Set up minimal test and close PR with
Public
Set up minimal test and close PR with
Agent
Pi
Model
Claude Opus 4.5 (latest)
Started
Jan 16, 10:22 PM
Messages111
Downloads0
Reviewed PR and existing tests.
UnderstandRemoved specified test file.
EditConfirmed approach preserves tool call context.
VerifyClosed 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:
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:
### Breaking Changes, ### Added, ### Fixed, etc.)Fixed foo ([#123](https://github.com/badlogic/pi-mono/pull/123) by [@user](https://github.com/user))### Breaking Changes, not just ### FixedOutput format per PR: PR: Changelog:
If no issues are found, say so under Bad and Ugly.
The user wants me to review a GitHub PR and then set up a minimal test for it in the ai package. Let me start by reading the PR.
>gh pr view 793 --repo badlogic/pi-mono --json title,body,comments,commits,files,author,labels,state,additions,deletions,changedFiles{"additions":152,"author":{"id":"MDQ6VXNlcjU1NzYxNjE=","is_bot":false,"login":"ferologics","name":"Fero"},"body":"The previous fix (b18f401d) converted unsigned tool calls to text like \"[Tool Call: name]\\nArguments: {...}\" to avoid API validation errors. However, this caused Gemini 3 to see this format in its context and mimic it as plain text instead of making actual function calls.\r\n\r\nThis fix:\r\n- Skips unsigned tool calls entirely for Gemini 3 models\r\n- Also skips the corresponding toolResults to avoid orphaned responses\r\n- Preserves signed tool calls with valid thoughtSignatures\r\n\r\nFixes tool call mimicry bug reported in sessions with cross-provider history (e.g. Claude via Antigravity -> Gemini 3).","changedFiles":2,"comments":[],"commits":[{"authoredDate":"2026-01-16T21:47:40Z","authors":[{"email":"[REDACTED]","id":"MDQ6VXNlcjU1NzYxNjE=","login":"ferologics","name":"ferologics"}],"committedDate":"2026-01-16T21:47:40Z","messageBody":"The previous fix (b18f401d) converted unsigned tool calls to text like\n\"[Tool Call: name]\\nArguments: {...}\" to avoid API validation errors.\nHowever, this caused Gemini 3 to see this format in its context and\nmimic it as plain text instead of making actual function calls.\n\nThis fix:\n- Skips unsigned tool calls entirely for Gemini 3 models\n- Also skips the corresponding toolResults to avoid orphaned responses\n- Preserves signed tool calls with valid thoughtSignatures\n\nFixes tool call mimicry bug reported in sessions with cross-provider\nhistory (e.g. Claude via Antigravity -> Gemini 3).","messageHeadline":"fix(ai): skip unsigned Gemini 3 tool calls to prevent mimicry","oid":"b6d56d639a8cb18634b5870633942a1c72863c1f"}],"deletions":30,"files":[{"path":"packages/ai/src/providers/google-shared.ts","additions":26,"deletions":17},{"path":"packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts","additions":126,"deletions":13}],"labels":[],"state":"OPEN","title":"fix(ai): skip unsigned Gemini 3 tool calls to prevent mimicry"}>gh pr diff 793 --repo badlogic/pi-monodiff --git a/packages/ai/src/providers/google-shared.ts b/packages/ai/src/providers/google-shared.ts
index deb920d4d..13a3c929c 100644
--- a/packages/ai/src/providers/google-shared.ts
+++ b/packages/ai/src/providers/google-shared.ts
@@ -71,6 +71,9 @@ export function requiresToolCallId(modelId: string): boolean {
export function convertMessages<T extends GoogleApiType>(model: Model<T>, context: Context): Content[] {
const contents: Content[] = [];
const transformedMessages = transformMessages(context.messages, model);
+ // Track tool call IDs that were skipped (e.g., unsigned Gemini 3 calls)
+ // so we can also skip their corresponding toolResults
+ const skippedToolCallIds = new Set<string>();
for (const msg of transformedMessages) {
if (msg.role === "user") {
@@ -134,26 +137,27 @@ export function convertMessages<T extends GoogleApiType>(model: Model<T>, contex
const thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.thoughtSignature);
// Gemini 3 requires thoughtSignature on all function calls when thinking mode is enabled.
// When replaying history from providers without thought signatures (e.g. Claude via Antigravity),
- // convert unsigned function calls to text to avoid API validation errors.
