[mountainowl]

Requested by Ajay Singh · project thread

What changed

• Before: a custom reviewer command falls back to plain-text token counters, and that parser had three bugs.
  • It took the first match of each counter, so review prose saying "input tokens: 5" beat the real summary of 10,000 printed after it.
  • It added cached tokens on top of input when it derived a total, so 10,000 input (8,000 of them cached) plus 500 output came out as 18,500 instead of 10,500.
  • Its input pattern also matched cached input tokens: lines.
• After:
  • The last match of each counter wins.
  • cached input isn't read as input.
  • A derived total is input plus output.
  • Codex and Claude runs are unaffected, because they report usage as JSON.

Verification

• tests/test_plain_text_token_usage.py drives the real parse_agent_output fallback.
• On main, the prose case read input as 5, and the cached case gave a total of 18,500. Both pass with this change.
• Two guard tests pass on both: a cached line printed after the input line, and a reported total being kept.
• The full suite on Python 3.14 gives 1931 passed. The only failures are the 2 container-only test_ui_export failures, which also fail on main.
• ruff check, ruff format --check and cz check pass.

Notes

• Output with only a cached counter, and no input or output counter, no longer gets a total equal to the cached count, because that number alone isn't a total.

---

Generated by Claude Code