fix: map empty Gemini tool-call arguments to {} instead of {"type":"object"}

[apex-mochen]

Fixes #43156.

Zero-arg OpenAI tool calls arrive with arguments="". _gemini_tool_call_invoke_helper injected a schema stub {"type":"object"} as runtime args instead of {}, diverging from the Anthropic/Bedrock path. Maps empty/missing arguments to {} and parses non-empty as before. Adds unit tests.

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Changes how empty tool arguments are formatted in Gemini calls.

The PR appears safe to merge; no outstanding finding or new actionable issue was established

Findings

1.  Mid-conversation system messages fail ▶
2.  Existing Gemini test fails ▶
3.  Template parameter loses its type ▶
4.  Duplicate thinking-block explanation ▶
5.  Factory file text corrupted ▶

Summary

The PR changes empty Gemini tool-call arguments from a schema-shaped placeholder to {} and updates the existing factory test, with added cases for missing and nonempty arguments

• The five previous findings are addressed in the current code

_{Reviews (2) · Last reviewed commit: "fix: map empty Gemini tool-call argument..."}

[greptile-apps]

Mid-conversation system messages fail Supported Anthropic models retain these messages, but the converter no longer handles them. It raises BadRequestError before the request reaches Anthropic.

Knowledge Base Used: Provider adapters and capabilities

[greptile-apps]

Existing Gemini test fails The new assertion expects {}, while the existing factory test still expects {"type": "object"}. Update that test; the repository guide requires bug fixes to extend the existing mapped file.

Context Used: AGENTS.md (source)

[greptile-apps]

Template parameter loses its type Both template helpers replace str | None with Any | None. The repository guide requires strongly typed parameters; restore that requirement before merging.

Context Used: AGENTS.md (source)

[greptile-apps]

Duplicate thinking-block explanation This docstring repeats the shared predicate’s explanation, creating text to maintain twice. The repository guide restricts unnecessary code comments; remove the duplicate before merging.

Context Used: AGENTS.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[greptile-apps]

Factory file text corrupted The line-ending rewrite makes the entire file appear changed and turns readable comment characters into 鈥?. Restore the encoding to keep source text readable and future changes reviewable.

[codspeed]

Merging this PR will not alter performance

⚠️ 1 benchmark spent significant time in system calls

System calls cannot be consistently instrumented, so they are not included in the measure, which understates the real cost. Please switch to the Walltime instrument to accurately measure system calls.

Measurement and system calls

• test_completion_streaming

✅ 31 untouched benchmarks

---

_{Comparing apex-mochen:fix/gemini-empty-arguments-empty-object (10511df) with main (ac98cab)¹}

Footnotes

1. No successful run was found on main (102619d) during the generation of this report, so ac98cab was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[lets-order-some-fries]

The substantive change here is right and matches the description — _gemini_tool_call_invoke_helper handing back {"type": "object"} as runtime args instead of {} is a real bug, and the three-line fix at factory.py:1144 is the correct shape.

The diff around it is the problem: +9456/−9479, with 5586 changed lines in factory.py. I pulled the head and normalised it against the merge-base (c19ce71bcd), and three separate things are riding along that I don't think you intended.

1. The file was converted to CRLF. Every line:

base: 5582 lines, 0 with CR
head: 5586 lines, 5586 with CR

Once \r is stripped, the real diff in factory.py is 62 lines, not 5586. Same in the test file: 52 real lines inside 3847.

2. Five comments were corrupted, and the pattern says the encoding is the cause. Em-dashes became 鈥?, which is UTF-8 — (E2 80 94) read back as GBK/CP936:

1631:  # translate to Anthropic document blocks, not image blocks 鈥?
2318:  # conversational-turn boundary 鈥?reset tracking.
2376:  # because there is no way to "pass through" an empty text block 鈥?the
3708:  # Fallback: no objects extracted 鈥?use empty dict.

