Skip to content
AI usage dashboard

synthetic-fork-inherited._expectations.md (SYNTHETIC)

All samples in this directory are synthetic, rather than extracted native sessions. This covers cross-file deduplication of inherited fork entries. Structure follows Pi session JSONL v3. Fixed session-manager.ts source copies verbatim all non-header entries into the new file, preserving id/parentId/timestamp; the new header records parentSession.

Scenario and expectations

  • Source ..._syn-sess-src.jsonl, four lines: session header id=syn-sess-src, model_change, assistant syn-fa-1 (100/50/0/0/150), assistant syn-fa-2 (200/60/40/0/300).
  • Fork ..._syn-sess-fork.jsonl, five lines: header id=syn-sess-fork, parentSession=syn-sess-src, verbatim source L2–L4, then new assistant syn-fa-3 (10/5/0/0/15; timestamp 2026-01-05T11:00:01.000Z).

Expectations

  • Two scans, source alone followed by the fork, give three calls: syn-fa-1, syn-fa-2, syn-fa-3. Copies do not count twice.
  • Summary (2026-01-05): input_total_known=350 (derived input+cacheRead+cacheWrite: 100+240+10), cache_read_known=40, output_total_known=115 (50+60+5), total_tokens_known=465 (150+300+15).
  • New syn-fa-3: session_id=syn-sess-fork, parent_session_id=syn-sess-src.
  • Event key: pi:message:<id>:<parentId>:<timestamp>; absent parentId uses “-”.

Known difference (measured)

Copied entries acquire the fork’s session_id/parent_session_id from its local header. They therefore have the same key but different content from stored source events: conflict, retaining the first-scanned event, rather than the capability declaration’s Keep for identical content. Statistics still count once and retain the first event; outcome.conflicts=2 and both stored events have conflict=1. See gap-test comments and the delivery report.