+ // skip the tool call entirely to avoid API validation errors.
+ // Previously we converted to text like "[Tool Call: name]\nArguments: {...}" but this caused
+ // the model to mimic that format instead of making actual function calls.
const isGemini3 = model.id.toLowerCase().includes("gemini-3");
if (isGemini3 && !thoughtSignature) {
- const argsStr = JSON.stringify(block.arguments, null, 2);
- parts.push({
- text: `[Tool Call: ${block.name}]\nArguments: ${argsStr}`,
- });
- } else {
- const part: Part = {
- functionCall: {
- name: block.name,
- args: block.arguments,
- ...(requiresToolCallId(model.id) ? { id: block.id } : {}),
- },
- };
- if (thoughtSignature) {
- part.thoughtSignature = thoughtSignature;
- }
- parts.push(part);
+ // Skip unsigned tool calls - they can't be replayed properly anyway
+ // Track the ID so we also skip the corresponding toolResult
+ skippedToolCallIds.add(block.id);
+ continue;
+ }
+ const part: Part = {
+ functionCall: {
+ name: block.name,
+ args: block.arguments,
+ ...(requiresToolCallId(model.id) ? { id: block.id } : {}),
+ },
+ };
+ if (thoughtSignature) {
+ part.thoughtSignature = thoughtSignature;
}
+ parts.push(part);
}
}
@@ -163,6 +167,11 @@ export function convertMessages<T extends GoogleApiType>(model: Model<T>, contex
parts,
});
} else if (msg.role === "toolResult") {
+ // Skip results for tool calls that were skipped (e.g., unsigned Gemini 3 calls)
+ if (skippedToolCallIds.has(msg.toolCallId)) {
+ continue;
+ }
+
// Extract text and image content
const textContent = msg.content.filter((c): c is TextContent => c.type === "text");
const textResult = textContent.map((c) => c.text).join("\n");
diff --git a/packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts b/packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts
index d0f4b43be..87f40397b 100644
--- a/packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts
+++ b/packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts
@@ -3,7 +3,7 @@ import { convertMessages } from "../src/providers/google-shared.js";
import type { Context, Model } from "../src/types.js";
describe("google-shared convertMessages", () => {
- it("converts unsigned tool calls to text for Gemini 3", () => {
+ it("skips unsigned tool calls for Gemini 3 to prevent mimicry", () => {
const model: Model<"google-generative-ai"> = {
id: "gemini-3-pro-preview",
name: "Gemini 3 Pro Preview",
@@ -51,19 +51,132 @@ describe("google-shared convertMessages", () => {
const contents = convertMessages(model, context);
- let toolTurn: (typeof contents)[number] | undefined;
- for (let i = contents.length - 1; i >= 0; i -= 1) {
- if (contents[i]?.role === "model") {
- toolTurn = contents[i];
- break;
- }
- }
+ // The assistant message with only an unsigned tool call should be completely skipped
+ // (parts.length === 0 causes the message to be omitted)
+ const modelTurns = contents.filter((c) => c.role === "model");
+ expect(modelTurns.length).toBe(0);
+ });
+
+ it("skips tool results for skipped unsigned tool calls", () => {
+ const model: Model<"google-generative-ai"> = {
+ id: "gemini-3-pro-preview",
+ name: "Gemini 3 Pro Preview",
+ api: "google-generative-ai",
+ provider: "google",
+ baseUrl: "https://generativelanguage.googleapis.com",
+ reasoning: true,
+ input: ["text"],
+ cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+ contextWindow: 128000,
+ maxTokens: 8192,
+ };
+
+ const now = Date.now();
+ const context: Context = {
+ messages: [
