fix(tools): 接上 tool_execution 三事件，并改用 id 关联工具卡
客户端此前**只映射了 `tool_execution.update`**——恰好是后端**没有生产者**的那一帧
（后端回执亦确认：`.start`/`.end` 有真实生产者，`.update` 运行期可能永不来）。
真正会发出的 `.start` / `.end` 落进 `else -> null` 被**静默丢弃**，所以
**工具执行状态从未到过界面**：工具卡能显示调用与参数，却永远不显示「执行中 / 完成 / 失败」。
后端 v0.50 已把 `tool_execution` 拆成三帧、载荷为 `{tool, phase, id?}`——刻意**不带
`index` / `streamId`**（WS bridge 注释明写「tool_execution.* 不属于某一次模型流」），
关联键是 **`id`**（与同轮 `tool_call.*` 同一个 provider 调用 id）。所以旧实现按 `index`
归属本身就是错的。
- data/ws/WsEvent.kt：三个事件类型承载 `tool` / `phase` / 可空 `id` / `delta`（仅 update）；
新增私有 `ToolExecutionData`（**无 `index`**）；三个帧名都进 `dispatchStreaming`
- ui/chat/ChatViewModel.kt：`ToolPhase { Running, Done, Error }` + `toolPhaseOf`；
`ToolCallState` 扩 `tool` 与可空 `phase`；新增两个纯函数
`toolExecutionTargetIndex` / `applyToolExecution`；三帧经既有 `isForSession` 会话闸
- ui/chat/StreamingBlocks.kt：工具卡标题与配色由 `phase` 驱动（`执行 <tool>…` /
`工具调用 · <tool>` / `工具调用失败 · <tool>`，失败用 error 色）
关联策略是「**不确定就不挂**」，绝不默认 0：`id` 非空 → 精确匹配，找不到返回 -1
（宁可不挂也不挂错）；`id` 缺失 → 仅当「恰好一张卡 id 为空」或「整个列表只有一张卡」
才挂，否则 -1。`applyToolExecution` 不建卡、不动别的卡；`tool` 空白保留已有名字、
`phase` 无法识别保留已有阶段、`delta` 为空不改输出。`.update` 那条路也实现并测了。
测试：218 全绿（194 → +24）。新增 10 条三帧解码（按真实 `toolExecutionData` 形状：
id 有/无、phase running/done/error、update 带/不带 delta、多余 `streamId` 被忽略、
旧单帧形状可容忍、会话闸）+ 14 条关联规则与 phase/name/delta 应用。
实机：只读 WS 探针抓到真实帧
`tool_execution.start {"id":"call_00_o9Q…","phase":"running","tool":"calculator"}` 与
`.end {…,"phase":"done","tool":"calculator"}`，`tool_call.*` 共享同一 id、执行帧确实
没有 `index`/`streamId`；屏幕录制抓到渲染出的「工具调用 · calculator」——工具名只可能
来自 `tool_execution.*`，故端到端成立。全程只见 `.start`/`.end`，印证 `.update` 无生产者。
未覆盖：`phase:"running"`（内置工具太快）与 `phase:"error"`（无法诱发）的瞬时态，
仅单测覆盖。
This commit is contained in:
AskaEth committed
1 parent
d12a920096
commit
4c14028c70
6 files changed
+522
-18
No files matched your search
@@ -0,0 +1,179 @@
|
||||
package top.yeij.cyrene.ui.chat
|
||||
|
||||
import org.junit.Assert.assertEquals
|
||||
import org.junit.Assert.assertNull
|
||||
import org.junit.Test
|
||||
|
||||
/**
|
||||
* 工具执行帧与工具卡的**关联规则**（契约 v0.50）：
|
||||
*
|
||||
* - 三帧 `tool_execution.*` **不带 `index`**（不属于模型流），关联键是 provider 的 tool call id；
|
||||
* - `id` 有值：按 id 精确匹配，匹配不上**不落位**（宁可丢状态，也不错挂）；
|
||||
* - `id` 缺席（provider 没给，后端整个省略该键）：只在**无歧义**时落位——唯一一张 `id` 为空的卡，
|
||||
* 或整轮只有一张卡；多张卡无法区分时**不落位**，绝不默认 `index == 0`。
|
||||
*
|
||||
* [toolExecutionTargetIndex] 是规则本体；[applyToolExecution] 是 VM 实际消费它的纯函数，
|
||||
* 这里一并验证 phase / 工具名 / delta 的落法。
|
||||
*/
|
||||
class ToolExecutionCorrelationTest {
|
||||
|
||||
private fun tool(
|
||||
id: String,
|
||||
index: Int = 0,
|
||||
tool: String = "",
|
||||
arguments: String = "",
|
||||
output: String = "",
|
||||
done: Boolean = false,
|
||||
phase: ToolPhase? = null,
|
||||
) = ToolCallState(
|
||||
id = id,
|
||||
index = index,
|
||||
tool = tool,
|
||||
arguments = arguments,
|
||||
output = output,
|
||||
done = done,
|
||||
phase = phase,
|
||||
)
|
||||
|
||||
// —— id 匹配 ——
|
||||
|
||||
@Test
|
||||
fun `a frame matches the tool card carrying the same id`() {
|
||||
val tools = listOf(tool("call_1"), tool("call_2"), tool("call_3"))
|
||||
|
||||
assertEquals(1, toolExecutionTargetIndex(tools, "call_2"))
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `a frame whose id matches no tool call is not attributed`() {
|
||||
val tools = listOf(tool("call_1"), tool("call_2"))
|
||||
|
||||
assertEquals(-1, toolExecutionTargetIndex(tools, "call_9"))
|
||||
assertEquals(-1, toolExecutionTargetIndex(emptyList(), "call_1"))
|
||||
}
|
||||
|
||||
// —— id 缺席的回落 ——
|
||||
|
||||
@Test
|
||||
fun `an id-less frame targets the only anonymous card`() {
|
||||
// provider 没给 id 时 tool_call.* 也带空 id，客户端那张卡的 id 就是空串。
|
||||
val tools = listOf(tool("call_1"), tool(""), tool("call_3"))
|
||||
|
||||
assertEquals(1, toolExecutionTargetIndex(tools, null))
|
||||
assertEquals(1, toolExecutionTargetIndex(tools, ""))
|
||||
assertEquals(1, toolExecutionTargetIndex(tools, " "))
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `an id-less frame targets the single card of the turn`() {
|
||||
// 这一轮只有一次工具调用：id 缺席也不会认错。
|
||||
val tools = listOf(tool("call_1"))
|
||||
|
