[Bug] Large workflow warning thresholds not triggered for multi-agent workflows exceeding token limits

Bug Description
Technical addendum — per-workflow measurements, session 2026-07-27/28

Support cited a "Large workflow" warning at >25 agents in a single workflow or

1.5M projected tokens. Neither threshold produced a warning at any point in

~24 hours. Measured per-workflow actuals below.

Workflow Agents Subagent tokens
investigation 11 1.55M
C1-C4 14 1.79M
C4b-C6 24 4.17M <- 1 under the agent cap
C6b-C9 16 2.77M
C10-C11 6 1.03M
full-diff sweep 6 1.61M
C12-C13 10 2.28M
C13 verify + C14 7 1.23M
C15 6 0.96M
TOTAL 100 17.38M

Two distinct gaps:

1. CHAINING. Each workflow individually respected the agent cap; ten were run

in sequence from one user request. Nothing aggregates across workflows
within a turn or a session.

1. PROJECTION ACCURACY. Six of nine workflows exceeded 1.5M in ACTUAL tokens

and none warned. The largest consumed 2.8x the token threshold while sitting
one agent under the agent threshold. If the warning keys off tokens
projected before execution, the projection appears to under-estimate badly
for workflows containing verify/fix loops of variable round count, where a
single round can re-run a ~2,000-test suite plus mutation batteries. A
projection made before the loop count is known cannot bound this.

Environment Info

• Platform: win32
• Terminal: windows-terminal
• Version: 2.1.220
• Feedback ID: 4ac69492-2882-408b-b5b0-65bd23b3a94a

Errors

[]