Compare commits
10 Commits
53a45ad7c4
...
4baa8e7a6b
| Author | SHA1 | Date | |
|---|---|---|---|
|
4baa8e7a6b | ||
|
|
b042b45ac7 | ||
|
|
3a8da51936 | ||
| 56612389ae | |||
|
|
141ffda1ee | ||
|
|
303f6d83e3 | ||
|
|
cde41e32a8 | ||
|
|
cea4bd3cfb | ||
|
|
c276381d6c | ||
|
|
58f461c953 |
82
Cargo.lock
generated
82
Cargo.lock
generated
@ -19,6 +19,18 @@ dependencies = [
|
||||
"cpufeatures 0.2.17",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "ahash"
|
||||
version = "0.8.12"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "5a15f179cd60c4584b8a8c596927aadc462e27f2ca70c04e0071964a73ba7a75"
|
||||
dependencies = [
|
||||
"cfg-if",
|
||||
"once_cell",
|
||||
"version_check",
|
||||
"zerocopy",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "aho-corasick"
|
||||
version = "1.1.4"
|
||||
@ -28,6 +40,12 @@ dependencies = [
|
||||
"memchr",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "allocator-api2"
|
||||
version = "0.2.21"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "683d7910e743518b0e34f1186f92494becacb047c7b6bf616c96772180fef923"
|
||||
|
||||
[[package]]
|
||||
name = "android_system_properties"
|
||||
version = "0.1.5"
|
||||
@ -93,6 +111,12 @@ version = "1.0.102"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
|
||||
|
||||
[[package]]
|
||||
name = "arraydeque"
|
||||
version = "0.5.1"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "7d902e3d592a523def97af8f317b08ce16b7ab854c1985a0c671e6f15cebc236"
|
||||
|
||||
[[package]]
|
||||
name = "async-trait"
|
||||
version = "0.1.89"
|
||||
@ -545,7 +569,7 @@ dependencies = [
|
||||
"libc",
|
||||
"option-ext",
|
||||
"redox_users 0.5.2",
|
||||
"windows-sys 0.59.0",
|
||||
"windows-sys 0.61.2",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
@ -616,7 +640,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
|
||||
dependencies = [
|
||||
"libc",
|
||||
"windows-sys 0.59.0",
|
||||
"windows-sys 0.61.2",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
@ -844,6 +868,27 @@ dependencies = [
|
||||
"weezl",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "gray_matter"
|
||||
version = "0.2.9"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "8666976c40b8633f918783969b6681a3ddb205f29150348617de425d85a3e3bd"
|
||||
dependencies = [
|
||||
"serde",
|
||||
"serde_json",
|
||||
"yaml-rust2",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "hashbrown"
|
||||
version = "0.14.5"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "e5274423e17b7c9fc20b6e7e208532f9b19825d82dfd615708b70edd83df41f1"
|
||||
dependencies = [
|
||||
"ahash",
|
||||
"allocator-api2",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "hashbrown"
|
||||
version = "0.15.5"
|
||||
@ -868,6 +913,15 @@ version = "0.17.0"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "4f467dd6dccf739c208452f8014c75c18bb8301b050ad1cfb27153803edb0f51"
|
||||
|
||||
[[package]]
|
||||
name = "hashlink"
|
||||
