dsh-messages-sanitizer · 自动修复孤立的 tool_calls，让 400 INVALID_REQUEST 不再卡死会话

[Leeminjing]

dsh-messages-sanitizer · 自动修复孤立的 tool_calls，让 400 INVALID_REQUEST 不再卡死会话 Automatically repair orphaned tool_calls so 400 INVALID_REQUEST errors no longer break the session.

一句话

工具调度崩溃后，会话历史里会留下「没有配对结果的 tool_calls」（或孤立的 tool 消息），下一轮请求就被 DeepSeek 接口反复用 400 INVALID_REQUEST 拒绝、会话卡死。本插件在运行时自动把 messages 数组修回合法，让对话继续——不改 harness 源码。

解决什么问题

OpenAI 兼容协议要求 tool_calls 与 tool 结果成对且紧邻出现。一次崩溃（例如工具调度在「记录 assistant tool_calls 之后、产出 tool 结果之前」中断）就会留下孤儿调用：

[..., assistant{tool_calls:[write]}, user{...}]        ← 非法

接口返回 400，重试时历史原封不动，于是被反复拒绝。这个故障类在官方仓库也有跟踪：#4843（残缺或孤立的 tool_calls 记录导致接口返回 400）。

怎么工作（四层防御）

1. 自动续跑（agent/status，主路径）—— 崩溃后把错误原样上报给 LLM 并唤醒它，下一步换工具 / 重试 / 向用户报告由 LLM 自己决定，保证 AI 消息始终是最后一条、对话不挂起。
2. 预防（agent/pre-step）—— 下一轮请求构建前，把合成 error tool-result 插到该步消息最前面，从源头杜绝 400。
3. 治愈（agent/request-error）—— 若仍因配对/紧邻违规返回 400，用 surface 替换修复历史后强制重试一次（幂等，不会无限重试）。
4. 兜底（llm/stream）—— 对每个请求做纯数组矫正（配对 + 紧邻重排 + 孤儿/重复/空消息丢弃）。

安装(已收录于 awesome-dsh-plugin 精选榜，已上架插件市场)

dsh plugin --profile web add dsh-messages-sanitizer

与 #4843 的关系：互补，而非替代

#4843 在 harness 源码层根治（agent-loop 剥离、llm-deepseek 合成 fallback id、compaction 按 callId 而非计数配对）。本插件是一个不改 DSH 源码、在运行时自动矫正 messages 数组的安全网：已中毒的旧会话、或跑在仍带该 bug（或工具调度崩溃）的 harness 版本上的用户，都能被自动救回并继续，不必等发版。

链接与状态

• 源码：https://github.com/Leeminjing/dsh-messages-sanitizer
• npm：dsh-messages-sanitizer → https://www.npmjs.com/package/dsh-messages-sanitizer
• 已收录于 awesome-dsh-plugin 精选榜
• MIT · 49 个测试（含真实 cordis + Session 集成）

欢迎反馈；尤其如果你有真实崩溃日志，欢迎用它来验证。

---

dsh-messages-sanitizer · auto-repairs orphaned tool_calls so 400 INVALID_REQUEST never stalls your session

In one line

After a tool-dispatch crash, session history is left with tool_calls that have no paired result (or orphaned tool messages); the next request is then rejected by the DeepSeek API with 400 INVALID_REQUEST, and the session stalls. This plugin auto-repairs the messages array at runtime so the conversation continues — without touching the harness source.

The problem

The OpenAI-compatible protocol requires tool_calls to be paired and adjacent to their tool results. A crash (e.g. tool dispatch interrupted after recording the assistant tool_calls but before producing the tool result) leaves an orphaned call:

[..., assistant{tool_calls:[write]}, user{...}]        ← invalid

The API returns 400, and because the history is unchanged on retry, it is rejected again and again. This class of failure is tracked upstream: #4843 ("Incomplete or isolated tool_calls records cause the DeepSeek API to return 400").

How it works (four layers of defense)

1. Auto-resume (agent/status, primary path) — reports the error verbatim back to the LLM and wakes it; the LLM itself decides whether to switch tools, retry, or tell the user, so the AI message is always last and the conversation never hangs.
2. Prevention (agent/pre-step) — prepends a synthetic error tool-result to the next request, eliminating the 400 at the source.
3. Healing (agent/request-error) — if the API still returns 400 on a pairing/adjacency violation, repairs the history via surface replacement and forces one retry (idempotent — no infinite loop).
4. Last resort (llm/stream) — pure array correction on every request (pairing + adjacency reorder + orphan/duplicate/empty drops).

Install(Listed in the awesome-dsh-plugin curated list)

dsh plugin --profile web add dsh-messages-sanitizer

Relation to #4843: complementary, not a replacement

#4843 fixes the root cause in the harness source (agent-loop stripping, llm-deepseek fallback id, compaction pairing by callId rather than count). This plugin is a runtime safety net that never modifies the DSH source — already-poisoned sessions, or sessions on harness builds that still carry the bug (or a tool-dispatch crash), are repaired and continue without waiting for the next release.

Links & status

• Source: https://github.com/Leeminjing/dsh-messages-sanitizer
• npm: dsh-messages-sanitizer → https://www.npmjs.com/package/dsh-messages-sanitizer
• Listed in the awesome-dsh-plugin curated list
• MIT · 49 tests (incl. real cordis + Session integration)

Feedback welcome; if you have a real crash log, please try it against this plugin.