fix(core): fix tool_use/tool_result pairing for Anthropic-compatible providers - #6651
Merged
Conversation
Contributor
Code Coverage Summary
CLI Package - Full Text ReportCore Package - Full Text ReportFor detailed HTML reports, please see the 'coverage-reports-22.x-ubuntu-latest' artifact from the main CI run. |
Add mergeConsecutiveAssistantMessages and cleanOrphanedToolCalls to the Anthropic converter, mirroring the same-name functions already present in the OpenAI converter. When the Gemini history has consecutive model turns (e.g. from streaming chunk-level recording, max_tokens recovery, or adaptive thinking splits), processContent emits one Anthropic message per Content. The Anthropic API requires tool_use blocks to be immediately followed by tool_result blocks — consecutive assistant messages break this pairing and cause HTTP 400. The merge pass combines consecutive assistant messages into one (thinking blocks first, then non-thinking). The clean pass removes orphaned tool_use and tool_result blocks. A second merge pass fixes any alternation issues created by dropped messages.
yiliang114
changed the title
yiliang114
marked this pull request as ready for review
July 10, 2026 07:02
Collaborator
|
Re-run triggered by @wenshao · commit Thanks for the PR! Template looks good ✓ Problem: The Direction: Aligned. Fixing tool_use/tool_result pairing is essential for Anthropic-compatible provider support. The two-layer approach (targeted compaction fix + converter-level safety net) mirrors the proven pattern already in the OpenAI converter ( Size: Core paths touched. Production logic: 247 lines (converter.ts: 220, chatCompressionService.ts: 27). Test lines: 936. Well under the 500-line threshold — no size concerns. Approach: Scope feels right. Layer 1 alone wouldn't cover the streaming/max_tokens/adaptive-thinking paths; Layer 2 alone wouldn't fix the compaction timing window. Both layers are needed and both are minimal. The implementation faithfully mirrors the OpenAI converter's Moving on to code review. 🔍 中文说明由 @wenshao 触发的重新审查 · commit 感谢贡献！ 模板完整 ✓ 问题： 方向： 对齐。修复 tool_use/tool_result 配对对于 Anthropic-compatible 提供方支持至关重要。两层方案与 OpenAI 转换器中已验证的模式一致。 规模： 触及核心路径。生产逻辑 247 行。测试 936 行。远低于 500 行阈值，无规模顾虑。 方案： 范围合理，两层缺一不可，实现忠实镜像了 OpenAI 转换器的模式。无无关改动。 进入代码审查 🔍 — Qwen Code · qwen3.7-max |
Collaborator
|
Re-run triggered by @wenshao · commit Code ReviewIndependent proposal (before reading diff): Add post-processing to the Anthropic converter — Comparison with the PR: The implementation matches my proposal exactly. The Anthropic-specific adaptations are correct:
Reuse check: The functions mirror the existing No critical blockers or AGENTS.md violations found. The code is clean, well-documented, and follows project conventions. TestingUnit Tests187/187 tests pass. New tests cover: merging consecutive assistant messages, thinking block reordering, orphaned tool_use/tool_result cleanup, tool_use deduplication, pending tool result inclusion in compaction side-query, and proper tool_use/tool_result pairing in existing tests. LintSmoke Test (dev build)Dev build starts and responds correctly. No startup regressions. Real-Scenario E2ECannot run — reproducing the HTTP 400 error requires a live Anthropic-compatible API (Claude 4.6 with adaptive thinking) and specific trigger conditions (streaming chunk-level recording, max_tokens recovery, or compaction during tool loop). No Anthropic API credentials are available in this CI environment. However, the maintainer (@wenshao) ran a deterministic converter-level E2E verification (see comment above) that feeds four realistic Gemini histories — one per trigger path — through the real 中文说明代码审查独立方案（未读 diff 前）： 在 Anthropic 转换器中添加后处理—— 与 PR 对比： 实现完全匹配。Anthropic 特有的适配正确：thinking 块重排、tool_use 去重、无条件运行 cleanOrphanedToolCalls、compaction 中正确检查 pending tool result。 复用检查： 函数镜像了 OpenAI 转换器中已有的同名函数，因消息格式不同必须独立实现。无重复问题。 无关键阻塞或 AGENTS.md 违规。 测试单元测试187/187 通过（converter: 60, compression: 90, cli: 37）。 Lintexit 0，无告警。 冒烟测试Dev 构建正常启动并正确响应，无回归。 真实场景 E2ECI 环境中无法运行（需要真实 Anthropic API）。但 maintainer 已进行确定性转换器级别 E2E 验证：main 上 4 个场景中 3 个失败，本分支全部通过。 — Qwen Code · qwen3.7-max |
Collaborator
|
