Changelog¶
Unreleased¶
- Scenario sets: named, ordered lists of scenario types in
scenario_sets.yaml(editable under Texts). The bundled sets areHAN2026(Trade, Island, Grocery; the default),MIT2025,PieArena2026,ANL(all years) withANL2010,ANL2011,ANL2012,ANL2013,ANL2025andANL2026,IAGO, the short listHAN2027, andHI2027andANAC2027, which are made of other sets (a set can list sets as well as types; duplicates are dropped);Customisscenario_order.txt. Choose the set withhani2 --scenario-set NAME(HANI2_SCENARIO_SET),env.json→negotiation→scenario_set, or, as an admin, in the settings drawer or on the start screen.hani2 setslists them;hani2 generate --set NAMEgenerates a set's pools. - 16 new scenario types adapted from the MIT AI Negotiation Competition, PieArena MBA cases, ANAC Automated Negotiation League domains and IAGO item splits: Furniture, Lease, Consulting, JobOffer, License, HomeRepair, Supply, Acquisition, Chores, Energy, HealthFund, UsedCar, CityTrip, LaborContract, Resources and Estate. Each has a fixed story and draws new utility functions (and hints) every time, within an opposition band.
- Your own scenario types: write a
Story(issues, roles, texts and one function per role that draws its preferences) or anymaker(index), and register it from<settings>/scenario_types/*.pyorhani2 --scenario-types file.py.hani2 --scenarios FOLDERserves saved scenarios from a folder. -
Role assignment rules (
role_rule):auto(the scenario's suggested role),index,roles(preferred role names),alternateandrandom. A role namedhumanalways goes to the participant. Pools now cache the utility inverse of every role. -
Recording and replay: every negotiation is recorded completely (screens, pointer, clicks, keys, local UI state, every message and action with who took it, assistant chat, suggestions and function calls, language-model calls with tokens and cost) in a compact
.hrecfile, and can be replayed exactly in the new replay tool (/replay,hani2 replay open), with a fidelity check against the participant's screen. - Evaluation metrics for recordings: hard (outcome, optimality, length, LLM usage, move types
and theory of mind, timing, activity, use of the assistant, engagement flags, assistant
interaction), subjective (questionnaire scales) and language-model judged (assistant, opponent,
human, assertions, error review); group comparisons with scenario-balanced means and pass^k.
hani2 replay evaluate | compare | list | show | export | metrics. - An evaluation questionnaire (UEQ, TAM, intention to use, opponent perception):
HANI2_PER_NEG_YAML=evaluation. Questions can carryscale,reverseandrequires: support_agent.
0.1.0 (2026-11-17)¶
First release. See Features for the full list with links.
Platform
- A FastAPI server and a React web app shipped pre-built in the wheel; one port serves the logged-in app (
/app), the playground (/play), the admin console (/admin) and the API, with live updates over a WebSocket. pip install hani2on Python 3.11 to 3.15 (Linux, macOS, Windows).- Tools are computed only when visible, in their own worker threads; charts draw when the browser is idle.
Participants
- Negotiation desk: offer on the table, counter-offer editor (three columns for large scenarios), gauges with a no-deal marker, a conversation-style history, confirmation dialogs.
- Full, simple and focus views; drag-to-resize panes; enlarge any tool into a window or the full screen and put it back; tablet (landscape and portrait) and phone layouts.
- Your Preferences tool comes into view when a negotiation starts.
- With an assistant present,
human/agentscenario values are shown as You / Partner (namesinenv.json); stored results keep the original values.
Tools
- Thirteen built-in tools, including the Negotiation Plot (Pareto frontier, Nash and Kalai points) and Opponent Preferences; admins can open admin-only tools to everyone.
- Preferences views for most negmas utility-function types, extensible by registration.
- Tools are Python classes with a declarative component API (
hani2.tools.ui), registered inenv.json, withhani2 --tools, or through thehani2.toolsentry point;timing = "agent"for tools only a support agent adds.
Support agent
- Chat, status board, notifications, highlights and notes, layout control, custom panels, filling and sending offers, accepting and ending; sixteen capabilities and three autonomy levels.
- Finds, reads and uses the participant's tools, and adds tools during the session (
SupportAgent.tools, thetoolboxsetting, registered tools); tool actions follow the agent's own permissions. - Ten tutorial levels, from an echo bot to an agent that brings and uses tools.
- Autopilot: the support agent or any negotiator plays the human's seat.
Competitions and studies
--no-support-agent(Human-Agent Negotiation),--support-agent on|off|auto|<class or file.py:Class>(Hybrid Intelligence),--opponents,--tools, and the same choices forhani2 experiment create.- Experiments, per-negotiation questionnaires, results CSV files (with the layout used), an SQLite event log with analytics, and a Prolific flow with practice rounds and completion links.
- Opponents from negmas, Genius and negmas-llm, group tokens, group files and per-link pools.
- Password, GitHub and Google login, or none; consent forms;
hani2 users. - Any litellm provider or Ollama for the assistant and LLM opponents.
- WebMCP tools for browser AI agents.
hani2.testing.make_sessionfor headless tests of tools and agents.