18 KiB
Phase 07: Phase 06 验收后补丁归档(mention/upload语义与附件预览复用)- Research
Researched: 2026-04-15
Domain: 前后端 mention/upload 语义收敛、附件预览组件复用、memory 清理与验证归档
Confidence: HIGH
User Constraints (from CONTEXT.md)
07-phase-06-mention-upload 目录下不存在 *-CONTEXT.md,因此无可逐字拷贝的 Locked Decisions/Discretion/Deferred。 [VERIFIED: codebase grep .planning/phases/07-phase-06-mention-upload/*-CONTEXT.md]
基于本次 objective 的硬约束如下:将 Phase 06 已验收绕行改动正式纳入 Phase 07,范围必须覆盖 mention/upload 语义统一、附件预览复用、memory 清理、可验证提交路径。 [VERIFIED: user objective]
Summary
Phase 06 的代码层关键补丁已经在仓库内落地:前端通过 additional_kwargs.files 单一 envelope 发送 uploads + mentions,后端 UploadsMiddleware 已区分 ref_kind=mention 并单独注入 <mentioned_files>,且 new_files 不再错误吸收 mention。 [VERIFIED: codebase grep frontend/src/core/threads/hooks.ts, frontend/src/core/threads/submit-files.ts, backend/.../uploads_middleware.py]
memory 侧也已有清理链路:MemoryMiddleware 在入队前剥离 <uploaded_files>/<mentioned_files>,MemoryUpdater 在落盘前清除上传事件句子与 facts;对应回归测试存在且本地通过。 [VERIFIED: codebase grep backend/.../memory_middleware.py, backend/.../memory/updater.py, backend/tests/test_memory_upload_filtering.py; VERIFIED: test run uv run pytest -q tests/test_memory_upload_filtering.py]
Phase 07 的核心不是“再造新功能”,而是“归档与验证闭环”:统一术语契约、固定附件预览复用边界、补齐 E2E 选择器漂移、同步 UAT/Validation/Requirements 文档状态,形成可审计提交路径。 [VERIFIED: codebase grep .planning/phases/06-/06-VERIFICATION.md, .planning/phases/06-/06-UAT.md, .planning/REQUIREMENTS.md; VERIFIED: test run pnpm -s test:e2e --grep "DF-INPUT-007|DF-INPUT-008|DF-INPUT-009"]
Primary recommendation: Phase 07 按 docs/contract-fix -> test-fix -> re-verify -> archive 四段执行,禁止再扩展功能面。 [VERIFIED: repo state + phase goal]
Project Constraints (from CLAUDE.md)
项目根目录不存在 CLAUDE.md,无额外项目级强制约束。 [VERIFIED: filesystem check test -f CLAUDE.md]
Standard Stack
Core
| Library | Version | Purpose | Why Standard |
|---|---|---|---|
@radix-ui/react-dropdown-menu |
repo: ^2.1.16; npm latest: 2.1.16 (2025-08-13) |
mention 候选面板(键盘/焦点/定位) | 已在输入框实现且与现有 shadcn 体系一致,避免自定义浮层分叉。 [VERIFIED: codebase grep frontend/src/components/workspace/input-box.tsx; VERIFIED: npm registry npm view @radix-ui/react-dropdown-menu version time] |
sonner |
repo: ^2.0.7; npm latest: 2.0.7 (2025-08-02) |
stale/上限提示 | 现有错误提示已基于 toast 语义,便于保持软失败行为一致。 [VERIFIED: codebase grep toast.error in hooks.ts/input-box.tsx; VERIFIED: npm registry npm view sonner version time] |
PromptInputAttachment(内部组件) |
repo internal | 输入区附件/引用缩略预览 | 当前 reference 预览已复用该组件,是 Phase 07 应固化的复用基线。 [VERIFIED: codebase grep frontend/src/components/workspace/input-box.tsx, frontend/src/components/ai-elements/prompt-input.tsx] |
UploadsMiddleware + MemoryMiddleware(内部中间件) |
repo internal | upload/mention 注入与 memory 入队清理 | 语义分层已形成:uploaded_files 与 mentioned_files 分离,memory 过滤双重防线。 [VERIFIED: codebase grep backend/.../uploads_middleware.py, backend/.../memory_middleware.py, backend/.../memory/updater.py] |
Supporting
| Library | Version | Purpose | When to Use |
|---|---|---|---|
@playwright/test |
repo: ^1.48.0; CLI: 1.48.0 |
