Skip to content

Switch to stronger model in CI builds - #776

Merged
eb8680 merged 2 commits into
masterfrom
eb-ci-model
Sep 4, 2026
Merged

Switch to stronger model in CI builds#776
eb8680 merged 2 commits into
masterfrom
eb-ci-model

Conversation

@eb8680

@eb8680 eb8680 commented Sep 2, 2026

Copy link
Copy Markdown
Contributor

Resolves #773 (hopefully)

@eb8680
eb8680 requested a review from jfeser September 2, 2026 23:37
@eb8680

eb8680 commented Sep 2, 2026

Copy link
Copy Markdown
Contributor Author

Two empty responses in the CI failures, not sure what that's about. Maybe context limits?

@eb8680
eb8680 removed the request for review from jfeser September 2, 2026 23:55
@eb8680
eb8680 marked this pull request as draft September 2, 2026 23:55
@eb8680

eb8680 commented Sep 3, 2026

Copy link
Copy Markdown
Contributor Author

It seems the problem is gpt-5-mini is a reasoning model, so the token budgets are wrong.

@eb8680 eb8680 changed the title Switch to gpt-5-mini in CI builds Switch to stronger model in CI builds Sep 3, 2026
gpt-5-mini is a reasoning model, and reasoning tokens are charged against
max_tokens. Two tests size that budget for a non-reasoning model and broke:
test_agent_tool_names_are_valid_integration asks for 64 tokens, all 64 of
which went to reasoning, leaving empty content that hooks.py decodes as "no
final response"; and the [tuple-empty] serialization case exhausted its
three-attempt retry for the same reason. Accommodating that would have meant
either pinning reasoning_effort in the tests or inflating budgets against a
spend that isn't bounded by anything we control.

gpt-4.1-mini is non-reasoning, so those budgets mean what they say and both
tests pass untouched. It also carries a 1M context against gpt-5-mini's 272K
and gpt-4o-mini's 128K, which is the window that overflowed in #773, and it
is markedly faster on the job that gates pull requests: the basics examples
run in 33s against 3m7s, with flight_booking alone converging in 24s where
gpt-4o-mini ground through retries for ~4 minutes before failing.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@eb8680
eb8680 marked this pull request as ready for review September 3, 2026 01:19
@eb8680
eb8680 requested a review from datvo06 September 3, 2026 21:23
@eb8680
eb8680 merged commit e3d0840 into master Sep 4, 2026
35 checks passed
@eb8680
eb8680 deleted the eb-ci-model branch September 4, 2026 13:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

LLM example tests are flaky with weak model

2 participants