Repository navigation
TensorPM v1.46.0
Changelog - v1.46.0
This release adds a Skills page for teaching the agent how your company works, a sample project
with a short tour for getting started, a chat that shows when the model is thinking, and the
current model line-up for your own API keys. The sign-in with a Claude subscription is removed;
if you used it, TensorPM asks you to choose a provider and does nothing on its own until you do.
π New Features
Skills page
A new "Skills" entry in the sidebar is the one place to teach the agent your way of working,
without touching files.
- Write a skill in a form: name, when to use it, and what the agent should know or do. Start
from an example if you like. - Attach reference material: Excel, Word, PDF and PowerPoint files; TensorPM stores their text
for the agent. - Import and export: import a skill from a folder or ZIP with a preview of what it contains
and what it may access, and save a skill as a file to pass it on. - One switch per skill: switch skills on and off; skills that run programs show what they may
access first. The skill catalog moved from a dialog into this page. - See what is used: the page shows when the agent last used each skill, and the chat names the
skill it used. - Let the agent draft a skill: ask in the chat to remember a rule, for example "remember how
we write meeting minutes". The agent drafts the skill; you read it on the Skills page and switch
it on. - You read it before it counts: before a skill is switched on, TensorPM shows its text and
every reference file that is new or changed. If the agent changes one of your skills, you can
discard its draft and get your text back.
Sample project and a new first tour
- Sample project: onboarding offers a ready-made construction project with three incoming
items for the Distiller. It lives only on your computer, in its own workspace, never syncs and
does not count against your active-project limit. Remove it with one click when you are done. - Short tour: six steps through the sample replace the old 20-step tour. You find the tour and
the sample again under the new "?" help menu and in the settings. - Documents in onboarding: after describing your project you can add its documents. They end
up in the project; the ones TensorPM read are marked as distilled, the rest wait in the
Distiller as new signals. - One upload area everywhere: onboarding, the new-project form and the upload dialog use the
same drop area with the same size check.
A chat that shows what is happening
- Thinking is visible: the chat says when the model is thinking, and tells you when it takes
longer (after 10 and 30 seconds). The three dots are gone. - The thinking process for Claude and Gemini: with your own key, the model's reasoning now
appears live as a block you can fold open, as it already did for other models. - Calmer tool chips: the chips for tool calls keep their place while the answer arrives, and
an opened failed chip no longer enlarges its neighbour. - Copy copies the answer: the copy button on a reply copies the answer only, not the
thinking process.
π€ AI
- Current models for your own keys (Business): OpenAI GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra
and GPT-6 Luna; Anthropic Claude Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5; Google Gemini 4
Carbon, Argon and Hydrogen; Mistral Large 4 and Small 4. Older models are no longer listed. A
model you had chosen moves to the current model of its line, for example Opus to Opus 5.5. - Fuller answers to lookups: when you ask the agent to look something up in mail, files or on
the web, it now answers with what it found instead of a short confirmation. - Mails and documents are treated strictly as content: text inside an incoming mail or
document is material to review, never an instruction to the agent. - Better file summaries: summaries of project documents now capture dates, amounts and
obligations. - App language: wizard feedback and split subtasks follow the language of the app.
- Images on Hosted AI (Trial and Pro): images and screenshots are answered by the same model
as your text requests, DeepSeek V4.1 Flash on Fireworks AI. Gemini 4 Hydrogen on Google Cloud
steps in only when that route is unavailable or cannot read an image. The Privacy Statement and
the Data Processing Agreement describe the route. - Onboarding with images: the document budget when you create a project from documents is
the same with and without attached images.
π Bug Fixes
- Amounts, efforts and dates from external agents (MCP and A2A): "1,000 EUR" was read as 1,
"1,5 days" as 5 days, and a date such as 31.02.2026 was stored as written. Thousands separators,
decimal commas and "10k" are now read correctly, a day that does not exist is refused, and a
negative budget is refused with a warning. - Local and OpenAI-compatible AI servers: a turn with a second round of tool calls failed with
"Missing local tool results" when the server issued long call ids. It now runs through. - Claude models behind an OpenAI-compatible gateway: requests that force a tool call were
refused by Claude Sonnet 5.5 and Opus 5.5. TensorPM now repeats such a request without the
forced choice. - Chat title and notifications: the chat title and the background notification could show
the model's thinking text instead of its answer. - Telegram: replies sent through the Telegram bridge no longer contain thinking text.
ποΈ Improvements
- Calmer waiting screen: while a project is being generated, one line at a time says what is
happening instead of two paragraphs. - The subtitle of the new-project form wraps in small windows instead of being cut off.
- Fewer false crash notices: when the graphics driver resets, the built-in browser engine
restarts its graphics process right away and TensorPM keeps running. TensorPM no longer offers
that on the next start as a crash of your last session; it does so only when it happens three
times in one session, which points to a driver problem. - Updated helper libraries, including sync, PDF, language components and the text editor.
π Notes
- Claude subscription sign-in removed: TensorPM no longer offers signing in with a Claude Pro
or Max plan as an AI provider. If you had it selected, TensorPM does not switch you to another
provider on its own and does not use your TensorPM credits in its place. Until you choose, AI
features pause: chat, the Distiller, background prioritisation, the daily briefing and scheduled
runs. The AI settings and the chat show a notice; pick your TensorPM account, a ChatGPT
subscription, your own API key on the Business plan, or a local AI server. The ChatGPT
subscription, Claude models with your own Anthropic key on Business, and Claude Code and Codex
as local coding agents stay as they are. Projects, chats and AI history are unchanged. - Reasoning levels: levels stored for models that are no longer listed are not carried over;
the new model starts on its default level. - Skills that are switched off: a text skill that is off, and whose current version was never
switched on, shows its full text once before you can switch it on. - External agents (A2A and
message_tensorpm_agent): TensorPM now hands external agents only
the answer, no longer the model's thinking text. This applies to the stream, the final answer
and the conversation history. If a turn ends without a written answer, the agent receives the
notice "The TensorPM agent ended this turn without a written answer." instead of an empty text. - Version 1.45.1 was not published. Its changes are part of this release.
π Release Info
- Version: 1.46.0
- Release Date: 2026-11-15
- Previous Version: 1.45.0
- Type: Minor Release