v0.61.1

Added

• A reopened turn shows what it was actually charged, beside the estimate. The receipt's usd is priced from the catalogue row of the model id, but one id is served by several routes at different prices. The router's generation record carries total_cost, the real charge. The replay that already learns a turn's route now also sums that charge over every call of the turn.

  • The answers found so far are kept between reopens.
  • billed_usd is written only once every call has answered, so a partial sum never shows a cheaper turn than the one that happened. A record without a number is unknown, never free.
  • The route and the bill share the existing 2.5 s budget, so reopening a conversation is no slower.
  • The Code screen draws billed $X beside the estimate, in ten languages. It never draws it on an external turn. (#590)

• Study 24's judge read is finished: bench/review_judge reports arm C out of sample. The read was pre-registered on 2026-08-20. Arm C's lead over arm A in Youden's J is +9.7 pp [+1.7, +17.8] on the 814 out-of-sample items, above the in-sample +6.8, so the clause that would have reinterpreted arms A–E does not fire.

  • Arm C still does not ship: its false-rejection rate is 43.5%, against a 20% ceiling.
  • Unregistered finding: re-judging the same 105 items changed arm C's verdict on 23 of them. (#588)

Fixed

• execute_code runs a Python the machine has, instead of the literal name python. Any machine without a python alias answered exit 127 on every call. That includes Debian and Ubuntu without python-is-python3, and current macOS. Measured on the harness-bench traces: 743 of 748 failing calls were python: not found, so the tool never ran in the #453 factorial or the study-24 benches. The effect was the same in every arm, so those comparisons stand. The interpreter is now chosen by where the code runs:

  • on this machine, the interpreter that runs Chimera, by absolute path;
  • in a frozen build, the first of python3/python on PATH (python/py on Windows), with an error that names what it looked for when none is found;
  • in a running container, whichever of the two the container has. (#587)

• The loop breaker stops the same failure repeated under different args, even when the tool reports it only in its answer. Until now it recognised a failure only by the loop's error: flag. So a run_shell given its command as a list ([exit 127] /bin/sh: 1: [bash,: not found, whatever the list held) and a ZeroDivisionError repeated over four pieces of code ran until max_steps.

  • Now [exit N] followed by text, and a last line shaped like an exception, count as failures. A silent non-zero exit, such as a grep that found nothing, does not.
  • Replayed on about 2,500 recorded traces (bench/tool_loop_silent_failure, pre-registered, US$ 0): the rule stops exactly those three runs, brings back none of the 86 distinct-edit false alarms, and would stop none of the 76 runs that continued under the fixed breaker. (#586)

• When the taint ledger refuses a call, the refusal now says why and names the way out, as the policy kernel's refusal has since 0.58.0.

  • With no approver: the refusal says nobody could be asked, that retrying will be refused identically, and names the ways through: the Code screen's card, chimera solve at a terminal, a run started with pause-on-taint, and CHIMERA_TAINT_NARROW=0 on a deployment that must act on its own. These are written for the person, and the model is told to relay them, not to route around the gate.
  • With an approver who said no: the refusal stays "Nobody approved it." (#589)