History read path silently treats a valid aborted-turn session log as "no more history" — root cause + fix path

[yamingmou]

Self-review update — correction of our own submitted bug report / 自我修正：对本人提交的 bug report 的修正 (2026-08-21)

⚠️ This is a self-review of the bug report we submitted. The original report described the failure as "history paging freezes (blank pages)" — a UI-level symptom. After re-reading the engine source, we are upgrading the report to the layer where the defect actually lives: the session log and its history read contract. The UI symptom is real, but it is only the visible edge of a read-contract failure between a valid, first-class log state and the path that is supposed to serve it.

TL;DR — EN: We originally reported "history paging freezes (blank pages) when a session ends with aborted turns." Upgrade: this is not a rendering bug. A session log whose tail consists of aborted turns (turn/end with reason.kind: "aborted" + trailing session/end-seed) is structurally valid and fully contiguous — every event is present on disk. Yet the history read path (session.history) cannot serve the earlier region: reads return empty pages even though the events exist, and the failure is silently degraded to "no more history" (dsh-client-runtime/lib/client.js:7398-7419). Aborted tails are the routine ending of scheduled/unattended runs (timeout, worker restart, kill), so any deployment that schedules agents produces this state routinely; the session's durable progress record becomes unreachable, which operators read as lost work.

一句话（中文）： 这是对本人提交的 bug report 的自我修正（升级）。原报告把故障描述为"历史分页冻结（空白页）"——那只是 UI 症状。重审源码后升级到缺陷真正所在的层：会话日志及其历史读取契约。尾部 aborted 区间（turn/end reason=aborted + 尾部 session/end-seed）在日志中结构合法、seq 完全连续、每个事件都在磁盘上；但历史读取路径（session.history）无法为更早区域提供服务——读取返回空页，且失败被静默降级为"没有更多历史"（dsh-client-runtime/lib/client.js:7398-7419）。aborted 尾部是调度/无人值守运行的常规结尾（超时、worker 重启、被杀），任何把 agent 调度化的部署都会例行产生；会话的持久化进度记录变得不可达，运维看到的就是"活儿没了"。

What we originally claimed / 初版表述 (EN): history paging freezes (blank pages) when a session ends with aborted turns; data is not lost on disk, only the UI cannot reach it.
What we now state / 现修正 (EN): upgraded to the log/read-contract layer — a structurally valid, first-class log state (trailing aborted turns) cannot be served by the history read path; the failure is silently degraded to "no more history"; the session's durable progress record becomes unreachable with no error signal; the defect is a read-contract gap, not a rendering issue.

初版表述（中文）： 会话以 aborted 轮次结束时，"加载更早"历史分页冻结（空白页）；磁盘数据未丢，只是 UI 够不着。
现修正（中文）： 升级到日志/读取契约层——一个结构合法、一等公民的日志状态（尾部 aborted 轮次）无法被历史读取路径服务；失败被静默降级为"没有更多历史"；会话的持久化进度记录变得不可达、无任何报错；这是读取契约缺口，不是渲染问题。

---

EN / English

Observed behavior

A session log whose tail consists of one or more aborted turns (turn/end with reason.kind: "aborted", followed by session/end-seed) is structurally valid: seq is contiguous from 0 and every event is present on disk. Yet the history read path (session.history) cannot serve the earlier region:

• reads for the earlier region return empty pages even though the events exist in the log,
• and the failure is silently degraded to "no more history" — the read path treats "cannot serve" as "history ends here" (dsh-client-runtime/lib/client.js:7398-7419, where empty pages continue silently, page discontinuities are logged and treated as end-of-history, and any thrown error is swallowed with a console-only log).

The observable symptom (blank paging in the conversation view) is real; the defect is a read-contract failure: a valid log state that its own consumer cannot serve, with no way to tell the user the difference between "no history" and "history that failed to load".

Log / read-path mechanism (with source references)

• The host serves history from the full contiguous event log — historyCutOf (dsh-host-apiproxy/lib/index.js:2084-2098) returns the complete events array; nothing is missing on the serving side.
• Paging is message-count based with a group cut — paginate (dsh-host-apiproxy/lib/index.js:1030-1049) — so a page always carries all events in its range, including the aborted-turn region.
• The client loadOlder() (dsh-client-runtime/lib/client.js:7388-7424) then swallows any inability to fold/render that region: empty page → continue silently (:7398-7404); discontinuity → console.error + hasMore = false + empty prepend (:7406-7411); thrown error → console.error only (:7418-7419).

The events are all present; the read path simply has no first-class handling for a trailing aborted region, and its silent-degradation design hides that.

