Skip to content
This repository was archived by the owner on Nov 9, 2026. It is now read-only.

v0.9.0-rc.2 — messages 端面不再被一条畸形历史打死 + 图片边界说法纠错 (preview)

Pre-release
Pre-release

Choose a tag to compare

@xia-sc xia-sc released this 16 Sep 05:25

v0.9.0-rc.2 — messages 端面不再被一条畸形历史打死 + 图片边界说法纠错

Pre-release(预览版)。修掉 v0.9.0-rc.1 之后发现的一处「一条持久历史打死整会话」
(#3 同一类),并订正上一版对图片
体积的描述。

修复

  • messages 端面:历史里的畸形 tool-call 不再让之后每一轮都失败。
    持久化的 assistant tool-call 只要 arguments 不是合法 JSON,toAnthropicMessages
    就抛 PROVIDER_PROTOCOL_ERROR,于是该会话在 messages 系模型上之后每一轮都失败,
    且错误里不说是哪一条历史。触发者甚至不一定是本插件:chat / responses 两个端面把
    arguments 原样透传(线上本来就是字符串),所以一条畸形调用在 chat 端面落盘后,
    要等会话切到 messages 系模型才引爆。
  • 降级而不是拒绝:这个调用**早就执行过、也有 tool_result**了,模型只需要看见
    「它当时调过这个工具」,所以拒发它毫无收益、代价却是整段会话。新增 toolInput(),
    历史 input 一律退成 {}:
    • 非法 JSON(not-json{、空串);
    • 能解析但不是对象——"5" / "[1]" / "null" / '"text"'。Messages API 要求
      input 是对象,所以这类与非法 JSON 一样会被对端拒绝,而上一版是原样发出去的,
      表现为一条看不出问题来源的上游 400。
  • 创作侧保持严格:pumpMessages 对刚刚流出来的调用仍然抛错。参数不是 JSON 是
    上游违反协议,一个无法执行的调用不该被保留或派发——宿主的 max-tokens 与中断规则也是
    这个取向(宁可丢掉被截断的 tool-call)。新增断言锁住这个方向,防止以后「统一成宽容」
    把两边都改掉。
  • 宿主自己也容忍这种畸形(dsh-agent-loop 的工具参数解析器保留原文、dsh-llm-pi-ai
    的回放解析器退成 {}),只有 dsh-llm-deepseek 的 messages 实现选择抛错——而它没有
    跨端面回放这一层要照顾。

文档

上一版 notes 与 README 把图片体积说大了,本版改准(只改说法,没有改行为):

  • 单图大小的真实边界是 attachment 的准入归一化,不是本插件:本部署在准入时就把每张图
    压到 ≤ 4 MiB / ≤ 2048×2048 像素(单边 ≤ 8192;上传侧单图 ≤ 20 MiB、单消息 ≤ 200 MiB),
    而 imageHostPath 给的就是那份归一化副本——所以本插件从不内联原始上传。
  • 插件自己那条 20 MiB 单图检查只对不做归一化的 attachment provider 生效,是兜底;
    真正管用的请求级保护是 maxRequestImageBytes,而且它按张数才会撞线:归一化后单图
    base64 约 5.4 MiB,64 MiB 预算约合十余张同请求。
  • lib/models.js 的常量注释、README.md / README.en.md 的多模态与
    maxRequestImageBytes 条目、AGENTS.md(约定 13 补边界事实、新增约定 14 记下
    「回放降级、创作严格」)同步更新。

验证

  • node test/smoke.mjs:56 项,55 pass + 1 skip(无 key 的活探针)。本版新增:
    6 种畸形历史载荷(非法 JSON、空串、数组、null、数字、字符串)都退成 {} 且
    id / name 照旧回放;以及 pumpMessages 对刚流出的非 JSON 调用仍然抛
    PROVIDER_PROTOCOL_ERROR。
  • node test/host-compat.mjs:8 项 8 pass,0 skip(真 dsh-llm + 真流语法不变量)。
  • npm run build:client:客户端半本版没有改动,lib/client.js 与 src/client/* 仍同步
    (测试里有断言守卫),所以浏览器侧无需刷新也不必重启。
  • 无联网断言:本版改动是纯转译层,不涉及上游行为,故未消耗额度做真机探针。

已知 / 未覆盖

  • 消息端面的 tool-call 参数只做「可发送」保证,不保证语义保真:input 降级成 {} 后,
    模型看到的是「调过这个工具」但参数为空。这是刻意的取舍(pi-ai 同款):保住会话
    可用性优先于历史参数保真,而真正的结果在紧随其后的 tool_result 文本里。
  • imageRequestPricing 仍未实现:图片 token 计入走 token meter 的启发式估值(不报错,
    只是不够准)。上游没有公布视觉计费口径,编一个公式会比不实现更差——token meter
    一旦拿到定价就整体替换掉自己的启发式。
  • 仍不参与 attachments.readImageRequest 的请求版降采样(那需要另一组只能靠猜的像素/字节
    目标,且有重编码的画质取舍);因此请求体侧唯一的保护仍是 maxRequestImageBytes。
  • 上游网关的请求体上限未公布,maxRequestImageBytes 的默认值是部署取值而非实测值。
  • 目录里没有 .github/,没有 CI;门禁就是这两个测试入口 + 构建脚本的语法检查。

升级

dsh plugin --profile web add github xia-sc/dsh-opencode-go
# 重启 dsh web(宿主侧改动必须重启才生效;客户端半本版没变)
  • 运行环境下限仍是 DeepSeek Harness 0.1.6-alpha.1(peer 依赖 ^0.1.6-alpha.1)。
  • 无需改任何配置:本版没有新增或改动配置字段,maxRequestImageBytes 的默认值不变。
  • 如果你的会话里有 messages 系模型(minimax / qwen)一直报
    stored tool-call arguments are not valid JSON,本版之后它就会正常继续——不需要清历史。

v0.9.0-rc.2 — the messages surface is no longer killed by one malformed history entry + image-bound corrections

Pre-release. Fixes another instance of "one durable history entry kills the session" (the same
class as #3) found after
v0.9.0-rc.1, and corrects the previous release's description of image sizes.

