conversation-flush-jsonl-transcript-pipeline

Claude Code session transcripts are stored as JSONL at ~/.claude/projects/<project-hash>/<session_id>.jsonl with standard message.role + message.content structure. The transcript-capture Stop and PreCompact hooks read session_id and transcript_path from stdin (hook context), spawn conversation-flush.py as a detached background process, which uses the Agent SDK to extract 3-10 insights and submit them to vault. Cost is ~$0.02-0.05 per session.

Related

• clawteam-openclaw-multi-agent-swarm-evaluation
• 2026-04-04-oracle-001-self-architecture-analysis
• enterprise-capability-expansion-5-pillars-from-digital-employee-analysis
• cache
• claude-code-to-nova-20260405-163418 (archived)
• vault-hygiene-four-lane-architecture
• jsonl-transcript-path-and-structure-for-flush
• transcript-capture-detached-background-process-pattern
• vault-lifecycle-four-lane-not-three-lane
• precompact-hook-as-transcript-capture-safety-net
• transcript-capture-dedup-window-120s