+ { role: "user", content: "Hi", timestamp: now },
+ {
+ role: "assistant",
+ content: [
+ {
+ type: "toolCall",
+ id: "call_1",
+ name: "bash",
+ arguments: { command: "ls -la" },
+ // No thoughtSignature
+ },
+ ],
+ api: "google-gemini-cli",
+ provider: "google-antigravity",
+ model: "claude-sonnet-4-20250514",
+ usage: {
+ input: 0,
+ output: 0,
+ cacheRead: 0,
+ cacheWrite: 0,
+ totalTokens: 0,
+ cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+ },
+ stopReason: "toolUse",
+ timestamp: now,
+ },
+ {
+ role: "toolResult",
+ toolCallId: "call_1",
+ toolName: "bash",
+ content: [{ type: "text", text: "file1.txt\nfile2.txt" }],
+ isError: false,
+ timestamp: now,
+ },
+ ],
+ };
+
+ const contents = convertMessages(model, context);
+
+ // Both the tool call and its result should be skipped
+ const modelTurns = contents.filter((c) => c.role === "model");
+ const functionResponses = contents.filter((c) => c.parts?.some((p) => p.functionResponse !== undefined));
+
+ expect(modelTurns.length).toBe(0);
+ expect(functionResponses.length).toBe(0);
+ });
+
+ it("preserves signed tool calls for Gemini 3", () => {
+ const model: Model<"google-generative-ai"> = {
+ id: "gemini-3-pro-preview",
+ name: "Gemini 3 Pro Preview",
+ api: "google-generative-ai",
+ provider: "google",
+ baseUrl: "https://generativelanguage.googleapis.com",
+ reasoning: true,
+ input: ["text"],
+ cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+ contextWindow: 128000,
+ maxTokens: 8192,
+ };
+
+ const now = Date.now();
+ // Valid base64 signature (at least looks valid)
+ const validSignature = "AAAA";
+ const context: Context = {
+ messages: [
+ { role: "user", content: "Hi", timestamp: now },
+ {
+ role: "assistant",
+ content: [
+ {
+ type: "toolCall",
+ id: "call_1",
+ name: "bash",
+ arguments: { command: "ls -la" },
+ thoughtSignature: validSignature,
+ },
+ ],
+ api: "google-generative-ai",
+ provider: "google",
+ model: "gemini-3-pro-preview",
+ usage: {
+ input: 0,
+ output: 0,
+ cacheRead: 0,
+ cacheWrite: 0,
+ totalTokens: 0,
+ cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+ },
+ stopReason: "toolUse",
+ timestamp: now,
+ },
+ ],
+ };
+
+ const contents = convertMessages(model, context);
- expect(toolTurn).toBeTruthy();
- expect(toolTurn?.parts?.some((p) => p.functionCall !== undefined)).toBe(false);
+ const modelTurns = contents.filter((c) => c.role === "model");
+ expect(modelTurns.length).toBe(1);
- const text = toolTurn?.parts?.map((p) => p.text ?? "").join("\n");
- expect(text).toContain("[Tool Call: bash]");
- expect(text).toContain("ls -la");
+ const functionCall = modelTurns[0]?.parts?.find((p) => p.functionCall !== undefined);
+ expect(functionCall).toBeTruthy();
+ expect(functionCall?.functionCall?.name).toBe("bash");
+ expect(functionCall?.thoughtSignature).toBe(validSignature);
});
});Let me read the current files on main to compare, and also check for changelog entries and any linked issues.
Now I have a complete picture. Let me look for the tests in the ai package to understand the testing patterns, and then understand what the user wants: set up a minimal test for the cross-provider tool call issue (Claude via Antigravity -> Gemini 3).
Let me check the existing tests to understand the testing patterns for live API tests:
>ls -la /Users/badlogic/workspaces/pi-mono/packages/ai/test/ | head -40total 624
drwxr-xr-x@ 26 badlogic staff 832 Jan 16 21:37 .
drwxr-xr-x@ 12 badlogic staff 384 Jan 16 22:35 ..
