PlanGate article: switch AI sessions by work boundary, not usage

The piece is built on two named failure records from the author's own repositories. In April 2026, a one-off handoff.md written at task completion twice reached the next session in an unfinished state, on 2026-04-24 and 2026-04-26, and a retrospective noted the same problem had appeared on 2026-04-24 without the fix sticking. In September 2026, all three premises in a handoff memo turned out to be wrong: 12 of 13 items listed as awaiting a decision had already been closed 13 days earlier, a stated ordering constraint was reversed, and two issues described as linked had no such link outside a body-text reference.

The author also reports that a source-of-truth file is not immune. A PlanGate index file, INDEX.md, drifted from reality three times inside a single session on 2026-07-31, tracked in issue #945: after plan approval it still read as awaiting approval, after a re-edit it still read as approved when re-approval was needed, and after execution it read as a clean pass when the result was a conditional pass with a warning. The fix was to add a rule requiring updates at every phase change rather than only at creation.

A separate record, issue #1061 from 2026-08-12, describes an orchestrator AI dropping a timeout detail it had just written from a delegation message; the worker then stalled for 600 seconds. This is tied to an absent mechanism for enforcing the delegation procedure, not to the choice between summaries and state files. Its prescribed reading order for a new session is the entry index, then current state, then details only when a specific question exists, and it notes that without PlanGate the same routine reduces to maintaining a single current-state file and pointing new sessions at it.