opencode

opencode

The adapter accepts one native OpenCode session export as a single JSON document:

{ "info": { "...": "..." }, "messages": [{ "info": {}, "parts": [] }] }

This is a whole document even when a corpus stores it with a .jsonl extension. info.id, info.directory, and info.time.created provide the session identity, working directory, and creation time. Message metadata provides role, model, and message time.

Message parts decode as follows:

• text becomes user or assistant prose.
• reasoning becomes a reasoning record.
• tool becomes one call and, when the state is terminal, its linked result. callID and tool are preserved; state.input is serialized as arguments; state.output or state.error supplies result content.
• completed and error states map to ok: true and ok: false. A running call without output remains an unfinished call.

Native part IDs are the preferred canonical source identity. Older or hand-built exports fall back to message IDs and then whole-document part ordinals. Part timestamps take precedence over message timestamps; a later part without its own time inherits the latest time in that message so the native part order remains stable.

step-start, step-finish, patch, snapshot, file, and subtask parts are transport/state records and are dropped. Other unknown part types are dropped with a diagnostic.

Local-store discovery is intentionally not provided. The caller supplies an already assembled OpenCode session export to normalizeTranscript({ source: "opencode", transcript }).