Re-run triggered by @wenshao · commit ReflectionStepping back, this PR does exactly what it should: it ports a battle-tested pattern from the OpenAI converter to the Anthropic converter. The The two-layer approach is well-justified. Layer 1 (compaction side-query) fixes a specific timing window where the pending tool result is missing from the API request. Layer 2 (converter post-processing) is a safety net that catches all paths that create consecutive model Content entries. Neither layer alone is sufficient; together they provide comprehensive coverage. The compaction fix is surgical — ~20 lines that check for a pending tool result and include it in the side-query. No over-engineering. What gives me confidence: 187 unit tests passing, covering the new merge/clean logic thoroughly including edge cases. The maintainer's deterministic converter-level E2E verification reproduces the failure on What I can't verify: live Anthropic API behavior. No E2E run against a real Anthropic-compatible provider was possible in CI. But the converter logic is well-tested, the pattern is proven in the OpenAI converter, and the fix targets a documented API constraint. The double This is a clean, focused bugfix that solves a real problem. Ships it. ✅ 中文说明反思这个 PR 将 OpenAI 转换器中经过实战验证的模式移植到 Anthropic 转换器。两层方案有充分理由：第一层修复压缩时机窗口，第二层是转换器层面的兜底方案。 信心来源：187 个单元测试全部通过，maintainer 的确定性 E2E 验证在 main 上复现了失败（4 个场景中 3 个失败），本分支全部通过。ESLint 干净，dev 构建冒烟测试通过。 无法验证的部分：真实 Anthropic API 行为。但转换器逻辑测试充分，模式已在 OpenAI 转换器中验证。 双重 这是一个干净、聚焦的 bugfix，解决了真实问题。可以合并。✅ — Qwen Code · qwen3.7-max |
qwen-code-ci-bot
approved these changes
Jul 10, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
LGTM, looks ready to ship. ✅
qwen-code-ci-bot
previously approved these changes
Jul 10, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
No issues found. LGTM! ✅
— qwen3.7-max via Qwen Code /review
qqqys
reviewed
Jul 10, 2026
qqqys
left a comment
qqqys
left a comment
Collaborator
There was a problem hiding this comment.
No review findings. Downgraded from Approve to Comment: CI still running.
— GPT-5 via Qwen Code /review
qwen-code-ci-bot
requested changes
Jul 10, 2026
Collaborator
Suggestions — commit
|
wenshao
requested changes
Jul 10, 2026
qwen-code-ci-bot
requested changes
Jul 10, 2026
Collaborator
✅ Local build & test verification (maintainer)Built and tested this PR locally as a merge reference. Everything passes, and I added an end-to-end converter check that reproduces the exact
Environment — tested at commit 1. Build & typecheck — ✅ clean2. Unit tests — ✅ 187 / 187 passed
3. Lint — ✅ clean4. End-to-end converter verification (before → after)Beyond the unit tests, I fed four realistic Gemini histories — one per trigger path in the PR's table — through the real BEFORE ( AFTER (this PR) — all invariants hold: The Verdict
LGTM from a build/test standpoint — recommend merge. 👍 Caveat, matching the PR's own scope note: this verifies converter output against Anthropic's documented pairing rules deterministically; it does not include a live Anthropic-compatible API E2E round-trip. The converter-level check above is what actually gates the HTTP 400, so I consider it sufficient coverage for the fix. 🇨🇳 中文版本（点击展开）✅ 本地构建与测试验证（维护者）作为合并参考，我在本地完整构建并测试了本 PR。全部通过，并额外补充了一个端到端转换器验证：在
环境 — 测试于 commit 1. 构建与类型检查 — ✅ 通过2. 单元测试 — ✅ 187 / 187 通过
3. Lint — ✅ 干净对两个改动的源文件运行 4. 端到端转换器验证（修复前 → 修复后）除单元测试外，我按 PR 表格中的四条触发路径各构造一个真实 Gemini 历史，喂入真实的 修复前（ 修复后（本 PR）— 全部满足： 4/4 场景通过。 结论
从构建/测试角度 LGTM，建议合并。 👍 说明（与 PR 自身范围一致）： 此验证以确定性方式校验转换器输出是否符合 Anthropic 文档的配对规则，不包含真实 Anthropic-compatible API 的 E2E 往返请求。由于上述转换器层面的检查正是实际拦截 HTTP 400 的环节，我认为对本修复而言覆盖已足够。 |
wenshao
previously approved these changes
Jul 10, 2026
Collaborator
|
@qwen-code /triage |
qwen-code-ci-bot
approved these changes
Jul 10, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
LGTM, looks ready to ship. ✅
wenshao
approved these changes
Jul 10, 2026
Collaborator
Author
Post-release verification: 0.19.9Verified the published The isolated test configuration intentionally omitted
No run emitted Some model-generated parallel Conclusion: |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this PR does
Fixes the
tool_use ids were found without tool_result blocks immediately afterHTTP 400 error that occurs with Anthropic-compatible providers (notably Claude 4.6 with adaptive thinking). The fix has two complementary layers:Layer 1 — Compaction side-query (first commit): When automatic compaction runs during a tool loop before the tool result is committed to chat history, the compaction side-query now includes the pending tool result so Anthropic-compatible providers see it immediately after the preceding tool_use block.