version = "0.8.4"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "e8094feaf31ff591f651a2664fb9cfd92bba7a60ce3197265e9482ebe753c8f7"
|
||||
dependencies = [
|
||||
"hashbrown 0.14.5",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "hashlink"
|
||||
version = "0.11.1"
|
||||
@ -1485,7 +1539,7 @@ version = "0.50.3"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
|
||||
dependencies = [
|
||||
"windows-sys 0.59.0",
|
||||
"windows-sys 0.61.2",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
@ -1596,6 +1650,7 @@ dependencies = [
|
||||
"dotenv",
|
||||
"encoding_rs",
|
||||
"futures-util",
|
||||
"gray_matter",
|
||||
"http",
|
||||
"iana-time-zone",
|
||||
"image",
|
||||
@ -2123,7 +2178,7 @@ dependencies = [
|
||||
"bitflags",
|
||||
"fallible-iterator",
|
||||
"fallible-streaming-iterator",
|
||||
"hashlink",
|
||||
"hashlink 0.11.1",
|
||||
"libsqlite3-sys",
|
||||
"smallvec",
|
||||
"sqlite-wasm-rs",
|
||||
@ -2188,7 +2243,7 @@ dependencies = [
|
||||
"errno",
|
||||
"libc",
|
||||
"linux-raw-sys",
|
||||
"windows-sys 0.59.0",
|
||||
"windows-sys 0.61.2",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
@ -2247,7 +2302,7 @@ dependencies = [
|
||||
"security-framework",
|
||||
"security-framework-sys",
|
||||
"webpki-root-certs",
|
||||
"windows-sys 0.59.0",
|
||||
"windows-sys 0.61.2",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
@ -2656,7 +2711,7 @@ dependencies = [
|
||||
"getrandom 0.4.2",
|
||||
"once_cell",
|
||||
"rustix",
|
||||
"windows-sys 0.59.0",
|
||||
"windows-sys 0.61.2",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
@ -3347,7 +3402,7 @@ version = "0.1.11"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
|
||||
dependencies = [
|
||||
"windows-sys 0.59.0",
|
||||
"windows-sys 0.61.2",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
@ -3657,6 +3712,17 @@ version = "0.6.3"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "1ffae5123b2d3fc086436f8834ae3ab053a283cfac8fe0a0b8eaae044768a4c4"
|
||||
|
||||
[[package]]
|
||||
name = "yaml-rust2"
|
||||
version = "0.8.1"
|
||||
source = "registry+https://github.com/rust-lang/crates.io-index"
|
||||
checksum = "8902160c4e6f2fb145dbe9d6760a75e3c9522d8bf796ed7047c85919ac7115f8"
|
||||
dependencies = [
|
||||
"arraydeque",
|
||||
"encoding_rs",
|
||||
"hashlink 0.8.4",
|
||||
]
|
||||
|
||||
[[package]]
|
||||
name = "yoke"
|
||||
version = "0.8.2"
|
||||
|
||||
@ -41,6 +41,7 @@ rustls = { version = "0.23", features = ["ring"] }
|
||||
wechatbot = { path = "vendor/wechatbot" }
|
||||
encoding_rs = "0.8"
|
||||
libc = "0.2"
|
||||
gray_matter = { version = "0.2", default-features = false, features = ["yaml"] }
|
||||
|
||||
[target.'cfg(windows)'.dependencies]
|
||||
windows-sys = { version = "0.59", features = [
|
||||
|
||||
17
README.md
17
README.md
@ -498,7 +498,7 @@ tools 配置示例：
|
||||
- shell - 执行 shell 命令（Windows PowerShell/Cmd）
|
||||
- http_request - HTTP 请求
|
||||
- web_fetch - 网页抓取
|
||||
- task - 创建和管理子代理，支持内置类型（general/explore）和用户自定义类型