||||
assertEquals(0, toolExecutionTargetIndex(tools, null))
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `an id-less frame is not attributed when several cards are present`() {
|
||||
// 两张卡都带真实 id：无法判定这一帧属于谁 → 不落位，绝不默认 0。
|
||||
val tools = listOf(tool("call_1"), tool("call_2"))
|
||||
|
||||
assertEquals(-1, toolExecutionTargetIndex(tools, null))
|
||||
// 两个匿名卡（provider 全程没给 id）同样无歧义 → 不落位。
|
||||
assertEquals(-1, toolExecutionTargetIndex(listOf(tool(""), tool("")), null))
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `an id-less frame on an empty list is not attributed`() {
|
||||
assertEquals(-1, toolExecutionTargetIndex(emptyList(), null))
|
||||
}
|
||||
|
||||
// —— 应用帧（VM 消费的纯函数） ——
|
||||
|
||||
@Test
|
||||
fun `running done and error phases drive the card state`() {
|
||||
val initial = listOf(tool("call_1"))
|
||||
|
||||
val running = applyToolExecution(initial, "call_1", "bash", "running")
|
||||
assertEquals(ToolPhase.Running, running.single().phase)
|
||||
|
||||
val done = applyToolExecution(running, "call_1", "bash", "done")
|
||||
assertEquals(ToolPhase.Done, done.single().phase)
|
||||
|
||||
val error = applyToolExecution(done, "call_1", "bash", "error")
|
||||
assertEquals(ToolPhase.Error, error.single().phase)
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `an update frame appends the delta and keeps the tool name`() {
|
||||
val tools = listOf(tool("call_1", tool = "bash"))
|
||||
|
||||
val once = applyToolExecution(tools, "call_1", "bash", "running", "line 1\n")
|
||||
val twice = applyToolExecution(once, "call_1", "bash", "running", "line 2\n")
|
||||
|
||||
assertEquals("line 1\nline 2\n", twice.single().output)
|
||||
assertEquals("bash", twice.single().tool)
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `the execution frame fills in the tool name the call frames do not carry`() {
|
||||
val tools = listOf(tool("call_1"))
|
||||
|
||||
val named = applyToolExecution(tools, "call_1", "read_file", "running")
|
||||
|
||||
assertEquals("read_file", named.single().tool)
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `an empty tool name or blank delta leaves existing state untouched`() {
|
||||
val tools = listOf(tool("call_1", tool = "bash", output = "already"))
|
||||
|
||||
val same = applyToolExecution(tools, "call_1", "", "running", "")
|
||||
val nullDelta = applyToolExecution(tools, "call_1", "", "running", null)
|
||||
|
||||
assertEquals("bash", same.single().tool)
|
||||
assertEquals("already", same.single().output)
|
||||
assertEquals("already", nullDelta.single().output)
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `an unrecognised phase keeps the previous phase`() {
|
||||
val tools = listOf(tool("call_1", phase = ToolPhase.Running))
|
||||
|
||||
val same = applyToolExecution(tools, "call_1", "bash", "paused")
|
||||
|
||||
assertEquals(ToolPhase.Running, same.single().phase)
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `other cards are never touched`() {
|
||||
val tools = listOf(tool("call_1", output = "a"), tool("call_2", output = "b"))
|
||||
|
||||
val after = applyToolExecution(tools, "call_2", "bash", "done", "!")
|
||||
|
||||
assertEquals("a", after[0].output)
|
||||
assertNull(after[0].phase)
|
||||
assertEquals("b!", after[1].output)
|
||||
assertEquals(ToolPhase.Done, after[1].phase)
|
||||
}
|
||||
|
||||
@Test
|
||||
fun `an unattributable frame returns the list unchanged`() {
|
||||
val tools = listOf(tool("call_1"), tool("call_2"))
|
||||
|
||||
// id 有值但不认识 → 不改任何卡。
|
||||
assertEquals(tools, applyToolExecution(tools, "call_9", "bash", "done"))
|
||||
// id 缺席且有多张卡 → 不改任何卡。
|
||||
assertEquals(tools, applyToolExecution(tools, null, "bash", "done", "x"))
|
||||
}
|
||||
|
||||
// —— phase 字符串映射 ——
|
||||
|
||||
@Test
|
||||
fun `phase strings map to the backend tool node enum`() {
|
||||
assertEquals(ToolPhase.Running, toolPhaseOf("running"))
|
||||
assertEquals(ToolPhase.Done, toolPhaseOf("done"))
|
||||
assertEquals(ToolPhase.Error, toolPhaseOf("error"))
|
||||
assertNull(toolPhaseOf("paused"))
|
||||
assertNull(toolPhaseOf(""))
|
||||
}
|
||||
}
|
||||
Reference in new issue
Block a user