-rw-r--r--@ 1 badlogic staff 7691 Jan 13 22:01 abort.test.ts
-rw-r--r--@ 1 badlogic staff 2634 Jan 13 22:01 bedrock-models.test.ts
-rw-r--r--@ 1 badlogic staff 552 Jan 13 22:01 bedrock-utils.ts
-rw-r--r--@ 1 badlogic staff 26414 Jan 13 22:01 context-overflow.test.ts
drwxr-xr-x@ 3 badlogic staff 96 Aug 30 18:31 data
-rw-r--r--@ 1 badlogic staff 21968 Jan 13 22:01 empty.test.ts
-rw-r--r--@ 1 badlogic staff 2890 Jan 13 22:01 google-gemini-cli-claude-thinking-header.test.ts
-rw-r--r--@ 1 badlogic staff 2638 Jan 13 22:01 google-gemini-cli-empty-stream.test.ts
-rw-r--r--@ 1 badlogic staff 1645 Jan 13 22:01 google-gemini-cli-retry-delay.test.ts
-rw-r--r--@ 1 badlogic staff 1527 Jan 13 22:01 google-gemini-cli-session-id.test.ts
-rw-r--r--@ 1 badlogic staff 1929 Jan 16 00:45 google-shared-gemini3-unsigned-tool-call.test.ts
-rw-r--r--@ 1 badlogic staff 1739 Jan 12 23:33 google-thinking-signature.test.ts
-rw-r--r--@ 1 badlogic staff 18761 Dec 25 00:35 handoff.test.ts
-rw-r--r--@ 1 badlogic staff 45401 Jan 13 22:01 image-limits.test.ts
-rw-r--r--@ 1 badlogic staff 15531 Jan 13 22:01 image-tool-result.test.ts
-rw-r--r--@ 1 badlogic staff 2512 Dec 25 04:09 oauth.ts
-rw-r--r--@ 1 badlogic staff 11171 Jan 12 23:33 openai-codex-stream.test.ts
-rw-r--r--@ 1 badlogic staff 38525 Jan 13 22:01 stream.test.ts
-rw-r--r--@ 1 badlogic staff 9181 Jan 13 22:01 tokens.test.ts
-rw-r--r--@ 1 badlogic staff 10581 Jan 13 22:01 tool-call-without-result.test.ts
-rw-r--r--@ 1 badlogic staff 21221 Jan 13 22:01 total-tokens.test.ts
-rw-r--r--@ 1 badlogic staff 22311 Jan 13 22:01 unicode-surrogate.test.ts
-rw-r--r--@ 1 badlogic staff 2215 Dec 25 00:35 xhigh.test.ts
-rw-r--r--@ 1 badlogic staff 654 Jan 12 23:33 zen.test.tsGood, now I understand the testing pattern. The user wants a minimal test that:
Let me look at how the handoff tests work since this is essentially a cross-provider handoff test:
Now I have a good understanding. Let me look at how the Antigravity models work:
Let me check how to get the Antigravity models and stream properly:
>cd /Users/badlogic/workspaces/pi-mono && grep -r "google-antigravity" packages/ai/src --include="*.ts" -lpackages/ai/src/providers/google-gemini-cli.ts
packages/ai/src/cli.ts
packages/ai/src/utils/oauth/types.ts
packages/ai/src/utils/oauth/index.ts
packages/ai/src/types.ts
packages/ai/src/models.generated.ts>cd /Users/badlogic/workspaces/pi-mono && grep -r "gemini-3" packages/ai/src/generated/models.ts | head -10grep: packages/ai/src/generated/models.ts: No such file or directoryLet me check the models.generated.ts and get the model setup for the Antigravity tests:
>cd /Users/badlogic/workspaces/pi-mono && grep -A 20 "google-antigravity" packages/ai/src/models.generated.ts | head -40"google-antigravity": {
"claude-opus-4-5-thinking": {
id: "claude-opus-4-5-thinking",
name: "Claude Opus 4.5 Thinking (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 5,
output: 25,
cacheRead: 0.5,
cacheWrite: 6.25,
},
contextWindow: 200000,
maxTokens: 64000,
} satisfies Model<"google-gemini-cli">,
"claude-sonnet-4-5": {
id: "claude-sonnet-4-5",
name: "Claude Sonnet 4.5 (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: false,
input: ["text", "image"],
cost: {
input: 3,
output: 15,
cacheRead: 0.3,
cacheWrite: 3.75,
},
contextWindow: 200000,
maxTokens: 64000,
} satisfies Model<"google-gemini-cli">,
"claude-sonnet-4-5-thinking": {
id: "claude-sonnet-4-5-thinking",
name: "Claude Sonnet 4.5 Thinking (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",Now I understand the setup. Let me check what models are available on Antigravity:
>cd /Users/badlogic/workspaces/pi-mono && grep -E "gemini-3.*pro" packages/ai/src/models.generated.ts | head -10"gemini-3-pro-preview": {
id: "gemini-3-pro-preview",
"gemini-3-pro-preview": {
id: "gemini-3-pro-preview",
"gemini-3-pro-high": {
id: "gemini-3-pro-high",
"gemini-3-pro-low": {
id: "gemini-3-pro-low",
"gemini-3-pro-preview": {
id: "gemini-3-pro-preview",>cd /Users/badlogic/workspaces/pi-mono && grep -B 2 -A 15 '"gemini-3-pro-preview"' packages/ai/src/models.generated.ts | head -30maxTokens: 64000,