|
||||
- task - 创建和管理子代理，支持内置类型（general）和用户自定义类型
|
||||
|
||||
注意：bash 和 shell 是同一个工具在不同平台上的名称，运行时自动检测。
|
||||
|
||||
@ -509,7 +509,8 @@ PicoBot 支持通过 `task` 工具创建子代理来处理复杂多步骤任务
|
||||
### 8.1 内置子代理类型
|
||||
|
||||
- **general**：通用型子代理，适合处理复杂多步骤任务。可以使用读写文件、执行命令、HTTP 请求等完整工具集。
|
||||
- **explore**：探索型子代理，用于代码库探索和信息收集。只使用只读工具，禁止任何写操作。
|
||||
|
||||
> 如需只读探索型子代理，可通过自定义子代理配合 `allowed_tools` 白名单实现（见下文）。
|
||||
|
||||
### 8.2 自定义子代理
|
||||
|
||||
@ -538,7 +539,8 @@ prompt_template: |
|
||||
3. 完成后给出简洁的总结
|
||||
|
||||
注意: 你是一个只读代理，禁止执行任何修改操作。
|
||||
allowed_tools: [read, bash, web_fetch] # 可选，覆盖默认工具白名单
|
||||
allowed_tools: [read, bash, web_fetch] # 可选，工具白名单（仅这些工具可用）
|
||||
denied_tools: [task] # 可选，工具黑名单（这些工具被禁用）
|
||||
max_execution_secs: 600 # 可选，覆盖默认执行时间
|
||||
---
|
||||
|
||||
@ -555,9 +557,12 @@ max_execution_secs: 600 # 可选，覆盖默认执行时间
|
||||
| `name` | string | 否 | 子代理名称，默认取目录名 |
|
||||
| `description` | string | 是 | 简短描述，用于 agent 选择 |
|
||||
| `prompt_template` | string | 是 | 提示词模板，支持变量插值 |
|
||||
| `allowed_tools` | array | 否 | 工具白名单，不指定时使用默认列表 |
|
||||
| `allowed_tools` | array | 否 | 工具白名单，指定后仅这些工具可用；不指定则不限制 |
|
||||
| `denied_tools` | array | 否 | 工具黑名单，指定后这些工具被禁用；在白名单之后应用 |
|
||||
| `max_execution_secs` | integer | 否 | 最大执行时间（秒） |
|
||||
|
||||
> **工具过滤语义**：`allowed_tools` 与 `denied_tools` 可共存。生效顺序为：先应用白名单（取交集），再扣除黑名单。两者都不指定时，子代理使用完整工具集。当子代理嵌套深度达到 `max_nesting_depth`（默认 2，即孙代理）时，始终移除 `task` 工具以防无限嵌套。白名单中未注册的工具名会被静默跳过。
|
||||
|
||||
#### 模板变量
|
||||
|
||||
`prompt_template` 支持以下变量插值：
|
||||
@ -615,7 +620,7 @@ PicoBot 的 Agent 是围绕工具调用构建的。当前默认注册的工具
|
||||
- bash / shell：执行 shell 命令（同一工具，Unix 下名称为 bash，Windows 下名称为 shell）
|
||||
- http_request：发起 HTTP 请求
|
||||
- web_fetch：抓取网页正文
|
||||
- task：创建和管理子代理，支持内置类型（general/explore）和用户自定义类型
|
||||
- task：创建和管理子代理，支持内置类型（general）和用户自定义类型
|
||||
|
||||
其中：
|
||||
|
||||
@ -625,7 +630,7 @@ PicoBot 的 Agent 是围绕工具调用构建的。当前默认注册的工具
|
||||
- skill_activate 负责把具体技能正文注入当前任务上下文
|
||||
- skill_manage 整合了技能列出与管理功能，支持运行时创建、更新、删除和批量禁用
|
||||
- bash / shell / http_request / web_fetch 让 Agent 具备更强的外部交互能力（bash 和 shell 是同一工具在不同平台的名称）
|
||||
- task 允许 Agent 创建独立上下文的子代理来处理复杂多步骤任务，支持内置类型（general/explore）和用户自定义类型
|
||||
- task 允许 Agent 创建独立上下文的子代理来处理复杂多步骤任务，支持内置类型（general）和用户自定义类型
|
||||
|
||||
### 9.1 MCP 工具集成
|
||||
|
||||
|
||||
@ -2453,6 +2453,57 @@ mod tests {
|
||||
// Verify no tool messages remain
|
||||
assert!(messages.iter().all(|m| m.role != "tool"));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_sanitize_strips_tool_calls_when_tool_results_not_immediately_following() {
|
||||
// [assistant(tool_calls=[A]), user, tool(A)]
|
||||
// Tool result exists but is NOT immediately after assistant → strip tool_calls
|
||||
// This is the scenario that triggers DeepSeek 400 "insufficient tool messages
|
||||
// following tool_calls message" — the reverse scan sees tool(A) as resolved,
|
||||
// but the API requires it to be IMMEDIATELY after the assistant.