Impact scope (upgraded — log / session / business)

• Session layer: the complete event log exists, yet the session history cannot be fully restored — "the data is there, but the session is not reproducible." For a long session this is the entire accumulated record of the work.
• Business layer: the progress/audit record — especially the transcript of scheduled/unattended runs — is silently hidden. Operators cannot distinguish "no history" from "history unreadable", so a recoverable session reads as lost work, and recovery requires manual log truncation (an operator-only operation).
• Log layer: the trigger state (trailing aborted turns) is a first-class, legitimate durable state, and it is the routine ending of scheduled/background runs — any deployment that schedules or backgrounds agents will produce it. A normal ending permanently hides a session's progress.

Expected behavior

The read contract must satisfy: a valid log state must be fully reproducible — or, when it cannot be, the read path must say so. Aborted turns are a normal first-class event and must be served (rendered as resumable/recoverable, or paged past); and "history failed to load" must be distinguishable from "no more history" — a failure must surface as an error, never as a silent blank.

Repro

1. Open a session, start a turn, interrupt it (aborted).
2. Repeat 2–3 times so the session ends with several consecutive aborted turns + trailing session/end-seed.
3. Reopen the session, read earlier history → empty pages, no error.
4. Verification of causality: truncate the log at the last completed turn/end → earlier history is served correctly again.

Environment

• Engine: @deepseek-ai/dsh 0.1.0-rc.6 / 0.1.0-rc.7
• Session format: multi-frame zstd JSONL (session.jsonl.zstd), strict contiguous seq from 0

Evidence (sanitized event sequence + source)

turn 40: turn/end reason=completed   <- last good turn
turn 41: turn/end reason=aborted    <- reads for earlier history fail from here
turn 42: turn/end reason=aborted
turn 43: turn/end reason=aborted
session/end-seed

• dsh-client-runtime/lib/client.js:7388-7424 — loadOlder(): silent degradation on empty page / discontinuity / error (console.error only).
• dsh-host-apiproxy/lib/index.js:2084-2098 — historyCutOf: the full contiguous event log is served, so the data is present.
• dsh-host-apiproxy/lib/index.js:1030-1049 — paginate: message-count paging with group cut.

---

Authorship note: reproduced and documented by me in real-world usage; root-cause tracing and drafting assisted by AI; verification and publication by me. I have no engineering background — if any technical claim reads wrong, please call it out; I will re-verify against the toolchain and correct.

中文版 / ZH

现象

尾部由一个或多个 aborted 轮次（turn/end 的 reason.kind: "aborted"，随后 session/end-seed）组成的会话日志结构合法：seq 从 0 连续、每个事件都在磁盘上。但历史读取路径（session.history）无法为更早区域提供服务：

• 对更早区域的读取返回空页——尽管事件都在日志里；
• 且失败被静默降级为"没有更多历史"——读取路径把"无法服务"当作"历史到此为止"（dsh-client-runtime/lib/client.js:7398-7419：空页静默继续、分页不连续仅 console 后当作历史尽头、抛错被吞掉仅 console）。

界面上的空白页（可观察症状）真实存在；但缺陷是读取契约失败：一个合法日志状态，其自身消费方却无法服务，且用户无法区分"没有历史"与"历史加载失败"。

日志 / 读取路径机制（附源码引用）

• host 从完整连续事件日志提供历史——historyCutOf（dsh-host-apiproxy/lib/index.js:2084-2098）返回完整 events 数组，服务端没有任何缺失。
• 分页按消息数切页并做 group cut——paginate（dsh-host-apiproxy/lib/index.js:1030-1049）——每页携带其区间内的全部事件，包括 aborted 区间。
• 客户端 loadOlder()（dsh-client-runtime/lib/client.js:7388-7424）随后把任何"无法 fold/渲染该区间"的情况吞掉：空页→静默继续（:7398-7404）；不连续→console.error + hasMore=false + 空 prepend（:7406-7411）；抛错→仅 console.error（:7418-7419）。

事件全部都在；读取路径只是对尾部 aborted 区间没有一等公民的处理，而它的静默降级设计掩盖了这一点。

影响范围（升级 —— 日志 / 会话 / 业务）

• 会话层： 完整事件日志存在，但会话历史无法完整还原——"数据在，但会话不可复现"。对长会话而言，这是工作的全部累积记录。
• 业务层： 进度/审计记录——尤其是调度/无人值守运行的转录——被静默遮蔽。运维无法区分"没有历史"与"历史不可读"，于是可恢复的会话被当作工作丢失，恢复只能靠手工截断日志（运维专属操作）。
• 日志层： 触发状态（尾部 aborted 轮次）是一等公民的合法持久化状态，且是调度/后台运行的常规结尾——任何把 agent 调度化/后台化的部署都会产生它。一个普通结尾永久遮蔽了会话的进度。

期望行为

读取契约必须满足：合法日志状态必须可完整还原——若无法还原，读取路径必须明说。aborted 轮次是正常的一等事件，必须被服务（渲染为可恢复/可继续，或跳过）；"历史加载失败"必须与"没有更多历史"可区分——失败必须呈现为错误，绝不允许静默空白。

