[Bug] /compact fails forever after a plugin writes a marker with null turn/step

[yamingmou]

EN / English

Found earlier in our daily usage; re-verified against the current npm line (through 0.1.3-alpha.2) before posting — still present. We publish one topic per report so each can be tracked independently; more findings will follow in the same form.

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.

Summary

/compact (manual compaction) fails hard on sessions that contain a plugin-authored assistant/message marker with turn: null, step: null (a surface replace event, the same primitive built-in compaction uses). Error: token meter: assistant/message at seq N has no matching step/start event. The event is fully legal for the persistence read path (foldSurface accepts it), so the session loads and works — but every pressure measurement and manual compaction after that point throws. We hit this on 7 sessions across our workspace; /compact stayed broken until the marker was removed.

This is another instance of "one log, two consumers, two verdicts" (see #3632): foldSurface (persistence) and the token meter (measurement) disagree about the same event.

Trigger Condition

A session log contains a surfaceOp: replace assistant/message event whose data.turn and data.step are null, written outside of any open step (e.g. between turn/end and the next turn/start). Then:

• dsh session compact / the /compact command → throws token meter: assistant/message at seq N has no matching step/start event;
• any manual pressure check on that session throws the same.

Minimal Repro

1. In any session, append (via a plugin, e.g. an edit/recall marker) an event:

   {"type":"assistant/message","seq":N,"time":T,"data":{"turn":null,"step":null,"message":{...}},
    "sourceEventSeqs":[...],"surfaceOp":{"op":"replace","start":A,"end":B}}

2. Open the session in the GUI, run /compact.
3. Observed: command/done carries kind:"error", text:"token meter: assistant/message at seq N has no matching step/start event"; no checkpoint is written.
4. Remove the marker (or rewrite it with a valid turn/step inside a step) → /compact works again.

Analysis (root cause)

dsh-token-meter/lib/index.js:590 (v0.1.1-rc.2; → :727 on 0.1.2-rc.1, → :753 on 0.1.3-alpha.2, verbatim) requires every assistant/message to match the currently open step:

// @deepseek-ai/dsh-token-meter lib/index.js:590 (v0.1.1-rc.2) → :727 (0.1.2-rc.1) → :753 (0.1.3-alpha.2), verbatim
if (stepStart === void 0 || stepStart.turn !== event.data.turn || stepStart.step !== event.data.step) throw new Error(`token meter: assistant/message at seq ${event.seq} has no matching step/start event`);

The meter folds the whole durable log (_sync consumes session.events from seq 0; index.js:556), so a single marker written between turns (no open step → stepStart === void 0) makes every subsequent measure() call throw. The persistence layer, by contrast, validates the same event as legal: foldSurface only requires the replace's start/end to exist in the surface and sourceEventSeqs to cover the shadowed nodes (dsh-session/lib/index.js assertProvenance / planSurfaceEvent). One consumer accepts, the other rejects.

Compaction depends on the meter: dsh-compaction-basic calls ctx.tokenMeter for pressure and for re-pricing after the replacement. When the meter throws, the compaction transaction fails before any checkpoint is written (the log records compaction/start … command/done:error, surface unchanged). Manual /compact reports summary/commit-class failures; automatic step-boundary pressure checks merely log a warning and continue (as the compaction README states), which is why sessions keep working for normal use while /compact stays permanently broken.

Attribution

• Primary: measurement consumer (token meter) rejects an event the persistence consumer (foldSurface) accepts — a contract gap between consumers, surfaced by the plugin extension point (markers). Not a crash in the read path; a hard failure of the compaction/pressure path.
• Secondary: the edit plugin chose boundary markers with turn/step = null for boundary markers, which is M1-legal for the persistence layer ("null turn/step assistant/message must be replace, never append") but invisible to the meter's step-pairing rule. We own this design debt; the ecosystem-level fix belongs upstream (see proposal).

Proposed fix

1. Upstream (recommended): in dsh-token-meter._foldEvent, skip the step-pairing check for assistant/message events whose data.turn/step are null and whose surfaceOp is a replace (the marker shape foldSurface already legitimizes). Fold their surface tokens for pricing; do not require an open step. One-line guard; keeps the meter strict for real assistant messages.
2. Plugin-side (complementary): write markers inside an open step where possible (carry the step's turn/step), and/or provide a pre-compact "marker cleanup" pass so users can compress after editing.
3. Tooling-side (ours): offline dsh-log-contract check now includes a token-meter fold (rule T1) so sessions that would fail /compact are flagged before the user hits it.