3. It removes live code that is currently on main. These are deletions in the normalised diff, not additions:

• from litellm.litellm_core_utils.prompt_templates.mid_conversation_system import anthropic_system_messages, together with its call site new_messages.extend(anthropic_system_messages(messages[msg_i]))
• is_unsignable_thinking_block from the imports and both of its call sites, including return [block for block in thinking_blocks if not is_unsignable_thinking_block(block)]
• TypeAlias from the typing import and _AnthropicMessageList: TypeAlias = list[AllAnthropicPassThroughMessageValues]
• two signatures weakened from chat_template: str | None to chat_template: Any | None

anthropic_system_messages is still referenced on main, so merging this as-is would drop that handling and the unsignable-thinking-block filtering.

All three are consistent with one cause: the file being opened and rewritten by an editor with a non-UTF-8 default codepage and CRLF endings, on a base that had drifted.

What I would suggest, rather than fixing it in place: branch fresh from current main, apply only the arguments_dict change and the new tests/unit/test_gemini_tool_call_convert.py, and leave factory.py's other 5580 lines untouched. Setting git config core.autocrlf input and making sure the editor writes UTF-8 should stop it recurring. The fix itself is worth landing — it just needs to be a three-line diff so a reviewer can see that it is one.

I have not run the test suite against this head; this is a read of the diff against the merge-base.

[apex-mochen]

@lets-order-some-fries Thank you — you were right about the corrupted, drifted diff. I rebuilt this PR from current main rather than trying to repair that branch in place. The current head (d5f4912) is +25/-6 across three files only: the three-line empty-object correction, the existing expectation now updated to an empty object, and three focused empty/missing/non-empty argument tests. It contains no CRLF/BOM or encoding rewrite, and it leaves the Anthropic system-message handling, thinking-block filter, and type annotations untouched. The repository workflows have been triggered on this rebuilt head; I will address any CI result that identifies a remaining issue.

[apex-mochen]

@mateo-berri @Shivam - small ping for review when you have a moment. This fixes #43156: when a Gemini/Vertex tool call has zero arguments, the helper currently emits {"type":"object"}, which differs from the Anthropic/Bedrock path (which passes {}) and can trip Gemini's tool schema validation. Diff is ~25 lines with regression coverage. Happy to adjust anything.

[lets-order-some-fries]

Re-checked at bf7d9d610e. This looks right now. The only change to factory.py is line 1144, the existing expectation is updated, and nothing else comes along.

What I ran:

• The repro from #43156 through convert_to_gemini_tool_call_invoke returns args={} for the arguments="" call at this head, and {'type': 'object'} on main (14f4c34c61).
• With only this PR's test files applied to main, test_empty_arguments_maps_to_empty_object, test_missing_arguments_maps_to_empty_object and the updated test_vertex_ai_transform_empty_function_call_arguments fail with assert {'type': 'object'} == {}. The non-empty case passes on both, as it should. All four pass here.
• tests/unit/litellm_core_utils/prompt_templates/ plus the Vertex Gemini transformation tests give the same results as main, apart from the three new passes.

The red misc / Run tests is not caused by this PR. Its only failures are tests/unit/test_model_prices_schema.py::test_price_map_has_no_duplicate_keys[main] and [backup], both on a duplicate openrouter/perceptron/perceptron-mk1.5 key in the price maps, which this PR does not touch. main failed the same job on the same two tests at your base, 99655b6f86. #43273 removed the duplicate at 06:15Z, after this run had started, and main has passed that job on every commit since, through 14f4c34c61. A re-run checks out the same merge commit (71a4e6a), so the check will stay red until the branch is updated from main.

One correction to my earlier review: I suggested keeping a separate test file, but AGENTS.md (line 30) asks bug fixes to extend the existing mapped test file. test_empty_arguments_maps_to_empty_object also repeats test_vertex_ai_transform_empty_function_call_arguments (test_litellm_core_utils_prompt_templates_factory.py:595), with the same empty arguments and the same assertion. Moving the missing-key and non-empty cases into that file and dropping test_gemini_empty_arguments.py would fit the guide. That is not a blocker, and it's a maintainer's call.

Also, the @shivam above resolves to an account with no activity on this repo. The maintainer you meant is likely shivamrawat1.