} satisfies Model<"openai-completions">,
"gemini-3-pro-preview": {
id: "gemini-3-pro-preview",
name: "Gemini 3 Pro Preview",
api: "openai-completions",
provider: "github-copilot",
baseUrl: "https://api.individual.githubcopilot.com",
headers: {"User-Agent":"GitHubCopilotChat/0.35.0","Editor-Version":"vscode/1.107.0","Editor-Plugin-Version":"copilot-chat/0.35.0","Copilot-Integration-Id":"vscode-chat"},
compat: {"supportsStore":false,"supportsDeveloperRole":false,"supportsReasoningEffort":false},
reasoning: true,
input: ["text", "image"],
cost: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
},
contextWindow: 128000,
--
maxTokens: 65536,
} satisfies Model<"google-generative-ai">,
"gemini-3-pro-preview": {
id: "gemini-3-pro-preview",
name: "Gemini 3 Pro Preview",
api: "google-generative-ai",
provider: "google",
baseUrl: "https://generativelanguage.googleapis.com/v1beta",
reasoning: true,
input: ["text", "image"],Now I need to find Gemini 3 Pro Preview on Antigravity:
>cd /Users/badlogic/workspaces/pi-mono && grep -B 1 -A 20 '"google-antigravity":' packages/ai/src/models.generated.ts | grep -A 18 "gemini-3-pro" | head -20(no output)Let me search more broadly for Gemini 3 Pro on Antigravity:
>cd /Users/badlogic/workspaces/pi-mono && grep -A 20 '"google-antigravity":' packages/ai/src/models.generated.ts | grep -A 15 "gemini"api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 5,
output: 25,
cacheRead: 0.5,
cacheWrite: 6.25,
},
contextWindow: 200000,
maxTokens: 64000,
} satisfies Model<"google-gemini-cli">,
"claude-sonnet-4-5": {
id: "claude-sonnet-4-5",
name: "Claude Sonnet 4.5 (Antigravity)",>cd /Users/badlogic/workspaces/pi-mono && grep "gemini-3" packages/ai/src/models.generated.ts | grep -i antigravity(no output)
Command exited with code 1Interesting - there's no Gemini 3 Pro on Antigravity in the models. Let me check which models are available on Antigravity:
>cd /Users/badlogic/workspaces/pi-mono && grep -A 200 '"google-antigravity":' packages/ai/src/models.generated.ts | head -250"google-antigravity": {
"claude-opus-4-5-thinking": {
id: "claude-opus-4-5-thinking",
name: "Claude Opus 4.5 Thinking (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 5,
output: 25,
cacheRead: 0.5,
cacheWrite: 6.25,
},
contextWindow: 200000,
maxTokens: 64000,
} satisfies Model<"google-gemini-cli">,
"claude-sonnet-4-5": {
id: "claude-sonnet-4-5",
name: "Claude Sonnet 4.5 (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: false,
input: ["text", "image"],
cost: {
input: 3,
output: 15,
cacheRead: 0.3,
cacheWrite: 3.75,
},
contextWindow: 200000,
maxTokens: 64000,
} satisfies Model<"google-gemini-cli">,
"claude-sonnet-4-5-thinking": {
id: "claude-sonnet-4-5-thinking",
name: "Claude Sonnet 4.5 Thinking (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 3,
output: 15,
cacheRead: 0.3,
cacheWrite: 3.75,
},
contextWindow: 200000,
maxTokens: 64000,
} satisfies Model<"google-gemini-cli">,
"gemini-3-flash": {
id: "gemini-3-flash",
name: "Gemini 3 Flash (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 0.5,
output: 3,
cacheRead: 0.5,
cacheWrite: 0,
},
contextWindow: 1048576,
maxTokens: 65535,
} satisfies Model<"google-gemini-cli">,
"gemini-3-pro-high": {
id: "gemini-3-pro-high",
name: "Gemini 3 Pro High (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 2,
output: 12,
cacheRead: 0.2,
cacheWrite: 2.375,
},
contextWindow: 1048576,
maxTokens: 65535,
} satisfies Model<"google-gemini-cli">,
"gemini-3-pro-low": {
id: "gemini-3-pro-low",
name: "Gemini 3 Pro Low (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 2,
output: 12,
cacheRead: 0.2,
cacheWrite: 2.375,
},
contextWindow: 1048576,
maxTokens: 65535,
} satisfies Model<"google-gemini-cli">,
"gpt-oss-120b-medium": {
id: "gpt-oss-120b-medium",
name: "GPT-OSS 120B Medium (Antigravity)",
api: "google-gemini-cli",
provider: "google-antigravity",
baseUrl: "https://daily-cloudcode-pa.sandbox.googleapis.com",
reasoning: false,
input: ["text"],
cost: {
