fix(core): Anthropic 4.6+ assistant-prefill 400 + thinking.display silently defaults to 'omitted'

[netbrah]

Description

Two related, verified bugs affecting every Claude Opus/Sonnet 4.6+ model
(and every 5.x family — Fable 5, Mythos 5, Sonnet 5, etc.) on the
Anthropic wire:

Bug 1: assistant-turn "prefill" 400s with no mitigation

When Gemini-format history ends on a model turn with no follow-up
(e.g. context trimming drops the next user turn, or a subagent
transcript is replayed mid-turn), the Anthropic converter forwards a
request whose last message has role: 'assistant'. Anthropic rejects
this outright on every 4.6+ model:

This model does not support assistant message prefill. The
conversation must end with a user message.

Per Anthropic's own migration documentation, this is a model-generation
behavior change (introduced at 4.6, applies through every subsequent
family), not a backend-specific quirk — confirmed live below across
both the affected and unaffected model boundary.

Bug 2: thinking.display silently defaults to omitted on newer models

Anthropic's adaptive-thinking display field controls whether thinking
content is returned as readable text (summarized) or redacted
(omitted, with only a continuity signature). The default value
changed between model generations: Sonnet 4.6 defaults to
summarized, but Opus 4.7+ and every 5.x family (including Sonnet 5)
silently default to omitted. qwen-code never sets display
explicitly, so any caller surfacing reasoning to users sees thinking
blocks stream back with empty text on newer models — no error, just
a long pause before output with nothing shown, where the same code
worked fine on Sonnet 4.6 or Opus 4.6.

Reproduction

Bug 1 — verified live against claude-sonnet-5, claude-opus-4.6,
claude-opus-4.7, claude-opus-4.8, and claude-sonnet-4.6 via the
Anthropic Messages API (all 400 identically):

curl -s -X POST https://api.anthropic.com/v1/messages \
  -H content-type:application/json \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model":"claude-sonnet-5",
    "max_tokens":16,
    "messages":[
      {"role":"user","content":"hi"},
      {"role":"assistant","content":"Sure, here is"}
    ]
  }'

{"type":"error","error":{"type":"invalid_request_error","message":"This model does not support assistant message prefill. The conversation must end with a user message."}}

The same request against claude-haiku-4.5 (a genuinely pre-4.6-generation
model — confirmed not a proxy alias to a 4.6+ model) succeeds normally,
confirming this is a real model-generation cutoff, not a universal
Anthropic restriction:

{"content":[{"type":"text","text":"\"Hi!\"\n\nIs there anything else..."}],...}

Bug 2 — verified live against claude-sonnet-5 with a task that
warrants visible reasoning: without display set, thinking blocks
return with empty thinking text; with display: 'summarized' set
explicitly, the same request returns full, non-empty reasoning text
(confirmed by direct comparison of both request variants against the
live API).

Expected Behavior

1. A conversation that would end on an assistant message should be
   made safe for 4.6+ models before it reaches the wire — either by
   dropping a genuinely empty trailing assistant message, or (when it
   carries real content) appending a synthetic user turn so nothing
   the model already said is discarded.
2. thinking.display should be set explicitly to 'summarized' when
   emitting the adaptive thinking shape, so reasoning text is visible
   regardless of which model generation's silent default applies.

Suggested Fix

• Add a stripTrailingAssistantPrefill converter option, gated on the
  existing modelSupportsAdaptiveThinking() check (the same 4.6+
  cutoff already used for the adaptive-thinking shape) — this reuses
  existing, already-correct version-detection infrastructure rather
  than introducing a new one.
• Set display: 'summarized' unconditionally on the { type: 'adaptive' } thinking shape returned by buildThinkingConfig(). This is a
  no-op on models whose default is already summarized and required
  on models whose default silently changed to omitted.

Both are minimal, additive changes localized to
anthropicContentGenerator/converter.ts and
anthropicContentGenerator/anthropicContentGenerator.ts, following the
same options-object pattern already used for the file's other
converter passes (dropUnsignedAssistantThinking,
injectThinkingOnToolUseTurns, etc.).

Happy to open a PR for this — will link it here once ready.

