模型已经开始生成工具参数，流也在继续返回，Agent 却在真正调用工具之前失败了。排查到最后，触发拒绝的不是错误工具名，也不是非法参数，而是后续碎片里一个 id: null。
这类修复很容易走向两个极端：继续拒绝所有 null，或者看到 null 就宽松放行。我们最后采用的是第三种办法：只在同一索引已经建立非空调用身份时，把后续 null 解释为“不更新身份”。
这是 open-etf-agent v0.15.2 的一个窄范围兼容修复。它没有发明工具 ID，没有改变工具授权，也没有跳过完整调用的最终校验。
一、工具调用是增量事件，不是每帧一个完整对象
对流式 Chat Completions，一次工具调用可能分散在多个事件中。应用按本次响应里的 index 找到正在组装的调用，再拼接函数名和参数字符串，最后把完整提议交给校验与执行链。
下面是简化的合成碎片，省略了外层 SSE、choices 和结束事件，不是原供应商载荷：
{"index":0,"id":"call-a","function":{"name":"read_page","arguments":"{\"page\":"}}
{"index":0,"id":null,"function":{"arguments":"1}"}}
第二片段仍然属于索引 0。函数参数从 {"page": 继续变成 {"page":1}；真正需要决定的是，后面的 null 是否应该推翻前面已经确定的 call-a。
旧实现的判断接近“只要提供了 id 字段，就必须是字符串，且不能冲突”。这对完整对象校验很直观，却把增量事件中的“没有身份更新”也当成了新的身份值。
关键不在 null 本身，而在 输入值与已有状态的组合。
二、先画清允许的状态转移
修复的前提是相同响应内的同一工具索引，而不是碰巧工具名相同，也不是上一轮请求曾经用过某个 ID。
原实现允许早期空字符串在后续被有效字符串补齐，这条既有行为不因本次修复改变。但空字符串不算“已建立的非空身份”，不能拿它为 null 续片开绿灯；完整提议仍须通过最终验证。
把新增例外单独表达，判断就很小：
const hasEstablishedId = typeof call.id === 'string' && call.id.length > 0;
const keepsEstablishedId = part.id === null && hasEstablishedId;
if (part.id !== undefined && !keepsEstablishedId) {
validateOriginalIdRules(call.id, part.id);
call.id = part.id;
}
这是提炼后的教学伪代码，validateOriginalIdRules 代表原有的类型和冲突检查。它不是可单独替换整个流解析器的实现，也没有承担参数、权限或执行验证。
三、兼容发生在组装层，信任不能跟着扩大
如果把“允许 null”写成通用清洗规则，例如给缺失 ID 自动补随机值，程序可能不再报错，却会制造一个供应商没有提供过的调用身份。后续结果关联、去重和诊断就失去了原始依据。
另一个危险做法是拿上一条调用的 ID 填补当前片段。并行工具可能交错返回，这会把不同索引的参数拼到错误调用上。
正确的位置是本次响应、当前索引对应的组装状态：只保留已建立的那个 ID，不跨索引借用，不跨请求继承，也不在冲突时选择一个“看起来更合理”的值。
流程依次定位本次响应索引、检查已有身份：已建立身份的 null 续片保留该身份；其他输入继续原身份校验，非法身份被拒绝。通过后才累积参数、校验完整提议，并交由原 Host 授权；非法提议不派发。
只有后续 null 且已有非空 ID，才走保留身份分支；其余情况继续原检查。最终校验仍检查组装后的参数与调用身份，随后由原 Host 决定工具是否获准执行。兼容片段不是执行许可。
流是否完整结束，也是另一条条件。这个修复没有把半段响应、异常结束或超出支持范围的协议伪装成一个可执行调用。
四、怎样证明修好的是问题，而不是删掉了保护
最小单元测试可以直接喂两个片段，断言 ID 仍为 call-a。但这样还不够：我们真正关心的是，这条调用能不能通过原网关进入运行时，并完成原本的用户任务。
因此，回归使用一条合成的完整链路：读取资料，保存报告，再通过原运行时结束任务。对照组的续片省略 ID，目标组的续片携带 null；两边最终保留相同调用身份，并得到相同的报告内容。
同时要覆盖拒绝分支：首次 null、新索引 null、不同字符串、非法类型，不能因为正例修好了就开始被接受。即使 ID 续片合法，参数拼接后不是有效 JSON，或者最终出现重复调用 ID，也仍然必须失败。
这是一类非常适合用状态转移测试的兼容问题。单测输入不只是“字段有哪些值”，还要包含“这一值到达时已经积累了什么状态”。
后续独立的真实模型任务验证了集成后的正常读取、发布与恢复路径，但不能据此证明那一批流里实际出现过 null 续片。目标边界由受控的 null 反例与正例证明，真实任务验证的是集成没有破坏其他行为。 两种证据不能互相冒充。
五、错误诊断不必保留原始参数
之前多个身份问题都可能被归到一个复合错误里。只看到 TOOL_FRAGMENT_ID，还无法分清是首片段没有身份、后续 ID 冲突，还是类型不受支持。
但为了区分这些情况，并不需要把真实工具参数、用户资料或完整模型输出写进日志。受限诊断可以只保留错误类别、输入 ID 类型、此前是否已有非空 ID，以及字符串是否为空等有限信息。
例如，“输入类型为 null，已有身份为真”有助于定位本次兼容问题；“输入为字符串，已有身份为真且冲突”则应继续走原拒绝路径。它们不是同一个修复理由。
诊断越能描述状态关系，越不需要依赖泄露原始内容来还原现场。
六、一个修复不能包办所有历史故障
这次调查拿到了一次真实资格验证中“已有 ID 后出现 null”的受限诊断证据，足以提出并验证这个具体修复。
更早的复合身份错误没有同样完整的分类信息，不能回头宣布它们也都是 null 引起的。旧失败仍是旧失败，不因为新版本任务通过就改写记录。
同样，某个供应商出现过这种流形态，不等于所有供应商的协议都承诺 null 续片合法。这里实现的是一个有前置条件、保持既有身份的兼容规则，不是对任意响应的普遍放行。
结语：好的兼容修复，增加的是一个有条件的解释
这次源码变更很小，真正需要想清楚的是哪种状态转移可以保留不变量：已有身份不被清除，不同调用不被混合，缺失身份不被伪造，完整提议仍要校验。
遇到下一次协议差异，可以先问：它是完整对象字段，还是增量更新？什么已有状态使它安全？在哪一层解释它，才不会把权限、执行或恢复规则一起放宽？
兼容不是把校验关掉，而是在不改变原事实的前提下，准确接受一种原先未覆盖的表示。
实现与回归可阅读公开仓库中的流组装代码、null 续片回归及版本验证说明。
延伸阅读：从工具调用到模型输入 · 可观测全景中的 Trace 完整性
欢迎为 open-etf-agent 点一个 Star，结合源码阅读这些边界。
