Patch workflow & FEEDBACK

Related guides:

Patch is the center of gravity for the AI loop. This is where requirements are refined, code is generated, and the commit message is drafted.

Accepted patch kinds

STAN supports two patch payload kinds:

File Ops (structural operations like move/delete/mkdir)
Unified diff (content edits)

You must not mix both kinds in a single payload.

Unified diffs (single file per invocation)

Plain unified diffs with diff --git, ---/+++, and @@ hunks.
Paths are relative to the repo root, with POSIX separators (/). Prefer a/ and b/ prefixes.
≥ 3 lines of context per hunk. LF line endings. No base64.
Hard rule: stan patch enforces exactly one target file per patch payload. For multi-file changes, apply multiple times (one diff per file).

Apply from clipboard (default), argument, or file:

stan patch
stan patch --check
stan patch -f changes.patch

File Ops (structural changes)

Use a ### File Ops block for safe, repo-relative operations:

### File Ops
mkdirp src/new/dir
mv src/old.ts src/new/dir/old.ts
rm docs/obsolete.md

Then follow with a unified diff for any content edits (in a separate payload).

On failure: diagnostics envelope

When a patch fails or partially applies, STAN prints a compact diagnostics envelope and (when possible) copies it to your clipboard. Paste it into chat as-is. It includes:

which apply strategies were attempted,
exit summaries and jsdiff reasons (when applicable),
the declared target file(s) (when detectable from the patch headers).

Assistants should respond by generating a corrected unified diff that applies cleanly, including Full Listings only for the failed file when necessary.

Tips

Keep hunks small and anchored; avoid large reflows in Markdown.
For multi-file changes, emit one diff per file and apply them one at a time.
For docs: preserve LF; minimal whitespace changes improve reliability.