[forkrul]

IMP-1 from the Antigravity review.

Before: hone reviewed one unit per task, and every round re-ran the full panel on every task, including tasks that had already had two clean reviews and that no fix touched.

Now (skills/hone/SKILL.md, Review Units and Convergence):

• Packing. Consecutive tasks, in quench-log.md order, are packed into review units of at most ~400 changed lines. A single task over the budget is still split by commit and logged as anvil feedback. The units are listed at the top of code-review.md.
• Per-unit review. The panel, the judge and the overlap signal run per unit. A unit is done after two consecutive clean reviews. Round 1 reviews every unit; later rounds review only units not yet done.
• Fixes re-open units. A fix makes every unit whose files it touches not done again, and its diff joins that unit (re-split if over budget). So "round N+1 reviews round N's fixes" still holds.
• Feature rating. Each round still logs a feature - Rating: line, the lowest unit rating (a done unit counts as A++). Smithy's state machine, the example gate table and CI all read that line, so nothing downstream changes. The 3-round cap is unchanged.

Worked example (examples/…/code-review.md) updated, with the same ratings (C → A+ → A++). Units are U1 = T001+T002 (246), U2 = T003–T005 (393) and U3 = T006 (127). Round 1's fix touches router.py, which U3 also changes, so U3 is re-opened. With U1's carried UNCONFIRMED check, all three units are reviewed in every round. The example shows the rules applied honestly rather than a saving: here the panel goes from 6 per round to 3 per round (9 critic reports instead of 18). The review's "-61%" estimate was a hypothetical and isn't claimed.

quench's log description is updated to match. Hone stays at 257 lines. Local: lint, sphinx zero warnings, smoke, dogfood, eval 100.00, example trail check.

Together with #33, this changes the hone contract, so the next release is 0.2.0.

🤖 Generated with Claude Code