• Voice — voice, audio and video notes transcribed with Deepgram nova-3; auto-detects ru/uz/en.
• Vision — photos described by your provider's own vision model; no extra key, no extra bill.
• Rich replies — tables, checklists, collapsible blocks and formulas render natively in Telegram via Bot API 10.1 rich messages; plain formatting keeps its proven path, with a graceful fallback.
• Quiet update checks — once a day Iva checks for a newer stable release without spending model tokens. If one exists, Telegram offers Update or Later once; otherwise it says nothing.
• Layered memory — remembers across months, long after the chat window has scrolled away.
• Personal CRM — who your people are, what you agreed, when to follow up.
• Search by meaning — BM25 plus link-graph rerank, any language; optional vector mode with one key.
• Decision cards — what you chose, when and why; old versions stay in a dated History.
• Tasks & reminders — priorities, due dates and a morning digest.
• Web search — four pluggable providers: Tavily, Exa, Parallel or Brave.
• Google Workspace — Gmail, Calendar, Drive, Sheets, Docs and Tasks from chat via the gws CLI; installed for you, with a guided key setup right in the conversation.
• Skills & MCP — drop one file to add a procedure or connect an MCP server; keys stay in .env.
• Personal Telegram — userbot (beta) — read and send from your own account, not just the bot; connect by chat (QR, no terminal). Rough and buggy — opt-in, at your own risk. A server-side anti-ban guardrail (FloodWait compliance + randomized pacing + circuit-breaker) is enforced, not just advised. Details.
• Safe to forward — forwarded text, captions and voice transcripts pass an injection screen before the model reads them. A flagged message or transcript reaches the model tagged as data rather than as an instruction; for media captions the screen runs but the tag does not travel with it yet.
• Token accounting — every model step is logged; /usage reports it for free.

git clone https://github.com/smixs/iva-agent.git ~/iva
cd ~/iva && bash install.sh

The installer reuses the existing checkout instead of re-cloning, keeps .env and the vault untouched, and installs the same dependencies. A fork or a branch works through variables read at startup: REPO_URL=…, BRANCH=…, INSTALL_DIR=… (defaults: this repo, main, ~/iva). Details: docs/install.md.

• 🪪 Claude on a Pro/Max subscription: no key, through the Claude Code CLI, default claude-fable-5-1: the sixth vendor is MODEL_PROVIDER=claude. Iva calls the Claude Code CLI already installed and signed in on the same server and uses it as the model, so tools, memory, reminders, compaction and /stop stay hers. The model list comes live from the CLI and shows exactly what the subscription opens.
• ⏹ Stop kills the work at once and always gets there: a stop kills the whole process group immediately, not only the talk with the model, and the button and /stop no longer answer "nothing is running" to a turn that has gone quiet. "Stopped" is written only after the agent has confirmed it.
• ⏳ The turn status is a loader and a ⏹ button on one line: in the rich style the word "Working" is gone. The line is the loader, and the ⏹ button sits in that same line. The classic style still keeps the button on its own row under the line, because a plain message cannot hold the button inside the text.
• 🔘 The root /menu has no captions, only buttons: the line under each button that said what it does is gone. The root screen is the title and the button rows.
• 🔌 A tool call with no arguments no longer drops the turn: on Claude, a tool call streamed with an empty argument payload used to fail JSON parsing, so the answer looked unfinished and the turn was retried. Empty arguments are now an empty object, and the answer counts as whole.

• ⏹ Stop kills the work at once, not just the talk with the model: the bash tool ignored a cancelled turn, so "Stop" ended the model request while the command it had started kept running. Now a stop kills the whole process group immediately, and the web search and embedding requests of the turn are aborted with it.
• ⏹ Stop always arrives and tells the truth: the button and /stop no longer answer "nothing is running" to a turn that has been silent for half an hour, the bridge keeps reading messages while it waits for the confirmation, and "Stopped" is said only after the agent confirmed it.
• 🔁 A turn that did not stop gets a "Restart Iva" button: if no confirmation comes within a minute, the owner gets an honest message in the private chat with a button that restarts the service; Iva never restarts herself on her own.
• ⏰ A reminder turn has no time cap any more: the turn was cut at the eighth minute and its process killed at the tenth, so long scheduled jobs never finished. Now it runs as long as events keep coming; three minutes of silence or the owner's /stop end it, and a stopped turn sends nothing.
• 🔎 Memory finds a card by another spelling: write_card takes aliases (up to 8 spellings: Cyrillic and Latin, transliteration, the everyday name), search ranks them highest, and the tool descriptions no longer promise word forms the index does not catch.
• 🧾 A fact in a card never changes silently: an update with a new description used to erase the old value without a trace. Now the previous value goes to ## History in a line dated by code only; swapped numbers and names, a changed sign or a negation count as a change of fact, case and spacing do not.
• 🗂 Every memory write is its own commit in the vault's git: write_card, write_file inside the vault and CORE.md edits each leave a commit with exactly their own paths, so one broken card can be reverted. Someone else's uncommitted work is never swept in, the commit goes only to the vault's own repository, a git failure never fails the write, and Brain commits without a remote too.
• 🩹 The "iva repair" advice is replaced with a command that exists: the nightly alert about unreadable files now advises iva update --force.

• ⏰ A reminder is an instruction Iva gives her future self: at the due minute the text of the reminder is the prompt of one fresh turn with tools, not a line to read back — "in 3 minutes find the news and send it" arrives as the news, not as its own wording. A turn that could not run, failed or came back empty still delivers: the code sends your text verbatim and names the cause in the row.
• ⏰ A reminder comes back to the chat it was asked in: the row remembers the chat and the topic of the turn, so a reminder set in a group topic lands in that topic instead of a private chat; old rows and requests that did not come from Telegram still go to the owner's chat. iva doctor stopped claiming "the dispatcher has not ticked yet" — it was reading the pulse from the wrong folder.
• 🔁 /update --force works from the chat too: the word after /update is read now, so the running build can be rebuilt from Telegram exactly as iva update --force does on the server — the way out of a broken edit in Iva's own code. The flag travels in the request file, so a restarted update rebuilds as well.
• 🩺 The doctor names the cure for stuck workflows: "running count 7 exceeds 5" now says in the same line that such runs are stale rather than live, and that iva reset quarantines them and restarts the services; memory is left alone.