Related

• [Bug] [Update] Orphan agent/inbox/spliced splits log validity from session readability — impact upgrade: a routine post-incident repair can produce a durably intact but unreadable session (self-review of our own bug report) #3632 (one log, two consumers, two verdicts) — same family: consumers disagreeing about a legal event.
• [Analysis] Why a valid session log silently becomes "no more history" — a read-contract gap, not one bug (corroborated by #4767/#4704/#4750) #4795 (read-contract gap) — adjacent: valid-but-unreadable states on the read path; this one is valid-but-uncompactable on the measurement path.
• [Architecture] [Update 2026-09-17] All persistence formats hard-refuse non-current versions with zero migration path — every on-disk session log becomes unreadable on the first bump #4910 (format v0, no migration) — the "no migration for plugin-authored event shapes" theme.

Environment

• DeepSeek Harness 0.1.1-rc.2 (Desktop app bundle @deepseek-ai/dsh@0.1.1-rc.2)
• dsh-compaction-basic (default), dsh-command-compact mounted
• Edit plugin under test (recall/edit markers)
• macOS 26.3 (arm64); 7 sessions reproduced

---

---

Reported by the OfferKuai Team — Founder: Zhaofeng (Yaming). Website: https://www.offerkuai.com/ | Contact: contact@offerkuai.com

中文版 / ZH

此为我们在日常使用中较早发现的问题，发布前已对照 npm 最新线（至 0.1.3-alpha.2）复核，确认仍存在。我们按"一帖一主题"发布以便独立跟踪，后续发现也将以同样形式陆续发出。

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。

摘要

凡会话日志中包含插件写入的、turn/step 为 null 的 assistant/message marker（surfaceOp: replace，与内置压缩同原语），该会话的 /compact 与手动压力测量必然失败，报 token meter: assistant/message at seq N has no matching step/start event。该事件对持久化读路径完全合法（foldSurface 接受），会话能正常加载使用——但从该 marker 起，每一次压力测量/手动压缩都抛错。我们在工作区 7 个会话上复现；移除 marker 前 /compact 一直不可用。

这又是「one log, two consumers, two verdicts」（见 #3632）的一个实例：foldSurface（持久化）与 token meter（计量）对同一合法事件结论相反。

触发条件

会话日志包含一条 surfaceOp: replace 的 assistant/message，其 data.turn 与 data.step 为 null，且写在任何打开的 step 之外（如 turn/end 与下一个 turn/start 之间）。此后：/compact 与手动压力检查均抛同一错误。

最小复现

1. 通过插件（如编辑/撤回 marker）在任意会话追加：

   {"type":"assistant/message","seq":N,"time":T,"data":{"turn":null,"step":null,"message":{...}},
    "sourceEventSeqs":[...],"surfaceOp":{"op":"replace","start":A,"end":B}}

2. GUI 打开该会话，输入 /compact。
3. 结果：command/done 携带 kind:"error", text:"token meter: assistant/message at seq N has no matching step/start event"，不产生检查点。
4. 移除该 marker（或改写为 step 内的合法 turn/step）→ /compact 恢复可用。

具体分析（根因）

dsh-token-meter/lib/index.js:590（v0.1.1-rc.2；0.1.2-rc.1 移至 :727、0.1.3-alpha.2 移至 :753——同一 throw 逐字保留）要求每条 assistant/message 匹配当前打开的 step：

// @deepseek-ai/dsh-token-meter lib/index.js:590（v0.1.1-rc.2）→ :727（0.1.2-rc.1）→ :753（0.1.3-alpha.2），逐字
if (stepStart === void 0 || stepStart.turn !== event.data.turn || stepStart.step !== event.data.step) throw new Error(`token meter: assistant/message at seq ${event.seq} has no matching step/start event`);

meter 折叠整份持久化日志（_sync 从 seq 0 消费 session.events，index.js:556），因此任何一条写在轮次之间的 marker（无打开 step → stepStart === void 0）都会让此后的每次 measure() 抛错。而持久化层判定同一事件合法：foldSurface 只要求 replace 的 start/end 存在于 surface、sourceEventSeqs 覆盖被遮蔽节点（dsh-session/lib/index.js 的 assertProvenance/planSurfaceEvent）。一个消费者接受，另一个拒绝。

