[zolotarev]

Frictions (seen on #36, owner's report: it stayed in rework for 5 rounds):

1. The PR title and body arrived in Russian. contribute sends the cluster title and --why, which harness-retro writes in the project's language. The review asked for English every round, but the rework agent has no way to edit a PR, so it skipped the item every time.
2. Budgets: main measures code_lines 6024 against a 5906 budget, and prompt_tokens 7035 against 6898. Both sit at the edge of the 2% tolerance, so any PR adding a few lines crosses it. The rework agent raised the budget and the review rejected that, round after round.

Change

• skills/harness-retro: titles and --why in English (+13 prompt tokens).
• review.yml rework step: Bash(gh pr edit:*) plus one prompt line, so the agent can fix the title and body itself.
• tools/budgets.json: code_lines 5906 → 6024 and prompt_tokens 6898 → 7035, i.e. main's measured values, not padding. The ratchet starts from reality again: nightly lowers it, and the tolerance absorbs small PRs.

VERSION, CHANGELOG. 224 OK.

🤖 Generated with Claude Code

[github-actions]

VERDICT: accept

• Real friction (PR decide --plan duplicates a decision already written in the plan #36 stuck 5 rounds): non-English contribute title/body and a rework agent that could not edit them. The change addresses both causes; the harness-retro line is +13 tokens and fits the tolerance.
• tools/budgets.json: the raise is justified. It re-bases to main's measured values (6024 / 7035) rather than adding padding. The nightly ratchet lowers them again.
• .github/workflows/review.yml: Bash(gh pr edit:*) is a small trust widening. PR content is untrusted, so the rework agent could be steered into rewriting the title or body. That is bounded to this one PR and the agent already has Edit/Write, so I find it acceptable.
• Nit: lib/dpcli/__init__.py:9 and CHANGELOG.md:3 jump from 42 to 44 and skip 43. Harmless, but the next bump will not be contiguous.
• Nit: no test covers the change. It touches only a skill text, CI config and budgets, so a test adds little. I did not run tools/metrics.py --check or the test suite here; the PR reports 224 OK.
• Nit: this edits a workflow file, so a rework agent cannot push follow-ups to this branch (CI token limit). A human will need to handle any rework.