AgentExecution uses one ordered journal for accepted execution meaning. The journal remains small enough to explain from the Entity model: accepted messages, delivery attempts and outcomes, accepted observations, state changes, owner effects, and bounded evidence references.

Context

ADR-0006.08 defines the AgentExecution interaction journal. ADR-0006.09 defines semantic operations. ADR-0006.10 defines structured-first interaction with optional terminal capability. Together they provide the required ownership model:

  • AgentExecution owns process state and the journal.
  • AgentExecutionRegistry resolves active AgentExecution instances and live process handles.
  • AgentAdapter translates provider launch, delivery, and output.
  • Terminal and TerminalRegistry own PTY transport and terminal recordings.
  • Owning Entities accept scoped meaning through their Entity behavior.

The journal vocabulary stays inside that ownership model.

Decision

AgentExecution journal records use these families as the baseline vocabulary:

  • accepted messages.
  • delivery attempts and outcomes.
  • accepted observations, including daemon-observed observations from semantic operations.
  • state changes.
  • owner effects when the owning Entity accepts them.
  • bounded evidence references when raw transport material needs audit linkage.

Daemon-observed reads or tool events are recorded as bounded AgentExecution observations. They share the same journal ordering, idempotency, replay, and projection path as other accepted observations.

Raw terminal and provider output remain evidence until a structured AgentExecution path accepts a record into the journal. Replay depends on accepted journal records, not on parsing raw output.

Consequences

  • The AgentExecution journal stays explainable without a separate taxonomy map.
  • Semantic operations leave bounded audit records as AgentExecution observations.
  • Replay remains deterministic because accepted journal records carry meaning.
  • Terminal recordings and provider streams remain inspectable evidence outside semantic replay.

Implementation Rules

  • Keep journal schemas in the AgentExecution Entity module.
  • Keep semantic operation audit records as bounded AgentExecution observations.
  • Keep raw transport material as evidence unless a structured AgentExecution path accepts it.
  • Keep projection and replay behavior derived from the same journal record schemas.