Context & Skills

This section covers how mono-agent assembles the prompt for every turn: a required identity document, an optional soul document, the running conversation history, and any explicitly selected skills. All of it is declared in the context block of mono-agent.config.json and is coverage type config (with a few auto and code escape hatches noted below).

What goes into a turn

Each turn the agent builds a prompt from these layers:

Layer	Source	Coverage	Key
Identity	IDENTITY.md (or path)	config	context.identityPath
Soul (optional)	SOUL.md (or path)	config	context.soulPath
Session	host-owned delivery and callback-safety guidance; physical route withheld	auto	none (auto-generated)
Conversation history	owner-only durable store	auto / code	64 messages per exact conversation id, independent of runtime.maxTurns
Selected skills	<skillsRoot>/<name>/SKILL.md	config	context.skillsRoot, context.selectedSkills

Identity is the only required piece of context — context.identityPath is the one field in this section you cannot omit. Everything else is opt-in.

Note

Recalled memory is not a prompt layer. When memory is enabled, recalled entries are appended to the user message each turn (not assembled into the system prompt), so they survive session resume on runtimes that drop the system prompt. See Context assembly → Memory recall.

Configuring the context block

{
  "context": {
    "identityPath": "./IDENTITY.md",
    "soulPath": "./SOUL.md",
    "skillsRoot": "./skills",
    "selectedSkills": ["research"],
    "skillMaxBytes": 48000,
    "skillDisclosure": "index"
  }
}

Every field has a matching MONO_AGENT_* env var that overrides the JSON (env > JSON > defaults):

Field	Env var	Default
context.identityPath	MONO_AGENT_IDENTITY_PATH	./IDENTITY.md
context.soulPath	MONO_AGENT_SOUL_PATH	unset (no soul)
context.skillsRoot	MONO_AGENT_SKILLS_ROOT	./skills
context.selectedSkills	MONO_AGENT_SELECTED_SKILLS	none selected
context.skillMaxBytes	MONO_AGENT_SKILL_MAX_BYTES	48000
context.skillDisclosure	MONO_AGENT_SKILL_DISCLOSURE	full

Paths are resolved relative to the agent folder. mono-agent init scaffolds an IDENTITY.md for you; see Folder Layout.

Note

There is no auto-selection of skills — a skill is loaded only if its exact name appears in context.selectedSkills and a SKILL.md exists at <skillsRoot>/<name>/SKILL.md.

How each layer behaves

• Identity & soul are plain markdown loaded into every prompt. Identity carries role and boundaries; soul is an optional secondary voice/guardrail document. See Identity & Soul.
• Conversation history assembly is automatic (coverage auto): an owner-only, disk-backed store retains the latest 64 messages for each exact conversation id, independent of runtime.maxTurns, and survives process restarts. To swap in a custom store you must use the programmatic path — createConfiguredAgentResponder({ historyStore }) (coverage code); see Programmatic.
• Selected skills are loaded explicitly and each is capped at context.skillMaxBytes bytes (default 48000, valid range 256–1,000,000). See Skills.

Caution

The default history window is bounded independently from the provider turn limit. Pair it with runtime compaction (runtime.context-compaction) so longer provider-owned sessions stay within the model window. See Sessions & Concurrency.

In this section

• Identity & Soul — authoring IDENTITY.md and the optional SOUL.md.
• Skills — the SKILL.md format, selectedSkills, and per-skill byte caps.
• Assembly — the full prompt-assembly order and how the layers combine each turn.

Related

• Memory — recalled memory is appended to the user message each turn (not the system prompt).
• Tools — tool definitions and policy live next to context in each turn.