Skip to content

Admin console

The admin console at /admin lets administrators watch live sessions, read results and event logs, manage users, and edit the settings files from the browser. Administrators also get extra controls on the negotiation page itself, described at the end of this page.

The admin console

Access

Open http://localhost:5006/admin, or choose Administration in the account menu of the negotiation page. If you are not logged in, the console sends you to the login page and back. A logged-in user who is not an administrator sees "This page is for administrators."

A user is an administrator when any of these is true:

  • the username is admin;
  • the account has the role admin (hani2 users add --admin, the Admin box when adding a user, or Make admin in the Users section);
  • the account's email is listed in admin.emails in env.json (useful for OAuth sign-in).

See Authentication. The ?admin=1 query parameter on the negotiation page does not open the console; see the interface.

The console has a navigation column on the left (a scrolling bar at the top on narrow screens), a theme button and a Back to negotiating link. The current section is kept in the URL fragment, so /admin#results opens the results directly. Section fragments are overview, results, analytics, users, server, llm, agent, autopilot and texts.

How saving works

The settings sections (Server and studies, Language model, Support agent, Autopilot, Texts) load the current effective settings, let you edit them, and save with the Save button at the bottom. "Unsaved changes" appears next to it while you have edits.

Saving writes the file in the settings folder ($HANI2_HOME/settings, default ~/negmas/hani2/settings). It always writes there, even while an experiment is running. If the active experiment has its own copy of the same file, that copy takes precedence and your saved change has no visible effect until you edit the experiment's copy or remove it. See Configuration reference for the lookup order.

Saved settings apply to sessions created after saving. Sessions already open keep their settings.

Live sessions

A table of every session the server holds in memory, refreshed every five seconds. Above it, four figures: open sessions, sessions negotiating now, sessions recorded in the event log, and logged events.

Column Meaning
user The user name (guest_<id> for playground visitors, prolific_<PID> for Prolific participants)
mode main (/app) or guest (/play)
phase idle, ready, starting, running, ended, questionnaire or finished
round Negotiations started in this session
scenario The loaded scenario
partner The opponent type of the latest negotiation
connections Open browser tabs
idle Seconds since the last message from the browser

The table is searchable. The server removes sessions that have no open tab, are not negotiating, and have been idle for more than six hours (it checks every ten minutes).

Results

Every finished negotiation from every participant, newest first, with:

  • a Download CSV button (/api/admin/results.csv, saved as hani2_results.csv);
  • a Leaderboard with user, score, negotiations and mean_utility, when there is anything to rank;
  • Choose columns: a collapsible picker of the columns to show. It starts with the default columns plus user;
  • a searchable table. Utilities and the human advantage are shown as percentages.

Studies and results describes the results files and their columns.

Event analytics

Charts and a filterable log of the events the server records:

  • Events by type: a bar chart of event counts.
  • Activity over time (per hour).
  • A filter row: user, event type, from date, until date, and Apply.
  • A searchable event table (up to 2000 rows) with timestamp, user, event_type, component, action, value, scenario_id and utility_value.
  • Recorded sessions: start and end time, user, mode, experiment and browser of the latest 200 sessions.
  • Export CSV downloads every event (/api/admin/events.csv, saved as hani2_events.csv).

See Studies and results for the event log and its event types.

Users

  • Add user: username, password, optional email, and an Admin box. Users created here are marked as having consented.
  • For every account: the name, an admin badge, the full name, email, OAuth provider and consent status, and the buttons Make admin / Remove admin, Set password, and delete. Deleting asks for confirmation and keeps the user's results.

The admin account cannot be demoted or deleted, and you cannot delete your own account. The same operations are available on the command line with hani2 users (see Command line).

Server and studies

Edits env.json.

Field Key
Title, Subtitle title, subtitle
Sign-in: username and password, OAuth only, password and OAuth, or no sign-in auth.mode (password, oauth, dual, none)
People can create accounts auth.allow_registration
Require consent auth.enforce_consent
Playground playground.enabled
Admin emails admin.emails
OAuth redirect base oauth.redirect_uri
GitHub OAuth, Google OAuth: client id and secret oauth.providers.github.key / .secret, oauth.providers.google.key / .secret

Negotiation defaults:

Field Key
Rounds negotiation.n_steps
Time limit (seconds), empty for none negotiation.time_limit
Opponent agents, comma separated opponents

Under the partner pool, Available groups lists the opponent groups and how many negotiators each contains. See Opponents.