|
||||
let mut messages = vec![
|
||||
ChatMessage::assistant_with_tool_calls(
|
||||
"calling tool",
|
||||
vec![ToolCall {
|
||||
id: "call_A".to_string(),
|
||||
name: "search".to_string(),
|
||||
arguments: serde_json::json!({}),
|
||||
}],
|
||||
),
|
||||
ChatMessage::user("interrupting message"),
|
||||
ChatMessage::tool("call_A", "search", "result"),
|
||||
];
|
||||
|
||||
let removed = crate::bus::message::sanitize_incomplete_tool_call_sequences(&mut messages);
|
||||
// The assistant should be removed (tool_calls stripped via removal)
|
||||
// and the orphaned tool(A) should also be removed
|
||||
assert!(removed >= 2, "should remove both the assistant and orphaned tool message, got {}", removed);
|
||||
assert_eq!(messages.len(), 1, "only the user message should remain");
|
||||
assert_eq!(messages[0].role, "user");
|
||||
assert!(messages.iter().all(|m| m.tool_calls.as_ref().map_or(true, |c| c.is_empty())),
|
||||
"no assistant should have tool_calls remaining");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_sanitize_preserves_tool_calls_when_immediately_followed() {
|
||||
// [assistant(tool_calls=[A]), tool(A), user] — valid, tool result is immediate
|
||||
let mut messages = vec![
|
||||
ChatMessage::assistant_with_tool_calls(
|
||||
"calling tool",
|
||||
vec![ToolCall {
|
||||
id: "call_A".to_string(),
|
||||
name: "search".to_string(),
|
||||
arguments: serde_json::json!({}),
|
||||
}],
|
||||
),
|
||||
ChatMessage::tool("call_A", "search", "result"),
|
||||
ChatMessage::user("next message after tool result"),
|
||||
];
|
||||
|
||||
let removed = crate::bus::message::sanitize_incomplete_tool_call_sequences(&mut messages);
|
||||
assert_eq!(removed, 0, "should not remove anything — tool result immediately follows");
|
||||
assert_eq!(messages.len(), 3);
|
||||
}
|
||||
}
|
||||
|
||||
#[derive(Debug)]
|
||||
|
||||
@ -316,6 +316,86 @@ pub(crate) fn sanitize_incomplete_tool_call_sequences(messages: &mut Vec<ChatMes
|
||||
}
|
||||
}
|
||||
|
||||
// Phase 1.5: Forward-order check — verify tool messages IMMEDIATELY follow
|
||||
// the assistant(tool_calls). If any non-tool message appears between the
|
||||
// assistant and its tool results, the API rejects with
|
||||
// "insufficient tool messages following tool_calls message".
|
||||
//
|
||||
// The reverse scan in Phase 1 only checks existence (tool result appears
|
||||
// somewhere after assistant), NOT immediacy. This pass catches cases like:
|
||||
// [assistant(tool_calls=[A]), user, tool(A)]
|
||||
// ^ Phase 1 sees tool(A) after assistant → "resolved"
|
||||
// but API requires tool(A) to be IMMEDIATELY after assistant
|
||||
{
|
||||
let mut pending_tool_ids: HashSet<String> = HashSet::new();
|
||||
let mut pending_assistant_idx: Option<usize> = None;
|
||||
|
||||
for (i, m) in messages.iter().enumerate() {
|
||||
// If we have pending tool_ids and encounter a non-tool message,
|
||||
// the assistant's tool results were NOT immediately following.
