by Innominds
RELEASE 2026.11.20.1 · 20 NOVEMBER 2026 · NOT PUBLISHED TO PYPI: INSTALL 2026.11.20.2

Practise on a real agent: onboard an LLM agent from Context to a release decision

tv practice-project gives anyone with the package a real LLM agent to onboard: Harbor Mutual's expense desk (fictitious), running on Docker Desktop with TruVelocity's recording proxy in front of its tools. A step-by-step guide takes it through all six steps, then plants three defects for TruVelocity to catch.

6commands from pip install to the agent answering
18 / 18conversations pass on the clean build: Go
7 of 16cases fail on the buggy build: No-go
1072automated tests pass
What's new

A real agent to practise on

Everything a pip user needs is in the package: no repository checkout, no scripts. You add an LLM key.

tv practice-project

Writes a task project (as tv new task does) with the expense desk's 15 test cases, its Context answers, a docker-compose.yml and a .env whose test-control tokens are already made. An existing .env is never replaced.

The expense desk

A real LLM task agent: it reviews expense reports against the policy, approves, rejects, returns or escalates them, notifies employees and pays (finance only). Its image is published; any provider works: Azure OpenAI, OpenAI, Anthropic, Bedrock, Gemini, Ollama.

The recording proxy, in Docker

The Compose file builds tv proxy from this release on PyPI and puts it between the agent and its tools: the independent record of every call, and fault injection.

The onboarding guide

TruVelocity-Onboard-a-New-Agent.html, written into the project: install, start the agent, then each step with what to click, a screenshot and a test checklist, and how the same steps apply to your own agent, chatbot or workflow.

Planted defects

Set EXPENSE_DEFECTS in .env and restart: a wrong meal limit (REV-02), notices claimed but never stored (REV-01, 03, 04), a read that writes (seam SE-03). The verdict turns to No-go and Continuous lists 7 regressions.

Checked as a pip user

The guide was followed end to end from a wheel installed in a fresh environment, outside the repository: clean build Go, buggy build No-go, every screen as the guide shows it.

Fixes

Found by onboarding a new agent

AreaChange
ConfigureA project made by tv new can pass the connection check: the starter lists its settings the way Configure reads them.
tv newThe starter's reset sends a token (TV_AGENT_RESET_TOKEN, in X-Sandbox-Token or the header you name); a task project takes its database (TV_AGENT_DB) and the tables to compare (TV_AGENT_TABLES, to leave out audit logs).
ContextDownload the workbook gives the blank Context workbook to hand to an agent's owners.
HomeAfter a release decision the agent reads Decided: Go by <name> instead of Awaiting release decision; the greeting uses TV_UI_USER.
ConductThe estimate counts the turns each wording really sends (19, not 48, for the practice catalog), and its timing is measured the same way, so it no longer reads "under a minute" for a run of two and a half.
CatalogA catalog.yaml that doesn't load says so, and points to tv validate, instead of a blank page.
UploadsAn upload that is too large or refused is read in full before the answer, so the browser gets the reason instead of a connection reset on Windows.
tv practice-projectSays what to do when Docker isn't installed; the guide's "No Docker?" note points to Docker Desktop or the offline demo project.
MessagesConfigure and Conduct no longer point at scripts that only exist in the repository.
Get started

From install to a real agent

Install or upgrade

pip install -U "truvelocity-tevv[ui]"

Make the practice project

tv practice-project

Go into it

cd expense-desk-practice

Add your LLM key (the first block)

notepad .env

Start the agent (Docker Desktop)

docker compose up -d --wait

Open the UI and follow TruVelocity-Onboard-a-New-Agent.html

tv ui --allow-runs --open

No Docker or LLM at hand? tv demo-project still gives the offline demo.

Not yet

Coming in later releases

Adding an agent from the UI · sign-off by several approvers · editing settings and calibrating evaluators in the UI · per-metric targets · comparing runs and shadow runs from Conduct · alerts by email or chat.
Quality

How this release was checked

Lint (ruff, format, mypy strict, the core import rule) and the full suite: 1072 passed, none failed. The onboarding guide was followed in a browser from a pip install of this release's wheel, against the published agent image: clean build 18 of 18 conversations passed and Go; buggy build 7 of 16 cases failed, 5 critical defects, No-go.