前端 @引用 回归 | 验证 DF-INPUT-007/008/009 与 testid 合同一致性。 [VERIFIED: frontend/package.json; VERIFIED: command pnpm exec playwright --version] |
pytest via uv run |
backend dev: pytest>=8.0.0 |
后端 middleware/memory 回归 | 本机无全局 pytest 时使用 uv run pytest。 [VERIFIED: backend/pyproject.toml; VERIFIED: env check command -v pytest; VERIFIED: test run] |
Alternatives Considered
| Instead of | Could Use | Tradeoff |
|---|---|---|
DropdownMenu |
自定义绝对定位浮层 | 自定义层更易与焦点管理/E2E 选择器漂移。 [VERIFIED: historical phase docs + current selector mismatch] |
PromptInputAttachment 复用 |
新建 mention-only 预览组件 | 会重复实现删除/图片缩略行为,增加 UI 行为分叉。 [VERIFIED: code comparison in input-box.tsx + prompt-input.tsx] |
Installation:
cd frontend && pnpm install
cd backend && uv sync
Architecture Patterns
Recommended Project Structure
frontend/src/components/workspace/input-box.tsx # mention candidate + 引用预览
frontend/src/core/threads/submit-files.ts # files envelope 归一化
frontend/src/core/threads/hooks.ts # 发送链路 + stale 软失败
backend/packages/harness/.../uploads_middleware.py # uploaded/mentioned 语义拆分
backend/packages/harness/.../memory_middleware.py # 入队前剥离标签
backend/packages/harness/.../memory/updater.py # 落盘前清理上传事件
backend/tests/test_uploads_middleware_core_logic.py # mention/upload 后端回归
backend/tests/test_memory_upload_filtering.py # memory 清理回归
frontend/tests/e2e/input-and-compose.spec.ts # DF-INPUT-007/008/009
[VERIFIED: codebase grep]
Pattern 1: 单一提交 Envelope + 语义位区分
What: 统一走 additional_kwargs.files,通过 ref_kind/ref_source 区分 mention 与 upload。 [VERIFIED: submit-files.ts, hooks.ts, uploads_middleware.py]
When to use: 所有消息级文件上下文(上传/引用)都应遵循。 [VERIFIED: current implementation]
Example:
// Source: frontend/src/core/threads/submit-files.ts
referenceFiles.push({
filename: reference.filename,
size: reference.size ?? 0,
path: reference.path,
status: "uploaded",
ref_kind: "mention",
ref_source: reference.ref_source,
});
Pattern 2: 输入区预览复用 PromptInputAttachment
What: 引用预览与上传附件预览统一使用同一渲染组件。 [VERIFIED: input-box.tsx + prompt-input.tsx]
When to use: 输入区顶部预览条(包含图片缩略图和删除动作)。 [VERIFIED: current UI structure]
Example:
// Source: frontend/src/components/workspace/input-box.tsx
<PromptInputAttachment
data={{ type: "file", id: `reference:${reference.ref_source}:${reference.path ?? reference.filename}`, filename, mediaType, url }}
onRemove={() => onRemoveReference(reference)}
/>
Pattern 3: 双层 memory 清理
What: 入队前去标签 + 落盘前清句子/事实。 [VERIFIED: memory_middleware.py, updater.py]
When to use: 任何会把会话瞬时文件路径写入上下文的中间件链路。 [VERIFIED: existing middleware design]
Example:
# Source: backend/packages/harness/deerflow/agents/middlewares/memory_middleware.py
stripped = _UPLOAD_BLOCK_RE.sub("", content_str).strip()
Anti-Patterns to Avoid
- 再开并行字段(如
mentions): 会破坏既有additional_kwargs.files消费链。 [VERIFIED:hooks.ts,message-list-item.tsx] - mention 进入
new_files: 会把引用误判为本次上传,污染<uploaded_files>。 [VERIFIED:uploads_middleware.pytests] - E2E 依赖不存在 testid:
reference-chip-remove当前无实现,导致回归假红。 [VERIFIED: grepreference-chip-removeonly in test files]
Don't Hand-Roll
| Problem | Don't Build | Use Instead | Why |
