Glossary
These definitions connect the product language used in the guides with the precise names exposed by the CLI, stored data, digest, and Task Review.
Task record
The complete user-facing record Orcaops builds while an agent works: the plan and its revisions, checkpoints, decisions, verification, evaluator results, uncertainty, summary, and related review evidence.
Artifact
The CLI and storage name for one captured task thread. An artifact has a stable ID and lifecycle. A branch or worktree can contain several artifacts, which a branch-level Task Review can consider together.
Plan
The intended outcome, ordered steps, acceptance criteria, boundaries, and decisions captured before implementation. Revisions extend the same artifact rather than erasing the earlier plan.
Checkpoint
A coherent unit of implementation recorded within an artifact. Its open and close boundaries connect changed code with the plan steps, decisions, verification, uncertainty, and completion evidence reported for that work.
Completion check
The user-facing last look at finished work before accepting it. Evaluators can compare the task record with its requirements, Task Review can independently inspect the code and captured account, and the digest can present the resulting evidence. It is not a single CLI command or an automatic merge verdict.
Evaluator
A check that runs at a defined lifecycle phase and records a pass, violation, informational result, or skip. Evaluators are distributed in packs and are optional; fresh configuration installs no packs.
Rule-based check
The product name for a command-engine evaluator. It runs deterministic program logic and does not need a language model.
Judgment check
The product name for an LLM-engine evaluator. It uses the selected coding-agent CLI and explicitly declared context for questions that require language reasoning.
Task Review
The branch-level human review assembled by orcaops-task-review and read in orcaops watch. Its forensic lane reviews code without the agent's account; its account lane organizes captured intent and evidence without inspecting code.
Story, Act, and Part
Task Review's reading structure. A Story is the causal account of the branch, an Act groups a major problem or transition, and a Part connects a specific unit of that account to captured checkpoints and owned code ranges.
Finding
A concrete lead produced by the capture-blind forensic review. Findings require human adjudication; their presence does not prove a defect, and their absence does not prove the branch is correct.
Digest
The reviewer-facing rendering of one captured artifact: its outcome, decisions, evaluator results, verification, open items, and usage attribution. It can also be used as the basis for a pull-request description.
Skill
An agent-facing workflow installed by Orcaops. A skill translates natural language intent into CLI calls, validates the responses, and handles the next step. The Skills guide lists every shipped skill ID.
Plan approval
The optional Cloud workflow in which a human reviews a plan on the web. The agent preserves the approved version as the task's source plan when capture begins.