Skip to content

fix(decisions): gemini-decision-1 needs a description on every choice and level - #117

Merged
VickyXAI merged 3 commits into
mainfrom
fix/gemini-decision-descriptions
Nov 18, 2026
Merged

VickyXAI merged 3 commits into
mainfrom
fix/gemini-decision-descriptions

Conversation

@VickyXAI

Copy link
Copy Markdown
Contributor

Problem

gemini-decision-1 launched on sol.blockrun.ai and blockrun.ai on 2026-11-18. It needs a description string on every choice and every score level. An empty string is fine.

Without one, the gateway answers 400 before payment, so nothing is charged. The call still fails, though. Every decisions() example in this SDK leaves description out, so copying an example and passing model="gemini-decision-1" fails.

Probed on sol, 2026-11-18. All of these are refused before payment, so they cost nothing.

Request Result
Gemini, choice without description 400 questions[0].choices[0].description must be a string (it may be empty)
Gemini, score level without description 400 questions[0].levels[0].description must be a string (it may be empty)
Gemini, every item has a description 402 (valid)
luna, with descriptions 402 (valid)

Change

  • build_decisions_body raises ValueError before sending when the model is gemini-decision-1 or google/gemini-decision-1 and any choice or level lacks a description string. The message names the exact item.
  • luna behaviour is unchanged.
  • The README example and the four decisions() docstrings give every choice and level a description. They now work on both models.
  • CHANGELOG: an Unreleased entry.

Tests

  • New TestGeminiDecisionDescriptions. Four of its cases failed before the fix:
    • a choice without a description, for both id forms;
    • a level without a description, for both id forms.
  • The cases that pass on both old and new code are controls:
    • every item has a description, and empty strings are accepted;
    • luna, with no model, bare id or openai/ id, does not require descriptions.
  • pytest tests/unit: 1612 passed.
  • ruff check and ruff format: clean.
  • Live run on sol from this branch, about $0.002:
    • the README example answered on gemini-decision-1 and on gpt-6-luna;
    • a Gemini choice with no description raised ValueError before anything was sent.

Noticed, not changed here

On the same input, Gemini returns score as a decimal (1.43) and luna returns an integer index (1). The README says the score is the "index on the rubric". The docs or DecisionsResponse may need a note. That is a separate decision.

🤖 Generated with Claude Code

… and level

Gemini Decision 1 refuses a choice or score level without a description string
(the gateway answers 400 before payment: "questions[0].choices[0].description
must be a string (it may be empty)", probed on sol.blockrun.ai 2026-11-18).
build_decisions_body now raises ValueError for it before sending, naming the
item. gpt-6-luna is unchanged. README and docstring examples give every choice
and level a description, so they work on both models (live-run on sol).
@VickyXAI
VickyXAI merged commit dc2268f into main Nov 18, 2026
6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant