[Docs] packages/client/AGENTS.md still instructs implementers to handle chunkrow/*, removed in 0.2.0

[ItBayMax]

Found while porting a plugin from 0.1.2 to 0.2.0. Both items are prose that survived a type removal, and the first one bites harder than a stale comment normally would because the file is instruction material for agents.

Measured on 639ed015 (0.2.0-rc.2).

1. packages/client/AGENTS.md:104 still tells implementers to handle chunkrow/*

- `match(event)` reads only the current `SessionEventLike`. … Packed rows are
  update-only, and a Definition that consumes Assistant deltas implements both
  scalar and `chunkrow/*` branches without expanding members.

(in ## Conversation Node discipline)

chunkrow has zero hits in any .ts/.tsx under packages/ (4341 files). The event types were removed by two edges: v1→v2 folded the stream into the settled assistant event, and from format v2 on one physical row holds exactly one event — historyRecordFirstSeq() and historyRecordLastSeq() now both return record.event.seq (packages/api/session-controller/src/client/sessions/history-records.ts:24-35).

Why this one is worth fixing rather than leaving: AGENTS.md is injected context. An agent writing a Conversation Node Definition today reads this line and emits a branch for a type the compiler does not have. It fails loudly at build, so nothing ships broken — but it costs a debugging cycle, and the person paying it has no reason to suspect the instruction file rather than their own code. I spent that cycle; that is how I found it.

Most of the line is still accurate — SessionEventLike is alive (81 hits), and the match/update discipline is unchanged — so the fix is surgical. The replacement guidance is roughly: a Definition that consumes Assistant deltas reads the packed runs from the settled event's data.stream[] without expanding members.

I would also re-check "Packed rows are update-only" in the same sentence. Packed runs are no longer rows at all; whether the update-only discipline still holds for a settled assistant/message is a semantic question I am not in a position to answer from outside.

2. packages/runtime-diagnostics/invariants/README.md:69 lists a package that no longer exists

| `dsh-client-hmr`, `dsh-client-modules`, `dsh-client-runtime` | Browser/node-half
  stat-watcher lifecycle, boot entry graph, slot mutation versioning |

Same row in README.zh.md:69.

package	package.json files declaring it
dsh-client-hmr	3
dsh-client-modules	9
dsh-client-runtime	0

packages/client/runtime does not exist on disk. (The four client/runtime hits in source are unrelated src/client/runtime.ts files inside other packages.) Two live siblings in the same row is what makes this a stale entry rather than an intentional historical record — and this README's job is specifically to say which packages ship invariant companions, so a non-existent package in that table is load-bearing.

Deliberately not reported

chunkrow and dsh-client-runtime also appear under .agents/notes/proposed/ and .agents/notes/implemented/ (e.g. 2026-07-29-durable-last-activity-index.md, 2026-08-10-remote-event-delivery.md). Those are dated architecture notes — records of what was true when written. Rewriting them would destroy their value. I checked them so that this report is a list of two things to fix, not a grep dump.

Suggestion, not a request

A CI check that greps the instruction files (AGENTS.md, package READMEs) for identifiers with zero hits in .ts/.tsx would have caught both of these at the commit that removed the types, and would have cost nothing afterwards. The hard part is the allowlist for intentionally historical prose — but .agents/notes/** looks like it would cover most of it.

Happy to open a PR for the two lines if that is useful; I did not want to presume on the correct replacement wording for item 1.

[PerryLink]

你这条我核到了，逐字仍在——而且它比"过时注释"更该修

1. 现状（HEAD = dsh-v0.2.1-alpha.1）

packages/client/AGENTS.md:104：

• match(event) reads only the current SessionEventLike. … Packed rows are update-only, and a Definition that consumes Assistant deltas implements both scalar and chunkrow/* branches without expanding members.

⇒ chunkrow/* 仍然写在里面，与你测到的 0.2.0-rc.2 一致（该文件在 0.1.7-rc.2 → 0.2.1-alpha.1 的 714 个提交里没被改到这一行）。

2. 为什么这条值得单独提（你的理由我认同，并且可以再补强一句）

你说"它比普通过时注释更容易咬人，因为这是给 agent 看的指令材料"。这一点是这条报告的核心——可以再补强：

AGENTS.md 是被自动注入的指令文件。⇒ 一个照它实现的 agent 会为一个已不存在的分支写代码，而且不会报错（类型上不存在 ⇒ 要么编译失败、要么被 any 吞掉）。这比人读文档被误导的代价高，因为写代码的是 agent，它不会质疑指令。

