Codex fan-out double-counts tokens. Each fork/subagent child file replays the parent's token_count history and the parser counts every copy.
One real day: 179 files (22 root, 157 children from 9 parents, one with 137 children). Raw ~14B, CLI reports ~12.9B, deduped real is ~1.7-1.9B. 88% of the mass is the same cumulative vectors repeated across 16-17 sibling files (identical 34-step sequence 40002 → … → 5,238,474). That tripped the daily cap on submit: Daily token total exceeds 10,000,000,000 ... 12,910,573,877.
Two guards miss it when the child doesn't re-embed the parent session_meta:
forked_child_turn_starts_own_session returns true on replay_session_id.is_none(), so the first turn_context closes the skip window and the replayed rows get counted.codex_token_count_dedup_key keys on the child's own session id, so sibling replays don't collapse.
Naive fix (dedup by cumulative vector / key under parent) breaks test_..._deduplicates_parent_replay_across_forks: two subagents doing distinct same-size turns share a vector and must stay separate. Probably needs a cross-file pass: look up the parent's total at fork time and skip child rows at/below it.
crates/tokscale-core/src/sessions/codex.rs, should_keep_deduped_message in lib.rs. Closed #678 (thought it was a strict cap). Can send a PR if you tell me the fork-log semantics you want.
Codex fan-out double-counts tokens. Each fork/subagent child file replays the parent's
token_counthistory and the parser counts every copy.One real day: 179 files (22 root, 157 children from 9 parents, one with 137 children). Raw ~14B, CLI reports ~12.9B, deduped real is ~1.7-1.9B. 88% of the mass is the same cumulative vectors repeated across 16-17 sibling files (identical 34-step sequence
40002 → … → 5,238,474). That tripped the daily cap on submit:Daily token total exceeds 10,000,000,000 ... 12,910,573,877.Two guards miss it when the child doesn't re-embed the parent
session_meta:forked_child_turn_starts_own_sessionreturns true onreplay_session_id.is_none(), so the firstturn_contextcloses the skip window and the replayed rows get counted.codex_token_count_dedup_keykeys on the child's own session id, so sibling replays don't collapse.Naive fix (dedup by cumulative vector / key under parent) breaks
test_..._deduplicates_parent_replay_across_forks: two subagents doing distinct same-size turns share a vector and must stay separate. Probably needs a cross-file pass: look up the parent's total at fork time and skip child rows at/below it.crates/tokscale-core/src/sessions/codex.rs,should_keep_deduped_messageinlib.rs. Closed #678 (thought it was a strict cap). Can send a PR if you tell me the fork-log semantics you want.