fix(opencode): fold schema items into gemini union branches

[Role1776]

Issue for this PR

Closes #43494

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

When a tool input is declared as a nullable array (type: ["null", "array"]), @ai-sdk/google's convertJSONSchemaToOpenAPISchema emits anyOf: [{ type: "array" }] but leaves a dangling sibling items at the parent level, and Gemini rejects the function declaration.

This PR adds a foldArrayItems helper in packages/opencode/src/session/llm/request.ts that recursively folds the sibling items into the array-typed branches of an anyOf or oneOf union. allOf is deliberately left alone, since folding into every branch there would change the intersection semantics. It works on a structuredClone of the schema rather than mutating it in place, so tool schemas shared across providers are not affected by a Gemini specific request. Non-array members of a multi-type union are preserved. It only runs for models whose npm is @ai-sdk/google or @ai-sdk/google-vertex.

How did you verify your code works?

A tool whose input schema is type: ["null", "array"] now produces a Gemini function declaration with items nested inside the array branch.

Added test/session/llm-fold-array-items.test.ts, 10 cases covering nullable array unions, multi-type unions, nested schemas, non-schema objects and idempotency.

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[Enough1122]

AI code review — automated review for reference; please use your judgment.

The workaround direction is right — fixing the schema before @ai-sdk/google's OpenAPI conversion instead of patching generated output — and correctly scoped to google/google-vertex only. Some concerns:

• packages/opencode/src/session/llm/request.ts:184-193 — foldArrayItems mutates the tool's inputSchema object in place, and tool objects are typically defined once at module scope. Why it matters: the first request through a Google model permanently rewrites the shared schema for every later request, including other providers — they'll now see the anyOf form instead of their preferred type:["null","array"] shape. It happens to be idempotent on re-entry, but cross-provider contamination by hidden global mutation is fragile. Suggestion: structuredClone the plain schema per request and return a shallow-copied tool ({ ...tool, inputSchema: clone }) from the map.

• packages/opencode/src/session/llm/request.ts (foldArrayItems, type-array branch) — A multi-type union like type: ["string","array"] collapses to anyOf:[{type:"array",items}] (plus null branch), silently discarding the "string" alternative. Rare in practice, but zod-style pipelines can emit it, and the tool then rejects valid string inputs at the model/API boundary. Suggestion: carry non-array, non-null members over as extra branches.

• packages/opencode/src/session/llm/request.ts (combiner tail) — When a combiner exists but no branch is array-typed, the final delete schema.items silently drops the constraint, weakening validation with no signal. Suggestion: keep the items (or log) when nothing matched, rather than quiet data loss.

• packages/opencode/src/session/llm/request.ts:185-187 — The extraction heuristic isRecord(schema.jsonSchema) ? schema.jsonSchema : schema will happily treat a raw Zod instance (or any object lacking those keys) as a plain JSON schema and traverse/mutate its internals. Suggestion: verify the candidate actually looks like JSON Schema (has one of type/properties/items/$ref/anyOf…) before folding.

• Testing — This recursive transformer has exactly the kind of edge cases that regress silently (nested nullable arrays, anyOf branches that already carry items, double-folding idempotency, multi-type unions) and the PR adds no tests. Suggestion: add unit tests covering at least the nullable-array fix case plus idempotency.

• Nit: folding items into every branch of an allOf changes intersection semantics; consider restricting the combiner fix to anyOf/oneOf.

No blocking issues found.

[Role1776]

Addressed the review points in f6a0c55:

foldArrayItems now works on a structuredClone of the schema instead of mutating in place, so schemas shared across providers stay untouched.

Non array members of a multi-type union are kept instead of being dropped.

Added test/session/llm-fold-array-items.test.ts, 10 cases covering nullable array unions, nested schemas, non-schema objects and idempotency.

I also updated the PR description above, it still described the old in-place version.

@kitlangton you are assigned on #43494, could you take a look when you have a moment?

[axiom-of-choice]

With this applied, the Kaggle MCP tools still fail on gemini-3.8-flash: 22 properties: only allowed for OBJECT type errors from nullable objects (type: ["object","null"]). #49392 moves properties/required as well as items, and gets that request to 200.