input: 0.09,
output: 0.36,
cacheRead: 0,
cacheWrite: 0,
},
contextWindow: 131072,
maxTokens: 32768,
} satisfies Model<"google-gemini-cli">,
},
"google-gemini-cli": {
"gemini-2.0-flash": {
id: "gemini-2.0-flash",
name: "Gemini 2.0 Flash (Cloud Code Assist)",
api: "google-gemini-cli",
provider: "google-gemini-cli",
baseUrl: "https://cloudcode-pa.googleapis.com",
reasoning: false,
input: ["text", "image"],
cost: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
},
contextWindow: 1048576,
maxTokens: 8192,
} satisfies Model<"google-gemini-cli">,
"gemini-2.5-flash": {
id: "gemini-2.5-flash",
name: "Gemini 2.5 Flash (Cloud Code Assist)",
api: "google-gemini-cli",
provider: "google-gemini-cli",
baseUrl: "https://cloudcode-pa.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
},
contextWindow: 1048576,
maxTokens: 65535,
} satisfies Model<"google-gemini-cli">,
"gemini-2.5-pro": {
id: "gemini-2.5-pro",
name: "Gemini 2.5 Pro (Cloud Code Assist)",
api: "google-gemini-cli",
provider: "google-gemini-cli",
baseUrl: "https://cloudcode-pa.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
},
contextWindow: 1048576,
maxTokens: 65535,
} satisfies Model<"google-gemini-cli">,
"gemini-3-flash-preview": {
id: "gemini-3-flash-preview",
name: "Gemini 3 Flash Preview (Cloud Code Assist)",
api: "google-gemini-cli",
provider: "google-gemini-cli",
baseUrl: "https://cloudcode-pa.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
},
contextWindow: 1048576,
maxTokens: 65535,
} satisfies Model<"google-gemini-cli">,
"gemini-3-pro-preview": {
id: "gemini-3-pro-preview",
name: "Gemini 3 Pro Preview (Cloud Code Assist)",
api: "google-gemini-cli",
provider: "google-gemini-cli",
baseUrl: "https://cloudcode-pa.googleapis.com",
reasoning: true,
input: ["text", "image"],
cost: {
input: 0,
output: 0,I see. So the models available on Antigravity are:
The user wants to test:
Now let me create the test. I need to:
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:
[Unreleased] section of packages/ai/CHANGELOG.md. This is required before merge.Good:
toolResult messagesBad:
Ugly:
Questions or Assumptions:
Change summary:
convertMessages() in google-shared.ts to skip unsigned tool calls for Gemini 3 models (instead of converting to text)toolResult messagesTests:
Now let me create the minimal integration test:
Now let me run the test to make sure it works:
>cd /Users/badlogic/workspaces/pi-mono/packages/ai && npm test -- test/antigravity-cross-provider-tool-call.test.ts 2>&1> @mariozechner/pi-ai@0.47.0 test
> vitest --run test/antigravity-cross-provider-tool-call.test.ts
RUN v3.2.4 /Users/badlogic/workspaces/pi-mono/packages/ai
stdout | test/antigravity-cross-provider-tool-call.test.ts > Antigravity Cross-Provider Tool Call > should handle Claude tool call followed by Gemini 3 without mimicry
Step 1: Sending request to Claude via Antigravity...
stdout | test/antigravity-cross-provider-tool-call.test.ts > Antigravity Cross-Provider Tool Call > should handle Claude tool call followed by Gemini 3 without mimicry
Claude made tool call: echo { message: [32m'Hello from Claude'[39m }
Step 2: Sending request to Gemini 3 Pro via Antigravity...
stdout | test/antigravity-cross-provider-tool-call.test.ts > Antigravity Cross-Provider Tool Call > should handle Claude tool call followed by Gemini 3 without mimicry
Gemini response stopReason: toolUse
Gemini response content: [ [32m'toolCall'[39m, [32m'text: ...'[39m ]
SUCCESS: Gemini made actual tool call: echo { message: [32m'Hello from Gemini'[39m }
✓ test/antigravity-cross-provider-tool-call.test.ts (1 test) 8272ms
✓ Antigravity Cross-Provider Tool Call > should handle Claude tool call followed by Gemini 3 without mimicry 8271ms
Test Files 1 passed (1)
Tests 1 passed (1)
Start at 23:24:08
Duration 9.47s (transform 247ms, setup 0ms, collect 802ms, tests 8.27s, environment 0ms, prepare 50ms)The test passes. Let me verify the test works with the current PR changes. The output shows:
This confirms the fix is working. Let me clean up by printing a summary.