Fixes

  • messages surface: a malformed tool call in history no longer fails every later turn.
    Whenever a persisted assistant tool-call had arguments that were not valid JSON,
    toAnthropicMessages threw PROVIDER_PROTOCOL_ERROR, so the conversation failed on every
    later turn
    with a messages-surface model, in an error that named no history entry. The
    culprit need not be this plugin: the chat and responses translations pass arguments
    through untouched (the wire takes a string), so a malformed call persisted on one of those
    surfaces only detonates once the session is switched to a messages-surface model.
  • Degrade instead of refusing. The call already ran and already has its tool_result; the
    model only needs to see that it made the call, so refusing to re-send it bought nothing and
    cost the whole conversation. A new toolInput() maps a historical input to {} for:
    • unparseable JSON (not-json{, the empty string);
    • anything that parses to a non-object — "5", "[1]", "null", '"text"'. The
      Messages API requires input to be an object, so these failed exactly as hard, while the
      previous release sent them verbatim: an upstream 400 naming no source.
  • Creation stays strict. pumpMessages still throws for a freshly streamed call whose
    arguments are not JSON. That is an upstream protocol violation, and an unexecutable call
    must not be retained or dispatched — the harness's own max-tokens and interruption rules take
    the same line (dropping truncated tool calls rather than keeping them). A new assertion locks
    that direction against a future "let us be tolerant everywhere" edit.
  • The harness itself already tolerates this malformation (dsh-agent-loop's tool-argument
    parser keeps the raw text, dsh-llm-pi-ai's replay parser returns {}); only the
    dsh-llm-deepseek messages implementation throws, and it has no cross-surface replay to
    worry about.

Documentation

The previous release's notes and the README overstated image sizes. This release corrects the
wording (no behavior change):

  • The real bound on one image is the attachment store's admission-time normalization, not
    this plugin
    : this deployment reduces every image to <= 4 MiB and <= 2048x2048 px (<= 8192
    per side; uploads are capped at 20 MiB each and 200 MiB per message), and imageHostPath
    returns that normalized copy — so nothing here ever inlines the original upload.
  • The plugin's own 20 MiB per-image check therefore only fires for an attachment provider that
    does no normalization: it is a fallback. The guard that matters is maxRequestImageBytes,
    and it triggers by image count — a normalized image is ~5.4 MiB once base64-encoded, so
    the 64 MiB default is roughly a dozen in one request.
  • Updated together: the constant's comment in lib/models.js, the multimodal and
    maxRequestImageBytes entries in README.md / README.en.md, and AGENTS.md (rule 13
    gains the bounds; new rule 14 records the replay-degrades / creation-strict asymmetry).

Verification

  • node test/smoke.mjs: 56 tests, 55 pass + 1 skip (the key-gated live probe). New: six
    malformed history payloads (invalid JSON, empty, array, null, number, string) all degrade
    to {} while id / name still replay; and pumpMessages still rejects a freshly
    streamed non-JSON call with PROVIDER_PROTOCOL_ERROR.
  • node test/host-compat.mjs: 8 tests, 8 pass, 0 skip (real dsh-llm + real stream
    invariant).
  • npm run build:client: the browser half is unchanged in this release and
    lib/client.js stays in sync with src/client/* (asserted by the suite), so no page
    refresh and no client-side restart is needed.
  • No live probe: the change is pure translation and involves no upstream behavior, so no quota
    was spent.

Known / not covered

  • Tool-call arguments are only guaranteed sendable, not semantically faithful: degrading
    input to {} tells the model that the call happened with no arguments. That is a
    deliberate tradeoff (same as pi-ai) — keeping the conversation usable outranks history
    fidelity, and the real outcome lives in the tool_result text that follows.
  • imageRequestPricing is still unimplemented: image tokens are priced by the token meter's
    heuristic (no failure, just less exact). The operator publishes no vision pricing, and
    inventing a formula would be worse than not implementing it — the meter replaces its own
    heuristic wholesale once pricing exists.
  • Still no attachments.readImageRequest request-version downscaling (it needs another set of
    pixel/byte targets that could only be guessed, plus a re-encode fidelity tradeoff), so
    maxRequestImageBytes remains the only request-body guard.
  • The upstream gateway's request-body limit is not published, so maxRequestImageBytes's
    default is a deployment choice, not a measured value.
  • There is no .github/ and no CI: the gate is these two test entries plus the build script's
    syntax check.

Upgrade

dsh plugin --profile web add github xia-sc/dsh-opencode-go
# restart dsh web (host-side changes need a restart; the browser half is unchanged)
  • The runtime floor is still DeepSeek Harness 0.1.6-alpha.1 (peer ^0.1.6-alpha.1).
  • No configuration change: no new or altered config field, and maxRequestImageBytes keeps its
    default.
  • If a session with a messages-surface model (minimax / qwen) kept failing with
    stored tool-call arguments are not valid JSON, it simply continues after this release — no
    history cleanup needed.