|---|---|---|---|
| mention 候选浮层 | 自定义定位/焦点层 | DropdownMenu* 组件族 |
避免键盘焦点与收起时机出现分叉。 [VERIFIED: input-box.tsx] |
| 引用缩略预览 | 新写一套 chip/thumbnail | PromptInputAttachment |
已含图片/文件两类渲染与 remove 交互。 [VERIFIED: prompt-input.tsx] |
| memory 上传清理 | 单点字符串替换 | memory_middleware + updater 双层过滤 |
一层漏掉仍可在另一层兜底。 [VERIFIED: code + test_memory_upload_filtering.py] |
Key insight: Phase 07 的价值在“收口”,不是“扩面”。任何新造轮子都会重新引入 Phase 06 已解决的不一致。 [VERIFIED: phase artifacts + current code]
Common Pitfalls
Pitfall 1: 测试选择器漂移导致误判回归
What goes wrong: E2E 断言 reference-chip-remove 失败,但功能未必失效。 [VERIFIED: test run output]
Why it happens: 预览组件复用后删除按钮 testid 未对齐旧用例。 [VERIFIED: grep results]
How to avoid: 在复用组件上补稳定选择器,或更新用例改查 aria-label。 [ASSUMED]
Warning signs: DF-INPUT-007 单点失败且 reference-chip 仍可见。 [VERIFIED: test run output]
Pitfall 2: mention/upload 语义回退
What goes wrong: mention 被算成 uploaded_files。 [VERIFIED: historical issue + tests]
Why it happens: _files_from_kwargs 未过滤 ref_kind=mention。 [VERIFIED: uploads_middleware.py]
How to avoid: 保持过滤并用 mixed-list 测试守护。 [VERIFIED: test_uploads_middleware_core_logic.py]
Warning signs: <uploaded_files> 出现 source=mention 的条目。 [VERIFIED: middleware behavior]
Pitfall 3: 会话瞬时文件路径被写入长期 memory
What goes wrong: 后续会话反复检索不存在的旧路径。 [VERIFIED: updater.py docstring/comments]
Why it happens: 上传标签/句子未在 memory pipeline 剥离。 [VERIFIED: memory_middleware.py, updater.py]
How to avoid: 保留双层清理并跑 test_memory_upload_filtering.py。 [VERIFIED: test pass]
Warning signs: memory facts 出现 /mnt/user-data/uploads/。 [VERIFIED: regex intent]
Code Examples
mention 与 upload 分流(后端)
# Source: backend/packages/harness/deerflow/agents/middlewares/uploads_middleware.py
if f.get("ref_kind") == "mention":
continue
构建单一 files envelope(前端)
// Source: frontend/src/core/threads/hooks.ts
const { files: filesForSubmit, staleCount } = buildFilesForSubmit(
uploadedFileInfo,
normalizedReferences,
);
memory 标签剥离(中间件)
# Source: backend/packages/harness/deerflow/agents/middlewares/memory_middleware.py
_UPLOAD_BLOCK_RE = re.compile(
r"<(?:uploaded_files|mentioned_files)>[\\s\\S]*?</(?:uploaded_files|mentioned_files)>\\n*",
re.IGNORECASE,
)
State of the Art
| Old Approach | Current Approach | When Changed | Impact |
|---|---|---|---|
| mention 与 upload 同池处理 | ref_kind/ref_source 明确区分并分块注入 |
Phase 06 后段(2026-04-15) | 消除“引用被当上传”副作用。 [VERIFIED: git log + middleware code] |
| memory 仅靠提示词约束不记上传 | middleware + updater 双层代码过滤 | 已在当前工作树 | 减少长期 memory 污染。 [VERIFIED: memory_middleware.py, updater.py, tests] |
Deprecated/outdated:
- 仅依赖文档状态判断 Phase 06 完成度(未同步会误判)。 [VERIFIED:
06-VERIFICATION.mdvs06-UAT.md/REQUIREMENTS.md状态差异]
Assumptions Log
| # | Claim | Section | Risk if Wrong |
|---|---|---|---|
| A1 | 通过补 data-testid 或改为 aria 断言即可稳定 DF-INPUT-007 |
Common Pitfalls | 可能需要更深层 UI 结构调整。 |
Open Questions
-
Phase 07 是否要“改代码”还是“仅归档文档+测试修正”?
- What we know: 语义与 memory 主链路代码已到位。 [VERIFIED: code + tests]
- What's unclear: 你是否接受只修测试契约与文档闭环,不再动功能实现。
- Recommendation: 先锁定“最小变更原则”,避免 Phase 07 再引入行为漂移。 [ASSUMED]
-
E2E 断言口径是否改为可访问性语义?
