fix: DeepSeek tool call parsing - nested objects & JSON repair

[valkryhx]

Summary

修复 DeepSeek 工具调用解析问题，支持嵌套 JSON 对象和缺失数组括号的自动修复。

Problem

DeepSeek 在返回 tool calls 时有时会输出不规范的 JSON：

• 缺失数组方括号：{"todos": {"content": "task1"}, {"content": "task2"}}
• 嵌套对象中的方括号丢失：{"input": {"q": "value"}}, {"input": {"path": "file"}}
• 未加引号的键名：{tool_calls: [...]}

这些情况导致工具调用被当作普通文本返回，客户端无法识别和执行。

Solution

1. 升级正则表达式支持单层嵌套

// 修复前：无法处理嵌套 {}
var missingArrayBracketsPattern = regexp.MustCompile(`(:\s*)(\{[^{}]*\}(?:\s*,\s*\{[^{}]*\})+)`)

// 修复后：支持单层嵌套
var missingArrayBracketsPattern = regexp.MustCompile(`(:\s*)(\{(?:[^{}]|\{[^{}]*\})*\}(?:\s*,\s*\{(?:[^{}]|\{[^{}]*\})*\})+)`)

2. 添加 RepairLooseJSON 函数

• 修复未加引号的键名：{key: -> {"key":
• 修复缺失的数组括号：{"a":1}, {"b":2} -> [{"a":1}, {"b":2}]

3. 增强关键词检测

支持多种 tool call 语法：

• tool_calls
• function.name:
• [tool_call_history]

4. 添加 OOM 保护

• 限制回溯搜索范围
• 限制 JSON 对象扫描长度

Files Changed

File	Change
internal/util/toolcalls_parse.go	新增 RepairLooseJSON 函数和正则修复
internal/util/toolcalls_test.go	新增 10+ 个测试用例
internal/util/toolcalls_candidates.go	增强关键词检测和 OOM 保护
internal/adapter/openai/tool_sieve_core.go	支持多关键词模式匹配
internal/js/helpers/stream-tool-sieve/sieve.js	JS 版本同步修复
TESTING.md	新增调试命令文档
README.md	新增测试快速参考

Test Cases Added

1. 单层嵌套对象 - 2个元素
2. 3个单层嵌套对象
3. 混合嵌套 - 对象和原始值混合
4. 4个嵌套对象
5. 无空格逗号分隔
6. 对象内包含数组
7. DeepSeek 8皇后真实输出
8. 简单无嵌套对象
9. 复杂单层嵌套
10. 5个嵌套对象

Verification

go test -v -run 'TestParseToolCalls|TestRepair' ./internal/util/
# All 33 tests pass

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7318d1f4a8

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Restrict capture start to payloads consumeToolCapture can parse

Including function.name: and [tool_call_history] in findToolSegmentStart makes the sieve enter capture mode for text-kv style output, but consumeToolCapture still only looks for tool_calls, so these captures never become ready until stream finalization; in practice, any chunk containing function.name: can cause the rest of the response to be buffered and emitted late as plain text instead of structured tool calls.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Escape \n/\t when repairing Windows-style path backslashes

The repair routine treats \n, \t, \r, etc. as already-valid escapes and leaves them unchanged, so inputs like C:\new\tools (model intended literal backslashes) are decoded as newline/tab characters after JSON unmarshal; this silently mutates tool arguments and can execute the wrong path/command instead of preserving the original Windows path text.

Useful? React with 👍 / 👎.

[vercel]

@valkryhx is attempting to deploy a commit to the cjack's projects Team on Vercel.

A member of the Team first needs to authorize it.