原文阅读：一个 null 如何让 Agent 工具调用失败。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The model has begun to generate tool parameters, the flow continues to return, and Agent has failed before the real caller. Finally, it was not the wrong tool name or the illegal parameter that triggered the rejection, but the id: nulul.
Such repairs can easily go to two extremes: continuing to reject all the Nulls, or seeing the Nulls loosely. We ended with a third approach: to interpret the follow-up Null as “no updated identity” only when the same index has established non-air conditioned identity”.
This is a narrow range compatible restoration for open- etf-agent v015.2. It does not invent a tool ID, does not change the tool authorization, and does not skip the final verification of the full call.
I. tool call is an incremental event, not a complete object per frame
tool call may be scattered in multiple incidents. Apply the index of this response to find the call that is being assembled, then spell the function name and parameter string, and then hand the complete proposal over to the check and execute chain.
The following are simplified synthetic fragments, omitting outer layers SSE, choices and closure events, not original supplier loads:
{"index":0,"id":"call-a","function":{"name":"read_page","arguments":"{\"page\":"}}
{"index":0,"id":null,"function":{"arguments":"1}"}}
The second is still index 0. Function parameter from {"page": continues to be {`page":}; what really needs to be decided is whether the left should overturn the previously identified call-a.
The old realization judgement is close to "as long as id field is provided, it must be a string and cannot conflict". This provides a visual view of the whole object, but also considers the “no identity update” of the incremental event as a new identity value.
The key is not nulla per se, and the input value is combined with the existing state.
II. Elaborate permitted state transfers
The restoration is premised on an index of the same tool in the same response, not on the coincidence that the tool name is the same or that an ID has been used in a previous request.
The original realization allows the early empty string to be followed by a valid string completion, which is not changed by this restoration. However, an empty string does not count as a " created non-empty identity " and cannot be used to turn on a green light for a Null sequel; the complete proposal must still be finally validated.
If the new exception were to be expressed separately, the judgement would be small:
const hasEstablishedId = typeof call.id === 'string' && call.id.length > 0;
const keepsEstablishedId = part.id === null && hasEstablishedId;
if (part.id !== undefined && !keepsEstablishedId) {
validateOriginalIdRules(call.id, part.id);
call.id = part.id;
}
It's a refined teaching code. Validate OrganalIdRules represents the original type and conflict check. It is not a stand-alone replacement for the realization of the entire flow solver, nor does it carry parameters, permission or perform authentication.