压缩依赖 meter：dsh-compaction-basic 用 ctx.tokenMeter 做压力判断与替换后重新计价。meter 抛错 → 压缩事务在任何检查点写入前失败（日志只留 compaction/start…command/done:error，surface 不变）。手动 /compact 报 summary/commit 类失败；自动 step 边界压力检查记 warning 后继续（dsh-compaction-basic/README.zh.md），所以会话日常使用正常，唯独 /compact 永久失效。

归因

• 主因：计量消费者（token meter）拒绝持久化消费者（foldSurface）认可的合法事件——消费者间契约缺口，由插件扩展点（marker）暴露。不是读路径崩溃，而是压缩/压力路径的硬失败。
• 次因：marker 的编写方（第三方编辑插件）选择边界 marker 用 turn/step=null——对持久化层 M1 合法（"null turn/step 的 assistant/message 只能 replace 不能 append"），但 meter 的 step 配对规则看不见它。这是我们的设计债；生态级修复应在上游（见建议）。

建议

1. 上游（推荐）：dsh-token-meter._foldEvent 对 data.turn/step 为 null 且 surfaceOp 为 replace 的事件（foldSurface 已认可的 marker 形态）跳过 step 配对检查，仅按其 surface tokens 计价。一行守卫；对真实 assistant 消息仍保持严格。
2. 插件侧（互补）：marker 尽量写入打开的 step 内（携带该 step 的 turn/step）；并提供压缩前的 "marker 清理" 通道，让编辑后仍可压缩。
3. 工具侧（我方）：离线 dsh-log-contract check 已加入 token-meter 折叠（规则 T1），在用户触发 /compact 前就标记出会失败的会话。

相关

• [Bug] [Update] Orphan agent/inbox/spliced splits log validity from session readability — impact upgrade: a routine post-incident repair can produce a durably intact but unreadable session (self-review of our own bug report) #3632（one log, two consumers, two verdicts）——同一家族：消费者对合法事件结论相反。
• [Analysis] Why a valid session log silently becomes "no more history" — a read-contract gap, not one bug (corroborated by #4767/#4704/#4750) #4795（read-contract gap）——相邻：读路径的 valid-but-unreadable；本篇是计量路径的 valid-but-uncompactable。
• [Architecture] [Update 2026-09-17] All persistence formats hard-refuse non-current versions with zero migration path — every on-disk session log becomes unreadable on the first bump #4910（格式 v0 无迁移）——"插件事件形态无迁移"主题的延伸。

---

本报告由 OfferKuai（Offer快）团队提交 —— 创始人：Zhaofeng（Yaming）。官网：https://www.offerkuai.com/ | 联系：contact@offerkuai.com

[yamingmou]

Evidence chain — why a turn/step-null replace marker permanently breaks the token meter (and thus /compact)

EN / English

Six steps, each with source file:line @ version, verbatim snippet and note.

Step 1 — the meter folds the entire durable log, no way to skip a bad event

dsh-token-meter/lib/index.js:556 (v0.1.1-rc.2) _sync(session) — now :705 on 0.1.3-alpha.2:

while (state.consumedEvents < session.events.length) {
    const event = session.events[state.consumedEvents];
    this._foldEvent(session, state, event);
    state.consumedEvents += 1;
}

session.events is the full restored log. A single problematic event anywhere in history poisons every later measurement. There is no partial-fold escape.

Step 2 — every assistant/message must match the currently open step

dsh-token-meter/lib/index.js:590 (verbatim on v0.1.1-rc.2; :753 on 0.1.3-alpha.2):

else if (event.type === "assistant/message") {
    const stepStart = state.stepStart;
    if (stepStart === void 0 || stepStart.turn !== event.data.turn || stepStart.step !== event.data.step) throw new Error(`token meter: assistant/message at seq ${event.seq} has no matching step/start event`);

state.stepStart is only set by a step/start (index.js:574 on v0.1.1-rc.2; :737/:769 on 0.1.3-alpha.2) and cleared by a matching step/end (index.js:582; :744-745 on 0.1.3-alpha.2). A marker appended between turns has no open step → stepStart === void 0 → throw.

Step 3 — the marker shape is fully legal for the persistence consumer

dsh-session/lib/index.js planSurfaceEvent (replace branch) validates only: surfaceOp.start/end exist in the surface and sourceEventSeqs cover every shadowed node (assertProvenance, lib/index.js:320). Nothing about data.turn/step. foldSurface replays the same event without error. So: persistence ✅, meter ❌ — the "two verdicts" on one event.

