dsh-qwen38-local-qol — local Qwen3.8 compaction reasoning fix for DSH

[Yunado]

dsh-qwen38-local-qol：仓库 · 设置 tab 文档 · awesome-dsh-plugin 页面

为 DeepSeek Harness（DSH）本地跑 Qwen3.8 的 QoL 插件，免 core patch、免 pi-ai patchfile：

• 模型：Qwen3.8-27B · Qwen3.8-Flash-Next
• 服务引擎：llama.cpp llama-server · NInfer · TabbyAPI (ExLlamaV3) · oMLX (Apple Silicon MLX)
• 说明：llama.cpp 线走标准 OpenAI 兼容 /v1，任何说这套 API 的服务都应可用（Strata、Unsloth Desktop、LM Studio）

1. 逐请求 thinking 预算：每个请求带所选 reasoning effort + 对应硬帽；四线字段各自分道；漏配档位回退 xhigh，thinking 开的请求永远带帽。NInfer 线的真实帽是启动参数 --default-thinking-budget，tab 有申报格对齐它。
2. 压缩精确落地，不再烧帽、不再撞墙：设置页滑杆 = 窗口的真实触发百分比，输出帽双向自动联动（线性墙护 = 帽的 1/4），任何窗口尺寸每一格精确落地；宿主内置触发按 1M 级云模型尺寸调校，本地窗口上要么过早、要么永不落地直撞上下文墙，这里全部修正。summarize 前先裁 prefill（只留最近 5 轮 reasoning、图片 strip、工具结果 2000 字符帽），摘要调用 thinking 关闭，checkpoint 拿满整个输出帽，不再 "incomplete checkpoint"（即 discussion 3465 根因修复的插件化 + 触发精确化）。
3. 设置 tab（Qwen3.8 本地）：四线各自记忆连接 / 窗口 / 预算 / 触发百分比 / 裁剪旋钮 / 各自的 API key（掩码 + 眼睛）；新装默认 = 128K 上下文 + 16K 输出帽起步值（任何本地服务开箱可扛，按服务端真实 ctx 上调）；一次填好默认值，热生效免重启。

安装：一条命令 + 重启（钉最新 tag，兼容矩阵见 Releases 与本贴兼容性报告评论）：

dsh plugin --profile web add github:Yunado/dsh-qwen38-local-qol#v0.3.1

重启 dsh web。qwen38 preset 由插件的 bundle patch 声明（启动零写盘，升级同 id 自动接管）。已知限制与全部 wire 细节见 README（中英双语）。

English

A QoL plugin for DeepSeek Harness (DSH) for running Qwen3.8 locally, no core patches, no pi-ai patchfile:

• Model support: Qwen3.8-27B · Qwen3.8-Flash-Next
• Serving engines: llama.cpp llama-server · NInfer · TabbyAPI (ExLlamaV3) · oMLX (Apple Silicon MLX)
• Note: the llama.cpp lane is the standard OpenAI-compatible /v1 API. Any server speaking it is expected to work there (Strata, Unsloth Desktop, LM Studio).

1. Per-request thinking budgets: every request carries the selected reasoning effort plus its hard cap; the wire fields split per engine; a missing tier falls back to xhigh, so a thinking-on request is always capped. The NInfer line's real cap is its --default-thinking-budget startup flag, mirrored by a declared field in the tab.
2. Compaction that lands exactly, and stops burning the cap: the settings-tab slider is the true trigger percent of the window, and the output cap syncs with it bidirectionally (linear wall guard = a quarter of the cap), so every position lands at any window size. The host's built-in trigger is sized for 1M-context cloud models, which on local 128K-262K windows either fires far too early or never reaches the configured point and slams the context wall. All fixed here. The summarizer prefill is trimmed before the one-shot call (last 5 turns of reasoning kept, images stripped, tool results capped at 2000 chars) and thinking is forced off for the summary, so the checkpoint gets the whole output cap instead of an "incomplete checkpoint" (the discussion 3465 fix as a plugin, plus an exact trigger).
3. Settings tab (Qwen3.8 Local): each of the four lines remembers its own connection / window / budgets / trigger percent / trim knobs / its own API key (masked + eye); fresh-install defaults are a starter 128K context + 16K output cap, raised to your server's real context; hot changes without restart.

