[BUG] llm-deepseek 在网关发送"空字符串续块"时会丢失工具调用的 id/name（聚合中转站）

[jadeweilin]

概述

 当经由某个 OpenAI 兼容的聚合网关、且该网关在工具调用的续块里发送                                                                                                                                       
 空字符串 `id`/`name`（而非省略这两个字段）时，dsh 会把 `tool/call` 记录成                                                                                                                              
 空的 `name`，从而触发 `UNKNOWN_TOOL`，导致所有工具（bash / fs 等）无法分派。                                                                                                                           
 聊天与推理不受影响——只有工具调用受影响。

复现步骤

 1. 将 `llm-deepseek` 的 `baseURL` 指向一个聚合网关，且该网关的流式                                                                                                                                     
    工具调用块在续块中发送 `id: ""` / `function.name: ""`。                                                                                                                                             
 2. 让 agent 执行任意工具（例如 bash 跑 `openspec list`）。                                                                                                                                             
 3. 该会话的 `tool/call` 事件里 `name: ""`、`callId: ""`，工具注册表                                                                                                                                    
    报 `UNKNOWN_TOOL`。

根因

 `packages/llm/llm-deepseek/src/translate.ts` 用 `call.id !== undefined`                                                                                                                                
 来判断是否更新字段。官方 OpenAI/DeepSeek 流式协议在续块中省略 `id`/`name`，                                                                                                                            
 但部分网关改为发送 `""`。空字符串能通过 `!== undefined` 判断，于是被当作                                                                                                                               
 "有值"，把首块里真实的 `id`/`name` 覆盖成了 `""`。`arguments` 走的是                                                                                                                                   
 `+=` 累加路径，不受影响——所以日志呈现典型的"参数完整、名字为空"。

建议修复（已验证，可直接合入）

 // 修复前                                                                                                                                                                                              
 if (call.id !== undefined) block.callId = call.id                                                                                                                                                      
 if (call.function?.name !== undefined) block.name = call.function.name                                                                                                                                                                                                                                                                
 // 修复后                                                                                                                                                                                              
 if (call.id) block.callId = call.id                                                                                                                                                                    
 if (call.function?.name) block.name = call.function.name                                                                                                                                                                                                                                                                                                                                                   
 真值判断同时跳过 `undefined` 和 `""`，因此空串续块不再清除已捕获的值，                                                                                                                                 
 而官方的"续块省略"格式不受影响。这是纯增强，不改变既有行为。

验证情况

 - 在 `tests/translate.spec.ts` 新增 3 个回归用例，覆盖两种线上格式                                                                                                                                     
   （省略式 / 空串式）以及参数累加混合场景。                                                                                                                                                            
 - `translate.spec.ts` 全套通过（33/33），无回退。                                                                                                                                                      
 - `typecheck` 与 `lint` 通过。                                                                                                                                                                         
 - 针对真实聚合网关做了端到端验证：工具调用现在能带上正确的                                                                                                                                             
   `name`/`callId`，工具成功执行。

[liyangbing]

I verified the same translator boundary in both rc.2 (b150a551) and tagged alpha.1 (cd5ef814): both update block.callId / block.name whenever the wire field is not undefined, so explicit "" can still overwrite the valid first delta.

The important regression gap is the sequence, not just an empty fixture. Alpha.1 already tests (a) normal continuation chunks that omit identity and (b) a hypothetical stream that never supplies identity. It does not test:

1. first delta with non-empty id and name;
2. continuation with id: "", name: "", plus arguments;
3. optional continuation with name: null at a lenient boundary;
4. terminal assertion that the first identity survives and every argument fragment is retained.

A truthy update guard should be paired with terminal validation: preserve a prior valid identity, but reject a call that never had one before policy and Session persistence. Otherwise the two cases remain ambiguous.

The cross-version source links, Session-containment steps, and full regression gates are here: https://sandbaseai.github.io/deepseek-harness-handbook/streamed-tool-call-identity.html

Disclosure: I maintain this independent handbook with SandBase; it is not an official DeepSeek AI project.

[jadeweilin]

Thanks — the cross-version confirmation against rc.2 (b150a55) and alpha.1
(cd5ef81) is useful: it shows the !== undefined identity guard is the common
root and the empty-string overwrite spans versions, not just one build.
Two notes on the four-point list — worth flagging that rc.2/alpha.1 predate the
fix, so they don't carry the new regression tests:
● (a) first delta with non-empty id/name, (b) continuation with id:""/name:""
plus arguments, and (d) terminal assertion that the first identity survives and
every argument fragment is retained — these sequences are already covered on
the fix branch (fix/gateway-toolcall-id-extraction). See
packages/llm/llm-deepseek/tests/translate.spec.ts: the gateway-variant,
official-variant, and split-arguments tests each feed a first delta with real
id/name, then continuation deltas carrying id:""/name:"" (or omitted), and
assert the closing block-end retains the original identity with the full
arguments reassembled intact. The pre-existing "never carries id or name"
defensive case covers the always-empty stream.
● (c) name: null at a lenient boundary — agreed, this is a genuine gap. The
truthy guard already treats null as falsy so behavior is correct, but nothing
pins it; we'll add a name: null continuation case.
On pairing the truthy guard with terminal validation: it's a valid robustness
goal, but it lands at a different boundary. The guard fixes identity assembly in
the translator; "reject a call that never had identity" is a policy/persistence-time
decision with its own error semantics, and folding both into this fix would blur
each. I'd keep this change scoped to the overwrite bug and open the
identity-less-call rejection as a separate change — happy to follow that up if the
maintainers agree it's in scope.