Concepts

The normative terminology of the Runne runtime. Each term has a definition and — where the source specifies it — an explicit "it is not".

Execution model

Term	Definition
Session	The context of a user interacting with the runtime. Stores message history and references to TaskRuns. A Session is not the unit of execution and does not own execution.
TaskRun	The independent unit of execution of an AI coding workflow — a unique ID, a lifecycle, and a state machine, kept separate from the Session.
Attempt	One execution attempt of a TaskRun. Each Attempt has a unique ID and a full audit trail; a TaskRun can have several Attempts. Retry happens inside one Attempt, not across Attempts.
Step	One logical step inside an Attempt — a model call, a tool call, a verification, etc.
Turn	One agent↔model interaction inside a Step (prompt + response).
Agent	The logical entity executing a TaskRun, configured with a model, system prompt, tools, skills, and policies.
Primary Agent / Subagent	The primary agent is directly tied to the TaskRun; a subagent is created by it for a subtask and receives a delegated TaskRun.
Agent Loop	The main execution loop — model call → tool call → observation → … → acceptance gate.

Durable state and recovery

Term	Definition
Durable Event	An event written to persistent storage before the associated action runs — not an in-memory log line.
Event Log	Append-only source of truth for TaskRun state.
Checkpoint	Saved TaskRun state sufficient to continue execution (resume).
Resume	Continuing a TaskRun from the last checkpoint after a controlled stop or crash.
Retry	Re-running an operation (model call, tool call) inside an Attempt after an error — distinct from resume.
Recovery	Restoring a TaskRun after a runtime crash — load the last checkpoint, replay durable events, reconcile uncertain side effects.
Audit Trail	The complete chronology of significant events, sufficient to reconstruct a TaskRun's execution.

Verification evidence

Term	Definition
Verification	The lifecycle stage where the runtime checks the acceptance criteria.
Acceptance Criterion	One checkable condition of success.
Acceptance Gate	The set of acceptance criteria that must pass for a TaskRun to reach completed.

Capability grants and approvals

Term	Definition
Capability	A named ability granted to an agent (fs:read, fs:write, process:exec, network:outbound, …).
Capability Grant	The act of granting a specific capability with a scope — not a global permission.
Permission Rule	The rule deciding when a grant is enough and when an approval is required.
Approval	The user's explicit permission for an operation, with a scope, grantee, and TTL. An approval is not a grant.
Sandbox	An isolated environment constraining the agent's filesystem, process, and network capabilities.
Secret	Confidential material (API key, token, password) resolved from an external store, never held by the runtime itself.

Budgets

Term	Definition
Budget	A limit on TaskRun resource consumption — time, cost, tokens, tool calls, retries, subagents, processes, network, storage. Budgets are enforceable, not advisory.
Budget Ledger	The tracked consumption of each active budget.

Control and execution

Term	Definition
Control Plane	The part of the runtime managing TaskRuns — creation, configuration, monitoring.
Execution Plane	The part that actually executes TaskRuns — agent loop, tool calls, model calls.

These definitions are normative in the runtime repository (docs/00-foundation/TERMINOLOGY.md). See the repository for the full glossary and per-term source status.