Edit in small, well-formed diffs — and verify each edit before the next.

The single biggest jump in measured edit success (Aider: 20%→61%) came from edit format: targeted search/replace diffs that apply cleanly, not whole-file rewrites that drift. A small model loses the thread on a full-file rewrite; a scoped diff keeps it honest.

Rules

1. Prefer diff / search-replace edits over whole-file rewrites. Change the minimum lines needed. Hermes' native edit tool renders inline_diffs and runs a file_mutation_verifier — use that native edit path; do NOT paste an entire regenerated file when a few lines change.
2. One coherent change per edit. Don't bundle unrelated edits — each should be independently correct and reviewable.
3. Well-formed: the old text must match exactly (indentation included) and the new text must be syntactically complete (balanced brackets/quotes, valid indentation). A malformed diff that half-applies is worse than no edit.
4. Incremental verify after EACH edit: call mcp_hermes_max_verify_quick_check(path) — fast lint + typecheck, no tests — to confirm the edit is well-formed before moving on. Cheap enough to run every edit. Run the full mcp_hermes_max_verify_verify(path) (adds the test stage) once at subtask end (workflow-subtask-loop), then checkpoint.
5. On a quick_check failure: fix THAT edit immediately — do not pile another edit on top of a malformed one. If two fixes don't make it well-formed, revert the edit and rethink.

Graceful degradation

If mcp-verify is down, quick_check is unavailable — fall back to the native file_mutation_verifier

• your own re-read of the changed hunk, and rely on the subtask-end full verify. Keep edits small regardless: small diffs are the protection, the verifier just confirms them.