improve

asks the executor to commit its work, then checks scope with a command that only shows unstaged changes. At , lines 30 and 55, a completed, clean executor worktree produces an empty even when its commit modifies a file outside the plan. The template's criterion has the same limitation. Reproduced with Git 2.45.1 on Windows in a disposable repository and separate executor worktree: the baseline contains , the plan permits only that file, and the executor commits a new . The results are: Could the review instructions capture the executor worktree's starting commit before execution and compare that base with its final HEAD for both scope and full-diff review? Keep the status check separately for staged, unstaged, and untracked leftovers; apply the same committed-change check to the plan template's scope criterion. Related work: closed, unmerged #16 included a base-to-HEAD comparison as part of its worktree-base fix; #20 changes review order but retains the bare diff; #26 adds a base comparison for batch mode while leaving single-plan execution unchanged. This report concerns the existing single-plan path and includes a concrete reproduction of the remaining gap.

The problem tells the advisor to "dispatch a cheaper executor subagent" — but that's an adjective, not a value it can pass. The only concrete default () lives in , a file the skill says to read "before the first dispatch." In most agent hosts, an omitted model parameter inherits the parent session's model. So an advisor dispatching from alone passes nothing, and the "cheaper executor" runs on the advisor's own model — the exact inversion of the economics the README describes (). Verified empirically on Claude Code: The override mechanism works fine. The instruction to use it is what's missing. The failure is also intermittent and invisible: whether you get the downgrade depends on whether the advisor happened to read the reference file before dispatching. Both paths narrate "dispatching a cheaper executor subagent" in the transcript. Nothing distinguishes them except the bill. The change — the bullet now states the requirement where the instruction actually lives, instead of depending on a reference file having been read first. — the bare default becomes a mandate, names the failure mode explicitly, and adds a hedge for hosts that can't select a model per subagent. Wording is deliberately vendor-neutral ("most hosts inherit the advisor's model"), since the README says the skill works in any Agent Skills host. Concrete model names stay a scoped parenthetical example, matching the existing convention at ("in Claude Code: Explore agents"). gains no vendor-specific model names. Docs-only: 2 files, +10 / −2. Nothing outside the dispatch path changes. Not included The audit subagents in Phase 2 () have the same underlying defect — six mandatory items listed for every subagent prompt, none of them a model, and fans out up to 8 concurrent agents that all inherit the advisor's model. Left out to keep this PR single-purpose. Happy to follow up if you'd like it fixed the same way.