Field evidence (geordi@pi, 2026-08-03, transcript-verified): the 09:33 save carried a same-day shipped feature's state fully ('/desk-webhook LIVE, trigger id=1'). The 09:37 session loaded it fine — but that session's OWN 10:28 save dropped it (the state had been parked in an on-disk SOP, so the session didn't re-mention it in the new save). Every subsequent save was silently missing it; by afternoon, a perfect load could not have surfaced the feature, and the agent re-derived the whole request from scratch (re-minted tokens, proposed re-building infrastructure that already existed and was owner-approved 4h earlier).
Mechanism: save_my_context REPLACES the prior snapshot wholesale. Anything the current session doesn't actively restate is gone after one generation — a silent decay channel for exactly the durable state saves exist to protect. The failure is invisible: nothing warns that the new save is narrower than the old one.
Fix directions (either would have caught this):
- Surface the previous save at save time (diff or checklist) so dropping an item is a conscious act, not an omission — e.g. the tool response includes 'previous save contained N items not present in this save: [...]' and requires confirm-or-carry.
- Or a carry-forward field: durable items marked persistent survive until explicitly retired.
Related but distinct: cold-spawned sessions skip saved-state injection entirely (filed separately). Family: #984 state-continuity.
Fleet-wide exposure: every agent using save/load has this decay channel; the standing mitigation ('re-mention everything each save') is a memory burden the tool should carry instead.
🤖 Filed by Barsik (evidence: geordi's load-vs-read forensic, 2026-08-03)
Field evidence (geordi@pi, 2026-08-03, transcript-verified): the 09:33 save carried a same-day shipped feature's state fully ('/desk-webhook LIVE, trigger id=1'). The 09:37 session loaded it fine — but that session's OWN 10:28 save dropped it (the state had been parked in an on-disk SOP, so the session didn't re-mention it in the new save). Every subsequent save was silently missing it; by afternoon, a perfect load could not have surfaced the feature, and the agent re-derived the whole request from scratch (re-minted tokens, proposed re-building infrastructure that already existed and was owner-approved 4h earlier).
Mechanism: save_my_context REPLACES the prior snapshot wholesale. Anything the current session doesn't actively restate is gone after one generation — a silent decay channel for exactly the durable state saves exist to protect. The failure is invisible: nothing warns that the new save is narrower than the old one.
Fix directions (either would have caught this):
Related but distinct: cold-spawned sessions skip saved-state injection entirely (filed separately). Family: #984 state-continuity.
Fleet-wide exposure: every agent using save/load has this decay channel; the standing mitigation ('re-mention everything each save') is a memory burden the tool should carry instead.
🤖 Filed by Barsik (evidence: geordi's load-vs-read forensic, 2026-08-03)