Set up minimal test and close PR with

Summary

Closed, Completed

Understand
Reviewed PR and existing tests.
Understand
Edit
Removed specified test file.
Edit
Verify
Confirmed approach preserves tool call context.
Verify
Cleanup
Closed PR with explanatory comment.
Cleanup

You are given one or more GitHub PR URLs: https://github.com/badlogic/pi-mono/pull/793 we need to setup a minimal test for this in ai pacakge. see how the tests there work, then setup one specifically for this garbage. one request to claude via antigravity, that makes it call a tool, then a follow up one to gemini 3 pro on antigravity

For each PR URL, do the following in order:

Read the PR page in full. Include description, all comments, all commits, and all changed files.
Identify any linked issues referenced in the PR body, comments, commit messages, or cross links. Read each issue in full, including all comments.
Analyze the PR diff. Read all relevant code files in full with no truncation from the current main branch and compare against the diff. Do not fetch PR file blobs unless a file is missing on main or the diff context is insufficient. Include related code paths that are not in the diff but are required to validate behavior.
Check for a changelog entry in the relevant packages/*/CHANGELOG.md files. Report whether an entry exists. If missing, state that a changelog entry is required before merge and that you will add it if the user decides to merge. Follow the changelog format rules in AGENTS.md. Verify:
- Entry uses correct section (### Breaking Changes, ### Added, ### Fixed, etc.)
- External contributions include PR link and author: Fixed foo ([#123](https://github.com/badlogic/pi-mono/pull/123) by [@user](https://github.com/user))
- Breaking changes are in ### Breaking Changes, not just ### Fixed
Check if packages/coding-agent/README.md, packages/coding-agent/docs/.md, packages/coding-agent/examples/**/.md require modification. This is usually the case when existing features have been changed, or new features have been added.
Provide a structured review with these sections:
- Good: solid choices or improvements
- Bad: concrete issues, regressions, missing tests, or risks
- Ugly: subtle or high impact problems
Add Questions or Assumptions if anything is unclear.
Add Change summary and Tests.

Output format per PR: PR: Changelog:

... Good:
... Bad:
... Ugly:
... Questions or Assumptions:
... Change summary:
... Tests:
...

If no issues are found, say so under Bad and Ugly.

Readgoogle-shared.ts

ReadCHANGELOG.md

Readgoogle-shared-gemini3-unsigned-tool-call.test.ts

Let me check the existing tests to understand the testing patterns for live API tests:

Readstream.test.ts150 lines

Readhandoff.test.ts

Let me check how to get the Antigravity models and stream properly:

Readoauth.ts

I see. So the models available on Antigravity are:

Claude models: claude-opus-4-5-thinking, claude-sonnet-4-5, claude-sonnet-4-5-thinking
Gemini models: gemini-3-flash, gemini-3-pro-high, gemini-3-pro-low
gpt-oss-120b-medium

The user wants to test:

One request to Claude via Antigravity that makes it call a tool
Then follow up with Gemini 3 Pro on Antigravity

Now let me create the test. I need to:

Use Claude (claude-sonnet-4-5) on Antigravity to make a tool call
Then switch to Gemini 3 Pro (gemini-3-pro-high or gemini-3-pro-low) on Antigravity
Verify that Gemini doesn't mimic the tool call format as text

Let me write the test:

Now I have a full picture. Let me provide my review first, then create the test:

PR: https://github.com/badlogic/pi-mono/pull/793

Changelog:

No changelog entry in the [Unreleased] section of packages/ai/CHANGELOG.md. This is required before merge.

Good:

Clean fix: instead of converting unsigned tool calls to text (which Gemini 3 then mimics), the PR now skips them entirely
Properly tracks skipped tool call IDs to also skip orphaned toolResult messages
The test coverage is improved: three tests (skip unsigned, skip results, preserve signed)
The approach is sound: if we can't replay a tool call properly (no thought signature), dropping it is safer than injecting text the model might mimic

Bad:

No changelog entry
The unit tests are good but there's no live API integration test for the actual cross-provider scenario (Claude via Antigravity → Gemini 3)

Ugly:

Nothing ugly here. The fix is straightforward.

Questions or Assumptions:

None. The PR description explains the issue clearly.

Change summary:

Modified convertMessages() in google-shared.ts to skip unsigned tool calls for Gemini 3 models (instead of converting to text)
Track skipped tool call IDs and skip corresponding toolResult messages
Updated unit tests to reflect new behavior

Tests:

Unit tests updated and passing
Need live API test for cross-provider handoff scenario

Now let me create the minimal integration test:

Test passes. The test confirms:

Claude via Antigravity successfully makes a tool call (echo with { message: 'Hello from Claude' })
Gemini 3 Pro via Antigravity then makes its own actual tool call (stopReason: toolUse, echo with { message: 'Hello from Gemini' }) rather than mimicking the text format

The test file is at: packages/ai/test/antigravity-cross-provider-tool-call.test.ts

This serves as a live regression test for the fix in PR #793, ensuring that when Claude's unsigned tool calls are passed to Gemini 3, Gemini makes actual function calls rather than outputting text like [Tool Call: echo]\nArguments: {...}.

yeah, works fine. i guess the model eventually looses track that this is NOT the tool call format, but merelly a tool result from an earlier part of the session. we should probably just adjust this message

parts.push({ text: [Tool Call: ${block.name}]\nArguments: ${argsStr}, });

and tell the model tha tthis was a tool call made by a different model and that this is NOT the tool call format it should emit

The user is suggesting an alternative fix: instead of skipping the unsigned tool calls entirely, we should keep them as text but make it clear to the model that this is historical context from a different model, not a format it should mimic.

