1. What cost the most: the shared checkout, not the code. The one test run took about 10 minutes. Three of its failures came from another session rebuilding target/fast mid-run, so I spent a rerun and a dig to clear them. Clippy was red on another agent’s uncommitted view_changes.rs, so I needed a second clippy run with that lint allowed. lowered.rs held a foreign hunk, so commit.nika couldn’t be used: I built a patch of my hunk and committed by hand. Then the badge needed a throwaway worktree to tell my share change from theirs. The move itself took two lower-std refusals and one build.
2. What the brief could have handed me:
   • Text can’t be cut at a fixed number (NK1214), and Bytes.text() can fail (NK2605); the answer is s.scalars.take(n).collect().
   • fs::write is async, so trace lowers async and the adapter needs block_on.
   • A String passes where an fs::Path is expected.
   • std’s IoError prints different text from std::io::Error, which matters for the stop condition on changed messages.
   • Which adapter it meant: refusal (plain anyhow) or refused (a refusal diagnostic).
3. Checks:
   • Useful: lower-std with the old binary (found both refusals in seconds), cargo check, one build cycle, the filtered it run plus the guard set, --lib project::, the rerun of the three failures, pre-push-check.
   • Partly wasted: the first full-tree clippy (red on foreign code). The where.nika lookup found no tests for these functions, so the it filters were broad guesses.
4. Change that would save the most: builder.md should give a recipe for shared-tree collisions, by name:
   • commit a single hunk of lowered.rs (git apply --cached), or let commit.nika take a patch;
   • run clippy with a named foreign lint allowed and report it;
   • measure the badge on HEAD plus your own diff;
   • copy the test binary before the it run.

   Each of these cost me a detour this run.

This site is open source. Improve this page.