Layer 2 — Anthropic converter merge/clean (second commit): Adds
mergeConsecutiveAssistantMessagesandcleanOrphanedToolCallsto the Anthropic converter, mirroring the same-name functions already present in the OpenAI converter. When the Gemini history has consecutive model turns (from streaming chunk-level recording, max_tokens recovery, or adaptive thinking splits),processContentemits one Anthropic message per Content. The Anthropic API requires tool_use blocks to be immediately followed by tool_result blocks — consecutive assistant messages break this pairing. The merge pass combines consecutive assistant messages into one (thinking blocks first, then non-thinking). The clean pass removes orphaned tool_use and tool_result blocks. A second merge pass fixes any alternation issues created by dropped messages.Why it is needed
Anthropic-compatible providers require every
tool_useblock to be followed immediately by its correspondingtool_result. Multiple code paths can create consecutive model Content entries that violate this requirement:Layer 1 is a targeted fix for the compaction timing window. Layer 2 is a converter-level fallback that covers all trigger paths. Together they provide comprehensive coverage.
Reviewer Test Plan
How to verify
Trigger a tool-heavy session with an Anthropic-compatible model (e.g. Claude Opus 4.6) and confirm that tool loops, compaction, and subagent spawning all complete without a missing-tool_result 400 error. Specifically: spawn a subagent that makes multiple tool calls in a single response, then verify the continuation request succeeds.
Evidence (Before & After)
N/A. This is a non-UI change.
Tested on
Environment (optional)
macOS.
cd packages/core && npx vitest run src/core/anthropicContentGenerator/converter.test.ts src/services/chatCompressionService.test.ts— 145 tests passed. ESLint clean.Risk & Scope
@lydell/node-ptytypecheck error on clean branch is outside this change.Linked Issues
Closes #6654
中文说明
What this PR does
修复 Anthropic-compatible 提供方（尤其是 Claude 4.6 adaptive thinking）出现的
tool_use ids were found without tool_result blocks immediately after400 错误。包含两层互补修复：第一层 — 压缩 side-query（第一个提交）： 自动压缩在工具循环中、工具结果尚未写入历史时触发，side-query 现在会携带待发送的工具结果。
第二层 — Anthropic 转换器合并/清理（第二个提交）： 在 Anthropic 转换器中添加
mergeConsecutiveAssistantMessages和cleanOrphanedToolCalls，与 OpenAI 转换器中已有的同名函数对齐。当 Gemini 历史中出现连续 model turn（流式 chunk 级别存储、max_tokens 恢复、adaptive thinking 分裂等），转换器逐 Content 生成 Anthropic 消息导致连续 assistant 消息，违反 API 的 tool_use → tool_result 配对要求。合并步骤将连续 assistant 消息合并为一条（thinking 块在前），清理步骤移除孤立的 tool_use/tool_result 块。Why it is needed
Anthropic-compatible 提供方要求每个
tool_use紧随对应的tool_result。多个代码路径会产生违反此要求的连续 model Content：第一层是压缩时机窗口的针对性修复，第二层是转换器层面的兜底方案，覆盖所有触发路径。
Reviewer Test Plan
How to verify
使用 Anthropic-compatible 模型（如 Claude Opus 4.6）进行工具密集型会话，确认工具循环、压缩、子代理启动均不出现缺少 tool_result 的 400 错误。具体：启动一个在单次响应中发起多个工具调用的子代理，验证后续请求成功。
Evidence (Before & After)
N/A，非 UI 变更。
Tested on
Environment (optional)
macOS；
cd packages/core && npx vitest run src/core/anthropicContentGenerator/converter.test.ts src/services/chatCompressionService.test.ts，145 tests passed；ESLint clean。Risk & Scope
@lydell/node-ptytypecheck 错误不在本次修改范围。Linked Issues
无关联 issue — 此 bug 在 Claude 4.6 测试中发现。如需跟踪可由 maintainer 创建 issue。