Three tests in test/agent/loop/autonomous_goal_loop_test.exs fail on main:
:462 — "with the budget at 0 the turn made 2 round-trips — the cap is not actually capping" (expected 1):310 — "an unfinished goal drove 8 round-trip(s); expected 1 answer + the shared budget of 3" (expected 4):476 — "an anchored goal drove 8 round-trips against a budget of 3" (expected 4)
Every goal turn burns roughly 2.6x the intended provider round-trips, in real money and latency. This is the same failure class as the previously-fixed inert per-subagent max_iterations.
Likely related: react_loop.ex:1160 halts via TerminalSource.halt/3 after a full generation, bypassing the message/accounting work every sibling clause in that cond performs.
Shipped knowingly in v1.0.177; the Elixir suite is red on main as a result.
Three tests in
test/agent/loop/autonomous_goal_loop_test.exsfail onmain::462— "with the budget at 0 the turn made 2 round-trips — the cap is not actually capping" (expected 1):310— "an unfinished goal drove 8 round-trip(s); expected 1 answer + the shared budget of 3" (expected 4):476— "an anchored goal drove 8 round-trips against a budget of 3" (expected 4)Every goal turn burns roughly 2.6x the intended provider round-trips, in real money and latency. This is the same failure class as the previously-fixed inert per-subagent
max_iterations.Likely related:
react_loop.ex:1160halts viaTerminalSource.halt/3after a full generation, bypassing the message/accounting work every sibling clause in thatcondperforms.Shipped knowingly in v1.0.177; the Elixir suite is red on
mainas a result.