Step 4 — compaction is a client of the meter; when the meter throws, no checkpoint is written

dsh-compaction-basic drives pressure via ctx.tokenMeter.measure(...); on throw the compaction transaction aborts before compaction/summary/checkpoint. Observable residue in the session log: compaction/start … command/done with kind:"error", text:"token meter: assistant/message at seq N has no matching step/start event", surface unchanged (verified on our sessions, e.g. session-e7b583f9 seq 243624).

Step 5 — the failure is durable, not transient

The marker event stays in the log (append-only). Every /compact retry after it throws again. We reproduced on 7 sessions; the only recovery was removing/rewriting the marker (our dsh-log-contract fix --drop-turnnull).

Step 6 — automatic pressure checks degrade quietly, which hides the issue

Per dsh-compaction-basic/README.zh.md: step-boundary pressure-check runtime failures "log a warning and continue". So sessions keep working for normal turns while /compact fails forever — the failure mode is silent until the user explicitly compresses.

Version note (2026-09-08): re-verified against @deepseek-ai/dsh-token-meter@0.1.3-alpha.2 — the package was restructured (782 lines: BlockAssembler import replaced by assembleAssistantStream / lastAssistantStreamChunk; _sync now at :705, step/start handling at :737/:769, step/end check at :744-745), but the core throw is verbatim: assistant/message at seq ${event.seq} has no matching step/start event at :753 (was :590 on 0.1.1-rc.2, :727 on 0.1.2-rc.1). The claim of this report — a turn/step-null replace marker permanently breaks /compact — holds on the npm line through 0.1.3-alpha.2.

---

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.

Reported by the OfferKuai Team — Founder: Zhaofeng (Yaming). Website: https://www.offerkuai.com/ | Contact: contact@offerkuai.com

中文版 / ZH

六步证据链，每步含 文件:行号 @ 版本、逐字代码片段与说明。

第 1 步 — meter 折叠整份持久化日志，无"跳过坏事件"的逃生舱

dsh-token-meter/lib/index.js:556（v0.1.1-rc.2）_sync(session)——0.1.3-alpha.2 为 :705：

while (state.consumedEvents < session.events.length) {
    const event = session.events[state.consumedEvents];
    this._foldEvent(session, state, event);
    state.consumedEvents += 1;
}

session.events 是完整恢复日志。历史中任何一条问题事件都会污染此后所有测量；没有部分折叠的逃生舱。

第 2 步 — 每条 assistant/message 必须匹配当前打开的 step

dsh-token-meter/lib/index.js:590（逐字，v0.1.1-rc.2；0.1.3-alpha.2 为 :753）：

else if (event.type === "assistant/message") {
    const stepStart = state.stepStart;
    if (stepStart === void 0 || stepStart.turn !== event.data.turn || stepStart.step !== event.data.step) throw new Error(`token meter: assistant/message at seq ${event.seq} has no matching step/start event`);

state.stepStart 仅由 step/start 设置（index.js:574，v0.1.1-rc.2；0.1.3-alpha.2 为 :737/:769）、由匹配的 step/end 清除（index.js:582；0.1.3-alpha.2 为 :744-745）。写在轮次之间的 marker 没有打开 step → stepStart === void 0 → 抛错。

第 3 步 — 该 marker 形态对持久化消费者完全合法

dsh-session/lib/index.js planSurfaceEvent（replace 分支）只校验：surfaceOp.start/end 存在于 surface、sourceEventSeqs 覆盖每个被遮蔽节点（assertProvenance，lib/index.js:320）。完全不涉及 data.turn/step。foldSurface 重放同一事件不抛。于是：持久化 ✅、meter ❌——同一事件的"两种判决"。

第 4 步 — 压缩是 meter 的客户端；meter 抛错则检查点不落盘

dsh-compaction-basic 通过 ctx.tokenMeter.measure(...) 驱动压力；抛错时压缩事务在 compaction/summary/检查点之前中止。会话日志中的残留：compaction/start … command/done 带 kind:"error"、text:"token meter: assistant/message at seq N has no matching step/start event"，surface 不变（我方会话实测，如 session-e7b583f9 seq 243624）。

第 5 步 — 失败是持久的，不是瞬态

marker 事件留在日志（append-only）。此后每次 /compact 重试仍抛错。我们在 7 个会话上复现；唯一恢复手段是移除/改写 marker（我们的 dsh-log-contract fix --drop-turnnull）。