|
||||
if !pending_tool_ids.is_empty() && m.role != "tool" {
|
||||
if let Some(idx) = pending_assistant_idx {
|
||||
if !remove_indices.contains(&idx) {
|
||||
tracing::warn!(
|
||||
message_index = idx,
|
||||
interrupted_by_index = i,
|
||||
interrupted_by_role = %m.role,
|
||||
pending_tool_call_count = pending_tool_ids.len(),
|
||||
"Removing assistant with tool_calls — tool results \
|
||||
not immediately following (interrupted by non-tool message)"
|
||||
);
|
||||
// Remove this assistant's tool_call_ids from with_parent
|
||||
// so Phase 2 cleans up the now-orphaned tool messages
|
||||
if let Some(calls) = messages[idx].tool_calls.as_ref() {
|
||||
for tc in calls.iter() {
|
||||
with_parent.remove(&tc.id);
|
||||
}
|
||||
}
|
||||
remove_indices.push(idx);
|
||||
}
|
||||
}
|
||||
pending_tool_ids.clear();
|
||||
pending_assistant_idx = None;
|
||||
}
|
||||
|
||||
if m.role == "assistant"
|
||||
&& m.tool_calls.as_ref().map_or(false, |calls| !calls.is_empty())
|
||||
{
|
||||
let already_marked = remove_indices.contains(&i);
|
||||
if !already_marked {
|
||||
pending_tool_ids = m.tool_calls.as_ref().unwrap()
|
||||
.iter().map(|tc| tc.id.clone()).collect();
|
||||
pending_assistant_idx = Some(i);
|
||||
}
|
||||
} else if m.role == "tool" {
|
||||
if let Some(ref tc_id) = m.tool_call_id {
|
||||
pending_tool_ids.remove(tc_id);
|
||||
if pending_tool_ids.is_empty() {
|
||||
pending_assistant_idx = None;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Handle trailing assistant with unresolved immediate tool results
|
||||
if !pending_tool_ids.is_empty() {
|
||||
if let Some(idx) = pending_assistant_idx {
|
||||
if !remove_indices.contains(&idx) {
|
||||
tracing::warn!(
|
||||
message_index = idx,
|
||||
"Removing trailing assistant with incomplete immediate tool results"
|
||||
);
|
||||
if let Some(calls) = messages[idx].tool_calls.as_ref() {
|
||||
for tc in calls.iter() {
|
||||
with_parent.remove(&tc.id);
|
||||
}
|
||||
}
|
||||
remove_indices.push(idx);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Remove in descending index order to avoid shifting
|
||||
for &idx in &remove_indices {
|
||||
messages.remove(idx);
|
||||
|
||||
@ -76,6 +76,9 @@ impl CommandHandler for ListTodosCommandHandler {
|
||||
id: r.id,
|
||||
content: r.content,
|
||||
status: r.status,
|
||||
priority: r.priority,
|
||||
created_at: r.created_at,
|
||||
updated_at: r.updated_at,
|
||||
created_by_message_id: r.created_by_message_id,
|
||||
})
|
||||
.collect();
|
||||
|
||||
@ -248,8 +248,6 @@ pub struct TaskConfig {
|
||||
pub enabled: bool,
|
||||
#[serde(default = "default_task_max_execution_secs")]
|
||||
pub max_execution_secs: u64,
|
||||
#[serde(default = "default_task_explore_max_execution_secs")]
|
||||
pub explore_max_execution_secs: u64,
|
||||
#[serde(default = "default_task_ttl_hours")]
|
||||
pub ttl_hours: u64,
|
||||
#[serde(default = "default_task_allowed_tools")]
|
||||
@ -266,10 +264,6 @@ fn default_task_max_execution_secs() -> u64 {
|
||||
3600 // 60分钟
|
||||
}
|
||||
|
||||
fn default_task_explore_max_execution_secs() -> u64 {
|
||||
3600 // 60分钟
|
||||
}
|
||||
|
||||
fn default_task_ttl_hours() -> u64 {
|
||||
24
|
||||
}
|
||||
@ -300,7 +294,6 @@ impl Default for TaskConfig {
|
||||
Self {
|
||||
enabled: default_task_enabled(),
|
||||
max_execution_secs: default_task_max_execution_secs(),
|
||||
