ZCode session store: add schema_version — and a structured session-summary facility for manager-plane sessions (field data: 222 GLM sessions / 7 days)

Positive feedback first

The SQLite session store (per-tool-call status/error/exit_code rows) is the best-designed of the three AI CLIs we orchestrate: it let us build accurate, fail-closed session analytics with zero text parsing. Thank you.

Field data

222 GLM sessions / 7 days in a multi-vendor orchestration (GLM as the manager plane; Codex and Claude Code as writer/verifier planes). Role mix: engineering 45%, coordination 20%, design review 16%, verification 7%; among evidence-bearing sessions, completion ~70%.

Requests

1. schema_version table row in db.sqlite so third-party analytics can detect store evolution (we currently fingerprint tables/columns on every run to compensate).
2. A structured session-summary facility for manager-plane sessions: manager-style sessions coordinate, write receipts and run audits — they produce no machine-checkable artifact, so artifact-presence outcome scoring is structurally blind for them. A session-summary (deliverables touched, receipts written, decisions made) emitted at turn end would fix this.

Full field-report package (aggregate tables; tokenized structural data only — no session content):
https://github.com/oguzhanpisgin/ai-session-analytics

Happy to share more aggregate data or answer questions about our measurement setup.

Start by reviewing the SQLite session store in db.sqlite and the linked ai-session-analytics field-report package. Clarify the schema-version contract and the required manager-plane session-summary fields before identifying implementation and test entry points. Done means third-party analytics can detect schema evolution and manager sessions expose machine-checkable summaries.