Little documentation

Glossary

Shared terms across Little Harness, Little Workflow, Little Compute and LittleDB, and where each product uses the same word differently.

Each product also has its own glossary with more detail: Harness, Workflow and Compute concepts.

Units of work

These are different things. A Compute Task can host a Harness Session, and a Harness agent can start a Workflow Run, but none of them is another name for the others.

TermProductMeaning
Harness SessionHarnessOne agent conversation or job, with a private /session/ workspace. The caller owns the transcript and resends messages each turn.
TurnHarnessOne generateHarness() or streamHarness() call within a session.
Workflow RunWorkflowOne durable execution of a workflow version, recorded as events in a World.
OperationComputeOne user intent and its bounded agent graph.
TaskComputeOne logical unit of agent work. It may have zero or more retry Attempts.
AttemptComputeOne durable try at a Task. It keeps lineage and a File Snapshot generation even while it holds no compute.
Sandbox LeaseComputeA time-bounded attachment to one isolated execution environment. An Attempt may hold sequential Leases, never more than one writable attachment.

Storage and state

TermProductMeaning
Local HostHarnessThe local host for sessions, traces, artifacts and Persistent Dirs, stored under a data directory you choose.
Persistent DirHarnessA directory checked out into each session and committed back at a commit boundary.
WorldWorkflowWhere Workflow stores run events and state. Local World is the default and records events in an embedded SQLite store.
File SnapshotComputeLittle-owned durable workspace state. A provider or VM snapshot is only an optional acceleration.
ArtifactAllA stored file outside the model's context. In Harness, a file under /artifacts/ such as a large tool result; in Workflow, a run output with an ID and manifest; in Compute, a content-addressed blob with retention rules.
MemoryHarness, WorkflowTwo different features. Harness memory is opt-in, model-authored memory backed by a Persistent Dir. Workflow Memory Stores are filesystem collections mounted at /mnt/memory/... for planners.

Definitions and formats

TermProductMeaning
HarnessHarnessThe runtime createHarness() builds. In Workflow pages, "harness" also names the runtime each workflow role (planner, worker, fixer) runs in.
Agent folderHarnessA folder with agent.ts plus tools/ and skills/, loaded by the little-harness CLI or loadHarness().
LWIRWorkflowThe JSON workflow format the planner drafts and the validator checks.
WorkflowVersionWorkflowAn immutable, validated workflow that runs execute.
VersionLittleDBOne state of a harness or workflow as plain JSON content, such as { system, model }. An app resolves which version to run; the config in its code is the bootstrap that becomes version 1. Private source.
ProposalLittleDBA candidate version with a rationale and evidence, gated by an experiment or a trial and approved by policy or a person. Private source.
Eval workerLittleDBThe loop in your app that claims LittleDB's experiment runs, runs a version on a dataset item and reports scores. Private source.
User feedbackLittleDBWhat an app's end users said about a turn (a thumbs, a rating, a comment, survey answers), stored as scores with source user and weighed by reviews, trials and rollback next to the judges. Private source.

Integration and status

TermMeaning
PortAn optional Harness integration point you can supply yourself, such as the execution environment or the durable session log. Little Compute implements several in its unpublished source. None is required.
Agent GroupA parent agent's bounded set of child agents. The Harness agents option that starts one is UNRELEASED; Little Compute's Agent Groups are unpublished source.
UNRELEASEDExists only in the current source. No npm version includes it, and no version number has been assigned. See Releases.

On this page