[tao-hpu]

Chat.execute() and AsyncChat.execute() called completion.usage.model_dump() without checking that usage is set. OpenAI-compatible servers can omit the usage key or send "usage": null, and every --no-stream prompt against them failed with AttributeError: 'NoneType' object has no attribute 'model_dump'.

This applies the guard the streaming branches already use (if chunk.usage:) to both non-streaming branches. usage is already initialised to None in both methods and set_usage() already accepts that, so the response is logged with empty token counts.

Changes:

• llm/default_plugins/openai_models.py: 2 guarded lines, one in Chat.execute() and one in AsyncChat.execute().
• tests/test_cli_openai_models.py: test_chat_completion_without_usage_sync_and_async, 4 cases (missing key and null, sync and --async). All 4 fail on main with the AttributeError and pass with the change. The test also checks that the turn is logged with input_tokens and output_tokens set to NULL.

Checks run locally (macOS arm64, Python 3.13.5):

• uv run pytest: 1175 passed
• uv run black . --check: 53 files would be left unchanged
• uv run cog --check ... README.md docs/*.md: no changes
• uv run mypy llm: no issues found in 19 source files
• uv run ruff check .: all checks passed

The repro script from the issue raises the AttributeError on llm==0.35 from PyPI and prints text: hello on this branch.

Closes #1697

AI assistance (Claude Code) was used for this change.

[chrikrah]

@tao-hpu the production change is right and should land, and the test as written breaks on current main. Your four parameters do reproduce for me on Linux and Python 3.12.3.

blocking: 0b51c61 ("Close SQLite connections after CLI commands and tests", 2026-09-22) dropped import sqlite_utils from tests/test_cli_openai_models.py and added a db_factory fixture at tests/conftest.py:40. Your branch appends below that import block, so git merges clean and line 779 then raises NameError. Rebase on main first, since ae8624e has no such fixture, then line 779 becomes db = db_factory(str(log_db)) with db_factory taken as a test argument.

Head ae8624e, merge base 508b4f8, main at 764dc38, pytest 9.1.1, exported trees under /var/tmp:

$ PYTHONPATH=$HEAD python -m pytest tests/test_cli_openai_models.py -q
51 passed in 29.76s

# same tree, openai_models.py replaced by its 508b4f8 copy, your tests kept
$ PYTHONPATH=$BASE python -m pytest tests/test_cli_openai_models.py -q -k without_usage
4 failed, 47 deselected in 11.22s
E   AttributeError: 'NoneType' object has no attribute 'model_dump'

# git merge-tree --write-tree ae8624e origin/main -> 94fbcd0, rc=0, no conflict
$ PYTHONPATH=$MERGED python -m pytest tests/test_cli_openai_models.py -q
4 failed, 47 passed in 28.63s
E   NameError: name 'sqlite_utils' is not defined   # tests/test_cli_openai_models.py:779

# not run: the rest of tests/, and no live OpenAI-compatible server

@simonw the guard applies to 764dc38 unchanged, where openai_models.py:1490 and :1607 are both still a bare completion.usage.model_dump(), and only Read the Docs has run here. If tao-hpu stays quiet until 2026-10-08, shall I open the rebased one-file version?

[tao-hpu]

Rebased onto 764dc38 and switched the test to db_factory as suggested. Full suite: 1208 passed. Thanks for the careful review.

[chrikrah]

@tao-hpu I would merge this at b9e8af4. The rebase onto 764dc38 clears the NameError, and the four parameters still fail with openai_models.py put back.

# Python 3.12.3, pytest 9.1.1, tree from git archive b9e8af4
$ PYTHONPATH=$HEAD python -m pytest tests/test_cli_openai_models.py -q -p no:randomly
51 passed in 96.32s

# llm/default_plugins/openai_models.py from 764dc38, tests kept
4 failed, 47 passed in 46.53s
E   AttributeError: 'NoneType' object has no attribute 'model_dump'
# not run: the rest of tests/, and no live OpenAI-compatible server

@simonw tao-hpu did the rebase, so my 2026-10-08 offer is off. The Test and Run Cog workflows on b9e8af4 are waiting for your approval to run.

[tao-hpu]

Hi @simonw, gentle ping on this one when you have a moment. It's rebased on 764dc38, approved by @chrikrah, and the full suite passes locally (1208 passed). The CI runs expired waiting for workflow approval, so they'd need an approve to run. Thanks!