explore_max_execution_secs: default_task_explore_max_execution_secs(),
|
||||
ttl_hours: default_task_ttl_hours(),
|
||||
allowed_tools: default_task_allowed_tools(),
|
||||
max_nesting_depth: default_task_max_nesting_depth(),
|
||||
|
||||
@ -795,11 +795,14 @@ fn load_experts_from_root(root: &Path, source: ExpertSource) -> Vec<Expert> {
|
||||
fn parse_expert_file(path: &Path, source: ExpertSource) -> Result<Expert, String> {
|
||||
let content = fs::read_to_string(path).map_err(|e| format!("failed to read file: {}", e))?;
|
||||
|
||||
let (frontmatter_raw, body) =
|
||||
split_frontmatter(&content).ok_or_else(|| "missing YAML frontmatter block".to_string())?;
|
||||
|
||||
let frontmatter: ExpertFrontmatter = serde_yaml::from_str(frontmatter_raw)
|
||||
.map_err(|e| format!("invalid YAML frontmatter: {}", e))?;
|
||||
let (frontmatter, body) = match crate::frontmatter::parse::<ExpertFrontmatter>(&content) {
|
||||
Ok(v) => v,
|
||||
Err(err) => {
|
||||
let bytes = content.len();
|
||||
let crlf = content.contains('\r');
|
||||
return Err(format!("{} (bytes={}, crlf={})", err, bytes, crlf));
|
||||
}
|
||||
};
|
||||
|
||||
let description = frontmatter.description.trim();
|
||||
if description.is_empty() {
|
||||
@ -823,26 +826,6 @@ fn parse_expert_file(path: &Path, source: ExpertSource) -> Result<Expert, String
|
||||
})
|
||||
}
|
||||
|
||||
fn split_frontmatter(content: &str) -> Option<(&str, &str)> {
|
||||
// 兼容 CRLF（Windows）和 LF（Unix）行尾符
|
||||
let rest = content
|
||||
.strip_prefix("---\n")
|
||||
.or_else(|| content.strip_prefix("---\r\n"))?;
|
||||
let marker = "\n---\n";
|
||||
let marker_crlf = "\n---\r\n";
|
||||
if let Some(idx) = rest.find(marker) {
|
||||
let frontmatter = &rest[..idx];
|
||||
let body = &rest[idx + marker.len()..];
|
||||
Some((frontmatter, body))
|
||||
} else if let Some(idx) = rest.find(marker_crlf) {
|
||||
let frontmatter = &rest[..idx];
|
||||
let body = &rest[idx + marker_crlf.len()..];
|
||||
Some((frontmatter, body))
|
||||
} else {
|
||||
None
|
||||
}
|
||||
}
|
||||
|
||||
// ========== State file I/O ==========
|
||||
|
||||
fn load_expert_disable_state(cwd: &Path) -> ExpertDisableState {
|
||||
@ -992,11 +975,17 @@ mod tests {
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_split_frontmatter() {
|
||||
let input = "---\ndescription: demo\n---\nhello";
|
||||
let (fm, body) = split_frontmatter(input).unwrap();
|
||||
assert!(fm.contains("description"));
|
||||
assert_eq!(body, "hello");
|
||||
fn test_parse_expert_file_handles_crlf_endings() {
|
||||
let dir = tempfile::tempdir().unwrap();
|
||||
let expert_dir = dir.path().join("demo");
|
||||
fs::create_dir_all(&expert_dir).unwrap();
|
||||
let expert_md = expert_dir.join("EXPERT.md");
|
||||
fs::write(&expert_md, "---\r\ndescription: demo expert\r\n---\r\nStep A\r\nStep B").unwrap();
|
||||
|
||||
let expert = parse_expert_file(&expert_md, ExpertSource::Project).unwrap();
|
||||
assert_eq!(expert.name, "demo");
|
||||
assert_eq!(expert.description, "demo expert");
|
||||
assert_eq!(expert.body, "Step A\nStep B");
|
||||
}
|
||||
|
||||
#[test]
|
||||
|
||||
86
src/frontmatter.rs
Normal file
86
src/frontmatter.rs
Normal file
@ -0,0 +1,86 @@
|
||||
use gray_matter::engine::YAML;
|
||||
use gray_matter::Matter;
|
||||
use serde::de::DeserializeOwned;
|
||||
|
||||
/// Parse a markdown document with YAML frontmatter into `(frontmatter, body)`.