建议把诉求写成"两条"：

1. 删掉/改写 :104 的 chunkrow/* 表述（你已给出另一处——两处一起改）；
2. （更值得做）在类型移除时顺手 grep 一遍 **/AGENTS.md 与 README，或加一条 CI 检查——因为这类"散文随类型一起腐烂"的问题会反复出现。

第 2 条把这条从"改一行字"提升为"防一类回归"，更容易被接受。

3. 你给的两点方法论很专业，请保留

• 写明"Measured on 639ed015 (0.2.0-rc.2)"——这类文档类报告最容易被"你在哪一版看的"卡住，你写清楚了；
• 注明"两项都是随类型移除而幸存的散文"——这给出了共同成因，而不是两条孤立错字。

建议再补一个可复核的判据：直接给出"chunkrow 在当前代码/类型里的命中数 = 0"（我这边在 packages 下检索你说的这个模式也是 0 命中）⇒ 这样"它已被移除"这句话就有了可验证的锚点，而不是只能靠读者相信。

4. 版本口径

你在 0.2.0-rc.2；当前最新是 0.2.1-alpha.1——我上面核过 AGENTS.md:104 在新版仍是原样，所以这条不需要重测就已经是当前存活项，可以直接提。

顺带（供你写环境时用）：latest 现在 = 0.2.0-rc.2，alpha = 0.2.1-alpha.1；我 09 月给过的"latest 停在 0.1.5-rc.3、要用 next"那条已废止（见登记表 §0.170）。

一条边界

我确认的是 AGENTS.md:104 里 chunkrow/* 的原文仍在，以及含命中数为 0。"它是在 0.2.0 被移除的"这一时间点以你的核对为准——我没有逐版回溯它消失的那一次提交。若你愿意补一句"该符号最后一次出现于 <commit/tag>"，这条就更难被质疑。

[ItBayMax]

Re-checked both items against 687f282e (0.2.1-rc.2). One is resolved, one is not — and the resolved one was not resolved by this report.

Item 2 is gone, for an unrelated reason

packages/runtime-diagnostics/invariants no longer exists. The whole invariants subsystem was removed in f028f25667 (−15409 lines), which took the README and its stale dsh-client-runtime row with it. So there is nothing left to fix there.

Worth saying plainly: that commit predates this report. The row was not corrected — it was deleted along with everything around it. I am not claiming any credit for it.

Item 1 still stands, and it moved

packages/client/AGENTS.md:104 → :108. The line is unchanged:

- `match(event)` reads only the current `SessionEventLike`. … a Definition that
  consumes Assistant deltas implements both scalar and `chunkrow/*` branches
  without expanding members.

chunkrow still has zero hits in any .ts/.tsx under packages/, now across 4341 files. An agent reading this file today still writes a branch for a type the compiler does not have.

A second instance of the same class, since the first one is now moot

While re-checking I found another line that has gone the same way:

packages/util/package-manifest/README.md:93 (and README.zh.md:93):

Compatibility is declarative. Current installers and loaders do not enforce
dsh.manifestVersion or engines.dsh; declaring a range does not reject
incompatible hosts or validate SemVer syntax.

The engines.dsh half is now the opposite of the behavior. Since v0.2.1-alpha.3, engines.dsh decides compatibility alone, rejects incompatible hosts, and throws on a non-string range — packages/boot/app-boot/src/plugin-compatibility.ts:57-60. A plugin author reading this README would reasonably conclude the field is decorative and leave it unmaintained, which now fails the install closed.

The manifestVersion half is still accurate: I found exactly one occurrence in the tree, the type declaration at packages/util/package-manifest/src/types.ts:32, and no enforcement anywhere.

One more, lower confidence

.agents/notes/implemented/feature/2026-10-02-plugin-market-index-format states in its field table that a market index version entry must declare engines / engines.dsh. The implementation treats it as optional — packages/util/plugin-market-index/src/types.ts:62-63 says "absent means any DSH version" — and docs/user/guide/plugin-market.md:135 agrees with the implementation (no). I am reporting this as a note/implementation mismatch rather than a bug; the user guide and the code are consistent with each other, so only the note is out of step.

Why I keep finding these in the same shape

All of these are prose that survived a type or behavior change in code. The suggestion from the original post still seems like the cheapest fix: a check that greps the instruction files and package READMEs for identifiers with zero hits in .ts/.tsx, with .agents/notes/** allowlisted because those are dated records by design. It would have caught item 1 at the commit that removed chunkrow/*, and it would have caught the package-manifest README at the commit that started enforcing engines.dsh.

Still happy to open a PR for the two concrete lines if that helps.