Anthropic 400: tools.N.custom.input_schema.type Field required (tool schema missing top-level `type`)

[seotk0319]

요약 / Summary

Anthropic(Claude) 백엔드로 사용할 때, Codex가 서브에이전트/브라우저/코덱스 툴 등 추가 툴을 붙이는 순간 아래 400이 발생합니다.

Provider error 400: {"type":"error","error":{"type":"invalid_request_error","message":"tools.26.custom.input_schema.type: Field required"},"request_id":"req_011CcqiYNzcinns1gsUkV9Cg"}

When using an Anthropic (Claude) backend, requests fail as soon as extra tools are attached (sub-agent, browser use, Codex tools). Anthropic rejects a tool whose input_schema lacks a top-level type.

환경 / Environment

• Package: @bitkyc08/opencodex@2.6.31-preview.20260707 (same code path exists upstream)
• OS: Windows
• Backend: Anthropic (Claude Opus)
• Trigger: only when the tool list grows (sub-agent / browser / Codex tools). Basic chat without those tools works fine.

원인 / Root cause

src/adapters/anthropic.ts 의 toolsToAnthropicFormat 가 input_schema 로 t.parameters 를 그대로 전달합니다. 일부 Codex/MCP/서브에이전트/브라우저 툴의 파라미터 스키마에는 최상위 type: "object" 가 없습니다. OpenAI 와이어는 이를 허용하지만 Anthropic Messages API 는 input_schema.type 을 필수로 검증하므로 400 을 반환합니다.

toolsToAnthropicFormat passes t.parameters straight through as input_schema. Some Codex/MCP/sub-agent/browser tools ship a parameter schema without a top-level type: "object". The OpenAI wire tolerates this, but the Anthropic Messages API requires input_schema.type, so it 400s. The failing index (e.g. tools.26) is one of the extra tools, which is why plain chat works but adding tools breaks.

const converted = tools.map(t => ({
  name: toolNames.toWire(namespacedToolName(t.namespace, t.name)),
  description: t.description,
  input_schema: t.parameters, // <- schema without top-level `type` is forwarded verbatim
}));

제안 수정 / Proposed fix

input_schema 를 Anthropic 요구사항에 맞게 정규화합니다 (최상위 type 누락 시 object 보정 + properties 보장).

// Anthropic requires every tool's `input_schema` to be a JSON Schema object with an
// explicit top-level `type` (it validates `tools.N.custom.input_schema.type`). Codex/
// OpenAI Responses tool params frequently omit it, which the OpenAI wire tolerates but
// Anthropic 400s on. Normalize so a missing/invalid top-level `type` becomes an object
// schema with `properties`.
function normalizeAnthropicInputSchema(schema: unknown): Record<string, unknown> {
  if (!schema || typeof schema !== "object" || Array.isArray(schema)) {
    return { type: "object", properties: {} };
  }
  const s: Record<string, unknown> = { ...(schema as Record<string, unknown>) };
  if (typeof s.type !== "string") s.type = "object";
  if (s.type === "object" && (s.properties === undefined || s.properties === null)) {
    s.properties = {};
  }
  return s;
}

// ...in toolsToAnthropicFormat:
input_schema: normalizeAnthropicInputSchema(t.parameters),

재현 / Reproduction

1. Configure an Anthropic (Claude) backend.
2. In Codex, trigger a sub-agent, browser use, or Codex tools so the tool list grows.
3. Request 400s with tools.N.custom.input_schema.type: Field required.

로컬에서 위 패치 적용 후 서브에이전트/브라우저/코덱스 툴 호출이 정상 동작하는 것을 확인했습니다. Verified locally: after applying the patch, sub-agent/browser/Codex tool calls work. Happy to open a PR if the approach looks good.

[seotk0319]

정정 / Correction — the type-only fix is not enough

원래 제안한 "최상위 type 보정"만으로는 해결되지 않았습니다. Anthropic Messages API는 input_schema 루트에 oneOf/anyOf/allOf 콤비네이터가 있으면 그것도 거부합니다. Codex/OpenAI Responses 쪽 일부 툴(그리고 MCP 툴)이 루트에 union 스키마를 그대로 내보내는 경우가 있어서, type만 채워도 같은 400이 계속 납니다.

Adding a top-level type alone does not fix it. The Anthropic Messages API also rejects a top-level oneOf/anyOf/allOf on input_schema. Some Codex/OpenAI Responses (and MCP) tools emit a root-level union schema, so filling in type still 400s.

완전한 원인 / Full root cause

• input_schema 루트에 type 누락 → 거부
• input_schema 루트에 oneOf/anyOf/allOf 존재 → 거부 (프로퍼티 내부에 중첩된 anyOf 등은 허용)

완전한 수정 / Complete fix (src/adapters/anthropic.ts)

루트 콤비네이터를 평탄화합니다: 각 branch의 properties를 루트 object 스키마로 병합, required는 allOf에서만 병합(oneOf/anyOf는 과도 강제 방지), 중첩된 anyOf는 그대로 보존, 그 외 루트 키(title, additionalProperties 등)는 유지.

