Codex usage feed is STALE, not missing - repair the collector cadence so codex is tracked alongside claude
Elazar 2026-09-12: 'we need to track codex too'. This is a REPAIR, not a build - the feed already exists and already lands in the same Postgres store. MEASURED 2026-09-12 03:27Z via fleet-tokens.sh: codex_token_count newest row 1005 minutes old, collected 13 minutes ago, 15839 rows over 30d - so the COLLECTOR is running and the SOURCE has stopped producing. Per-host: whey/codex last update 2026-09-11 02:00 UTC (25.7h ago), venus/codex 2026-09-11 10:00 UTC (17.7h ago), both past the 6h staleness threshold. lezama/codex is MISSING outright - zero rows in 30 days, which is a different state from stale and must be diagnosed separately. Note fleet-tokens.sh is already honest about all three states and prints STALE/UNFED/MISSING rather than zeros; do not 'fix' this by making an absent feed render as 0. ALSO UNATTRIBUTED AND WORTH MORE THAN THE STALENESS: over the last 7 days a single '(unresolved)' codex row across venus+whey carries 15.98M net-new tokens over 3228 responses - rank 2 in the entire fleet, behind no named agent. A second '(unresolved)' claude row across lezama+venus carries 3.96M. So roughly 20M net-new per week is recorded without an agent name. Attribution is the point of the single-lane design, so find why those rows resolve to nothing: is the sender not setting its agent name, or is the join failing. Acceptance: name the root cause per host, and show a codex row landing with a resolved agent name. Diagnose before changing anything - a cadence fix that leaves attribution broken solves the smaller half.
Activity
-
PROMOTED FROM HOUSEKEEPING TO A PRECONDITION, 2026-09-12. Elazar has ruled that -ca (codex) agents will carry a share of fleet work specifically to split cost across the Claude Code and Codex subscriptions, with the codex plan resetting Sep 15 2026 12:16 PM GMT-3. That makes this row load-bearing rather than tidy. The decision it gates: how much work moves onto codex, and when. Right now that cannot be answered, because the codex feed is stale (whey 25.7h, venus 17.7h, lezama MISSING) and approximately 20M net-new tokens per week land as '(unresolved)' with no agent attached. Nobody in the fleet can currently state what is left on the codex plan, and no -ca agent is running anywhere, so there is no live pane to read a meter from either. Add to the acceptance: the fix must surface REMAINING CODEX ALLOWANCE and its reset boundary, not only per-agent token counts. A ranking that cannot answer 'how much codex is left before Sep 15 12:16' does not close this row, because that is the only question the cost-split decision actually needs. Same honesty requirement as before - STALE, UNFED and MISSING stay visible as distinct states, and an empty codex row must never render as zero burn.