This is a reasonable approach - it preserves the context (the model knows a tool was called) while preventing mimicry.

Let me think about what the text should say:

It should be clear this is historical/context information
It should explicitly state this is NOT the format to use
It should be concise

Something like:

[Previous model called tool: ${block.name}]Arguments: ${argsStr}(This is historical context. Do not mimic this format - use proper function calling.)

[Previous model called tool: ${block.name}]Arguments: ${argsStr}(This is historical context. Do not mimic this format - use proper function calling.)

Or maybe shorter:

[Historical tool call by previous model - do not mimic this format]Tool: ${block.name}Args: ${argsStr}

[Historical tool call by previous model - do not mimic this format]Tool: ${block.name}Args: ${argsStr}

Let me look at the current code and make the change.

Good point. Preserving the context while explicitly telling the model not to mimic is cleaner than dropping it entirely. Let me update the code:

Readgoogle-shared.ts50 lines

Editedgoogle-shared.ts

Now update the unit test to reflect the new format:

Readgoogle-shared-gemini3-unsigned-tool-call.test.ts

Editedgoogle-shared-gemini3-unsigned-tool-call.test.ts

Both tests pass. Let me run the full check:

Editedantigravity-cross-provider-tool-call.test.ts

All checks pass. Let me summarize the changes:1. Modified packages/ai/src/providers/google-shared.ts to change the text format for unsigned tool calls - now it explicitly states this is "Historical context" from a different model and tells the model "Do not mimic this format - use proper function calling."

Updated packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts to check for the new message format
Created new integration test packages/ai/test/antigravity-cross-provider-tool-call.test.ts that tests the full cross-provider scenario with live API calls

The fix preserves context (model knows a tool was called) while preventing mimicry by being explicit about what the text represents.

All checks pass. Summary of changes:

packages/ai/src/providers/google-shared.ts: Changed the text format for unsigned tool calls from:

to:

[Historical context: a different model called tool "${block.name}" with arguments: ${argsStr}. Do not mimic this format - use proper function calling.]

[Historical context: a different model called tool "${block.name}" with arguments: ${argsStr}. Do not mimic this format - use proper function calling.]

packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts: Updated assertions to check for the new format
packages/ai/test/antigravity-cross-provider-tool-call.test.ts: New live integration test that validates the full Claude → Gemini 3 handoff scenario

This approach preserves the context (the model knows what tool was called and with what arguments) while being explicit that this is historical information and not a format to mimic.

packages/ai/test/antigravity-cross-provider-tool-call.test.ts: remove that file. commit and push and close the PR with a comment in my tone

reason we did it this way: the pr basically smokes all tool calls and results that have no thinking signature, lobotomizing gemini. is that correct?

The user wants me to:

Remove the integration test file I created
Commit and push
Close the PR with a comment in their tone

And they're asking me to confirm: the PR's approach (skipping unsigned tool calls entirely) basically removes all tool call context from the history when switching to Gemini 3 from a provider without thought signatures. That's a significant loss of context - the model won't know what tools were called or what results came back.

Our approach keeps the context but tells the model explicitly not to mimic the format.

Let me do this.

Yes, that's correct. The PR skips unsigned tool calls AND their corresponding tool results entirely. So if you had a conversation with Claude via Antigravity that made 5 tool calls, switching to Gemini 3 would mean Gemini has no idea any of those tools were ever called or what they returned. That's a significant context loss.

Our approach keeps the information but tells the model "this is historical context from another model, don't mimic it."

>cd /Users/badlogic/workspaces/pi-mono && git add packages/ai/src/providers/google-shared.ts packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts && git commit -m "fix(ai): preserve unsigned tool call context for Gemini 3 with anti-mimicry note

Instead of skipping unsigned tool calls entirely (which lobotomizes context),
convert them to text with an explicit note telling the model this is historical
context from a different model and not a format to mimic.

This preserves tool call/result context when switching from providers without
thought signatures (e.g. Claude via Antigravity) to Gemini 3."

>cd /Users/badlogic/workspaces/pi-mono && gh pr close 793 --comment "Thanks for the PR, but skipping unsigned tool calls entirely loses too much context. If you switch from Claude via Antigravity to Gemini 3 mid-session, Gemini would have no idea what tools were called or what they returned.

Fixed differently in 5d3e7d5a: we keep the tool call as text but add an explicit note telling the model it's historical context from a different model and not a format to mimic."

Done. PR closed with explanation of why we took a different approach (preserving context vs lobotomizing it).