[wsj040303]

Adding an endpoint to an indented block list wrote the new entry at column zero, leaving extra-openai-models.yaml unparseable. Appending after an explicit ... document end marker caused the same failure, so subsequent model-loading commands stopped working.

Preserve the block list's indentation and insert new entries before its document end marker. Read indentation from the block-start token so sequence tags and anchors work too. Existing records, comments and quoting are preserved.

Regression tests cover zero/two/four-space indentation, missing final newlines, trailing comments, document end markers, tagged/anchored lists, empty explicit documents, and repeated additions followed by model loading.

Validation on Linux / Python 3.13:

• Existing endpoint baseline: 21 passed; new regressions against the original code: 11 failed, 23 passed.
• Fixed endpoint tests: 34 passed. Full pytest -q: 1,413 passed.
• Original CLI reproduction and document-end variant: llm models and endpoint list --json fail before the fix and succeed afterwards.
• Black, Ruff, mypy, Cog documentation checks, package build, and tests/test-llm-load-plugins.sh all passed.

Fixes #1864

[millionbillions]

Code Review Bot — PR #1865

0 patch(es) generated for simonw/llm.

llm/default_plugins/openai_models.py [LOW]

Fault: FATAL: Error code: 401 - Incorrect API key provided: sk-proj-*******************************************************************************************************************************************
Quality score: 0/5 —

tests/test_openai_endpoint_add.py [LOW]

Fault: FATAL: Error code: 401 - Incorrect API key provided: sk-proj-*******************************************************************************************************************************************
Quality score: 0/5 —

---

Generated by CodeReviewBot — GraphRAG-powered code analysis

[chrikrah]

@wsj040303 From my perspective this can be merged. It matches the indent it finds, which answers the question in #1864, and the repro from that issue now loads both models:

# git archive f605dcf673 (base 39d3254), Python 3.13.5, llm.__file__ from the export
$ llm openai endpoint add http://localhost:11434/v1 -m qwen3 --id local-b && llm models
  (base, list indented by 2) -> yaml.parser.ParserError: expected '<document start>', but found '<block sequence start>'
  (head, list indented by 2) -> OpenAI Chat: local-a, OpenAI Chat: local-b
  (base, list ending in ...) -> the same ParserError
  (head, list ending in ...) -> both models, `...` still the last line
$ pytest -q tests/test_openai_endpoint_add.py
  (head) -> 34 passed
  (head, openai_models.py from 39d3254) -> 11 failed, 23 passed
$ pytest -q
  (head) -> 1413 passed
  (base) -> 1400 passed
# _append_extra_model at head, reloaded with yaml.safe_load: a comment line above the list, `---` above an
# indented list and a model_name of "a\n\nb" all come back intact
# not run: Windows

[simonw]

Thanks for this!