Consistency takes place on the assembly floor, and trust cannot be expanded.
If “permissive nul” were written as a generic clean-up rule, such as auto-reduced random values for missing IDs, the program might no longer be wrong, but would create a call identity not provided by a supplier. Subsequent results are linked, weighed and diagnosed without the original basis.
Another dangerous approach is to fill the current segment with the last called ID. Parallel tools may be staggered back, and this will spell different index parameters to the wrong call.
The correct position is the current response, the current index corresponds to the assembly state: only the created ID is retained, is not borrowed from the index, does not request succession and does not select a value that appears to be more reasonable in case of conflict.
The process then locates the current response index and checks the existing identity: the Null sequel of the created identity retains that identity; the other entries continue to verify the original identity and the illegal identity is rejected. After adoption, the parameters are accumulated, the complete proposal is verified and submitted to the original Host; the illegal proposal is not circulated.
Only subsequent Null and existing non-empty IDs will keep the identity branch; the rest continues. The final verification still checks after assembly parameters and call identities, and then it is up to the original Host to decide whether the tool is approved for implementation. Compatible clips are not enforcement permits.
The complete end of the flow is another condition. The repairs did not disguise half-response, abnormal termination or agreements beyond the scope of support as an enforceable call.
IV. How to prove that it's a problem, not to delete protection.
The minimum unit test can feed two clips directly, asserting that ID is still called-a. But it's not enough: what we really care about is whether this call can go into operation through the original gateway and do the original user mission.
Therefore, a complete synthetic link is returned: read the information, save the report, and finish the mission when it is run. A follow-up to the control group omits ID, and a follow-up to the target group carries the nudity; the two sides eventually retain the same call identity and receive the same report.
The rejection branch is also to be covered: the first nulla, the new index nulla, the different strings, the illegal type, which cannot begin to be accepted because the normal rule is fixed. Even if the ID replica is legal, the arguments are not valid after the collage, or the duplicate call ID eventually appears, it must fail.
This is a type of compatibility issue that is well suited to the state transfer test. The single measurement input is not only "what is the value of field" but also "what was accumulated when the value arrived".
The subsequent independent real model mission validated normal post-integrated reading, publish and recovery paths, but could not prove that the stream actually had a nulple replica. The target boundary was proven by the controlled nuld inverse and formality that real task certified integration as non-destructive. The two types of evidence cannot pretend to each other.
Evident diagnosis does not have to retain original parameters
Many of the previous identity issues could be attributed to a compound error. Only Tool FRAGMENT ID was seen, and it was not clear whether the first episode was without identity, followed by an ID conflict or whether the type was not supported.
In order to distinguish these situations, however, real tool parameters, user information or complete model outputs need not be recorded in the log. Limited diagnostics can save only limited information such as error categories, type of input ID, whether non-empty IDs exist before, and whether the string is empty.
For example, " Enter type nuld, existing identity is genuine " helps to locate this compatibility problem; " Enter as string, existing identity is real and conflicting " should continue to follow the original rejection path. They are not the same cause of repair.
The more the diagnosis describes the state relationship, the less the need to rely on the disclosure of original content to restore the site.
VI. A repair cannot cover all historical failures
The survey was able to present and validate the specific restoration with restricted diagnostic evidence of the “nuld presence” of an existing ID in a genuine qualification.
Early composite identity errors do not have the same complete sort of information, and they cannot be reciprocated to be fully attributed. Old failures remain old failures, not because the new version of the assignment is passed.
Similarly, the fact that a supplier has experienced this flow pattern does not mean that all suppliers have agreed to commit themselves to the legality of the Null sequel. What is achieved here is a rule of compatibility with pre-conditions and preservation of an established identity, rather than universal release from arbitrary response.
Conclusion: good compatibility restoration, with a conditional explanation
This is a very small change in the source code, and what really needs to be figured out is what kind of state shift can be left unvariable: Existing identities are not cleared, different calls are not mixed, missing identities are not forged, and complete proposals are still subject to verification.
When the next agreement is different, the question can be asked whether it is a complete object field or an incremental update. What has made it safe? At what level would it be explained that the rules of permission, enforcement or restoration would not be relaxed?
Compatibility is not to turn off the verification, but to accept an expression that was not previously covered, without altering the facts.
Achieving and returning can read the current assembly code in open storage, the Null replica return and the version validation instructions.
Extension reading: from tool call to model input Trace integrity in a visible panorama
Welcome to a Star for an open-etf-ent point, which combines source code to read these boundaries.
Original reading: A Null how to make Agent tool call fail.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.