复现步骤

1. 打开会话，开始一轮对话后中断（aborted）。
2. 重复 2–3 次，让会话以多个连续 aborted 轮次 + 尾部 session/end-seed 结尾。
3. 重新打开会话，读取更早历史 → 空页，无报错。
4. 因果验证：把日志截断到最后一个 completed 的 turn/end → 更早历史恢复正常。

环境

• 引擎：@deepseek-ai/dsh 0.1.0-rc.6 / 0.1.0-rc.7
• 会话格式：多帧 zstd JSONL（session.jsonl.zstd），seq 从 0 严格连续

证据（脱敏事件序列 + 源码）

turn 40: turn/end reason=completed   <- 最后一个正常轮次
turn 41: turn/end reason=aborted    <- 更早历史的读取从这里开始失败
turn 42: turn/end reason=aborted
turn 43: turn/end reason=aborted
session/end-seed

• dsh-client-runtime/lib/client.js:7388-7424 —— loadOlder()：空页 / 不连续 / 抛错时均静默降级（仅 console.error）。
• dsh-host-apiproxy/lib/index.js:2084-2098 —— historyCutOf：提供完整连续事件日志，数据确实存在。
• dsh-host-apiproxy/lib/index.js:1030-1049 —— paginate：按消息数切页、带 group cut。

---

Original report (archived for transparency) / 初版报告（归档保留）

以下为初版表述（被上方"自我修正"升级）。保留以透明展示我们修正了什么。

EN

Observed behavior (original). When a session ends with one or more aborted turns, the conversation UI renders the recent messages normally, but "Load earlier" paging returns blank pages indefinitely — older history becomes inaccessible even though the data exists on disk.

Impact scope (original). Any session whose last turns were interrupted (very common in daily use) may lose UI access to all earlier history. Data is not lost on disk — only the UI cannot reach it, which makes it look like a data-loss bug to users. Workaround found: truncating the log so it ends on the last completed turn/end restores paging completely.

中文

现象（初版）。 会话以一个或多个 aborted 轮次结束时，对话 UI 最近的几条消息能正常显示，但"加载更早"分页永久返回空白页——更早的历史明明存在磁盘上，UI 却永远访问不到。

影响范围（初版）。 任何以中断轮次结尾的会话（日常使用非常常见）都可能失去 UI 对全部早期历史的访问。磁盘数据未丢失——只是 UI 够不着，给用户的感觉就像数据丢了。已找到临时方案：把日志截断到最后一个 completed 的 turn/end，分页即可完全恢复。

---

署名 / About this report

EN: Reported by the OfferKuai (Offer快) Team — an AI startup building full-lifecycle job-application services, guided by the belief that "users need results, not repeated conversations." Founder: Zhaofeng (Yaming). We use DeepSeek Harness as part of our daily development workflow; this report is our way of contributing back to the ecosystem. Website: https://www.offerkuai.com/ | Contact: <contact@offerkuai.com>

ZH: 本报告由 OfferKuai（Offer快）团队提交 —— 一家专注 AI 全流程求职托管的创业团队，核心理念是「用户要的是结果，不是重复对话」。创始人：Zhaofeng（Yaming）。我们将 DeepSeek Harness 用于日常开发工作流，这份报告是我们对社区的回馈。官网：https://www.offerkuai.com/ | 联系：<contact@offerkuai.com>

`[Bug] 会话以 aborted 轮次结束时，"加载更早"历史分页冻结（空白页）`

**EN / 英文版**

### Observed behavior

When a session ends with one or more `aborted` turns (user interrupted the agent, or the turn was cut off), the conversation UI:

- renders the recent messages normally,
- but "Load earlier" paging returns **blank pages indefinitely** — older history becomes inaccessible even though the data exists on disk.

### Expected behavior

Aborted turns are a normal, first-class event (`turn/end` with `reason.kind: "aborted"`). They must not break history paging. The UI should either render aborted turns as resumable/recoverable markers, or page past them — never freeze.

### Impact scope

- Any session whose last turns were interrupted (very common in daily use) may lose UI access to all earlier history.
- Data is **not lost on disk** — only the UI cannot reach it, which makes it look like a data-loss bug to users.
- Affects both the web UI and the desktop shell (same engine/UI packages).
- Workaround found: truncating the log so it ends on the last `completed` `turn/end` restores paging completely (data layer accepts it; seq stays contiguous).

### Repro

1. Open a session, start a turn, interrupt it (`aborted`).
2. Repeat 2–3 times so the session ends with several consecutive aborted turns + trailing `session/end-seed`.
3. Reopen the session, click "Load earlier" repeatedly → blank pages.

