docs
// Execution Management (EEM)

Glossary

37 terms across 7 sections. Vocabulary grouped by the part of the Erebine Execution Manager (EEM) the term describes; cross-references are clickable. Platform-wide vocabulary (projects, API keys, service tiers) lives in the router docs and is not duplicated here. Internal schema names are omitted by design.

Runtimes and Enrollment

The roles, identities, and lifecycle events of the runtimes that join the mesh.

EEM

Erebine Execution Manager. The headless runtime that runs tool and shell executions for workspaces, dispatched by an inference model. This glossary entry is the canonical expansion; other EEM pages link here.

EIM

Erebine Inference Microservice. The runtime that serves models on accelerators: it pairs a model with vLLM and produces tokens. One EIM drives one accelerator or many, and several EIMs can share a machine; the operator decides the split.

Execution Manager

The expansion of EEM. Also the desktop app's execution screen; see Desktop Execution Manager.

runtime

Any service enrolled against the Erebine router. Either an EIM (inference role) or an EEM (execution role).

registration_name

Operator-chosen handle for an enrolled EEM (for example eem-jumphost-01). The router records an EEM under its worker id; the EEM's identity across restarts is its enrollment file.

join key

A short-lived token minted by the router for one role, EIM or EEM. It admits up to its enrollment limit (10 by default) until it expires. Mint with erectl eems join-keys --create.

capability manifest

JSON document the EEM publishes at enrollment and on every lease_renewal, naming every workspace it serves and every tool it can execute.

manifest_hash

Digest of the capability manifest that the EEM recomputes on each renewal. The router uses the hash to detect drift between declared and actual capability.

manifest_version

Integer version field on the capability manifest. Bumped by the EEM when the declared tool set or workspace bindings change. Lets operators correlate manifest_hash drift with intentional capability revisions.

lease_renewal

Heartbeat frame the EEM emits by default every 10 seconds (configurable via lease_renewal_interval_ms). Carries the in-flight invocation list so the router can reconcile after an unplanned restart on either side.

conflict_pill alias: poison pill

EEM state entered when two distinct CURVE keys claim the same registration_name. The router parks the EEM already holding the name and refuses the new one until the conflict is resolved by an administrator.

config discovery

Boot-time probes the EEM can run to survey the host for workspace bindings an operator might declare. Advisory only: they record what they found and bind nothing. Off unless the EEM's discovery mode is set.

Workspaces

The named scopes that bind EEMs to the work they are allowed to perform.

workspace

A named operational-scope binding. Either of kind chat (knowledge only) or operational (bound to EEMs and capable of running tool calls).

EEM binding

Pins which EEMs serve a workspace. Create one with erectl workspaces bind-eem. Without one, the first online EEM that advertises the workspace is used.

operational workspace

A workspace of kind operational, bound to one or more EEMs. Its kind is set at creation and never changes.

three-model binding

Every operational workspace binds three model roles (planner plus two execution-side roles). Unavailability of any of the three fails the workspace closed.

workspace declaration

An entry in the EEM's workspaces.yaml. Its name must equal the router-side workspace name; see Workspaces.

Tool Execution

The lifecycle of a single tool call from creation to durable result.

tool

A named unit of execution declared in the capability manifest. An EEM exposes tools; the inference model calls them.

invocation

A single tool-call lifecycle: creation, optional approval, dispatch, execution on an EEM, completion.

idempotency_key

Client-supplied JSON body field on POST /v1/exec/invocations that lets the router deduplicate retries. The same key within the retention window returns the original invocation's result.

ambiguous mutation

A tool invocation whose success the router cannot confirm because the EEM went offline mid-operation. Requires explicit operator resolution.

learnings

Post-execution observations recorded on successful tool calls so future invocations can short-circuit known-good parameter sets. Stored by the router under the /v1/exec/learnings route family and read back into agentic loops at planning time.

request_subplan alias: sub-agentic

Agentic primitive that hands a goal to one nested model turn on the workspace's inference endpoint and returns its answer. It sends a requested tool scope and iteration count, which are not enforced yet: no tools run in the nested turn.

auto-fork-branch

Router behavior that forks the chat branch before any destructive or irreversible tool call; on failure, the operator can roll back to the pre-fork branch.

Approvals

The human-in-the-loop gates that protect destructive and irreversible operations.

auto-approval grant

Standing permission for one workspace and domain to run calls up to read or write risk without approval. It never covers a destructive or irreversible call, or the database domain.

approval policy

A per-workspace document declaring, for each risk level, whether approval is required and what the SLA and escalation chain look like.

approval request

A pending decision created by the router when an approval policy gates a tool invocation. Resolves to approved, rejected, timeout, or escalated. An escalated request releases the invocation slot the same way a rejected one does.

risk level

Per-tool classification: read, write, destructive, or irreversible. Drives the approval policy lookup.

Chat Templates

Reusable, versioned recipes for recurring procedures.

chat template

A JSON document scoped to platform, project, or workspace. Pins the system prompt, the allowed tool set, and the approval cadence for a given procedure.

template scope

The visibility tier of a chat template. platform templates are visible to every project; project templates are visible only within the owning project; workspace templates are visible only within the owning workspace. More specific scopes override less specific ones at resolution time.

Tiers

The service-tier classifications that set price, limits, and scheduling priority.

service tier

The price and limits class assigned to a model endpoint: its token rate, any hourly rate, its caps and timeouts, and its scheduling priority. A tier carries no hardware and no rate limit; see Service Tiers for the ladder.

self_hosted

The service tier assigned to endpoints backed by operator-owned EIMs. Billed an hourly rate per EIM plus a rate per uncached token, against compute the operator already pays for.

Security and Transport

The wire-level and access-control primitives that make EEM dispatch safe.

CurveZMQ

ZeroMQ's Curve25519-based authenticated encryption layer. Every frame between the router and an EIM or EEM is CurveZMQ-encrypted.

dispatch queue

The SQLite-backed outbound queue on the EEM. Persists frames until router acknowledgement so an unplanned restart does not lose in-flight work.

scope

Two distinct meanings.

  1. API-key scope (inference, execution, management, research) gating route access; see Authentication.
  2. Per-workspace scope narrowing what resources the workspace's tools may touch; see Workspaces.

trace_context

W3C trace-context value carried on every ZMQ frame (wire field trace_context) so distributed traces span router, EIM, EEM, and vLLM boundaries. The equivalent HTTP-surface header is traceparent.