|
||||
///
|
||||
/// Tolerates CRLF, CR, and LF line endings. A `---` appearing inside the body
|
||||
/// is not treated as a delimiter (only the leading frontmatter block is split).
|
||||
///
|
||||
/// - Returns `Err("missing YAML frontmatter block")` when no leading `---`
|
||||
/// delimiter is present.
|
||||
/// - Returns `Err("invalid YAML frontmatter: {e}")` when the block is present
|
||||
/// but the YAML fails to parse or deserialize into `T`.
|
||||
pub fn parse<T: DeserializeOwned>(content: &str) -> Result<(T, String), String> {
|
||||
let normalized = content.replace("\r\n", "\n").replace('\r', "\n");
|
||||
|
||||
if !normalized.starts_with("---\n") {
|
||||
return Err("missing YAML frontmatter block".to_string());
|
||||
}
|
||||
|
||||
let matter = Matter::<YAML>::new();
|
||||
let result = matter.parse(&normalized);
|
||||
let pod = result
|
||||
.data
|
||||
.ok_or_else(|| "invalid YAML frontmatter".to_string())?;
|
||||
let data: T = pod
|
||||
.deserialize()
|
||||
.map_err(|e| format!("invalid YAML frontmatter: {}", e))?;
|
||||
Ok((data, result.content))
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
use serde::Deserialize;
|
||||
|
||||
#[derive(Debug, Deserialize, PartialEq)]
|
||||
struct FrontMatter {
|
||||
description: String,
|
||||
#[serde(default)]
|
||||
name: Option<String>,
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn parses_lf_endings() {
|
||||
let input = "---\ndescription: demo\n---\nbody text";
|
||||
let (fm, body) = parse::<FrontMatter>(input).unwrap();
|
||||
assert_eq!(fm, FrontMatter { description: "demo".to_string(), name: None });
|
||||
assert_eq!(body, "body text");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn parses_crlf_endings() {
|
||||
let input = "---\r\ndescription: demo\r\n---\r\nbody text";
|
||||
let (fm, body) = parse::<FrontMatter>(input).unwrap();
|
||||
assert_eq!(fm.description, "demo");
|
||||
assert_eq!(body, "body text");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn parses_cr_endings() {
|
||||
let input = "---\rdescription: demo\r---\rbody text";
|
||||
let (fm, body) = parse::<FrontMatter>(input).unwrap();
|
||||
assert_eq!(fm.description, "demo");
|
||||
assert_eq!(body, "body text");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn missing_block_is_rejected() {
|
||||
let err = parse::<FrontMatter>("no front matter here").unwrap_err();
|
||||
assert_eq!(err, "missing YAML frontmatter block");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn invalid_yaml_is_rejected() {
|
||||
let err = parse::<FrontMatter>("---\n: : bad\n---\nbody").unwrap_err();
|
||||
assert!(err.starts_with("invalid YAML frontmatter"));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn body_with_inner_delimiter_is_preserved() {
|
||||