// Anthropic requires every tool's `input_schema` to be a JSON Schema object with an
// explicit top-level `type` (it validates `tools.N.custom.input_schema.type`). It also
// rejects top-level oneOf/allOf/anyOf, even though Codex/OpenAI Responses tools may emit
// them. Normalize only the root schema so missing/invalid `type` becomes an object and
// root compositions are flattened while nested property schemas stay intact.
function normalizeAnthropicInputSchema(schema: unknown): Record<string, unknown> {
  if (!schema || typeof schema !== "object" || Array.isArray(schema)) {
    return { type: "object", properties: {} };
  }
  const obj = schema as Record<string, unknown>;
  const compositionKeys = ["oneOf", "anyOf", "allOf"] as const;
  const hasRootComposition = compositionKeys.some(k => Array.isArray(obj[k]));

  if (!hasRootComposition) {
    const s: Record<string, unknown> = { ...obj };
    if (typeof s.type !== "string") s.type = "object";
    if (s.type === "object" && (s.properties === undefined || s.properties === null)) {
      s.properties = {};
    }
    return s;
  }

  const properties: Record<string, unknown> = {};
  const required = new Set<string>();

  if (obj.properties && typeof obj.properties === "object" && !Array.isArray(obj.properties)) {
    Object.assign(properties, obj.properties as Record<string, unknown>);
  }
  if (Array.isArray(obj.required)) {
    for (const item of obj.required) if (typeof item === "string") required.add(item);
  }

  for (const key of compositionKeys) {
    const variants = obj[key];
    if (!Array.isArray(variants)) continue;
    const mergeRequired = key === "allOf";
    for (const variant of variants) {
      if (!variant || typeof variant !== "object" || Array.isArray(variant)) continue;
      const v = variant as Record<string, unknown>;
      if (v.properties && typeof v.properties === "object" && !Array.isArray(v.properties)) {
        Object.assign(properties, v.properties as Record<string, unknown>);
      }
      if (mergeRequired && Array.isArray(v.required)) {
        for (const item of v.required) if (typeof item === "string") required.add(item);
      }
    }
  }

  const s: Record<string, unknown> = {};
  for (const [key, value] of Object.entries(obj)) {
    if (key === "oneOf" || key === "anyOf" || key === "allOf") continue;
    if (key === "type" || key === "properties" || key === "required") continue;
    s[key] = value;
  }
  s.type = "object";
  s.properties = properties;
  if (required.size > 0) s.required = [...required];
  return s;
}

toolsToAnthropicFormat 내부는 동일하게 input_schema: normalizeAnthropicInputSchema(t.parameters) 로 연결합니다.

검증 / Verification

파일에 적용한 실제 함수 본문을 그대로 실행하여 다음 케이스를 확인했습니다:

• type 누락 → {type:"object",properties:{}}
• root anyOf / oneOf → 각 branch properties 병합, required 강제 안 함
• root allOf → properties + required 병합 (["p","r"])
• nested anyOf (프로퍼티 내부) → 원형 그대로 보존
• root union + 여분 키(title, additionalProperties) → 여분 키 유지, 콤비네이터만 제거

이 상태에서 서브에이전트/브라우저/Codex 툴 호출이 모두 정상 동작합니다. Original title/body should be read together with this correction. Happy to open a PR with this complete version.

[Ingwannu]

Confirmed valid and opened a fix PR: #76

What I verified:

• Current dev was forwarding Anthropic input_schema: t.parameters verbatim.
• A schema with properties/required but no root type stayed missing type.
• A root anyOf schema also stayed at the root, matching the corrected report here.

Fix in #76:

• Normalize Anthropic tool input_schema to a root type: "object" schema.
• Ensure properties exists.
• Flatten root oneOf / anyOf / allOf while preserving nested property schemas.
• Merge required only for allOf, not oneOf / anyOf.

Validation:

• bun test tests/anthropic-tool-schema.test.ts tests/anthropic-empty-content.test.ts tests/kiro-adapter.test.ts -> 38 pass
• bun run typecheck -> pass
• bun test ./tests/ -> 1677 pass, 0 fail

The PR body includes Fixes #75, so this issue should close automatically when the PR is merged.

[lidge-jun]

Fixed and verified. The Anthropic adapter now normalizes tool input_schema roots so schemas without a top-level type are sent as object schemas, and the regression coverage is in tests/anthropic-tool-schema.test.ts.\n\nVerified locally with:\n- bun test tests/anthropic-tool-schema.test.ts tests/anthropic-empty-content.test.ts tests/kiro-adapter.test.ts\n- bun test tests/multi-agent-compat.test.ts tests/native-model-toggle.test.ts\n- npm run build (gui)\n- bun run typecheck\n\nAlso verified on dev Cross-platform CI: https://github.com/lidge-jun/opencodex/actions/runs/29034128441