Tools: one row per registered tool with its description, its pane, when it appears, and whether it is admin-only. Untick Offered to add the tool to tools.disabled; untick Admins only to give an admin-only tool to every participant (tools.overrides.<name>.admin_only). Participants see opponent utilities sets tools.show_opponent_utility. Extra tools edits tools.extra (class paths such as my_pkg.tools:MyTool). Enlarged tools open as chooses whether a tool's enlarge button opens it in a large window, fills the screen, or is not offered (tools.maximize: window, fullscreen or off; see The interface). See Tools.

The OAuth rows edit the key and secret only. The enabled flag of a provider is not shown; a provider is used when it has both a key and a secret and is not explicitly disabled.

Language model

Edits llm_settings.json, used by the LLM offer tool, message writing, offer extraction and LLM partners. A banner at the top shows whether the model is configured ("Ready: provider / model") or which API key variable is missing.

Fields: provider (ollama, openai, anthropic, gemini, azure, groq, mistral, openrouter), model, Ollama address (ollama_base_url), API key variable (api_key_env, the name of the environment variable that holds the key), temperature and max tokens. When the settings contain extraction_prompt, generation_prompt or instruction_prompt, each is shown in a text area. See Language models.

Support agent

Edits support_agent_settings.json.

Field Key
Who gets the assistant: everyone, admins only, nobody mode (on, auto, off)
Agent class (module:Class, empty for the built-in agent) agent_class
Provider and model provider, model
API base api_base
API key variable api_key_env
Temperature, max tokens, tool steps temperature, max_tokens, max_tool_iterations
Highest autonomy: Suggest, Semi, Full autonomy
Abilities granted capabilities.<name>
Speaks up unprompted when a negotiation starts, it is the participant's turn, a negotiation ends proactive.on_negotiation_started, proactive.on_action_requested, proactive.on_negotiation_ended
System prompt system_prompt

Participants can narrow these permissions in their assistant settings but never widen them. See Support agent.

The support agent section of the admin console

Autopilot

Edits autopilot_settings.json.

Field Key
Autopilot allowed allowed
Default driver driver
Participants choose the driver: playground only, yes, no user_can_choose (null, true, false)
Drivers offered (none ticked offers all) allowed_drivers
Pause between steps (s) step_delay

Texts

Plain-text editors for three files, each with its own Save button:

  • Welcome text: announcements.md, shown before the first negotiation.
  • Consent form: consent.md, shown on registration and on the consent page.
  • Scenario sets: scenario_sets.yaml, the named lists of scenario types (see Scenario sets).
  • Custom order: scenario_order.txt, one scenario type per line (the Custom set).

See Scenarios for the scenario order.

Admin controls on the negotiation page

An administrator negotiating on /app (or on /play while logged in) gets more than a participant:

  • The Admin badge and the partner's name in the header, and no time limit by default (time_limit starts as none for admins).
  • The Scenario picker before each negotiation: a scenario type and a scenario name, then Load.
  • The Opponent for the next negotiation picker, with suggestions from every opponent group. Leave it empty for a random partner from the pool.
  • In the counter-offer editor: Text only, Offer from message and Write message.
  • Admin-only tools in the tool panes.
  • Extra sections in the settings drawer.

The opponent picker and the partner's name are also available in admin mode (?admin=1) without an admin account. The other controls are not.

The admin settings drawer

The gear button opens the settings drawer. Below the display settings, administrators see four more sections. They change the current session only; most take effect at the next negotiation. "Server-wide defaults are in Administration."

Section Setting Key Default
Timing Rounds n_steps 100
Time limit (s) time_limit 300 (none for admins)
Time per step (s) step_time_limit none
Negotiator time (s) negotiator_time_limit none
End probability per step pend 0
End probability per second pend_per_second 0
Negotiation rules An empty offer ends the negotiation end_on_no_response off
Allow messages without an offer allow_text_only_offers on
Show messages from the AI partner allow_text_agent on
Let the participant write messages allow_text_human on
Text-only mode text_only_mode off
Write a message when none is given (uses the LLM) auto_generate off
Start the editor from my last offer init_with_last on
Otherwise start from my best outcome init_with_best on
Scenarios and partners Use saved scenario pools (off: generate a fresh scenario each time) load_existing on (off in the playground)
Follow the scenario order predefined_order on (off in the playground)
Participants may choose the scenario allow_scenario_selection off
Show the opponent agent's type show_opponent_type off (on for admins)
Opponent agents (comma separated) opponent_types empty
Tools Tools can be switched on and off allow_tool_switch on
Tools can be moved between panes allow_moving_tools on

Three further per-session keys exist but have no control in the drawer: human_index (which negotiator seat the participant takes, default 1), auto_extract (fill the offer from the message with the LLM before sending, default off) and generate_on_failure (generate a scenario when loading one fails, default on).

Every key in this table can also be given a server-wide default under "negotiation" in env.json. The bundled env.json sets n_steps to 100 and time_limit to 300. See Configuration reference.