ClaudeCodeTranscript

The Claude Code adapter: one agent tool's log, read as a transcript.

Only two of Claude Code's entry types carry conversation — user and assistant. Everything else in a transcript file is bookkeeping the reader never sees (file snapshots, mode changes, attachments), and is dropped.

The hard part is that a user entry is not the same thing as a person speaking. Claude Code writes the person's typed message, the expanded body of a slash command or skill, the results of every tool the agent ran, and its own notifications, all as user entries. Attributing any of those to the person would put words in their mouth, so each is recognised and shown as activity instead:

• isMeta is Claude Code's own marker for text it injected rather than received — expanded command and skill bodies, resumption prompts.
• isCompactSummary marks the summary written when a session runs out of context. It carries what came before and is worth reading, but it is the tool's account of the conversation, not a turn in it.
• a turn consisting solely of one of Claude Code's wrapper blocks is that wrapper, not a message; text that merely mentions a wrapper is not. One of those wrappers is how work left running reports that it finished, and it names the call it finished, so it is read as that call's answer. Scaffolding holds which elements are the tool's own, and takes them and the terminal's control codes off every activity body.
• tool results arrive as tool_result content blocks quoting the id of the call they answer. That id is the token the two rows share, so a reader can tell which answer belongs to which call even when other rows came between them. An answer whose call is not on this page — the page is a slice of a session — belongs to nothing a reader can see, and is shown as plain activity.

A call the agent made keeps the shape it was made in: this vendor hands over a tool's input as plain data of whatever shape the tool's own schema takes, and reading that as ToolInput is the last place the shape is known.

Sub-agent entries are a separate source with their own file, joined to a parent tool call; a sidechain entry appearing here is not part of this conversation and is dropped.