Install: one command + restart (pinned to the newest tag; the compatibility matrix lives in Releases and the compatibility-report comments below):

dsh plugin --profile web add github:Yunado/dsh-qwen38-local-qol#v0.3.1

Restart dsh web. The qwen38 preset is declared by the plugin's bundle patch (nothing generated at boot; upgrades take it over under the same id). Known limits and full wire notes are in the README (bilingual).

[Yunado]

Re-verified on the latest DSH (0.1.5-alpha.1, 2026-09-09); README now covers Update / Uninstall. main @ 55c8dcb1.

[Yunado]

插件兼容性报告 / Plugin Compatibility Report

插件：dsh-qwen38-local-qol（https://github.com/yunado/dsh-qwen38-local-qol)
dsh 版本：0.1.5-alpha.1
安装方式：dsh plugin add github:Yunado/dsh-qwen38-local-qol（bundle，web profile）
结果：通过
备注：DSH Web 面全功能跑通——① 逐请求 thinking 预算：llama.cpp 线按 effort 档发送
reasoning_budget_tokens（覆盖服务端 --reasoning-budget），NInfer 线发送
reasoning_effort（白名单档），off 档关 thinking；② 压缩（compaction）后端：长会话
摘要不再把输出帽烧在 thinking 上（摘要前裁剪 + wire 层强制 thinking-off 与完整输出
帽），不再出现被截断的 "incomplete checkpoint"；③ 设置页 Qwen3.8 Local tab：
llama.cpp / NInfer 两条线各自记忆连接与窗口/预算，保存即时生效；④ 启动自动接线：
生成 qwen38-qol preset，仅在未设默认时设默认。GUI 逐轮 reasoning token 显示正常。

Plugin: dsh-qwen38-local-qol (https://github.com/yunado/dsh-qwen38-local-qol)
dsh version: 0.1.5-alpha.1
Install: dsh plugin add github:Yunado/dsh-qwen38-local-qol (bundle, web profile)
Result: pass
Notes: All features verified on the DSH Web surface — (1) per-request thinking
budgets: the llama.cpp line sends reasoning_budget_tokens per effort level
(overrides the server's --reasoning-budget), the NInfer line sends
reasoning_effort (whitelisted levels), the off level disables thinking;
(2) compaction backend: long-session summarization no longer burns the output
cap on thinking (pre-summarization trim + wire-enforced thinking-off and full
output cap), no more truncated "incomplete checkpoint"; (3) the Qwen3.8 Local
settings tab: each line (llama.cpp / NInfer) keeps its own connection and
window/budget settings, changes apply on save; (4) boot auto-wiring: generates
the qwen38-qol preset and sets the default only when none is configured.
Per-turn reasoning-token display in the GUI works.

[Yunado]

0.1.6 兼容性报告（v0.2.0）

验证范围：

• 0.1.6-alpha.1 纯上游树：全功能实跑验证（设置 tab / 四线 wire / 按线预算与窗口 / preset 自动生成）
• 0.1.6-alpha.2：安装测试通过（dsh.so install 徽章）
• 0.1.5-rc.2：安装测试通过（同徽章体系）

0.1.5 到 0.1.6 之间的 host 差异已由插件内部适配，用户不需要改配置：

• token meter 对 imageRequestPricing 的无守卫调用：adapter 自带该方法的实现（旧宿主上闲置无害）
• 图片价格读取修正（NaN 不再毒化自动压缩）
• NInfer vision 预算 400 归类为上下文溢出并自愈（compact + retry）

升级到 v0.2.0：

pnpm dsh plugin --profile web update dsh-qwen38-local-qol

然后重启 dsh web。新装 = README 一条命令 + 重启。钉住本版本 = spec 加 #v0.2.0（见 Releases）。

[Yunado]

0.1.6 Compatibility Report (v0.2.0)

Verified on:

• 0.1.6-alpha.1 stock upstream tree: full-feature live run (settings tab / four-line wire / per-line budgets and windows / preset auto-generation)
• 0.1.6-alpha.2: install test passed (dsh.so install badge)
• 0.1.5-rc.2: install test passed (same badge setup)

The host-side changes between 0.1.5 and 0.1.6 are absorbed inside the plugin; nothing for users to change:

• token meter calls imageRequestPricing without a guard: the adapter ships its own implementation of the method (inert and harmless on older hosts)
• image pricing reads corrected (no more NaN poisoning auto-compaction)
• NInfer vision-budget 400s classify as context overflow and self-heal (compact below threshold, retry the step)

Upgrade to v0.2.0:

pnpm dsh plugin --profile web update dsh-qwen38-local-qol

then restart dsh web. Fresh installs: the single command in the README + restart. Pinning this exact version: add #v0.2.0 to the spec (see Releases).

