[Refactor] Modular pipeline architecture

docs/workflow.md

@@ -1,23 +1,85 @@
+# 单步模式工作流
+
 ```mermaid
 flowchart TD
-    A["ta-kickoff"]
-    A --> B["Phase 0: Initialize"]
-    B --> C["Phase 1: Detect Commits"]
-    C -->|no new commits| D["Done: Already Up-to-Date"]
-    C -->|new commits found| E["Phase 2A: Merge Upstream"]
-    E -->|conflicts| F["Phase 2B: AI Resolve Conflicts"]
-    E -->|no conflicts| G["Phase 2C: Build and Test"]
-    F -->|resolved| G
-    F -->|max retries| H["Failure: write FAILURE.md"]
-    G --> G1["Build: setup.py install"]
-    G1 -->|failed| I["AI Fix Code"]
-    G1 -->|passed| G2["Test: pytest"]
-    G2 -->|failed| I
-    G2 -->|passed| J["Commit Fixes"]
-    I -->|retry| G1
-    I -->|max retries| H
-    J --> K["Phase 2D: Finalize"]
-    K --> L["Success"]
-    L -->|push enabled| M["Push Branch and Create PR"]
-    L -->|push disabled| N["Done: work branch kept"]
-```
+    A["ta-kickoff"] --> B["Phase 0: Prepare"]
+    B --> C["Phase 1: Detect"]
+    C -->|无新 commit| D["Done"]
+    C -->|有新 commit| E["Phase 2: Plan"]
+    E --> F["Phase 3: Per-Step Loop"]
+
+    subgraph STEP["每个步骤"]
+        F1["Merge"] --> F2{冲突?}
+        F2 -->|是| F3["AI Resolve"]
+        F2 -->|否| F4{LLVM 变更?}
+        F3 --> F4
+        F4 -->|是| F5["IR Patch Pipeline"]
+        F4 -->|否| F6["Build + Fix"]
+        F5 --> F7{通过?}
+        F6 --> F7
+        F7 -->|否| FAIL["UpgradeFailed"]
+        F7 -->|是| F8["Test + Fix"]
+        F8 -->|失败| FAIL
+        F8 -->|通过| F9["Commit"]
+        F9 --> F10["current_step += 1"]
+    end
+
+    STEP --> G["Phase 4: Finalize"]
+    G -->|push| PR["Create PR"]
+
+    style FAIL fill:#d73
+    style D fill:#4a9
+```
+
+## IR Patch Pipeline（LLVM hash 变更时）
+
+```mermaid
+flowchart TD
+    MERGE["Merge 后发现 LLVM hash 变更"] --> P1["1. 切换到目标 LLVM commit<br/>确保工作区干净"]
+    P1 --> P2["2. 直接应用现有补丁<br/>llvm_patch_f6ded0b.patch"]
+    P2 --> P3{应用成功?}
+    P3 -->|否| P4["AI 分析失败原因<br/>根据当前 LLVM commit 调整补丁"]
+    P4 --> P2
+    P3 -->|是| P5["3. 编译 LLVM<br/>(编译失败则 AI 修复补丁)"]
+    P5 --> P6{编译成功?}
+    P6 -->|否| P7["AI 根据编译错误修复补丁"]
+    P7 --> P5
+    P6 -->|是| P8["4. 编译 TA + AI 修复编译错误"]
+    P8 --> P9["5. Run pytest"]
+    P9 --> P10{测试通过?}
+    P10 -->|是| DONE["Done"]
+    P10 -->|否| P11["AI 分类: IR 问题 vs 代码问题"]
+    P11 -->|IR 问题| P12["AI 在现有补丁基础上<br/>补充缺失 OP 的适配<br/>(不重新生成)"]
+    P12 --> P13["重新编译 LLVM"]
+    P13 --> P8
+    P11 -->|代码问题| P14["AI 修复代码 -> Rebuild"]
+    P14 --> P9
+
+    style DONE fill:#4a9
+```
+
+## 关键变化（vs 旧流程）
+
+| 旧流程 | 新流程 |
+|--------|--------|
+| 先做 OP 分析 + 变更分析 | 直接应用现有补丁 |
+| AI 从零生成补丁 | AI 在现有补丁上调整/补充 |
+| 分析阶段在补丁之前 | 分析阶段在测试发现问题后 |
+| 每次重新生成完整补丁 | 保留已有内容，只补充缺失 OP |
+
+## 修复代码评价机制
+
+```
+AI 修复
+  |
+Layer 1: AI 自检 (prompt.md step 6)
+  |-- 检查文件路径 -> 不在允许目录则自行回退
+  |
+Layer 2: 代码校验 (validate_fix)
+  |-- 硬检查 -> 不通过则 git revert + 反馈 + 不消耗 attempt
+  |
+Layer 3: 实际验证
+  |-- 编译/测试结果 -> 失败则继续 fix loop
+```
+
+详见 [fix-validation-flow.md](fix-validation-flow.md)

src/TA_main2main_workflow/agent/prompt.md

@@ -745,6 +745,46 @@ The active mode is: {mode}
   Output ONLY by modifying `{ascend_patch_file}` in-place — the existing
   patch file. Do NOT write analysis.md, step_summary.md, or review.md.
 
+── ir_fix_patch_apply mode ──────────────────────────────────────────────
+
+  Trigger: {mode} is "ir_fix_patch_apply" (fix patch that failed to apply).
+
+  The existing `{ascend_patch_file}` failed to apply to LLVM at
+  `{target_llvm_hash}`.  Analyze the error and adjust the patch:
+
+    1. Read the current patch file: `{ascend_patch_file}`
+    2. Read the apply error: {patch_error_msg}
+    3. For each failing hunk, use git show to see the target file:
+       git -C {llvm_project_path} show {target_llvm_hash}:<path>
+    4. Adjust the patch to match the target LLVM's actual code:
+       - Fix line numbers and context lines
+       - If the OP/func being patched was removed, remove that hunk
+       - If the API changed, adapt the patch code
+    5. Write the fixed patch to `{ascend_patch_file}`
+
+  Reference: {reference_dir}/05-ir-patch-generation-guide.md
+
+── ir_supplement_patch mode ────────────────────────────────────────────
+
+  Trigger: {mode} is "ir_supplement_patch" (add missing OPs to existing patch).
+
+  Tests revealed IR compatibility issues that the current patch does NOT
+  cover.  The current patch at `{ascend_patch_file}` already handles some
+  OPs.  You must SUPPLEMENT it — keep all existing content and ADD
+  adaptations for the missing OPs identified in the diagnosis.
+
+    1. Read the current patch to understand what's already covered
+    2. Read `{step_dir}/ir_diagnosis.json` for the IR issues found
+    3. For each missing OP, check its definition at the target LLVM:
+       git -C {llvm_project_path} show {target_llvm_hash}:<path>
+    4. ADD the new OP adaptations to the patch file (do NOT remove
+       existing content)
+    5. Apply the same patch patterns (per the guide) to the new OPs
+    6. Write the supplemented patch to `{ascend_patch_file}`
+
+  Reference: {reference_dir}/05-ir-patch-generation-guide.md
+            {reference_dir}/04-ir-compatibility-and-backend-adaptation.md
+
 ── For ir_diagnose mode ──
   Output ONLY `{step_dir}/ir_diagnosis.json` — the structured JSON specified
   in the ir_diagnose section above. Do NOT write analysis.md, step_summary.md,