### Environment

- Engine: `@deepseek-ai/dsh` `0.1.0-rc.6` / `0.1.0-rc.7`
- Session format: multi-frame zstd JSONL (`session.jsonl.zstd`), strict contiguous seq from 0
- Repro shell: `anywhere-labs/deepseek-harness-desktop` (Electron, bundles official engine unchanged)

### Evidence (sanitized event sequence)

```
turn 40: turn/end reason=completed   <- last good turn
turn 41: turn/end reason=aborted    <- paging freezes from here
turn 42: turn/end reason=aborted
turn 43: turn/end reason=aborted
session/end-seed
```

After truncating at turn 40's `turn/end`, "Load earlier" pages correctly through all 40 turns.

---

**ZH / 中文版**

### 现象

会话以一个或多个 `aborted` 轮次结束时（用户中断了 agent，或轮次被截断），对话 UI：

- 最近的几条消息能正常显示；
- 但点击"加载更早"**永久返回空白页** —— 更早的历史明明存在磁盘上，UI 却永远访问不到。

### 期望行为

`aborted` 轮次（`turn/end` 的 `reason.kind: "aborted"`）是**正常的一等事件**，不应破坏历史分页。UI 要么把 aborted 轮次渲染为"可恢复/可继续"的标记，要么直接跳过它们 —— 绝不应当冻结。

### 影响范围

- 任何以中断轮次结尾的会话（日常使用非常常见）都可能失去 UI 对全部早期历史的访问。
- **磁盘数据未丢失** —— 只是 UI 够不着，给用户的感觉就像数据丢了。
- 同时影响 Web UI 与桌面壳（同一套引擎/UI 包）。
- 已找到临时方案：把日志截断到最后一个 `completed` 的 `turn/end`，分页即可完全恢复（数据层接受，seq 保持连续）。

### 复现步骤

1. 打开会话，开始一轮对话后中断（`aborted`）。
2. 重复 2–3 次，让会话以多个连续 aborted 轮次 + 尾部 `session/end-seed` 结尾。
3. 重新打开会话，反复点"加载更早" → 空白页。

### 环境

- 引擎：`@deepseek-ai/dsh` `0.1.0-rc.6` / `0.1.0-rc.7`
- 会话格式：多帧 zstd JSONL（`session.jsonl.zstd`），seq 从 0 严格连续
- 复现壳：`anywhere-labs/deepseek-harness-desktop`（Electron，官方引擎包未改动）

### 证据（脱敏事件序列）

```
turn 40: turn/end reason=completed   <- 最后一个正常轮次
turn 41: turn/end reason=aborted    <- 分页从这里开始冻结
turn 42: turn/end reason=aborted
turn 43: turn/end reason=aborted
session/end-seed
```

截断到 turn 40 的 `turn/end` 后，"加载更早"可正确翻完 40 轮。

---

### 署名 / About this report

**EN:** Reported by the **OfferKuai (Offer快) Team** — an AI startup building full-lifecycle job-application services, guided by the belief that *"users need results, not repeated conversations."* Founder: **Zhaofeng (Yaming)**. We use DeepSeek Harness as part of our daily development workflow; this report is our way of contributing back to the ecosystem. Website: <https://www.offerkuai.com/> | Contact: <[contact@offerkuai.com](mailto:contact@offerkuai.com)>

**ZH:** 本报告由 **OfferKuai（Offer快）团队**提交 —— 一家专注 AI 全流程求职托管的创业团队，核心理念是「用户要的是结果，不是重复对话」。创始人：**Zhaofeng（Yaming）**。我们将 DeepSeek Harness 用于日常开发工作流，这份报告是我们对社区的回馈。官网：<https://www.offerkuai.com/> | 联系：<[contact@offerkuai.com](mailto:contact@offerkuai.com)>

---

*

*协作说明：本文由我在真实使用中复现并记录；根因与成文由 AI 协助完成，我负责校验与发布。我非工程背景，技术表述如有错误，欢迎指出——我会回去用工具链重新验证后更正。*

[yamingmou]

We posted a deeper root-cause analysis as a separate discussion, reframing this from "a single lost turn" into a read-contract gap: the projection treats a valid (aborted-turn) session log as exhausted, and proposes fix paths (distinguishable read failures, first-class aborted-turn handling, visible gap markers). Cross-linking for reference:

#4795 — Why a valid session log silently becomes "no more history": a read-contract gap, not one bug

---

中文版

我们把更深入的 root-cause 分析单独发了一帖，把这个 issue 从"丢了一轮"重定义为读契约缺口：投影把一份合法（含中断轮）的会话日志当作已耗尽，并给出修复路径（区分可读失败、把中断轮作为一等公民处理、可见的缺口标记）。在此交叉引用：

#4795 —— 为什么一份合法的会话日志会静默变成"没有更多历史"：这是读契约缺口，不是一个 bug