[qwen-code-ci-bot]

✅ Qwen Triage finished — view run. See the stage comments in this thread for the result.

✅ Qwen Triage 已完成 —— 查看运行。结果见本线程中的各阶段评论。

[qwen-code-ci-bot]

Triage

• Type: bug
• Labels: type/bug, category/core, scope/content-generation, priority/P1, welcome-pr
• Next: Confirmed via static analysis — both bugs are real; community PR welcome

Static Analysis

Bug 1 (trailing assistant prefill → 400): Confirmed. convertGeminiRequestToAnthropic() in converter.ts assembles messages, merges consecutive assistant turns, and cleans orphaned tool calls, but has no pass that ensures the final message is role: 'user'. If context trimming or subagent replay leaves history ending on an assistant turn, the request hits the Anthropic API as-is and 400s on every 4.6+ model.

Bug 2 (thinking.display defaults to omitted): Confirmed. buildThinkingConfig() in anthropicContentGenerator.ts returns { type: 'adaptive' } without setting display. On Opus 4.7+ and all 5.x models the server-side default is omitted, so thinking blocks stream back with empty text — no error, just invisible reasoning.

Affected area: packages/core/src/core/anthropicContentGenerator/ (converter.ts, anthropicContentGenerator.ts)

Fix direction:

1. Add a trailing-assistant-message guard in the converter (drop empty trailing assistant messages; append a synthetic user turn when content is non-empty), gated on modelSupportsAdaptiveThinking().
2. Set display: 'summarized' explicitly on the { type: 'adaptive' } thinking shape.

Both are minimal, additive, and follow existing converter-pass patterns (dropUnsignedAssistantThinking, injectThinkingOnToolUseTurns).

Reproduction not executed (unattended CI — static review only). The reporter's live-API curl reproduction is consistent with the code path analysis.

中文说明

• 类型: bug
• 标签: type/bug, category/core, scope/content-generation, priority/P1, welcome-pr
• 下一步: 静态分析已确认两个 bug 均真实存在，欢迎社区 PR

Bug 1（尾部 assistant 消息导致 400）： 已确认。converter.ts 中的 convertGeminiRequestToAnthropic() 没有任何步骤确保最终消息为 role: 'user'。当上下文裁剪或子代理回放导致历史以 assistant 消息结尾时，请求会原样发送到 Anthropic API，在所有 4.6+ 模型上返回 400。

Bug 2（thinking.display 默认为 omitted）： 已确认。anthropicContentGenerator.ts 中的 buildThinkingConfig() 返回 { type: 'adaptive' } 但未设置 display。在 Opus 4.7+ 和所有 5.x 模型上，服务端默认值为 omitted，导致思考内容以空文本返回。

修复方向： 在 converter 中添加尾部 assistant 消息保护；在 adaptive thinking 配置中显式设置 display: 'summarized'。

--- Qwen Code

[vansh2408]

/assign

[doudouOUC]

Confirmed both bugs against source — nice investigation.

Bug 1 (trailing assistant prefill → 400):
The conversion pipeline in packages/core/src/core/anthropicContentGenerator/converter.ts (convertGeminiRequestToAnthropic → processContents) maps Gemini role: 'model' to Anthropic role: 'assistant' but has no final pass that ensures the message array ends with a user turn. The existing post-processing steps (mergeConsecutiveAssistantMessages, cleanOrphanedToolCalls, mergeConsecutiveUserMessages) handle other edge cases but not this one. A guard that either appends a minimal user continuation or strips the trailing assistant turn (depending on the desired semantics) would fix it.

Bug 2 (thinking.display defaults to omitted):
buildThinkingConfig in packages/core/src/core/anthropicContentGenerator/anthropicContentGenerator.ts emits { type: 'adaptive' } for 4.6+ models without setting a display field. Since newer model families (Opus 4.7+, all 5.x) default display to omitted, thinking blocks arrive with empty text unless we explicitly pass display: 'summarized'.

Both are straightforward to fix. Welcome the PR — feel free to reach out if you need any pointers on the converter internals.

[vansh2408]

@doudouOUC PR - #8063