Glossary
The canonical vocabulary used throughout the rulvar documentation. These terms are used exactly as defined here on every other page; when a page and this glossary appear to disagree, this glossary wins.
Terms are grouped by the subsystem they belong to: the journal, the agent runtime, adaptive orchestration, and the platform surface around them.
Journal terms
The journal is rulvar's durability substrate: a content-addressed memoizing log, not an event-sourcing stream. The full mechanics live in the journal guide and the durability guide.
| Term | Definition |
|---|---|
| journal | The content-addressed memoizing log of completed effects; the single source of run truth under the never-pay-twice invariant. |
| journal entry | One record in the journal, identified by (scope path, content key, ordinal) and qualified by its hash version. |
| entry kind | The discriminator naming what a journal entry records (agent, rand, resolution, abandon, plan.revision, and so on), governed by the kinds registry. |
| content key | sha256 over the RFC 8785 canonical JSON of a call's identity input; the content-addressed part of entry identity. |
| scope path | The structural "/"-joined path locating a call site within a run's execution tree; part of entry identity. |
| ordinal | The repeat counter of an identical (scope, key) call within one scope, disambiguating repeated identical calls. |
| live call | A call actually executed (and paid for) against a provider or effectful resource, as opposed to being served from the journal. |
| replay | Serving a completed journal entry's result instead of performing a live call. |
| rerun | Executing a call live again despite an existing journal entry, as directed by the replay disposition. |
| replay disposition | The single canonical kernel predicate (exposed as replayDisposition) mapping (entry, fold state) to replay, rerun, or skip. Also called the replay predicate. |
| scoped forward-matching | The resume algorithm matching calls to entries forward within a scope; a miss does not move the cursor and does not suppress later hits. |
| fold (pure fold) | A pure derivation over already-journaled entries producing derived state (plan state, ledger, digests); never a source of new effects. |
| decision entry | The single journaled record of a dynamic decision, written strictly before any of its effects, per the decision-entry principle. |
| ref-entry | An entry (kind resolution or abandon) referencing an earlier entry by seq (ref < seq), closing or annotating it. |
| resolution | An attempt to close a suspended entry; all attempts pass through the ResolutionArbiter and are always appended. |
| abandon | A journaled decision to stop pursuing a subtree; its descendants become derived skipped and cost zero live calls. |
| derived skipped status | The skipped status computed by the abandon fold; never stored in an entry's status field. |
| first-closing-wins fold | Among racing ref-entries for one target, the first appended closing entry wins and alone carries the debit; later entries are superseded. |
| suspended entry | An entry whose completion awaits an external resolution (approval, external input, escalation decision), with an optional journaled deadline. |
| two-phase entry | An entry written as running before dispatch and completed by a terminal status, enabling at-least-once dispatch without double pay. |
| orphaned entry | A running entry whose completing write never arrived (crash); handled by recovery rules on resume. |
Replay is not rerun
Replay serves a completed result for free; rerun deliberately pays again because the disposition table says the stored outcome cannot be trusted for this resume. The two words are never interchangeable in these docs.
Agent terms
The agent runtime owns the model loop, its checkpoints, and the documents that teach an agent its surface. See the agents guide and the model routing guide.
| Term | Definition |
|---|---|
| turn | One model invocation cycle of an agent: one assistant response together with its tool calls. |
| turn-boundary checkpoint | The canonical-history checkpoint the agent runtime writes at each turn boundary under a durable store; resume continues from the same turn. |
| card | A compact rendered document teaching an agent a surface: the agent profile card, the planner API card, the model knowledge card. |
| model ladder | An ordered sequence of rungs an agent may escalate through, with journaled acceptance gates. |
| ladder rung | One level of the model ladder: a model plus optional effort and per-rung limits. |
Adaptive orchestration terms
Adaptive orchestration is the opt-in machinery for wide fan-out workloads: typed plans, admission, escalation, and bounded termination. See the adaptive orchestration guide, the planner guide, and the budgets guide.
| Term | Definition |
|---|---|
| admission | The AdmissionController check every spawn passes before any effect: budget reserve, structural limits, dedup and reuse, lineage. |
| admission verdict | The typed decision-entry union (AdmitVerdict) produced by admission: admit, reject codes, full reuse, and the other admitting arms. |
| escalation | A child agent's typed report that its task exceeds its scope or is blocked, requesting a decision. |
| plan revision (replan) | A journaled plan.revision changing the TaskPlan through the rebase algorithm. |
| park / unpark | Suspending a plan node while retaining its checkpoint (park), and resuming it later (unpark). |
| wake (wakeup) | An orchestrator turn triggered by a wait_for_events trigger firing. |
| wake digest | The coalesced, snapshot-pinned digest (WakeDigest) delivered on wake: digest ordinal, plan hash, completed task digests, escalations, termination and budget blocks, reuse stats. |
| lineage | The retry ancestry linking attempts of one logical task, carrying depth and approach signature; the basis of escalation caps and the single-live-attempt rule. |
| logical task | The stable identity of a task (LogicalTaskId) across retries, decompositions, and reuse, minted under the lineage rules. |
| oscillation guard | The guard detecting revision loops (A-B-A plan churn) and forcing a terminating fallback. |
| reuse-by-reference | The admission outcome linking a new plan node to a completed donor entry (a node.link) instead of respawning the work. |
| graft | Transplanting a donor subtree's results into the current plan via aliasing during reuse. |
| termination account | The frozen per-run vector of countable resources (spawns, revisions, escalations, wakeups) debited by decision entries; the basis of the termination guarantee. |
| run budget ceiling | The immutable dollar ceiling fixed at run start; no API, including human-in-the-loop decisions, can raise it. |
| overshoot | Spend beyond the ceiling, bounded by one turn per in-flight agent; the tightest bound possible since providers bill aborted streams. |
Where these live
Everything in this section is engine-owned typed data, never prose in a transcript. The core admission and termination machinery ships in @rulvar/core; the typed plan (TaskPlan, the plan runner, the run ledger) ships in @rulvar/plan.
Platform terms
Cross-cutting mechanisms of the runtime and its stores. See the tools guide, the stores guide, and the orchestration modes guide.
| Term | Definition |
|---|---|
| role quality floors | Per-role explicit model allowlists and denylists in engine config, keeping unsuitable models out of critical roles. |
| permission chain | The ordered tool-permission pipeline: hooks, then deny rules, then ask rules, then canUseTool, then the terminal default. |
| worker sandbox | The worker_threads sandbox executing machine-generated scripts with the seeded, journaled global set; a determinism and blast-radius boundary, not a security boundary. |
| lease with fencing epoch | The LeasableStore ownership mechanism for queue workers; appends carrying a stale epoch are rejected and invisible. |
Related pages
- Journal for entry identity, replay, and the disposition table in depth.
- Durability for crash recovery, two-phase entries, and resume.
- Agents for turns, checkpoints, and history projection.
- Adaptive orchestration for admission, plans, escalation, and termination.
- Budgets for the three-layer budget, the ceiling, and overshoot.
- Design principles for the invariants behind these terms.