Signet stores transcripts as JSONL under the workspace at memory/{harness}/transcripts/transcript.jsonl
. The file is appendable during live sessions, easy to copy from harnesses that already persist JSONL, and stable enough to become the input for transcript-based summary, fallback, and lineage flows.
The daemon accepts three transcript sources. If a hook provides a transcript path or inline transcript, the daemon normalizes the conversation into canonical JSONL records and replaces the prior slice for that agent, harness, and session. If a harness only calls prompt-submit with the current user turn, the daemon appends live user and previous-assistant turns so active sessions still appear before session-end. If an existing install only has legacy markdown artifacts or database transcript rows, the daemon backfills those records into the same JSONL location before new writes.
The canonical record schema is signet.transcript.v1
. Each line carries the agent id, harness id, session key/id, project, sequence number, role, content, capture timestamp, source format, optional source path, and source hash. This keeps one transcript substrate across Claude Code, Codex, OpenCode, OpenClaw, Hermes Agent, Gemini, Oh My Pi, and Pi without relying on any one harness’s native log shape.
Markdown transcript artifacts remain readable as legacy inputs, and the session_transcripts
table remains a compatibility/indexing surface while existing retrieval and FTS paths migrate. New transcript persistence must write JSONL first and treat markdown transcript artifacts as historical compatibility, not the forward source of truth.