The tool-layer release, in all three editions, and the part that matters most is argument validation.
Tool-call arguments are checked against the tool's own schema before anything runs. Small local
models send arguments that are almost right: "3" where an integer is wanted, "true" where a
boolean is wanted, a bare string where a list is expected, a key in the wrong place. Until this, those
reached the tool. Now they are coerced where coercion is unambiguous, and rejected where it is not,
with an error the model can act on. That is the difference between a 7B model being usable as a daily
driver and being a source of random failures: the model does not need to be perfect at JSON, it needs
a layer that knows a schema when it sees one.
Writes are atomic. write_file, replace_in_file and apply_changes stage a hidden temp file
beside the target, check and sync it, copy the target's permission bits (and owner where allowed), and
rename it over the target. A write that fails partway, a disk that fills, a crash mid-write, no longer
leaves a half-written config. The failure mode this removes is the worst kind: a file that is
syntactically valid, mostly right, and quietly wrong.
Editing a Windows file no longer converts it to LF. Files are read without newline translation, and
an exact-text edit that has no match as typed, contains \n and no \r, and whose file uses CRLF, is
retried with CRLF. The symptom was an edit tool that worked on Linux text and failed on Windows text
with "no match found", and then wrote LF line endings into a file whose other tools expect CRLF.
Plus replace_in_file gaining replace_all, so the multiple-match error tells you the flag that fixes
it.