Test passes. The test confirms:
echo with { message: 'Hello from Claude' })stopReason: toolUse, echo with { message: 'Hello from Gemini' }) rather than mimicking the text formatThe 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:
Something like:
[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}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:
Good, I see the current code on main. Now let me update it to include a clearer message that tells the model this is historical context and not a format to mimic:
Now I need to update the unit test as well since it was checking for the old format:
Now update the unit test to reflect the new format:
Now update the test to check for the new message format:
Now run both tests to make sure they pass:
>cd /Users/badlogic/workspaces/pi-mono/packages/ai && npm test -- test/google-shared-gemini3-unsigned-tool-call.test.ts 2>&1> @mariozechner/pi-ai@0.47.0 test
> vitest --run test/google-shared-gemini3-unsigned-tool-call.test.ts
RUN v3.2.4 /Users/badlogic/workspaces/pi-mono/packages/ai
✓ test/google-shared-gemini3-unsigned-tool-call.test.ts (1 test) 2ms
Test Files 1 passed (1)
Tests 1 passed (1)
Start at 23:40:34
Duration 439ms (transform 37ms, setup 0ms, collect 75ms, tests 2ms, environment 0ms, prepare 66ms)Unit test passes. Now let's run the live integration test to confirm the fix still works:
>cd /Users/badlogic/workspaces/pi-mono/packages/ai && npm test -- test/antigravity-cross-provider-tool-call.test.ts 2>&1> @mariozechner/pi-ai@0.47.0 test
> vitest --run test/antigravity-cross-provider-tool-call.test.ts
RUN v3.2.4 /Users/badlogic/workspaces/pi-mono/packages/ai
stdout | test/antigravity-cross-provider-tool-call.test.ts > Antigravity Cross-Provider Tool Call > should handle Claude tool call followed by Gemini 3 without mimicry
Step 1: Sending request to Claude via Antigravity...
stdout | test/antigravity-cross-provider-tool-call.test.ts > Antigravity Cross-Provider Tool Call > should handle Claude tool call followed by Gemini 3 without mimicry
Claude made tool call: echo { message: [32m'Hello from Claude'[39m }
Step 2: Sending request to Gemini 3 Pro via Antigravity...
stdout | test/antigravity-cross-provider-tool-call.test.ts > Antigravity Cross-Provider Tool Call > should handle Claude tool call followed by Gemini 3 without mimicry
Gemini response stopReason: toolUse
Gemini response content: [ [32m'thinking'[39m, [32m'toolCall'[39m, [32m'text: ...'[39m ]
SUCCESS: Gemini made actual tool call: echo { message: [32m'Hello from Gemini'[39m }
✓ test/antigravity-cross-provider-tool-call.test.ts (1 test) 5983ms
✓ Antigravity Cross-Provider Tool Call > should handle Claude tool call followed by Gemini 3 without mimicry 5982ms
Test Files 1 passed (1)
Tests 1 passed (1)
Start at 23:40:38
Duration 6.98s (transform 258ms, setup 0ms, collect 552ms, tests 5.98s, environment 0ms, prepare 77ms)Both tests pass. The integration test confirms that Gemini 3 Pro makes an actual tool call (stopReason: toolUse) instead of mimicking the format.
Let me run the full check to make sure everything compiles:
Both tests pass. Let me run the full check:
>cd /Users/badlogic/workspaces/pi-mono && npm run check 2>&1> pi-monorepo@0.0.3 check
> biome check --write . && tsgo --noEmit && npm run check -w @mariozechner/pi-web-ui
packages/ai/test/antigravity-cross-provider-tool-call.test.ts:16:8 lint/correctness/noUnusedImports FIXABLE ━━━━━━━━━━
! This import is unused.
