Glossary

Every page in these docs uses the terms below with these meanings. Where the engine uses a different name in its files or commands, that name follows in parentheses.

Admission

A signed record that one exact version of a program passed qualification and may run under the Standard or Regulated profile. qualify --certify signs a local admission with a test key, which production refuses.

Agent Skill

A SKILL.md file that tells an AI assistant what one workflow does, its inputs, and what to do after each outcome. Written by emit-skill. It holds no recorded values.

Approval mode

Whether a run waits for an answer before each save: each_item (wait for each save), problems_only (only problems go to a person; not yet available), or not_gated (no pause, recorded in the report).

Bundle

The folder a compile writes: the program (workflow.json, or workflow.json.enc when encrypted), screen crops it uses to find targets, and a manifest with a digest of every file. The docs also call it the program. openadapt flow seal writes an encrypted copy.

Canary

The first few runs of a repaired program after it's activated. Any wrong success or failed check switches back to the previous version.

Certify

Check a program against a policy. A property of the program alone.

Compile

Turn one recording into a program. induce does the same from two or more recordings.

Composition

A parent program that runs several programs in order, each on its own surface, and passes confirmed values between them (compose).

Demonstration

One performance of the task in the real application, by a person or an agent, while OpenAdapt records it. A program is learned from one or more demonstrations.

Deployment file

A YAML file (deploy.yaml) that says where the app is, where its records can be read, which profile and policy apply, and whether model help is allowed. It holds environment variable names, never secret values.

Durable run

A run that saves a checkpoint after each confirmed step and pauses when it stops, so it can resume without repeating a confirmed save (--durable, resume).

Execute

OpenAdapt's commercial API for submitting runs to a customer-controlled runner and getting back signed results.

Halt

A designed stop: the run refused to act because a check failed. The report says which check and why.

Model grounding

Asking an AI vision model where a target is when the recorded control can't be found. Off by default. When on, screenshots can leave the computer unless the model runs on your network.

Policy

A set of rules a program must meet to run, such as "every save has a right-record check". Built in: clinical-write and permissive.

Profile

The set of requirements a run is held to: demo, standard, or regulated. See Run a workflow.

Qualify

Run a program against the normal case and five faults in a test copy of the environment where it will run, and check it saves correctly on the normal case and stops on every fault. A property of the program in one environment.

Receipt

A small JSON summary of a VERIFIED run that's safe to share. It holds counts and digests, not record values (receipt.json). Only VERIFIED runs get one.

Reconciliation

A person checking the record after a run ended RECONCILIATION_REQUIRED, because a save may or may not have landed.

Record check (effect check)

A read of the saved record after a save, through a different path from the screen that did the work, such as the app's API, its database, or a separate signed-in session (effects).

Recording

The folder record writes: the steps, screenshots, and what each step acted on. Password fields and fields marked secret are never saved.

Repair candidate

A changed program proposed after a halt. It must pass review, a replay campaign, a fault campaign, approval by a person, and a canary before it becomes active.

Right-record check (identity check)

A check before an action that the screen shows the record the inputs name, for example the right patient's banner.

Run report

REPORT.md and report.json in the run directory: every step, how its target was found, every check and its result, and the number of model calls.

Seal

OpenAdapt Cloud's signature on the receipt of a run that Cloud Execute tracked. Anyone can check it offline with openadapt seal verify.

Surface

Where a program acts: a browser tab, a Windows, macOS, or Linux app window, or an RDP or Citrix client window. A program belongs to one surface.

System of record

The application's own store of the data, such as its database or API, as opposed to what its screen shows.

Teach

Record only the steps that fix a halt and add them to the program as a guarded branch for that situation (teach, teach-branch).

Worklist

A CSV or JSON list where each row is one run's inputs (for-each, run --worklist).

The outcomes a run can end with, such as VERIFIED and RECONCILIATION_REQUIRED, are listed in Run outcomes.