Messages 协议拒绝遗留的 reasoning 内容，导致旧会话永久打不开

[bozhang1214]

Environment

• dsh 0.1.6-alpha.1 (profile: web)
• Provider: deepseek-official, wire protocol messages — the default since b0641b83fc
• The affected session was created before 2026-09-14 and had used a background subagent

Symptom

A pre-existing session fails on every turn with:

DeepSeek Messages cannot represent user/tool-result content reasoning
code = UNSUPPORTED_CONTENT

Retrying, switching models, and restarting the process all fail. Setting
protocol: chat-completions makes the same session work again, which points at the protocol
rather than at the data.

Root cause: the two protocols disagree on identical durable history

packages/llm/llm-deepseek/src/protocols/messages/serialize.ts:11-13, 57-59 accepts only
text and image in a user/tool-result position and throws for anything else.

Scanning ~/.dsh/sessions/**/session.v3.jsonl.zstd found 7 user/message events in one
session whose data.content carries a reasoning block directly. All 7 come from the same
producer — the background-subagent settlement notice, which inlined the child's closing
content blocks verbatim into the parent's user message (timestamps 2026-09-03 and 09-09).

That producer was already fixed:

29debb8b24 fix(subagent): keep settlement notices text-only (2026-09-14) —
packages/subagent/subagent/src/continuation-messages.ts:143 now keeps text blocks only. But
history persisted before that date cannot be rewritten, and Messages fails closed on it.

Running the same legacy user message through both serializers:

--- chat-completions ---  OK    -> [{"role":"user","content":"Its closing message:"}]
--- messages ---          THREW -> DeepSeek Messages cannot represent user/tool-result content reasoning

chat-completions' flattenText() keeps only text blocks and silently drops the rest
(serialize.ts:95-100); its assertTextOnly() rejects only images. Messages throws for any
non-text/non-image block.

So this is not data corruption — it is a contract mismatch between the two protocols.
Because 0.1.6-alpha.1 made messages the default, that mismatch silently made a class of
existing sessions permanently unusable.

Why simply dropping every non-text block is wrong

Scanning all local sessions for non-text blocks in user/tool-result positions gives
reasoning (20), file (5), image (47). image is already supported. A file block is
real model-visible input — discarding it would silently change what the model sees — so it
must keep failing closed. Only reasoning carries no model-visible meaning in that position,
and chat-completions already drops it.

Suggested fix

Treat reasoning in a user/tool-result position the way chat-completions already does: drop
it and report the degradation through the existing onReplayDegrade channel, keeping every
other unrepresentable block fail-closed.

if (block.type === 'reasoning') {
  onReplayDegrade?.('discarded reasoning content from a user/tool-result turn')
  return []
}
if (block.type !== 'image') return unsupported(`user/tool-result content ${block.type}`)

Patch (1 commit, 2 files, +30/−1 — serialize.ts + its spec, on my fork; no PR per CONTRIBUTING):

bozhang1214/deepseek-harness@bff8e9a...225ffa1

Alternatives the team may prefer:

• Sanitize at derivation (Session.deriveMessages() / the surface fold) and keep the
  protocol serializers strict. Caveat: SessionMessageProjection is designed for
  plugin-owned event types, so intercepting the core user/message would replace the built-in
  derivation.
• A session migration that strips these blocks from persisted logs — most thorough, but it
  rewrites user history.

All three need the same underlying decision: legacy history must be allowed to degrade,
because the producer fix can only apply going forward.

Verification

• Implemented the serializer-alignment variant (patch linked above): reasoning in a
  user/tool-result position is dropped with a degradation diagnostic; text / image
  unchanged; every other type still fails closed.
• Added a regression test (drop in both user and tool-result positions) and converted the
  assertion that previously required a throw into a tool-call fail-closed assertion
  (packages/llm/llm-deepseek/tests/messages/serialize.spec.ts:299 on upstream master).
• Targeted specs 113/113 and the whole llm-deepseek package 577/577 pass; oxlint
  clean.
• Same-input check after the change: messages returns
  [{"type":"text","text":"Its closing message:"}] — identical to chat-completions — plus the
  degradation diagnostic.
• Scope note: that is a serializer-level check plus unit tests. I have not re-run the
  affected session end-to-end against the provider yet (that needs a process restart).

Impact

Any session persisted before 2026-09-14 that used a background subagent becomes permanently
unusable once the messages protocol becomes the default — and it fails on every single turn.
The error names only "content reasoning" and never hints that the data is legacy or that
chat-completions still handles it, so users are unlikely to find the cause.

Related, but not duplicate

Separate from the dsh_session_log first-upload issue (HTTP 413): that one is about request
body size, this one about representable content.