Glossary

The words these docs use with a meaning of their own, in alphabetical order. Each has one or two sentences and a link to the page that covers it. There are 64 terms.

Term	Meaning
Abandon	To end an open run by hand, with realm run abandon or the abandon_run tool. The run’s phase becomes abandoned. See realm run: commands that act.
Adapter	The code that calls one outside service for a workflow. Realm has 10 built in, and a project can add its own. See Adapters.
Agent profile	A text file of standing instructions that an agent step is given with its prompt. See Services, profiles and context.
Agent step	A step whose output comes from an AI model. Realm checks the answer against the step’s schema before it records it. See Step kinds.
agent_action	The field of a tool’s reply that tells the caller what to do after an error or a block. See Error codes.
Auto step	A step Realm runs itself, by calling a handler or an adapter. See Step kinds.
Batch	Several runs of one workflow started with one call to start_run_batch. See Idempotency and batches.
Choice	One of the answers a gate offers, such as approve or reject. See Gates.
Claim	The mark Realm puts on a step when it starts, so that no second caller runs the same step. See realm run: commands that act.
Cleanup step	Another name for a finalizer.
Condition	A comparison of a value from the run with a fixed value, used by when, preconditions and abort_unless. See Conditions.
Deployment manifest	The file realm.yaml, which holds a project’s settings. See Deployment manifest.
Drain	To run the finalizers that a run still owes after it has ended. See realm run: commands that act.
Drift	A difference between the project’s code or settings on disk and what a run recorded when it ran. See Project extensions.
Drive failure	A failed request from realm agent to a model provider. Realm keeps the last 5 on the run’s record. See realm agent.
Driver	Whatever calls Realm to move a run forward: an AI assistant over MCP, realm agent, realm listen, realm workflow run or realm run advance. See Who drives a run.
Engine	The part of Realm that decides which step can run, checks each answer and writes the record. The realm command and the MCP server are built on it. See Core library.
Evidence	What a run records about each step: what it was given, what it produced, and a hash of the output. See Evidence.
Evidence entry	One item of a run’s evidence: one attempt of one step, or one gate answer. See Run record and export bundle format.
Export bundle	The one file that realm run export writes, with a run’s record and its other files. See Run record and export bundle format.
Extensions	The code files of a project that give Realm adapters and handlers. A workflow names them with extensions. See Project extensions.
Finalizer	A step that runs after the run has ended, for cleanup or for a notice. It names the endings it runs for with on_outcome. See Step kinds.
Fixture	A file that describes one test of a workflow: the parameters, a stand-in reply for each service call and agent step, and the expected ending. See Test a workflow.
Gate	A stop at a step, where the run waits until a person picks one of the gate’s choices. A step has one when its trust is human_confirmed. See Human gates and trust levels.
Guard	A step that checks conditions and stops the whole run if one is false. See Step kinds.
Handler	A function from your project that an auto step or a finalizer runs. See Handlers.
Hash	A short fingerprint of a piece of data, which changes if the data changes. Realm records one for each step’s output.
Health finding	Something Realm notices is wrong or unusual about a run, such as a step whose caller has gone away. See Run state and health findings.
Idempotency key	A name given when a run is started, so that starting it twice with the same name gives one run. See Idempotency and batches.
Input map	The input_map of a step, which says where each of the step’s inputs comes from. See Input map and templates.
Lease	A short hold a process takes on a finalizer while it runs it, so that no other process runs it at the same time. See realm run: commands that act.
Loader	The part of Realm that reads a workflow file and refuses it if something is wrong. See Loader diagnostics.
MCP	The Model Context Protocol, the standard way an AI assistant calls outside tools. Realm is an MCP server with 11 tools (10 in version 0.46.0). See MCP tools.
Operation	One action of an adapter, such as read on the filesystem adapter. A step names it with operation. See Adapters.
Parameters	The values a run is started with. A workflow’s params_schema says which it accepts. See Workflow file: top-level fields.
Phase	Where a run stands: running, gate_waiting, completed, failed, aborted or abandoned. See Workflows, steps and runs.
Precondition	A condition a step states with preconditions. If it is false, the step is blocked and the caller is told why. See Conditions.
Processor	A third kind of thing a project’s code can give Realm, beside adapters and handlers. No workflow field uses one. See Project extensions.
Project	The folder a workflow belongs to: the nearest folder at or above it that holds a package.json or a .git. See Deployment manifest.
Provider	The company whose model realm agent calls: OpenAI or Anthropic, or code of your own. See realm agent.
Purge	To delete an ended run’s files, with realm run purge. See realm run: commands that act.
Reclaim	To take a step back from a caller that has gone away, so that it can be run again. See realm run: commands that act.
Register	To store a copy of a workflow under its ID in ~/.realm/workflows, where the MCP server finds it. See realm workflow.
Replay	A re-check of a run’s preconditions against the outputs it recorded, with the outputs you choose changed. It runs no step. See realm run: commands that read.
Retry	To try a step again after it fails. A step asks for it with a retry block. See Retry and timeouts.
Run	One execution of a workflow. It has an ID, its parameters, and a phase. See Workflows, steps and runs.
Run record	The JSON object that holds everything Realm knows about a run. See Run record and export bundle format.
Schema	A JSON Schema, which says what shape a piece of data must have. A workflow uses one for the run’s parameters and for each step’s answer. See JSON Schema blocks.
Secret reference	${secret:NAME} in realm.yaml, which stands for a secret’s value without writing it there. See Deployment manifest.
Service	An outside system a workflow calls, declared under services with the adapter that reaches it. See Services, profiles and context.
Settle	To record a step’s final result. After that the step cannot be started again in this run.
Skipped step	A step that will not run, because its when was false, or because the steps it depends on ended in a way its trigger_rule rules out. See Order and routing.
Stale claim	A claim whose deadline has passed, which suggests its caller has gone away. The step can then be reclaimed. See realm run: commands that act.
Step	One unit of work in a workflow. Each has a kind: auto, agent, guard or finalizer. See Step kinds.
Store	Where Realm keeps runs and workflows. The realm command uses files under ~/.realm. See Core library.
Stuck run	An open run with a health finding, as listed by realm run list --stuck. See realm run: commands that read.
Tool server	A separate MCP server whose tools an agent step’s model may call under realm agent. See Give an agent step tools.
Trace	Notes a caller sends about how it worked on an agent step, recorded with the step. See MCP tools.
Trigger	The trigger block of a workflow, which lets realm listen start a run from a webhook request. See Webhook trigger.
Trigger rule	A step’s trigger_rule, which says which results of the steps it depends on let it run. See Workflow file: step fields.
Trust	A step’s trust field, which says whether the step has a gate. See Human gates and trust levels.
Workflow	A file that lists the steps of a job and the order they run in. See Workflows, steps and runs.
Workflow context	Text files a workflow names with workflow_context, whose text agent steps can be given. See Services, profiles and context.
Writer nonce	An identifier a caller gives with its trace entries and its answer, so that Realm can tell which entries came from the caller that completed the step. See MCP tools.