Glossary

The terms these pages use, in the sense AOP gives them.

  • Harness: a coding agent's own command-line program, such as Claude Code (claude) or Codex (codex). AOP runs the harness and keeps its login, models, tools, and sessions; it does not reimplement any of them. --agent selects one.
  • Run: one invocation of a harness, by aop run or aop resume. It gets a UUID run ID and a directory .aop/runs/<run-id>/ holding its request, result, and logs.
  • Task: a named piece of work in a repository, such as task-a. Each task has its own Git worktree, and its runs share it. Sealed runs need no task name.
  • Session: the harness's own conversation. A run records its harness's session ID.
  • Exact resume: aop resume continues the same harness session with the same harness, model, effort, provider, and profile as the run it resumes, rather than starting a new conversation with similar settings.
  • Profile: the boundary a run gets, chosen with --profile: edit (default), review, sealed, or host. See Execution profiles.
  • Sealed run: a run in an empty private workspace that sees only the files passed with --input, and no repository or user instructions.
  • Input snapshot: a read-only copy of a file or folder passed with --input, taken when the run starts and mounted at /inputs/<name>.
  • Artifact: a file the run must produce in its output directory, declared with --artifact. AOP checks it exists and archives it with the run.
  • State root: a .aop folder that holds run records. Each run's result names its state root.
  • Inference provider and route: the service a harness sends model requests to, and the endpoint and credential it uses. Most harnesses have one; Codex, Hermes, and DeepSeek Harness can be pointed at others with --provider.
  • Projected credential: the single login AOP copies into a run's private harness home, out of everything the harness has stored, so the run can sign in.
  • --no-web: the policy that denies the model web access while keeping its inference connection; see CLI guide.
  • Participant mode: an experimental Hermes mode for one turn in a structured exchange, such as a game move, rather than open-ended work.
  • API-equivalent cost: what a run's tokens would cost at the provider's public API prices, from the model-prices data. It is a comparison figure, not your bill: subscription runs cost nothing extra per token.
  • Accounting status: whether a run's token usage is complete, partial, or unavailable; see Token usage and pricing.
Edit this page