第 6 步 — 自动压力检查静默降级，掩盖了问题

据 dsh-compaction-basic/README.zh.md：step 边界压力检查的运行故障"记 warning 后继续"。于是会话日常轮次照常工作，唯独 /compact 永远失败——直到用户显式压缩才暴露。

版本注（2026-09-08）： 已对照 @deepseek-ai/dsh-token-meter@0.1.3-alpha.2 复核——该包已重构（782 行：BlockAssembler 导入改为 assembleAssistantStream / lastAssistantStreamChunk；_sync 现位于 :705、step/start 处理在 :737/:769、step/end 校验在 :744-745），但核心 throw 逐字保留：:753 的 assistant/message at seq ${event.seq} has no matching step/start event（0.1.1-rc.2 为 :590、0.1.2-rc.1 为 :727）。本帖论点——turn/step 为 null 的 replace marker 会永久破坏 /compact——在 npm 线截至 0.1.3-alpha.2 仍成立。

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。

本报告由 OfferKuai（Offer快）团队提交 —— 创始人：Zhaofeng（Yaming）。官网：https://www.offerkuai.com/ | 联系：contact@offerkuai.com

[yamingmou]

Fix proposals — making the token meter and foldSurface agree on turn/step-null replace markers

EN / English

Three options, each with ✅/⚠️ critique. Goal: a plugin-authored replace marker with turn/step = null must not make every later pressure measurement and /compact throw.

Option 1 — upstream: meter skips step-pairing for turn-null replace events (recommended)

In dsh-token-meter/lib/index.js _foldEvent, before the assistant/message step check (:590), short-circuit when the event is a replacement marker shape:

if (event.type === "assistant/message") {
    const isMarker = event.surfaceOp?.op === "replace" && event.data?.turn == null && event.data?.step == null;
    if (!isMarker) { /* existing step-pairing check */ }
    // else: fold surface tokens only, no step requirement
}

• ✅ Minimal surface: one guard; the marker shape is already legitimized by foldSurface (the persistence consumer), so this aligns the two consumers instead of invalidating plugin data.
• ✅ Unblocks the whole plugin ecosystem: any plugin using the same replace primitive (compaction itself writes user-role checkpoints, but third-party markers are the affected class) stops breaking /compact.
• ✅ No data migration needed: existing logs with markers become compactable again.
• ⚠️ Semantics: a null-turn message has no step to attribute usage to; the meter must fall back to folding it into the running surface total (its existing foldSurfaceTokens path already handles surface pricing). Edge cases to define: what happens to the step anchor after such a marker — keep the current open step untouched.
• ⚠️ Severity scoping: only skip the pairing check; still run surface-token pricing and usage estimation, so no measurement is skipped, only the impossible step attribution.

Option 2 — plugin-side: write markers inside an open step

Change marker emission so a marker carries the turn/step of the step being edited (write it before step/end of that step).

• ✅ Keeps the meter strict everywhere; no upstream change.
• ⚠️ Not always possible: edits/recalls are issued between turns when no step is open; forcing a step context means the plugin would have to reopen a synthetic step — invasive and likely to violate other consumers.
• ⚠️ Does not fix already-written logs; migration still required.

Option 3 — plugin-side: pre-compact marker cleanup (/retrace-doctor)

A plugin command that, before compaction, scans for turn/step-null markers and either rewrites them or removes them (unshadowing the replaced content), then lets /compact proceed.

• ✅ Works with zero upstream change; gives users a deterministic path after editing.
• ⚠️ Removing markers reverts the edit visually (shadowed content becomes visible again) — a UX trade-off that must be explicit.
• ⚠️ Rewriting to a valid turn/step is only possible inside an open step (same limitation as Option 2).
• ⚠️ Operational: user must remember to run it; easy to forget until /compact fails.

Recommendation

Ship Option 1 upstream (it is the ecosystem-level fix), and implement Option 3 as the short-term plugin-side workaround plus Option 2 going forward for newly written markers. Offline checkers (ours: dsh-log-contract rule T1) should simulate the meter fold so affected sessions are flagged before the user compresses.

Questions for Maintainers

1. Is the turn/step = null replace marker shape a supported contract for third-party plugins, or should the meter's step-pairing be considered the intended gate?
2. If Option 1 is accepted, should the meter also tolerate step/end-without-step/start for the same marker class, or is the pairing check strictly for real assistant messages?
3. Is there an event-level flag we should use instead of inferring the marker shape from surfaceOp.op === "replace" && turn/step null?

