Configuration

WASP has three configuration entry points:

1. .env — startup-time environment variables (Pydantic-validated). See Environment Variables.
2. prime.md — operator override prompt injected at the top of every system prompt. Writable at runtime.
3. config:overrides (Redis) — runtime feature-flag overrides applied at startup.

This page covers (2) and (3).

prime.md

prime.md is mounted at /data/config/prime.md inside the container. It is injected at the top of every system prompt — before WASP's built-in identity block, before memory injection, before everything.

Use it to:

• Set persona / tone preferences
• Define operator-specific rules ("never auto-trade", "always quote sources", etc.)
• Override default behavior in edge cases

Editing

Two ways:

Dashboard: open /config. Left column is the prime.md editor (Ctrl+S to save). Changes take effect on the next message — no restart required.

Filesystem (inside the container):

docker exec -it agent-core nano /data/config/prime.md

The volume core-config is writable at runtime; no rebuild needed for .md content (HTML/Jinja templates and prime.md reload from disk on each use).

prime.default.md

prime.default.md is the canonical reference copy. The two files MUST be byte-identical at release time. Validate:

docker exec agent-core diff /data/config/prime.md /data/config/prime.default.md

Empty output is required. If you've edited prime.md, copy the changes back to prime.default.md so future installs ship in sync.

Sections

prime.md is structured into 10 numbered sections covering: identity, intent boundaries (safety), tool rules, scheduling honesty (safety), side-effect policy (safety), failure honesty, task rules, language rules, output format, and the operator override slot.

The Safety sections are load-bearing — they back the deterministic policy guards in src/policy/. Modifying them affects how the agent reasons about side-effects, scheduling, and grounding.

Feature flags

The /config page right column lists 12 boolean feature flags grouped into 4 categories.

Autonomy

Flag	Default	Effect
dream_enabled	true	Dream cycle (gated by inactivity)
autonomous_goal_enabled	true	Autonomous goal generator job
perception_enabled	true	Background crypto perception

Execution

Flag	Default	Effect
plan_critic_enabled	true	LLM validation of TaskGraphs
goal_meta_reflection_enabled	true	Per-goal post-mortem
opportunity_engine_enabled	true	Opportunity detection

Learning

Flag	Default	Effect
behavioral_learning_enabled	true	Behavioral rule learning loop
skill_evolution_enabled	true	Skill pattern synthesis

Memory & Safety

Flag	Default	Effect
vector_memory_enabled	true	Embedding-based memory retrieval
knowledge_graph_enabled	true	KG extraction + injection
cpi_monitor_enabled	true	Cognitive Pressure Index monitor
integrity_monitor_enabled	true	Self-Integrity Monitor

Persistence

Flags toggled in the UI are written to Redis key config:overrides (JSON). On every startup, all 13 supported boolean flags are read from this key and applied to settings via setattr(). Changes survive restarts.

Runtime parameters

The /config page also surfaces read-only runtime parameters:

• Active LLM model
• Active embedding provider
• Resource Governor caps (goals/day, agents, tasks/hour, LLM calls/min, API calls/min)
• Memory injection budgets
• Browser idle timeout
• Audit retention days

These are sourced from settings. To change them, edit .env and restart.

Models

Open /models to:

• See the active default model and provider
• See the per-provider catalog (no health-gating — full catalog always shown)
• Test connectivity to each provider (sends a 1-token ping)
• Set the default model

The model router (models/router.py) classifies each request (vision / code / quick / complex / default) and suggests a model when none is pinned. You can override the suggestion per-request by mentioning the model name in /model.

Integrations

Open /integrations to manage 44 named connectors:

• Slack, Discord, GitHub, Telegram (the bridge), Notion, Zapier, Webhook
• Gmail, Google Calendar
• Philips Hue, Sonos, Home Assistant
• Spotify, Shazam
• 1Password (secrets)
• Platform integrations (macOS, Linux, Windows, Android, iOS)
• 30+ more

Each connector exposes a ConnectorManifest with available actions and a risk_level. Calls are gated by the policy engine and circuit-broken on repeated failures. Circuit breaker state persists in Redis (cb:state:{integration_id}) and survives restarts.

See also

• Environment Variables — startup-time config
• Models — how providers are wired
• Resource Governor — the limits
• Dashboard → Identity — agent persona page