check_diff / check_list_files + fix should optionally allow running check to show errors

[astei]

This is just an idea I had for making hk check --all more useful on large repos, but it applies more generally to other development workflows.

We use hk on a repo with a lot of files (120,000+), and in particular, we tend to use linters that don't expose a clean check_list_files / check_diff interface (i.e. ruff-check). When hk processes a chunk of files for a given step and the checker exits with a non-zero status / finds errors, and only check + fix is provided, hk emits a fix command that is quite long (since it doesn't know which file emitted the error). We can work around this to some degree by wrapping ruff's output with a small script and emit only the unique files with violations, which neatly condenses the proposed fix command, but we lose the fidelity of error output as a result.

My proposal is to add a new option for Step (maybe call it check_verbose) that allows hk to run a file-scoped check command if check_list_files / check_diff ran and exited with a list of files that have problems.

This is particularly important for agentic development flows (a long fix command just wastes tokens for no good reason, and here, it would likely be looking at output from a failed CI run), but it's also a tangible improvement for humans who want to look at hk output without using the LLM.

[farfromexact]

The idea is useful, but I would name the option after its behavior rather than check_verbose. Verbose usually means �print more of the same output,� while this feature performs a second, file-scoped check.

Something like check_on_failure or check_failed_files would make the contract clearer:

1. Run check_list_files or check_diff as today.
2. Deduplicate the returned paths.
3. If the checker reports a failure, run the file-scoped check only for those paths.
4. Preserve the original checker output for diagnostics.
5. Use the focused file set when constructing the suggested fix command.

The option should be opt-in because not every check command accepts file paths, and some tools have expensive startup costs that would make a second pass worse. It should also define what happens when the file-list command succeeds but returns no paths, when paths are outside the repository, or when the file-scoped check returns a different status from the aggregate check.

For large agentic workflows, the biggest win is probably a structured result internally: failed paths, diagnostics, and a compact human-readable rendering. That avoids forcing the user or an agent to parse a giant command line. A regression test with a 120,000-file fixture can verify that the output stays bounded while the original error details remain available.

[jdx]

Implemented in PR #1123. It adds the opt-in check_failed_files setting, which uses check_diff or check_list_files to identify the affected paths and then runs check only on that focused set for full diagnostics. The implementation also bounds oversized fix suggestions and reapplies ARG_MAX batching before the focused command. Tests cover focused diagnostics, inconsistent exit statuses, validation, bounded suggestions, and large-file batching.

This comment was generated by Codex.