[Yunado]

0.2.0-rc.2 兼容性报告（v0.3.1）

验证范围：

• 0.2.0-rc.2 生产 web profile：全功能实跑（设置 tab / 四线 wire / 触发滑杆与输出帽联动 / 预算校验 / patch 声明 preset）
• 0.1.7-alpha.1 - 0.2.0-rc.1：应可用（同一批缝，未实测）
• 0.1.6-alpha.2 及更早：请钉 #v0.2.0

v0.3.1 重点 = 本地模型 compaction 修复：

• 宿主内置压缩的触发公式按 1M 级上下文云模型尺寸调校（比例 + 固定 65536 头寸），
  在本地 128K-262K 窗口上要么触发过早、要么永远到不了设置点，长会话一头撞上下文墙（400 / 被截断的 checkpoint）。
• v0.3.1 把触发点做精确：滑杆 = 窗口真实百分比，输出帽双向联动（线性墙护 = 帽的 1/4），每一格都精确落地；
  thinking 预算必须小于输出帽（红区点名 + 拒存，后端独立 fail-closed 校验）；
  NInfer 线新增「服务端 thinking 上限」申报格（对齐 --default-thinking-budget 启动参数）；
  新装默认降为 128K 窗口 + 16K 输出帽起步值，任何本地服务开箱可扛。

升级到 v0.3.1：

dsh plugin --profile web add github:Yunado/dsh-qwen38-local-qol#v0.3.1

然后重启 dsh web。已钉 #v0.3.0 的 = 重跑上面同一行换钉 + 重启；新装 = README 一条命令 + 重启。已声明的 patch 几何不受影响（新默认只补未设置的字段）。

[Yunado]

0.2.0-rc.2 Compatibility Report (v0.3.1)

Verified on:

• 0.2.0-rc.2 production web profile: full-feature live run (settings tab / four-line wire / trigger slider with output-cap sync / budget validation / patch-declared preset)
• 0.1.7-alpha.1 - 0.2.0-rc.1: should work (same seam batch, not exercised)
• 0.1.6-alpha.2 and older: pin #v0.2.0

The headline of v0.3.1 is a local-model compaction fix:

• The host's built-in compaction trigger is sized for 1M-context cloud models (ratio + a fixed 65536 headroom). On local 128K-262K windows it either fires far too early or never reaches the point you configured, and long sessions slam the context wall mid-request (400 / truncated checkpoint).
• v0.3.1 makes the trigger land exactly: the settings-tab slider is the true percent of the window, and the output cap syncs bidirectionally with it (linear wall guard = a quarter of the cap), so every position lands at any window size; thinking budgets must stay below the output cap (red checker + refused save, with an independent fail-closed backend validation); the NInfer line gains a declared server-thinking-ceiling field (mirrors --default-thinking-budget); fresh-install defaults drop to a starter 128K context + 16K output cap that any local server can host.

Upgrade to v0.3.1:

dsh plugin --profile web add github:Yunado/dsh-qwen38-local-qol#v0.3.1

then restart dsh web. Already pinned to #v0.3.0? Re-run the same line to re-pin, then restart. Fresh installs: the single README command + restart. Patch-declared geometries are untouched (new defaults only fill unset fields).

[Yunado]

0.2.1-alpha.2 兼容性报告（v0.3.3）
验证范围：

0.2.1-alpha.2 生产 web profile：实跑（设置 tab / 四线 wire / 会话与工具链，作者日常生产从升级起就跑在该宿主上）
0.2.0-rc.2：#v0.3.2 或 #v0.3.3 皆可（v0.3.3 在其上只做减法，不要求宿主新功能）
0.1.7-alpha.1 - 0.2.0-rc.1：应可用（同一批缝，未实测）
0.1.6-alpha.2 及更早：请钉 #v0.2.0