[apex-mochen]

@lets-order-some-fries Thanks for the re-check! I've rebased the branch onto the latest upstream main (d18fcb09) and folded the new missing-key / non-empty regression cases into the mapped test file (tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py), per the repo's AGENTS.md. The standalone test_gemini_empty_arguments.py is dropped — no duplicate file.

The core three-line fix in factory.py is unchanged. New head: bd92a0fd. Diff vs main is now just the two expected files. Happy to keep the current test structure if you'd rather the cases stay in their own file — just say the word and I'll revert the test-file reshuffle.

[lets-order-some-fries]

Checked at bd92a0fd. The diff against main is the one-line change in factory.py plus the two new cases in the mapped test file. With factory.py reverted to main, test_vertex_ai_transform_empty_function_call_arguments and test_gemini_missing_arguments_maps_to_empty_object fail with assert {'type': 'object'} == {}, and test_gemini_nonempty_arguments_is_parsed passes either way. Good as it is from my side, no need to revert the test move.

[apex-mochen]

Thanks for re-checking at bd92a0fd, and especially for running the revert control — with a one-line fix the tests are the only thing standing behind it, so that was the check I most wanted a second pair of eyes on.

To confirm the state you verified is the state we are leaving it in: no further changes are planned on this branch. The test move stays where it is, and the three assertions are deliberate — the two that flip pin the fix, and test_gemini_nonempty_arguments_is_parsed is there to catch the opposite mistake, a change that starts handing back {} for calls that do carry arguments.

CI is 97/97 at that head with nothing outstanding from our side.

[apex-mochen]

Hi @lets-order-some-fries, thanks for the re-check at bd92a0fd — glad the revert control confirmed the fix. Since all checks are green (97/97) and the branch is stable, could you approve when you get a chance? Happy to address anything else first if needed.

[lets-order-some-fries]

Approving at bd92a0fd, the head I checked on 09-26; nothing has been pushed since. The diff is still the one-line change at factory.py:1144 plus the updated expectation and two new cases in the mapped test file. In the revert control I ran then, with factory.py reverted to main, the empty and missing-arguments tests failed with assert {'type': 'object'} == {} and the non-empty case passed either way. CI is green at this head, and main at b370996b9d still has {"type": "object"} in that branch, so the fix is still needed.

[apex-mochen]

@mateo-berri @Shivam Could you take a maintainer look when convenient? The current head bd92a0fd has an external approval (September 30), all CI checks are green, and the reviewer independently confirmed that reverting the fix makes the two empty/missing-argument regression tests fail. The change remains limited to the Gemini empty tool-call argument mapping and its mapped test file. No further branch changes are planned unless review finds an issue. Thank you.

[apex-mochen]

Merge request, same state as #43293: head bd92a0fd carries the 09-30 approval (lets-order-some-fries) at the exact current SHA, mergeable=true, all checks green. Could a maintainer from @litellm-engineering hit merge when convenient? Thanks!

[apex-mochen]

Hi @litellm-engineering, gentle follow-up on this one. #43162 was approved by @lets-order-some-fries on Sep 30, and I've kept it rebased onto the latest main; mergeable is true and CI shows no PR-introduced failures. The change remains scoped to the cache-auth fix plus its regression test. When you have a moment, could this be merged? Thanks!

[apex-mochen]

Hi @mateo-berri — could you take a look at this PR for a final maintainer merge decision when convenient?

The current HEAD (bd92a0f) has community approval, CI is green, and the fix is limited to mapping empty Gemini tool-call arguments to {}, with regression coverage.

The five remaining review threads are historical CRLF/encoding artifacts from older revisions. Their underlying issues have been addressed, and the current HEAD has been confirmed clean.

For context, I previously contributed the cancellation-race analysis and shielded-cleanup suggestions acknowledged in the recently merged #43061. I appreciate your work getting that fix upstream.

Could you let me know if anything else is needed before this PR can be merged? Happy to address any remaining concerns promptly.

Thanks!

[apex-mochen]

Rebased onto latest upstream main. The two failing jobs were from an old run: test_component_allowlists (the /v1/logs allowlist, a repo-level baseline) and test_seed_tracing_fixtures (fixture freshness). Rebased so CI re-runs against the current baseline.