The Schema

A normalized session is an ordered list of events. Exactly one payload field is set per event, matching its kind:

Points worth knowing link

• The event vocabulary is shared. Event kinds and tool classifications follow the Agent Client Protocol’s session-update vocabulary where an analog exists. Token usage fields follow OTel GenAI semantic conventions. Every field in the schema is tagged with its provenance (acp, otel, or ext), enforced by a test, so the mapping cannot rot.
• Token fields carry provider semantics; totals carry a derived comparable field. The field names are shared, and the meanings follow the provider: OpenAI-style usage (codex) reports input_tokens as the total prompt with the cache fields as subsets of it (the native cache_write_input_tokens maps to cache_creation_input_tokens), while Anthropic-style usage (claude-code) reports it as only the uncached remainder alongside disjoint cache read and creation fields. Adapters preserve what the harness recorded, so input_tokens is never directly comparable across harnesses (and Copilot CLI records no per-message usage at all, so its totals are omitted; see Comparing Token Counts). Session totals therefore also carry total_prompt_tokens, derived from each harness’s documented convention: that is the number to compare. See the worked comparison for a real case where the naive input_tokens reading is wrong by four orders of magnitude, and Comparing Token Counts for the full field guide (caches, reasoning tokens, missing usage, and cost caveats).
• assistant_message is the accounting anchor. Harnesses may split one API message across many records with usage written as a growing snapshot. The adapter folds them and takes the final snapshot. Exactly one assistant_message is emitted per API message, even when all of its content became thinking or tool_call events, so token totals are always derivable. Events from the same API message share a message_id. When a harness interleaves a subagent’s conversation into the parent transcript (copilot), every event of that conversation carries the subagent’s agent_id; it is empty for the main agent.
• One session record covers one transcript; a task can span several. A harness that delegates to subagents writes each subagent conversation as its own transcript, and each parses to its own session with its own totals. Where the harness records it (claude-code, codex), a subagent’s meta shares the parent’s session_id and carries subagent_id with is_subagent: true, so grouping a task is a session_id match. Copilot CLI is the exception: its subagents write into the parent transcript, so one session record already covers the task, and agent_id is what separates the parent’s events from each subagent’s. Task-scope numbers come from stats --include-subagents (or agentminutes.Task), which gathers a session’s transcripts and reports an aggregate plus a by_agent split; a single session’s stats carries by_agent itself when its events span more than one agent. See Subagents.
• tool_call and tool_result stay separate, in stream order. Ordering is data. Interleaving, parallel tool execution, and retries are visible in the sequence. Session.ToolInteractions() provides the joined view, including unanswered calls and orphaned results.
• Results carry what the model saw, plus what the harness knew. tool_result.content is the post-pipeline content the model actually received. Harness sidecar data rides along verbatim in enrichment, with retrieval metrics promoted to fetch (URL, raw bytes fetched, status, duration) when present. For summarizing pipelines like Claude Code’s WebFetch, comparing fetch.raw_bytes against the content size measures the pipeline’s compression directly. Which field the model saw is a per-tool decision the adapters pin with a test: every long string in enrichment that the content does not account for is listed with its reason (the raw file behind a numbered rendering, a UI diff, the call’s own input, output the harness persisted instead of showing). An image the model saw is an image content block holding the harness’s own descriptor: the bytes inline for Claude Code, a path beside the transcript for Antigravity, and a reference to a separate asset record for Copilot CLI, so the block is always present and its size is not comparable.
• is_error means the call failed as far as the model could tell, and fetch failures are the exception to compare with care. Every adapter marks a tool that could not run (a denied or absent path, an edit whose target text is missing) and a shell command that exited nonzero as failed, including harnesses that record the nonzero exit as a successful tool run. Fetch failures diverge: Copilot CLI and Antigravity flag a 404, Claude Code does not (the status is in fetch.status_code and the content is prose about the failure), and Codex records nothing that distinguishes a failed fetch from a successful one. stats.tool_errors inherits these conventions, so a cross-harness error rate needs the fetch calls separated out; see Harness Support.
• system.text is the record’s model-visible text, when it carries one. Injected context surfaces as text, bare: the system prompt, a CLAUDE.md body, a reminder, a delivered skill body, a listing. The tags of any wrapper the harness delivered it in (Claude Code’s <system-reminder>, Copilot’s <skill-context>) stay in details, while content the wrapper carries of its own (Copilot’s names the skill’s base directory and lists the files under its directory) is in text; convert --text-form delivered (library: Options.TextForm) puts the wrapped form in text instead, where the transcript records it. text is also where a diagnostic’s message lands (an abort reason, an API error, a session notice), which the model never saw; subtype tells the two apart. What deliberately stays in details only is enumerated per adapter with a reason (bytes, paths, tool schemas, echoes of text surfaced by another event), and a test holds each adapter to that list over its fixtures and the local corpus.
• Every event points back at its source. provenance carries the 1-based line range in the native transcript, and optionally the verbatim records (--keep-raw).
• The report closes the loop. report counts skipped record types (harness UI bookkeeping with no model-visible content), unknown events, and orphaned tool results. Between events, skips, and errors, every input line is accounted for.

The JSON encoding of Session and Event is the cross-language output contract; the agentminutes_schema field identifies its revision (currently 0.2.0). The Go types in session are documentation for it.