- What we know:
reference-chip-removetestid 当前缺失。 [VERIFIED: grep + test output] - What's unclear: 团队更偏好稳定 testid 还是 aria 文案断言。
- Recommendation: 若追求跨重构稳定,优先 aria;若追求低改动,补 testid。 [ASSUMED]
- What we know:
Environment Availability
| Dependency | Required By | Available | Version | Fallback |
|---|---|---|---|---|
| Node.js | frontend tests/tooling | ✓ | v24.14.0 | — |
| pnpm | frontend scripts | ✓ | 10.32.1 | npm(不推荐,lockfile 不一致) |
| Playwright CLI | DF-INPUT E2E | ✓ | 1.48.0 | — |
| Python | backend tests | ✓ | 3.12.3 | — |
| uv | backend test runner | ✓ | 0.10.10 | — |
| pytest (global) | backend tests | ✗ | — | uv run pytest |
[VERIFIED: local command checks]
Missing dependencies with no fallback:
- None. [VERIFIED: local checks]
Missing dependencies with fallback:
- 全局
pytest缺失;使用uv run pytest。 [VERIFIED: local checks + successful runs]
Validation Architecture
Test Framework
| Property | Value |
|---|---|
| Framework | Node test runner + Playwright + pytest (via uv) |
| Config file | frontend/playwright.config.ts, backend/pyproject.toml |
| Quick run command | cd frontend && node --test src/core/threads/hooks.test.ts |
| Full suite command | `cd backend && uv run pytest -q tests/test_uploads_middleware_core_logic.py tests/test_memory_upload_filtering.py && cd ../frontend && pnpm -s test:e2e --grep "DF-INPUT-007 |
[VERIFIED: codebase files + executed commands]
Phase Requirements → Test Map
| Req ID | Behavior | Test Type | Automated Command | File Exists? |
|---|---|---|---|---|
| P7-SEM-01 | mention 不计入 new upload | unit | cd backend && uv run pytest -q tests/test_uploads_middleware_core_logic.py -k "mention or files_from_kwargs" |
✅ |
| P7-MEM-01 | memory 不保留上传事件 | unit | cd backend && uv run pytest -q tests/test_memory_upload_filtering.py |
✅ |
| P7-UI-01 | @候选/引用 chip 交互稳定 | e2e | `cd frontend && pnpm -s test:e2e --grep "DF-INPUT-007 | DF-INPUT-008 |
| P7-DOC-01 | 验收状态文档闭环 | docs check | `rg -n "ATREF-01 | ATREF-02 |
Sampling Rate
- Per task commit: 对应最小命令(前端 unit 或后端 targeted pytest)。 [VERIFIED: commit guide + current tests]
- Per wave merge: 跑后端双测 + 前端三条 E2E。 [VERIFIED: current phase scope]
- Phase gate: 三类测试全绿且文档状态同步后再进入 verify-work。 [VERIFIED: verification gaps]
Wave 0 Gaps
frontend/tests/e2e/input-and-compose.spec.ts与组件选择器合同未对齐(reference-chip-remove)。 [VERIFIED: test failure + grep].planning/phases/06-/06-UAT.md状态未回写到最新结果。 [VERIFIED: file content].planning/REQUIREMENTS.md中ATREF-01..04仍 Pending。 [VERIFIED: file content]
Security Domain
Applicable ASVS Categories
| ASVS Category | Applies | Standard Control |
|---|---|---|
| V2 Authentication | no | 本 phase 不新增 auth 面。 [VERIFIED: scope] |
| V3 Session Management | no | 不改会话机制。 [VERIFIED: scope] |
| V4 Access Control | yes | mention 候选限定当前 thread 数据源。 [VERIFIED: input-box.tsx + phase docs] |
| V5 Input Validation | yes | 后端 _files_from_kwargs 校验 filename/path。 [VERIFIED: uploads_middleware.py] |
| V6 Cryptography | no | 无加密实现变更。 [VERIFIED: scope] |
Known Threat Patterns for this phase stack
| Pattern | STRIDE | Standard Mitigation |
|---|---|---|
| 跨线程文件引用泄露 | Information Disclosure | 候选仅取当前 thread artifacts/uploads。 [VERIFIED: input-box.tsx] |
伪造 additional_kwargs.files 注入 |
Tampering | 后端校验 basename 与 /mnt/user-data/ 前缀。 [VERIFIED: uploads_middleware.py] |
| memory 泄露临时路径 | Information Disclosure | middleware + updater 双层过滤上传标签与句子。 [VERIFIED: memory code + tests] |
Sources
Primary (HIGH confidence)
- 本仓库代码:
frontend/src/components/workspace/input-box.tsx、frontend/src/components/ai-elements/prompt-input.tsx、frontend/src/core/threads/hooks.ts、frontend/src/core/threads/submit-files.ts。 [VERIFIED: codebase grep] - 本仓库代码:
backend/packages/harness/deerflow/agents/middlewares/uploads_middleware.py、memory_middleware.py、memory/updater.py。 [VERIFIED: codebase grep] - 本地执行结果:
node --test,uv run pytest,pnpm test:e2e --grep ...。 [VERIFIED: command output] - npm registry:
@radix-ui/react-dropdown-menu、sonner版本与发布时间。 [VERIFIED: npm view]
Secondary (MEDIUM confidence)
.planning/phases/06-/06-VERIFICATION.md、06-UAT.md、06-VALIDATION.md、.planning/REQUIREMENTS.md的状态交叉对比。 [VERIFIED: local docs]
Tertiary (LOW confidence)
- None.
Metadata
Confidence breakdown:
- Standard stack: HIGH - 基于当前仓库依赖与 npm registry 实查。
- Architecture: HIGH - 关键链路均有代码与测试证据。
- Pitfalls: MEDIUM - 一部分为当前失败现象,一部分为经验性防回退建议。
Research date: 2026-04-15
Valid until: 2026-05-15(30 天)