---

Version note (2026-09-08): re-verified against @deepseek-ai/dsh-token-meter@0.1.3-alpha.2 — the assistant/message step check referenced above still throws at :753 (was :590 on v0.1.1-rc.2, :727 on 0.1.2-rc.1), verbatim; the proposals in this comment target logic that is unchanged on the npm line through 0.1.3-alpha.2.

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.

Reported by the OfferKuai Team — Founder: Zhaofeng (Yaming). Website: https://www.offerkuai.com/ | Contact: contact@offerkuai.com

中文版 / ZH

三个方案，各带 ✅/⚠️ 批判。目标：插件写入的 turn/step=null 的 replace marker 不得让此后的压力测量与 /compact 全部抛错。

方案 1 — 上游：meter 对 turn-null replace 事件跳过 step 配对（推荐）

在 dsh-token-meter/lib/index.js 的 _foldEvent 中，于 assistant/message 的 step 检查（:590）之前短路 marker 形态：

if (event.type === "assistant/message") {
    const isMarker = event.surfaceOp?.op === "replace" && event.data?.turn == null && event.data?.step == null;
    if (!isMarker) { /* 既有 step 配对检查 */ }
    // else：只折叠 surface tokens，不要求 step
}

• ✅ 改动面最小：一行守卫；该 marker 形态已被 foldSurface（持久化消费者）认可，此举是让两个消费者对齐，而非宣告插件数据非法。
• ✅ 解锁整个插件生态：任何使用同一 replace 原语的插件（受影响的是第三方 marker 这一类）都不再破坏 /compact。
• ✅ 无需数据迁移：已有带 marker 的日志恢复可压缩。
• ⚠️ 语义：null-turn 消息无 step 可归因；meter 需回退到其既有的 foldSurfaceTokens 路径计入 surface 总量。需定义边界：遇到此类 marker 后 step 锚点如何处理——建议保持当前打开的 step 不动。
• ⚠️ 严重度界定：只跳过配对检查；surface 计价与 usage 估算照常，不跳过任何计量。

方案 2 — 插件侧：marker 写入打开的 step 内

改为 marker 携带被编辑 step 的 turn/step（在该 step 的 step/end 之前写入）。

• ✅ 处处保持 meter 严格；无需上游改动。
• ⚠️ 并不总可行：编辑/撤回通常在轮次之间发出、无打开的 step；强制 step 上下文意味着插件要重开一个合成 step——侵入性强，且可能违反其他消费者。
• ⚠️ 不修复已写日志；仍需迁移。

方案 3 — 插件侧：压缩前 marker 清理（/retrace-doctor）

插件命令：压缩前扫描 turn/step-null marker，改写或移除（解除遮蔽、恢复被替换内容），再执行 /compact。

• ✅ 零上游改动即可工作；给用户一条确定的编辑后路径。
• ⚠️ 移除 marker 会让编辑外观回退（被遮蔽内容重新可见）——必须显式说明的 UX 取舍。
• ⚠️ 改写为合法 turn/step 只在打开的 step 内可行（与方案 2 同限）。
• ⚠️ 运维：用户需记得执行；忘记则到 /compact 失败才发现。

建议

上游做方案 1（生态级修复）；插件侧短期做方案 3 兜底、长期按方案 2 写新 marker。离线检查器（我方 dsh-log-contract 规则 T1）应模拟 meter 折叠，在用户压缩前就标记受影响的会话。

给维护者的问题

1. turn/step=null 的 replace marker 形态是否为第三方插件受支持的契约？还是 meter 的 step 配对本就被视为预期闸门？
2. 若接受方案 1，同一 marker 类的 step/end 无 step/start 是否也应容忍，还是配对检查仅针对真实 assistant 消息？
3. 是否应提供一个事件级标记（而非靠 surfaceOp.op === "replace" && turn/step null 推断形态）？

版本注（2026-09-08）： 已对照 @deepseek-ai/dsh-token-meter@0.1.3-alpha.2 复核——上文引用的 assistant/message step 检查在 :753 仍抛（v0.1.1-rc.2 为 :590、0.1.2-rc.1 为 :727），逐字保留；本评论的方案针对的逻辑在 npm 线截至 0.1.3-alpha.2 未变。

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。

本报告由 OfferKuai（Offer快）团队提交 —— 创始人：Zhaofeng（Yaming）。官网：https://www.offerkuai.com/ | 联系：contact@offerkuai.com