Glossary

Little Harness keeps the durable conversation and file vocabulary even when a deployment uses Little Compute. Compute terms describe logical work and temporary execution capacity; they do not rename a Harness Session.

A

Agent Group — The public Harness-facing handle for a Little Compute Task Scope: one immutable, atomically sealed set of child Tasks with a declared join and reduction policy.

Approval — Authorization for one exact action digest. An Approval for one digest cannot authorize a modified action.

Artifact — An immutable file or payload with verified identity and provenance, usually too large or too durable to keep inline in model context. Examples include patches, logs, test reports, screenshots, and Continuation Capsules.

Attempt — One retryable execution of a Little Compute Task. An Attempt is durable independently of a sandbox and can use zero or more sequential Sandbox Leases.

C

Cell — The Little Compute correctness, database, scheduling, and blast-radius boundary that owns an Operation for one writable ownership epoch.

Change Set — An immutable code result relative to a Repository Snapshot. It records the producing Task and Attempt rather than mutating shared product state automatically.

Continuation Capsule — A credential-free Artifact containing the exact Harness event, File Snapshot, workspace generation, Task Scope, and pinned configuration cursors needed to resume an Attempt on another trusted runner. It is a derived checkpoint, not independent mutation authority.

E

Export — A deliberate app-level write that moves a harness artifact or result into product state.

Execution API — The public call surface for a harness. streamHarness() streams an AI SDK-compatible UI response; generateHarness() returns a completed result. Both should share the same Harness Session, input, tool, Artifact, persistence, and trace semantics.

Extra Body — App-defined request context passed to execution calls, tools, Persistent Dir callbacks, events, and trace metadata.

F

File Snapshot — An immutable, content-addressed root for the Session Files committed at one workspace generation. It is distinct from a provider filesystem checkpoint, rootfs snapshot, or memory snapshot.

H

Harness — The managed runtime boundary around an agent: model loop, tools, skills, files, artifacts, tracing, and persistence behavior.

Harness Input Type — A reusable definition for a non-chat type + input call. It can describe the input, validate it, translate it into model messages, and provide default output expectations.

Harness Dir — A directory path inside the Little Harness filesystem, such as /persistent/user. It is separate from the source directory, object-store prefix, or app storage record that backs it.

Harness Memory — Opt-in long-term memory: a Persistent Dir of plain files the model reads and edits with its file tools, listed in every turn's prompt, with a built-in memory skill on how to keep it.

L

Lease Resource Grant — A conserved Little Compute right for the CPU, memory, disk, browser, and related capacity required by one Sandbox Lease. The grant must exist before provisioning and remains accounted for until safe release; it is not the provider sandbox itself.

Local Host — The zero-service development host for Little Harness. It stores Harness Session folders, Artifacts, traces, local Persistent Dir sources, and locks under a local dataDir.

M

Message Staging — The pre-turn hook that lets a developer turn uploaded chat-message files into Session Files and return a small path-only notice for the model.

O

Operation — One root user intent managed by Little Compute, such as one UI comment or requested code change. It owns the root Task graph, budgets, convergence, and final result within one home Cell ownership epoch.

P

Persistent Dir — An opt-in developer-backed directory exposed inside a Harness Session at a harnessDir so selected files can persist across Harness Sessions.

Project — Little Compute's tenant-scoped configuration boundary for policies, repositories, Agent Definitions, Templates, and execution profiles.

Project Dir — An explicit local filesystem source reference that must resolve inside the configured project root before Local Host exposes it.

R

Repository Materialization — An ephemeral, private checkout or mount made from a Repository Snapshot and the current Attempt workspace generation. It is a disposable cache, never a second mutable file authority.

Repository Snapshot — An immutable, normalized repository tree and lineage point used as the base for Tasks and Change Sets.

S

Sandbox Lease — Temporary, fenced authority to attach one provider sandbox generation to an Attempt at an expected workspace generation. An Attempt can hold zero Leases while reasoning or waiting and later acquire sequential Leases; it has at most one writable attachment at a time.

Scheduler Lane — A fenced, high-throughput Little Compute admission partition that spends only the conserved rights delegated to its current epoch. It does not own Task truth or provider node placement.

Scheduling Domain — A Little Compute capacity and fairness boundary whose allocator delegates conserved rights to Scheduler Lanes.

Session / Harness Session — The durable isolated boundary for one conversation or job. It owns conversation replay, Session Files, Artifacts, trace, and continuation correlation. It is not a VM, container, checkout, Attempt, or Sandbox Lease.

Session Files — Files inside one Harness Session's isolated filesystem. They can be written by the app, message staging, tools, or the agent, and persist across turns in that Harness Session.

Shared Execution Context — An immutable Task-Scope base that pins the Repository Snapshot, Template, base File Snapshot, dependency or index inputs, and verification base shared by the Scope's members.

Skill — Procedural knowledge exposed read-only for the agent, usually as a SKILL.md file plus supporting material.

T

Task — One logical, durable unit of work in a Little Compute Operation graph. A Task can have retry Attempts, but has at most one committed successful winner. Harness may expose a short Harness Session-scoped handle such as task_1 while Compute stores the globally unique correlation.

Task Execution Context — An immutable per-Task overlay on a Shared Execution Context. It pins the managed Agent Definition or attached coordinator identity, input Artifacts, allowed narrower authority, and result schema.

Task Publication — A fenced immutable progress, finding, change_intent, change_preview, work_claim, check, evidence, or needs_input record from an Attempt. A Publication can inform a parent or UI, but is not a winning Task result by itself.

Task Scope — An immutable set of Little Compute Tasks created atomically with member order, dependencies, budgets, join predicate, and bounded reduction policy. It is exposed to Harness callers as an Agent Group; later fan-out creates a new Scope instead of mutating membership.

Tool — A host-provided capability exposed to the agent.

Trace — The host-facing record of model calls, tool calls, file changes, artifacts, commits, and failures.

W

Workspace Generation — A monotonic Attempt-local pointer to one immutable File Snapshot committed at a command or quiescence boundary. A Sandbox Lease must attach at the expected generation, and advancing the pointer is a fenced compare-and-set.

Workspace — Avoid this unqualified word in Little Harness docs when referring to the agent-visible filesystem. Prefer Session Files or a path such as /session/. In Little Compute, use the precise term workspace generation only for the Attempt's canonical File Snapshot pointer and Repository Materialization for a disposable checkout or mount.