Agent harness
The process or program that runs an agent. It might call LLM APIs, invoke tools, access files, run tests, or emit telemetry. Distinct from the workload identity it can be launched as.Annotation
A structured judgment recorded against a run — an eval score, a pass/fail label, a reviewer’s note. Annotations are telemetry events, so you filter and aggregate on them like anything else. See Annotations.Annotation schema
A named, versioned JSON Schema that annotation payloads are validated against at ingest. Schemas are immutable and versioned per organization.API edge
The boundary that authenticates every request, resolves the credential to an organization, user, and scope, and routes it to backend services. Services trust the identity the edge stamps; they don’t re-authenticate.Artifact
A stored output object produced by a runtime operation or command execution. Command stdout and stderr can be stored as artifacts.Capability
A runtime feature a deployment either serves exactly or refuses. A request naming a capability the deployment cannot serve is rejected at admission withunsupported_capability rather than
silently weakened.
Capture
Recording an agent’s telemetry from the outside — you wrap a command withhiloop run instead of
adding an SDK to the harness. Capture covers the harness’s HTTPS traffic, its own OpenTelemetry, and
its stdio, all keyed to the run that produced them. See
Capturing telemetry.
Content-addressed payload
A large event body — a prompt, a completion, an HTTP response — stored once in a blob store and referenced from the event by digest, so the event stream stays small and identical payloads are stored once.Data view
A named, reusable SQL query saved and run by name and re-validated on every run. See Saved and data views.Egress policy
The fixed outbound-network boundary of a sandbox. Cluster DNS, the public internet, and the API’s dedicated proof-authenticated capture listener are reachable, while other sandboxes, all other cluster services, private and link-local addresses, and cloud metadata are denied outside the guest. The workload has no capture proof. There are no per-sandbox network modes or destination lists. See Network egress.Execution
A command run inside a sandbox. Executions track state, exit code, sandbox id, and output artifact ids.Idempotency key
An optional, client-generated key used only on create-style mutations to make retries safe. Reuse the same key with the same body to retry without creating duplicate work; the SDKs generate one automatically.Lineage path
A run’s dotted, time-sortable path of run ids from the root of its tree, such as01H8A.01H8B. A subtree is addressable by prefix, so queries can scope to one run and all of its
descendants. Telemetry events line up by the lineage path of the run that produced them.
Operation
An asynchronous control-plane action. Sandbox create, execute, stop, resume, and delete calls return operations that can be polled.Organization
The top-level account and the isolation boundary. Projects, runs, sandboxes, workspace metadata, and telemetry all live in an organization, and a request acts only within the caller’s own. An organization carries its members (users) and its API keys.Project
An organization-scoped workspace for grouping runs, telemetry, sandboxes, and resources.Provider-neutral
Describes the public runtime API: you request resources and capabilities and the platform reports what it supports, without exposing the underlying runtime’s native ids or errors. The same contract holds across deployments.Run
A captured agent workflow — also called a session. Runs may carry logical parent/child lineage; every event is keyed to the run that produced it (and its lineage path). Don’t confuse it with an API session (the cached credential fromhiloop login).
Run lineage
The logical parent/child relationships between runs in a project, recorded on each run and resolvable throughparent_run_id and root_run_id. Lineage is recorded data you can query and
group by; hiloop does not ship a dedicated tree browser. Logical lineage does not promise runtime
filesystem fork.
Sandbox
An isolated execution environment for an agent harness or command workload.Sandbox base image
The reproducible starting environment a sandbox boots from, typically an exact OCI image or named immutable environment plan.Saved view
A namedSELECT over your telemetry, saved and run by name and re-validated on every run — the
bare { name, sql } primitive a data view is built on. See
Saved and data views.
Scope
What a credential acts as. An organization-scoped credential acts across the whole organization — administration, billing, and all runtime and telemetry work. A sandbox-scoped credential is further confined to one sandbox and its own run lineage, so it can never widen its own reach.hiloop whoami prints your current scope.
Secret binding
A requested credential mapping for a sandbox. Request-time release is proof-bound to the live sandbox, executing principal, exact grant, and destination. A deployment without the complete release path rejects such requests before launch. See Egress and secrets.Signal
The family a telemetry event belongs to —llm, net, log, exec, span,
metric, egress, runtime, or annotation. Queries commonly filter or break down by
signal. See the signal reference.
Snapshot
An immutable capture of one sandbox’s disk state, usable as the source for new sandboxes. Creating from a snapshot is restore, fork, and branch in one verb. See Snapshots and branching.Storage class
The backing-storage behavior chosen when a sandbox is created:standard (node-bound,
the default) or durable (deployment-backed storage that recovers /workspace after node loss). See
Sandboxes.
Subtree
A run and all runs descending from it in run lineage. A query can scope to a subtree; see querying telemetry.Telemetry event
A structured observation attached to a run. Events have a signal, name, typed attributes, and optional payload reference.Telemetry gateway
The service that ingests captured events, derives the organization from the credential, and serves SQL queries.hiloop run exports to it;
hiloop query reads from it.
Run-scoped observability
An observability model where every telemetry event is keyed to the run that produced it. Runs also record logical lineage, so a query can additionally scope to a related subtree when that is useful. See Run-scoped observability.Workload identity
A named, organization-scoped machine identity a run can be launched as — the identity an agent harness runs as, distinct from the harness itself. A workload has no login and no console; it is registered explicitly, carries an owner trail and a launch ACL, and gives attribution and policy (such as an egress policy) something to bind to per identity. Declare it withhiloop run --as workload/<name>. See
Workload identity.
Workspace
The/workspace directory inside a sandbox, where code and durable data belong. Its
lifetime is set by the storage class chosen at create. See
Snapshots and branching.