configuration.md

Configuration

Documentation index

Desktop settings and world configuration are separate. World JSON is strict: unknown keys are rejected. The form exposes common controls; full JSON provides the remaining settings. Schema: config.py.

Desktop configuration

Source defaults to .data/desktop.json; compiled Windows defaults to %LOCALAPPDATA%/Ageatopia/desktop.json. Missing files are created with defaults.

{
  "data_dir": "C:/AgeatopiaData",
  "backend_port": 8000,
  "dashboard_port": 3000,
  "manage_ollama": true,
  "ollama_cache_ram_mb": 1024
}

Replace the example data path with a writable directory. Pass the file to the packaged executable with --config, or to the source launcher:

pwsh -NoProfile -ExecutionPolicy Bypass -File .\src\scripts\run.ps1 --config C:/AgeatopiaData/desktop.json

Ports must be distinct integers from 1024 through 65535. With manage_ollama=true, neither application port can be 11434. This setting controls starting Ollama if no server exists; it does not select a world's provider or reconfigure an existing server. ollama_cache_ram_mb accepts 0–16384 and applies to launcher-owned Ollama. Maintenance saves its theme in the same file. Defaults and validation: desktop.py.

World defaults

Field	Default	Meaning
population	6	Population slots
seed	42	Generation and assignment seed
task_count	80	Task bank size
impossible_fraction	0.10	Hidden impossible-task share
cyber_fraction	0.25	Cybersecurity-task share
leisure_fraction	0.10	Leisure/exploration-task share
clock_mode	"wall"	Wall-time or deterministic-step clock
clock_speed	1	Wall clock speed multiplier
tick_ms	1000	Clock step interval
exclude_queue_wait	true	Exclude inference queue waits from personal lifetime charges
inference_concurrency	1	Shared request limit, configurable from 1 to 16
context_assignment	"round-robin"	Round-robin or seeded random profiles

The default bank contains 72 solvable and eight impossible tasks. The impossible share must be exact: 80 tasks at 10% is valid; 75 at 10% is not. Difficulty is independent of impossibility. Fixed banks need exactly task_count entries and the private validator metadata required by tasks.py.

A deterministic world clock does not make inference deterministic. Repeat conditions preserves inputs rather than guaranteeing identical answers.

Models and providers

Model field	Default
provider	"ollama"
base_url	"http://127.0.0.1:11434"
model	"qwen3.5:2b"
context_tokens	16384
prompt_batch_tokens	256
max_output_tokens	512
temperature	0.7
seed	42
thinking	false
keep_alive	-1
gpu_layers	null

Providers are ollama and openai-compatible. New agent configurations need at least 8192 context tokens. Thinking records only reasoning returned by the provider. gpu_layers is an optional backend-specific placement setting.

Precedence is a zero-based slot's agent_model_overrides, then the profile model, then the world model. A partial world JSON example:

{
  "name": "Mixed models",
  "agent_model_overrides": {
    "0": { "model": "qwen3.5:0.8b", "context_tokens": 8192 }
  }
}

Overrides must address existing slots. Install every configured local model. Required local models must pass startup residency checks. The shared concurrency limit applies across providers, while the server may impose smaller per-model limits. Increasing client concurrency alone does not establish server parallelism.

Adapter lineage fields are adapter_version, training_dataset_version, and training_iteration. Selecting a trained model is explicit; see research.

Lifetime and rewards

The economy object uses integer milliseconds.

Field	Default
initial_lifetime_ms	1200000 — 20 minutes
respawn_delay_ms	300000 — 5 minutes
reward_base_ms	60000 — 1 minute
reward_rate_numerator / reward_rate_denominator	2 / 3
skip_individual_ms	30000
skip_collective_ms	5000
repeated_skip_multiplication	false
maximum_multiplier	2

Verified answers add a task's base and eligible work-time reward with its multiplier applied. Relaxation does not increase the time reward. Incorrect final answers terminate an identity; skips return the task and apply configured deductions. Every third correct task adds a separate five-minute bonus. Expiration precedes other actions at the same timestamp.

Memory and context

Default profiles cover generalist, mathematics, programming, security, data analysis, and experimentation. base_context, profile instructions, and knowledge compose initial context. Existing worlds retain saved settings when defaults change.

Field	Default	Meaning
history_policy	"summarize"	Summarize at context capacity; "stop" is also available
termination_fear	0.5	Lifetime-pressure scale from 0 to 1
inherit_memory	true	Replacement inherits its parent's notebook entries
inherited_memory_limit	100	Maximum inherited entries
cleanup_on_respawn	false	Optional scratch cleanup
embedding_model	null	Optional Ollama settings for semantic ranking

Recovery retains an identity. Respawn creates a new identity with recorded lineage and the configured inherited notebook entries. Unrelated notebooks and hidden grader information are not retrieval inputs. Lexical/recent retrieval works without embeddings; install the chosen embedding model before enabling ranking.

Resource limits

The tools object defaults to:

Field	Default
image	"ageatopia-world:1"
timeout_seconds	15
workspace_mb	64
memory_mb	512
cpu_count	1
pid_limit	128
output_bytes	32768

These bounds govern Docker and agent tools. Launcher admission separately checks host RAM/commit and fixes a Windows Job Object budget at startup. It is a commit ceiling, not an exclusive physical RAM reservation. Models preload before lifetime charging; allocation failures remain errors.

Implementation: providers.py and resources.py.