From 60223d6c7100873bdf6580e093c6d872c9306ede Mon Sep 17 00:00:00 2001
From: wangzhenjia <wangzhenjia@ehz.cn>
Date: Tue, 22 Sep 2026 11:41:07 +0800
Subject: [PATCH] =?UTF-8?q?fix(trace):=20=E8=AF=84=E4=BC=B0=E8=BF=90?=
 =?UTF-8?q?=E8=A1=8C=E7=9A=84=20trace=20input=20=E4=B8=8D=E5=86=8D?=
 =?UTF-8?q?=E5=9B=9E=E9=80=80=E6=88=90=20answer?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

`projectTrace()` 组装 `FolioTrace.input` 时：

```ts
input: run?.input ?? evaluationRun?.answer ?? ''
```

`EvaluationRun` 没有 `input` 字段（它的输入在 benchmark case 的
`input.prompt` 里），于是评估类 trace 会把 agent 的**输出**当成**输入**
投影出去。TraceInspector 里 "Input" 一行因此显示成回答文本，而且永远
落不到 `t('trace.notRecorded')` 的兜底。

这与 trace-projection 的既定纪律直接冲突——文件头写明"NEVER reconstruct
historical context from current app state … anything else is not-recorded"，
且同文件 `contextFromCase()` 的注释明确"benchmark case input —
authoritative (spec §9)"。

修复：改为从权威来源取输入

```ts
input: run?.input ?? evaluationCase?.input.prompt ?? ''
```

- 会话运行：仍用 `run.input`（未变）
- 评估运行：用 case 的 `prompt`
- 两者都没有：空串，UI 显示"未记录"

范围：仅 `projectTrace()` 的 `input` 字段赋值，其余投影逻辑未动。
有可见 UI 变化：评估类 trace 的 Input 行从"回答文本"变为"case prompt"
（无 case 时显示"未记录"）。

测试：`packages/core/src/trace-projection.test.ts` 新增 2 个用例
（用 case prompt；无权威输入时为空串）。旧代码下 11 pass / 2 fail，
修复后 51 pass / 0 fail（core 全量）。
---
 packages/core/src/trace-projection.test.ts | 45 ++++++++++++++++++++++
 packages/core/src/trace-projection.ts      |  6 ++-
 2 files changed, 50 insertions(+), 1 deletion(-)

diff --git a/packages/core/src/trace-projection.test.ts b/packages/core/src/trace-projection.test.ts
index 4a34d385..d27ee39a 100644
--- a/packages/core/src/trace-projection.test.ts
+++ b/packages/core/src/trace-projection.test.ts
@@ -224,4 +224,49 @@ describe('projectTrace (V9.1)', () => {
     const toolStep = trace.steps.find((s) => s.kind === 'tool' && s.tool?.id === 'live-run');
     expect(toolStep?.status).toBe('running');
   });
+
+  it('projects the benchmark case prompt as the trace input, never the answer', () => {
+    const caseDef: EvaluationCase = {
+      id: 'case-1',
+      name: 'NVDA research',
+      category: 'tool-selection',
+      difficulty: 'golden',
+      input: { prompt: 'Research NVDA' },
+      expected: {},
+      tags: [],
+      source: 'hand-authored',
+    };
+    const evalRun: EvaluationRun = {
+      id: 'eval-run-1',
+      experimentId: 'exp-1',
+      caseId: 'case-1',
+      datasetId: 'ds-1',
+      status: 'completed',
+      startedAt: 1000,
+      completedAt: 9000,
+      answer: 'NVDA looks fine',
+      toolCalls: [],
+      failureModes: [],
+    };
+    const trace = projectTrace({ evaluationRun: evalRun, evaluationCase: caseDef });
+    expect(trace.input).toBe('Research NVDA');
+    expect(trace.input).not.toBe(evalRun.answer);
+  });
+
+  it('leaves the trace input empty when no authoritative prompt was recorded', () => {
+    const evalRun: EvaluationRun = {
+      id: 'eval-run-2',
+      experimentId: 'exp-1',
+      caseId: 'case-2',
+      datasetId: 'ds-1',
+      status: 'completed',
+      startedAt: 1000,
+      completedAt: 9000,
+      answer: 'NVDA looks fine',
+      toolCalls: [],
+      failureModes: [],
+    };
+    const trace = projectTrace({ evaluationRun: evalRun });
+    expect(trace.input).toBe('');
+  });
 });
diff --git a/packages/core/src/trace-projection.ts b/packages/core/src/trace-projection.ts
index 6997bbab..10bc573c 100644
--- a/packages/core/src/trace-projection.ts
+++ b/packages/core/src/trace-projection.ts
@@ -291,7 +291,11 @@ export function projectTrace(input: TraceProjectionInput): FolioTrace {
     startedAt,
     completedAt,
     latencyMs,
-    input: run?.input ?? evaluationRun?.answer ?? '',
+    // The evaluation run record carries no input of its own; the benchmark
+    // case prompt is the authoritative input for that run (spec §9). Falling
+    // back to `answer` would project the OUTPUT as the input, which is the
+    // exact reconstruction-from-run-state the projection must never do.
+    input: run?.input ?? evaluationCase?.input.prompt ?? '',
     answer,
     error: run?.error?.message ?? evaluationRun?.error?.message,
     completeness: deriveCompleteness(Boolean(run || evaluationRun), tools, traceEvents, traceRefForOutput),