一条 alpha 宿主升级注意：宿主 0.2.1-alpha.1 移除了内置 {{cwd}} 提示词变量（工作目录改由 working-directory 服务自动注入），
v0.3.3 之前的插件 tag 在 preset persona 里引用了它，在新宿主上每轮请求都会被 fail-closed 拒绝（unknown prompt variable）。
v0.3.3 已删除该引用，用户侧配置零改动。

v0.3.3 同时修了一个设置 tab 的 bug（issue #4）：工具结果字数帽与"保留最近 N 轮 reasoning"现在可以填 0 保存（各自的文档化关闭模式），
profile patch 里手写的 0 也不会再被下次保存覆盖。

升级（同宿主，无 schema 迁移）：
dsh plugin --profile web add github:Yunado/dsh-qwen38-local-qol#v0.3.3
之后重启 dsh web。
发布页：https://github.com/Yunado/dsh-qwen38-local-qol/releases/tag/v0.3.3

[Yunado]

0.2.1-alpha.2 Compatibility Report (v0.3.3)
Verified on:

0.2.1-alpha.2 production web profile: live run (settings tab / four-line wire / sessions and tools; the author's own production runs on this host since the upgrade)
0.2.0-rc.2: #v0.3.2 or #v0.3.3 (v0.3.3 only removes content on top, no new host requirements)
0.1.7-alpha.1 - 0.2.0-rc.1: should work (same seam batch, not exercised)
0.1.6-alpha.2 and older: pin #v0.2.0

One alpha-host upgrade note: host 0.2.1-alpha.1 removed the built-in {{cwd}} prompt variable (the working-directory service now supplies the directory itself).
Plugin tags before v0.3.3 referenced it in the preset persona, so on the new host every request failed fail-closed ("unknown prompt variable"). v0.3.3 drops the
reference; nothing to change in user config after updating.

v0.3.3 also ships a settings-tab bugfix (issue #4): the tool-result character cap and the keep-reasoning knob now accept 0 (their documented off modes), and a
hand-written 0 in the profile patch is no longer overwritten by the next save.

Upgrade (same host, no schema migration):
dsh plugin --profile web add github:Yunado/dsh-qwen38-local-qol#v0.3.3
Then restart dsh web.
Release: https://github.com/Yunado/dsh-qwen38-local-qol/releases/tag/v0.3.3

[Yunado]

0.2.1-rc.2 兼容性报告（v0.3.3，生产实测 10-18）

生产宿主从 0.2.1-alpha.2 升到 dsh-v0.2.1-rc.2（区间 990 commit），dsh-qwen38-local-qol 继续钉 #v0.3.3，零动作跟上了新版本：

验证范围

• 活动线接线：会话继续跑在插件设置 tab 配置的本地线上（rc.2 未动 provider/schema 缝）
• preset：qwen38 正常挂载；提示词变量白名单（provider/model）在 rc.2 上依然全量可用
• 版本门禁：本插件 peer 声明 ^0.2.0-rc.2 且未声明 engines.dsh，rc.2 新判法（engines.dsh 优先、无则回退 peer）下直接放行，无需 allow-version 豁免
• 压缩后端：与 alpha 系列同一批 summarize() 缝，rc.2 release notes 无缝变更信号，等自然首触发验证

插件侧不需要为 0.2.1-rc.x 发新版本，#v0.3.3 覆盖 0.2.1-alpha.1 及更新宿主，README 兼容表已把 0.2.1-rc.2 标注为生产实测。

---

0.2.1-rc.2 compatibility report (v0.3.3, production-verified Oct 18)

The production host moved from 0.2.1-alpha.2 to dsh-v0.2.1-rc.2 (990 commits) with dsh-qwen38-local-qol still pinned at #v0.3.3, and nothing needed doing:

Verified scope

• Active-line wiring: sessions keep running on the local line configured in the plugin settings tab (rc.2 touched no provider or schema seams)
• preset: qwen38 mounts cleanly; the prompt-variable whitelist (provider/model) stays fully in range on rc.2
• Version gate: this plugin declares peer ^0.2.0-rc.2 and no engines.dsh, so the new engines.dsh-first check falls back to the peer rule and passes, no allow-version exemption needed
• Compaction backend: same summarize() seam batch as the alpha series; rc.2 release notes show no seam change, first natural trigger pending

No plugin release is needed for any 0.2.1-rc.x host: #v0.3.3 covers 0.2.1-alpha.1 and newer. The README compatibility table now records 0.2.1-rc.2 as production-tested.