[mikkel]

Docs only, in docs/cli.md. The CLI docs described the smartapply model fallback incorrectly. They said --applymodel / GPTDIFF_SMARTAPPLY_MODEL fall back to --model / GPTDIFF_MODEL, but they don't.

What the code does (smart_apply_patch in patchling/core.py, used by --apply)

• Model: --applymodel, else GPTDIFF_SMARTAPPLY_MODEL if it is set and non-blank, else the hardcoded openai/gpt-4.1-mini. --model and GPTDIFF_MODEL are never used for the apply step. The existing --applymodel help text in core.py already says openai/gpt-4.1-mini.
• API key: GPTDIFF_SMARTAPPLY_API_KEY, else GPTDIFF_LLM_API_KEY.
• Base URL: GPTDIFF_SMARTAPPLY_BASE_URL, else GPTDIFF_LLM_BASE_URL, else https://nano-gpt.com/api/v1/.
• When the model is called: every file in the diff goes to the model, with no deterministic first pass in the CLI path. The exception is deleted files (+++ /dev/null), which are removed directly. Each request uses --max_tokens.
• Library smartapply() does not read the GPTDIFF_SMARTAPPLY_* variables. It uses its model argument, else GPTDIFF_MODEL, else DEFAULT_MODEL, and it tries a deterministic patch before calling the model. The docs now say this, so CLI and library behavior aren't confused.

openai/gpt-4.1-mini is on NanoGPT's public model list ($0.40 / $1.60 per 1M input / output tokens, 2026-10-21).

Verified locally (no network, no spend)

An offline probe ran smart_apply_patch with a stubbed apply call (probe-apply.out in the Ember pack):

Case	Apply model	Base URL
Nothing set (--model and GPTDIFF_MODEL set)	openai/gpt-4.1-mini	nano-gpt.com
Blank GPTDIFF_SMARTAPPLY_MODEL	openai/gpt-4.1-mini	nano-gpt.com
GPTDIFF_SMARTAPPLY_MODEL set	that model	nano-gpt.com
--applymodel and env both set	the --applymodel value	nano-gpt.com
GPTDIFF_SMARTAPPLY_BASE_URL set	openai/gpt-4.1-mini	the SMARTAPPLY URL
Only GPTDIFF_LLM_BASE_URL set	openai/gpt-4.1-mini	the LLM URL

In every case the deleted file was removed without a model call.

Other checks: pytest (offline, fake_tiktoken) gives 207 passed, 2 skipped, the same as main. mkdocs build --strict passes.

Notes

• Overlap with docs: replace delisted model names with ids on NanoGPT's model list #26: docs: replace delisted model names with ids on NanoGPT's model list #26 also edits the GPTDIFF_SMARTAPPLY_MODEL line in docs/cli.md (gpt5-mini → openai/gpt-5-mini). Whichever merges second will conflict on that one line and needs a rebase. This PR's wording drops the recommendation entirely, so it can win.
• Also fixed (Ember follow-up commit):
  • The README env table marks GPTDIFF_SMARTAPPLY_MODEL as CLI-only, with default openai/gpt-4.1-mini and no fallback to GPTDIFF_MODEL.
  • In docs/api.md, the smartapply variables paragraph is rewritten in two parts: the library (model argument, then GPTDIFF_MODEL, then DEFAULT_MODEL; key and base URL from arguments, then GPTDIFF_LLM_*) and the CLI-only GPTDIFF_SMARTAPPLY_* variables.
  • An offline probe of library smartapply() with a stubbed model call confirms that it ignores GPTDIFF_SMARTAPPLY_* (probe-lib.out in the pack).
  • More overlap with docs: replace delisted model names with ids on NanoGPT's model list #26: docs: replace delisted model names with ids on NanoGPT's model list #26 also edits the README model table and the docs/api.md GPTDIFF_SMARTAPPLY_MODEL bullet, so whichever PR merges second rebases those lines too. docs(cli): describe the real smartapply model fallback (openai/gpt-4.1-mini) #27's wording should win.

No code changes, version bump, tag, or publish.