v0.6.0

v0.6.0

Added

• Topic tags and verdicts now sync across your machines. Tags — the topic segments a session is split into, plus a routine verdict — used to live only in each machine's local index. They now ride the transcript archive: each machine writes its tags to per-machine files beside its transcripts, archive push/pull carry them, and a pull ingests every other machine's tags. If two machines tag the same session differently, ingest keeps one deterministic winner and surfaces the disagreement in rawclaw archive status — and every machine's tag file is retained, so nothing is lost. Agent tagging labor now travels with the archive the way raw transcripts already do.

Fixed

• Codex accepts the rawclaw discovery banner again. On Codex, setup's SessionStart hook printed a [rawclaw]… banner that Codex's hook parser rejected as invalid JSON (it treats stdout starting with [ as a JSON object), so the banner was silently dropped. The Codex hook now delivers the banner as a proper SessionStart hook-JSON object; Claude Code is unchanged. Re-run rawclaw setup to refresh the installed hook. (Uses python3 for the JSON encoding; if it's absent the banner is skipped rather than erroring the hook.)
• The tagging queue only accepts sessions that produced a transcript. Claude Code fires SessionEnd for ephemeral sessions too — opened and closed without a message ever landing — which flooded the queue with ids nothing could resolve. The SessionEnd hook now queues only when the transcript_path Claude Code hands it exists on disk. Re-run rawclaw setup to refresh the installed hook.