[aoreshkov]

Summary

• Cost estimate before a capped run. A needs-confirmation stop now carries estimate: the likely and worst-case agents and tokens for the pairs in wouldRun and for all pairs. next tells the caller to show it when asking the user.
  • At the defaults, 12 pairs come to about 32 agents and 1.1M tokens, at most 85 agents and 2.4M.
  • The constants are per-agent averages from a real 62-pair run: 43k tokens per reviewer, 26k per verifier, 1.3 findings per reviewer, 8.5% claimed critical or high. The result says they grow with the repository's size.
• Measured token budget. Sizing under a +500k-style budget uses measured output per agent (4k per reviewer, 2k per verifier, 25k fixed) and now follows votes, maxVerifyPerJob and verify. A job costs 16k at the defaults instead of a flat 25k, so +500k runs 29 jobs instead of 19.
• Shorter listing entry. whenToUse drops each argument's shape (the bad-args messages already give it) and points one-branch or one-PR reviews at /code-review.
  • The entry is 932 chars, down from 1,366 in 0.4.0. A test keeps it under 993, half of 0.2.2's 1,986.
  • On 2.1.312, claude plugin details reports ~152 tokens because it counts only the agents. /context puts the agents at 169 and an installed 0.3.1 entry (1,303 chars) at ~440–460 tokens. The 0.5.0 entry works out to ~320 tokens at the same ratio (derived, not measured).
• README. New "When to use which" section, the estimate, the re-derived token budget, measured run costs, and reviewer-model advice: model: { reviewer: "opus" } for a release gate, the defaults for routine sweeps.

Eval (2.1.312, one real run)

The shorter whenToUse could stop the workflow being picked, so the seeded-defects case ran once as a real session against this branch: 11/12 graders, workflow-used passed (one Workflow call with matching input), 5/5 defects and 1/4 decoys confirmed, 119 s, $1.84. The decoy count is within 0.4.0's range on the default models.

Test plan

• node --test (206 pass)
• claude plugin validate --strict for ., ./review-kit, ./token-audit, ./token-meter
• claude plugin test ./token-meter (40 pass)
• Seeded-defects real run as above

After merge: tag review-kit-v0.5.0 with notes from the changelog, then confirm the listing figure with /context on an installed 0.5.0.

🤖 Generated with Claude Code