Activity ledger

Each tool call is recorded as one or more effects, read from the call's own record (its tool, its redacted arguments and the gateway's reading of the action), never guessed.

• Verbs: read, search, create, write, delete, execute, send, transact, grant, deploy, call.
• Resource kinds: file, directory, repository, table, host, command, record, recipient, agent, secret, cloud, package, tool.

Code-hosting calls name the issue, comment or pull request they touch; only pushes, merges and branches count as repository changes. Setup messages, model requests and verification probes add nothing.

On each agent's page (Activity), on each session page, and under Runtime → Activity ledger for the whole organization, as a summary, a list of changes or every effect. The contents of edits and writes are shown only to people allowed to view raw prompts.

When the model proposes a tool call and the agent then runs it, that is one action, not two. A call that was held for approval and then made folds into the later record. A held call that never ran shows how its approval ended: denied, approved but not run, or expired. Two separate allowed runs stay two.