GlossaryEvery term in the Harness vocabulary, defined once, with the page that explains it.

The three nouns

Space — everything an agent has, as one addressable tree: ~/memory/ (its files) and ~/tools/ (its tools). Configuration is content. → The Space

Log — everything that happened in a thread: an append-only sequence of events, each stamped with an actor. A shared workspace log, not a chat transcript. → The Log

Runs — everything that executes. Every turn, child, and script is a row in a tree; the table doubles as the dispatch queue. → Runs

The five verbs

read — one verb, any address: ~/memory/…, ~/tools/…, hm://…, ipfs://…, https://…, activity:, attachment:…, thread:…, run:….
write — the mirror: memory files, authored tools, hypermedia publishing (gated by the publish grant), IPFS.
call — invoke a callable tool by name. Wrong input returns the tool's contract instead of an error.
delegate — spawn a child run: a model child (given a brief) or a script child (given a script).
plan — maintain the thread's visible checklist.
The Five Verbs

Core objects

agent — a definition (name, prompt, provider, model, grants, signing keys) plus a private state directory. → Agents and Grants

session (thread) — a conversation: a durable event log with a title and a status. One session accumulates many runs. → Sessions and Threads

run — one unit of execution, of kind: 'agent' | 'workflow'. An agent run has a session; a workflow run does not.

event — one durable row on the Log: message, tool_call, tool_result, or error.

actor — who did it: user, agent, system, or trigger. Every event carries one.

signing identity — an Ed25519 key the server signs hypermedia writes with, stored encrypted and referenced by name from an agent definition.

Tools

tool document — every tool is a content-addressed DAG-CBOR document in ~/tools/; its CID is its version. Either builtin (a runtime binding) or lambda (authored source). → Tools as Documents

contract — a tool's full model-facing spec: description plus input/output JSON Schemas. read ~/tools/<name> returns it.

Space index — the compact <space> block in every system prompt, budgeted at 2 KiB: one line per tool, memory's top level, active triggers.

touch-expand — a call to an unexpanded tool (or with wrong input) returns the contract instead of failing; the retry executes.

promotion — once a contract enters the transcript, that tool becomes a first-class provider tool for the rest of the thread — derived from durable events, so it survives restarts. Always intersected with the enabled callable set.

grants — per-agent permissions: the callable set (search, web_search, navigate, execute) and publish. Verbs are never grants; memory writes are never gated.

callable — a tool reached through call rather than exposed directly to the provider.

lambda — an agent-authored tool whose source runs in the sandbox. Requires the same execute grant its runtime needs.

Delegation and scripts

child — a run spawned by delegate. A model child gets its own session; a script child runs in QuickJS.

brief — the markdown briefing that becomes a model child's first message verbatim. The prompt is the interface.

typed result — a child spawned with an output schema must deliver through return_result; validation failures bounce back for self-correction.

detacheddelegate {await: false}: the parent does not park on the child, but the child stays in the run tree.

script / ctx — a script child's world: ctx.call, ctx.delegate, ctx.parallel, ctx.step, ctx.plan, ctx.sleep, ctx.waitForEvent, ctx.continueAsNew, ctx.now, ctx.log, ctx.progress. → Scripts and the Journal

journal — a script run's durable effect record. Resume replays the source against it; completed effects never re-execute.

content key — the deterministic key an effect is journaled under, so replay matches by content rather than order.

narration — the {description} on a ctx.call: display metadata that rides the journal entry but stays outside the content key.

Plans

step — one checklist item with a stable id and a human label. Models rewrite labels; ids persist.

attachment (plan) — a child spawned while a step is running belongs to that step, joined by planStepId.

batch step — one running step owning a whole parallel batch of children.

settlement — closing a step. The runtime derives it only from success: a step whose attached children all succeeded is closed resolvedBy: 'runtime'. → Plans

Time and continuation

park / wait — a run pausing without holding resources. Reasons: children, timer, event, budget-pause. → Time and Parking

wake source — whatever ends a park: children finishing, a timer, a signal, an activity event, or a human Resume.

SignalRun — the action that delivers a signal to a parked run; exactly-once, transactional. → SignalRun

answerWith — the signal name a park advertises, so the Answer button sends something the wait accepts.

continueAsNew — finalizing a run into a fresh successor that keeps its place but starts a clean journal. → continueAsNew

obligation — work a run owes before ending honestly: typed debt (undelivered result) or plan debt (unsettled steps). → Obligations

Triggers

trigger — standing configuration binding a source to a continuation. → Triggers and the Event Bus

source — schedule, document-comment, user-mention, site-update, or run-completed.

continuation — what to do when it fires: newThread or wake.

firing — one trigger activation, deduped exactly-once by an activity key.

loop guard — the check that stops run-completed chains from ping-ponging forever (8 hops).

watermark — the durable per-account cursor into the activity feed.

Platform

signed envelope — the DAG-CBOR wrapper carrying an Ed25519 signature; the signature is the authorization. → The Signed Protocol

idempotency keyclientRequestId/clientMessageId; same key and bytes replays the stored response.

subscription keyaccount/…, agents/…, sessions/…, or runs/… (root run id).

wrench palette — the composer's tool button: run read/write/call yourself; results land on the log with a "You" chip. → The Desktop Surface

run card — one component with two moments: pinned while live, frozen into the log when done.

Process

checkmark — a milestone branch, verified (suites, adversarial review, gates) before review is requested.

live gate — scripted scenarios against a real server and real model; deterministic properties are enforced, model behavior is reported. → Testing and Gates

STALE marker — the file that makes stale cassettes skip loudly instead of pretending to pass.

Related

Do you like what you are reading? Subscribe to receive updates.

Unsubscribe anytime