v0 codec refuses two released writer variants: permission/preset.origin and subagent/descriptor version 2

[mcknicks]

Two released v0 writers stamped payloads the frozen v0 codec refuses, so their sessions cannot be observed by any current dsh — the web gateway reports Failed to load history and resuming refuses before any turn runs.

1. permission/preset.origin

@deepseek-ai/dsh-permission-presets@0.1.1-rc.1 wrote permission/preset data as { preset, origin } with origin ∈ {default, inferred, selection}. 0.1.1-rc.2 removed origin and no later format carries it. The frozen v0 disposition is required: ['preset'], so the v0→v1 edge throws:

@deepseek-ai/dsh-session-format-v0-to-v1 refuses this format v0 Session:
permission/preset 0 data has unexpected member "origin"

2. subagent/descriptor version 2

@deepseek-ai/dsh-subagent 0.0.1-rc.3 through 0.1.1-rc.2 stamped subagent/descriptor version: 2; 0.1.2-rc.1 raised it to 3. The v0 edge refuses version 2 (unsupported descriptor version 2), so child sessions fail to restore and a parent's catalog degrades those children to unavailable.

Reproduction

A store written by 0.1.1-rc.1 reproduced both: 15 top-level sessions and 27 child sessions refuse exactly as above. The largest parent holds 36,642 events (99,074 physical rows) and 24 children; every one of them was unreadable.

Patch

I have a branch in my fork that adds two bounded historical normalizers beside the existing request/header.header.messagePrefix removal, both applied before assertReleasedEventPayload so the normal outcome validation still runs:

• permission/preset — drop the retired origin member. Effective permission is folded from preset plus sandbox/mode and approval/policy; origin was consumed only by the blank-session default-reuse path, which no later format has, and the affected sessions are non-blank. preset and every other fact are preserved.
• subagent/descriptor — advance version: 2 to 3. Version 3's schema is a superset that adds only the optional agentReasoningEffort, and the v3→v4 catalog interpreter already admits descriptor versions 1..3. The advanced marker is validated under the existing version 3 semantics, so a payload that version 2 accepted without a label still refuses.

A member no released writer produced still refuses; added negative cases cover that.

Branch: mcknicks/deepseek-harness → fix/v0-legacy-writer-variants (single commit). Since this repository does not accept external pull requests, this discussion is the report.

Verification

• packages/session/session-format-v0-to-v1: 113/113 tests pass, including two new cases.
• Mutation check: removing the two normalizers fails exactly the two new cases (2 failed, 20 passed).
• Downstream session-format packages: v1-to-v2 145/145, v2-to-v3 519/519, v3-to-v4 388/388 pass.
• End-to-end on the affected store: the largest session restores through the real catalog (with explicit child facts) to 36,642 v4 events and zero unavailable children, and a real dsh --profile headless --session-id resume loads the history through the production persistence stack.

Workaround for owners of affected stores

Until a release carries the normalizers, the two members can be removed from the stored v0 JSONL: strip data.origin from permission/preset events and set data.version to 3 on subagent/descriptor events with version 2. Keep the originals, and note that dsh-session-persistence-jsonl requires the first Zstandard frame to contain exactly the header line and every later frame to end on a record boundary, so a repair must preserve or rebuild the frame layout, not recompress the whole file as one frame.

[xiaoshenming]

你报的这两条形状我们一直在覆盖（#6151 / #6175 / #6189 / #6194 那一族），可以给你已经量过的结论，省得你在 fork 里重做一遍归一化。

1. permission/preset.origin —— 可离线修，且是无损的：origin 不被任何后续代际携带（{default, inferred, selection} 只是 0.1.1-rc.1 的中间状态），所以 repair 只做一件事：从 permission/preset 的 data 里删掉 origin、保留 preset，其余成员不动。我们的健康码是 v0-preset-extra-member。

2. subagent/descriptor version 2 —— 也可离线修，2 → 3。这条不是猜：v1→v2 与 v2→v3 完全不碰 subagent/descriptor，而 v3→v4 的 childCatalogFact 白名单是 [1,2,3]（还给 version 1 特判），并且 2 与 3 在那个函数里处理完全相同 —— 所以归一成 3 对下游无损。健康码 v0-descriptor-version（缺 version 或非整数的变体也归它，官方报 version must be a non-negative safe integer）。

一条你可能还没量到的爆炸半径（我们在 #7995 上实测的）：subagent/descriptor 那条闸门不只让那一条子会话打不开 —— persistence observer 会把每条已存 v0 会话都喂进这个 payload 闸门，于是一个坏工件让整个 store 的 session-search 查询全部失败（SESSION_QUERY_PERSISTENCE_FAILED），不是只坏它自己。你那台 15 + 27 条一起拒，正好是这个形状的放大版。

边界（我们刻意不碰的）：未知事件类型（如 session/imported、插件消息）迁移连 ignorable 都拒，我们只报告不删；成员集超出 released 形状的也一律不猜。

和你的补丁的关系：你的是写侧/编解码侧的历史归一化（让以后不再产生这种工件），我们的是离线工件修复（已经落盘的救回来）。两者互补，不冲突 —— 你已经有的那 42 条会话，不用等新版本也能开。

安装/更新：桌面端「插件」页 → dsh-session-surgeon → 更新；repair <id> 默认 dry-run，--apply 会先写 .bak.<utc>。先停写者再修。