14 │ import { complete } from "../src/stream.js";
15 │ import type { Context, Tool } from "../src/types.js";
> 16 │ import { StringEnum } from "../src/utils/typebox-helpers.js";
│ ^^^^^^^^^^^^^^
17 │ import { resolveApiKey } from "./oauth.js";
18 │
i Unused imports might be the result of an incomplete refactoring.
i Unsafe fix: Remove the unused imports.
14 14 │ import { complete } from "../src/stream.js";
15 15 │ import type { Context, Tool } from "../src/types.js";
16 │ - import·{·StringEnum·}·from·"../src/utils/typebox-helpers.js";
17 16 │ import { resolveApiKey } from "./oauth.js";
18 17 │
Skipped 1 suggested fixes.
If you wish to apply the suggested (unsafe) fixes, use the command biome check --write --unsafe
Checked 407 files in 366ms. Fixed 1 file.
Found 1 warning.
> @mariozechner/pi-web-ui@0.47.0 check
> biome check --write . && tsc --noEmit && cd example && biome check --write . && tsc --noEmit
Checked 73 files in 48ms. No fixes applied.
Checked 3 files in 8ms. No fixes applied.There's an unused import warning. Let me fix it:
Now let me run check again:
>cd /Users/badlogic/workspaces/pi-mono && npm run check 2>&1> pi-monorepo@0.0.3 check
> biome check --write . && tsgo --noEmit && npm run check -w @mariozechner/pi-web-ui
Checked 407 files in 381ms. No fixes applied.
> @mariozechner/pi-web-ui@0.47.0 check
> biome check --write . && tsc --noEmit && cd example && biome check --write . && tsc --noEmit
Checked 73 files in 52ms. No fixes applied.
Checked 3 files in 8ms. No fixes applied.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:
[Tool Call: ${block.name}]Arguments: ${argsStr}to:
[Historical context: a different model called tool "${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.
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:
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."
>rm /Users/badlogic/workspaces/pi-mono/packages/ai/test/antigravity-cross-provider-tool-call.test.ts(no output)Now commit and push. Let me check which files I modified:
>cd /Users/badlogic/workspaces/pi-mono && git statusOn branch main
Your branch is up to date with 'origin/main'.
Changes not staged for commit:
(use "git add <file>..." to update what will be committed)
(use "git restore <file>..." to discard changes in working directory)
modified: packages/ai/src/providers/google-shared.ts
modified: packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts
modified: packages/coding-agent/CHANGELOG.md
modified: packages/coding-agent/src/core/settings-manager.ts
modified: packages/coding-agent/src/modes/interactive/components/settings-selector.ts
modified: packages/coding-agent/src/modes/interactive/interactive-mode.ts
no changes added to commit (use "git add" and/or "git commit -a")I see there are other modified files that aren't mine (coding-agent stuff). I'll only commit the ai package files that I modified:
>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."Running formatting, linting, and type checking...
> pi-monorepo@0.0.3 check
> biome check --write . && tsgo --noEmit && npm run check -w @mariozechner/pi-web-ui
Checked 406 files in 349ms. No fixes applied.
> @mariozechner/pi-web-ui@0.47.0 check
> biome check --write . && tsc --noEmit && cd example && biome check --write . && tsc --noEmit
Checked 73 files in 45ms. No fixes applied.
Checked 3 files in 8ms. No fixes applied.
✅ All pre-commit checks passed!
[main 5d3e7d5a] fix(ai): preserve unsigned tool call context for Gemini 3 with anti-mimicry note
2 files changed, 6 insertions(+), 2 deletions(-)Good, committed. Now push:
>cd /Users/badlogic/workspaces/pi-mono && git pushTo https://github.com/badlogic/pi-mono.git
1405e304..5d3e7d5a main -> mainNow close the PR with a comment in the user's tone. The user's tone is direct, technical, no fluff.
>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."✓ Closed pull request badlogic/pi-mono#793 (fix(ai): skip unsigned Gemini 3 tool calls to prevent mimicry)Done. The PR is closed with a comment explaining why we went a different route.
Done. PR closed with explanation of why we took a different approach (preserving context vs lobotomizing it).
[Previous model called tool: ${block.name}]Arguments: ${argsStr}(This is historical context. Do not mimic this format - use proper function calling.)[Historical tool call by previous model - do not mimic this format]Tool: ${block.name}Args: ${argsStr}[Tool Call: ${block.name}]Arguments: ${argsStr}[Historical context: a different model called tool "${block.name}" with arguments: ${argsStr}. Do not mimic this format - use proper function calling.]