[lgrawet]

A restarted session adopts the executors a dead session of its strategy left RUNNING (CORR-726): _adopt_running_executors feeds them into the tick's snapshot, the stop gate allows them, and the risk gate counts their exposure. Three other reads still used only the provider's rows, which are fetched under this session's own controller_id — so they reported a book smaller than the one the gate protects:

• the journal snapshot recorded open=<provider only>;
• the journal summary (what the agent reads back) showed the same short count;
• the canvas nudge was told the short count too.

_book_totals() is now the one place the provider rows and the adopted rows are summed, and the gate, the snapshot, the summary and the prompt all read it.

The snapshot's exposure deliberately stays the provider's own. It is the denominator of get_drawdown_pct_checked, which divides the provider's PnL — a figure that excludes adopted executors — so counting adopted size there would dilute the percentage (a $100 loss on $1,000 of own positions against $9,000 adopted would read 1% instead of 10%) and configured pause/shutdown limits could stop firing. The book exposure still feeds the risk gate, where it is only compared against limits and never divided by a session-scoped PnL.

Tests: three cases in tests/test_executor_adoption.py pin the arithmetic (_book_totals ignores malformed rows), the snapshot's book count, and the scoped exposure. Full file: 40 passed. Full suite: 8050 passed; the 6 failures present are identical on a clean upstream/main and unrelated (conversation listing, host load, timing probes).

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; the previous loss-limit issue is fixed.

Summary

Counts adopted executors alongside this session's own executors through _book_totals.

• Restarted sessions now show adopted executors in their book counts.

_{Reviews (2) · Last reviewed commit: "fix(agents): keep drawdown's exposure in..." · Reviewed by Greptile}