshots become
first-class inputs. The optional layer grows to cover vision, OCR,
network probes, durable memory, visual watches, MCP servers, and
user-defined external skills. Routing splits into FAST + SMART LLM
profiles so a Raspberry Pi or Mac mini can keep classification cheap
while reasoning calls hit a bigger local model only when needed.

Nothing about the project's posture changes. openLight is still a
small, deterministic, local-first infrastructure tool — the LLM is
still optional, still allowlist-bound, and never expands the surface
on its own.

Good fit for

Raspberry Pi, Mac mini, and homelab operators who want one persistent
Telegram bot for status, logs, and safe restarts
Self-hosters who prefer local Ollama by default and want to layer
vision / OCR / browser as opt-in
Operators who want to drop a skill.yaml next to a Python or Bash
script and have it appear in /skills without rebuilding the binary

Highlights

Single openlight binary, four subcommands. agent, cli,
doctor, and a new skills (list / validate / reload) for
inspecting builtin + external skills.
Deterministic-first router, now with a generic registry stage.
Slash → explicit text → normalized shortcut → registry alias →
registry generic → semantic rules → optional LLM classifier. The
generic stage lets external skills with arguments resolve without
any LLM call.
FAST + SMART LLM profiles, with always-warm policy. The router
classifier runs on FAST; chat, log explanations, and final answers
run on SMART. Profiles listed in llm.warmup.profiles are loaded in
the background at startup with exponential backoff so users don't
wait on cold model loads.
Telegram UI. Inline home menu, per-group submenus, multi-step
skill input sessions (mutating skills now ask field by field), and a
callback router for action buttons. Watch alerts route through the
same UI layer.
Voice notes. Telegram voice / audio messages are downloaded,
resampled via ffmpeg, transcribed with whisper-cli, and fed back
into the regular routing pipeline. Optional reply with transcript.
Image inbox. Photos and image documents are dispatched to
vision_analyze (default) or ocr_extract (when the caption looks
like an OCR ask), with the result returned through the normal reply
path.
Vision and OCR skills. vision_analyze / vision_compare and
ocr_extract ship as optional groups with configurable providers
(Ollama VLM, OpenAI, Tesseract).
Network skills and watches. port_check, http_check,
cert_check, dns_check, plus two new watch kinds (port_down,
cert_expiring_soon) gated by network.enabled.
Visual watches. visual_watch_add / _list / _test /
_remove schedules periodic screenshot diffs with baseline storage,
threshold, cooldown, and optional keyword detection on OCR / HTML
text.
Durable memory. /remember, /memories, /forget skills
persist tagged facts in SQLite. Memory can live in its own SQLite
file (memory.db_path) if you want it separated from message
history.
MCP integration. Configured Model Context Protocol servers are
started as stdio JSON-RPC subprocesses; their tools register as
skills in a dedicated mcp group, routed exactly like builtins.
External skills. Drop a skill.yaml plus an executable into
external_skills.roots; the runtime spawns the process per
invocation, writes one line of JSON to stdin, reads one line back
from stdout. Builtins always win on duplicate names. See
docs/skills/EXTERNAL.md.
More watch packs. /enable tls, /enable homelab,
/enable mac, /enable pi join docker / system / auto-heal.
Mac mini support. Darwin-specific system provider (CPU, memory,
swap, memory pressure, uptime), launchd plist, deploy-macmini
scripts, brew-friendly dependency installers.
Storage retention. storage.retention_days prunes old
messages and skill_calls rows on startup; new indexes keep the
prune cheap on Pi-class hardware.
Doctor expanded. New probes for Telegram, Ollama, nodes,
filesystem, watches, voice, browser, vision, OCR, workbench, visual
watch, and a security-warnings pass that flags obvious foot-guns.

Routing pipeline, in one paragraph

router.Router.Route runs slash commands, explicit text (service tailscale), normalized shortcuts (статус → status), registry
aliases (exact name or alias), and registry generic (first word
resolves to a skill, remainder becomes the text argument) before
trying semantic rules. Only after every deterministic stage misses
does it consult the LLM classifier — and that classifier picks first a
group and then a skill inside that group, both from closed
lists. Confidence at or above execute_threshold (default 0.80)
runs; between clarify_threshold (default 0.60) and that, the
router asks the user to confirm; below clarify_threshold is no
match. The LLM never expands the surface.

Try it

export TELEGRAM_BOT_TOKEN=123456:replace-me
export ALLOWED_USER_IDS=111111111
export OPENLIGHT_REF=v0.2.0
curl -fsSL https://raw.githubusercontent.com/evgenii-engineer/openLight/v0.2.0/scripts/install.sh | bash

Once it starts, try:

/start
/enable system
/watch add service tailscale ask for 30s cooldown 10m
/remember Synapse listens on 8008 inside docker

For deterministic-only (no LLM at all), set LLM_ENABLED=false
before running the installer.

Migration notes from v0.1.0

You can stay on a v0.1.0-shaped config and pick up v0.2.0 features
incrementally. Nothing in the existing minimal config has been
removed.

Single binary command. Older flows that called
cmd/agent/main.go or cmd/cli/main.go now use
openlight agent / openlight cli. The Dockerfile, systemd unit,
and launchd plist already point at openlight.
nodes: is the canonical key. access.hosts: is still
accepted and is merged into nodes: at load. The agent and
openlight doctor log a deprecation note when it's in use.
files: replaces filesystem:. Both keys still load; new
configs should use files:.
LLM profiles. The legacy single-model setup keeps working — when
no llm.profiles.fast is defined, the SMART provider serves both
roles (logged as FastFallback). To split the workload, add a
profiles.fast block (qwen 1.5b on Mac mini, gemma 4b on Pi, etc.).
Warmup policy. llm.warmup is enabled by default with
profiles: ["smart"] and keep_alive: -1. Set
llm.warmup.enabled: false if you don't want background loads at
startup.
Storage retention. storage.retention_days defaults to 0
(keep everything, same as before). Setting a positive value triggers
a one-shot prune on startup. New SQLite indexes are applied
automatically via migration 0005.
Memory is enabled by default. memory.enabled defaults to
true with a 20-row list limit, sharing the main SQLite file. Set
memory.db_path to split it into its own file. No data is migrated
out of notes — memories and notes are separate stores.
Optional modules stay off. vision, ocr, voice,
visual_watch, network, mcp, workbench, accounts,
browser, and external_skills are all gated by their own
enabled flag (or non-empty servers / roots / providers map).
Doing nothing keeps the v0.1.0 surface area.
Watch packs are idempotent. Re-running /enable docker after
upgrading updates existing watches in place; it does not duplicate
them.

Known limitations

Metric watches still support notify only. Service-down watches
are the ones that take ask / auto and offer action buttons.
Running inside Docker still does not grant host-level file or
service access automatically. Mount what you need, or reach hosts
via the nodes: block.
The Telegram UI sessions and inline buttons are Telegram-only. The
CLI keeps the same skill surface but renders results as text.
mac watch pack omits temperature on Darwin (powermetrics /
SMC requires sudo). The pi pack includes it.
The project is still v0.x and intentionally narrow in scope.

Navigation Menu

v0.2.0

Choose a tag to compare

Sorry, something went wrong.