let input = "---\ndescription: demo\n---\nexcerpt\n---\nmore content";
|
||||
let (_fm, body) = parse::<FrontMatter>(input).unwrap();
|
||||
assert_eq!(body, "excerpt\n---\nmore content");
|
||||
}
|
||||
}
|
||||
@ -1,25 +0,0 @@
|
||||
<!--
|
||||
# 自定义 Agent 配置
|
||||
|
||||
在此文件添加您的个性化配置，这些内容会与系统默认提示词合并。
|
||||
|
||||
## 可用配置项示例
|
||||
|
||||
### 额外身份
|
||||
- 我是后端开发者
|
||||
- 我熟悉 Rust / Python
|
||||
|
||||
### 个人偏好
|
||||
- 代码注释使用英文
|
||||
- 优先使用中文回复技术概念
|
||||
|
||||
### 工作方式
|
||||
- 复杂任务请先给出整体方案
|
||||
- 性能优化建议优先
|
||||
|
||||
## 注意
|
||||
- 此文件内容会追加在系统默认提示词之后
|
||||
- 留空或删除内容则不会生效
|
||||
- 使用 HTML 注释 <!-- --> 可以保留说明文字而不生效
|
||||
- 删除本注释块并添加您的自定义配置即可生效
|
||||
-->
|
||||
@ -13,17 +13,21 @@ use super::session::Session;
|
||||
/// has already finished by this point there is no response-time impact, and
|
||||
/// the synchronous guarantee means the next execution always starts with
|
||||
/// freshly compacted history.
|
||||
///
|
||||
/// 按 topic_id 隔离：压缩只处理指定 topic 的历史，DB 替换也只影响该 topic。
|
||||
pub(crate) async fn schedule_background_history_compaction(
|
||||
session: Arc<Mutex<Session>>,
|
||||
chat_id: impl Into<String>,
|
||||
topic_id: impl Into<String>,
|
||||
) -> Result<(), AgentError> {
|
||||
let chat_id = chat_id.into();
|
||||
let topic_id = topic_id.into();
|
||||
|
||||
let mut session_guard = session.lock().await;
|
||||
session_guard.ensure_persistent_session(&chat_id)?;
|
||||
session_guard.ensure_chat_loaded(&chat_id)?;
|
||||
session_guard.ensure_chat_loaded(&chat_id, Some(&topic_id))?;
|
||||
|
||||
let history = session_guard.get_or_create_history(&chat_id).clone();
|
||||
let history = session_guard.get_or_create_history(&topic_id).clone();
|
||||
let compressor = session_guard.compressor().clone();
|
||||
|
||||
if !compressor.should_compress(&history) {
|
||||
@ -36,6 +40,7 @@ pub(crate) async fn schedule_background_history_compaction(
|
||||
|
||||
tracing::info!(
|
||||
chat_id = %chat_id,
|
||||
topic_id = %topic_id,
|
||||
msg_count = history.len(),
|
||||
"Starting synchronous two-segment compression"
|
||||
);
|
||||
@ -47,19 +52,20 @@ pub(crate) async fn schedule_background_history_compaction(
|
||||
.compress_two_segment(&history, &provider_config)
|
||||
.await?;
|
||||
|
||||
// Replace the entire history with the compressed result.
|
||||
// Since we hold the lock, no concurrent modifications can occur.
|
||||
// Replace only this topic's history in DB (not the entire session).
|
||||
// This avoids clobbering other topics' messages during compaction.
|
||||
store
|
||||
.replace_active_history(&session_id, &compressed)
|
||||
.map_err(|e| AgentError::Other(format!("replace_active_history error: {}", e)))?;
|
||||
.replace_topic_history(&session_id, &topic_id, &compressed)
|
||||
.map_err(|e| AgentError::Other(format!("replace_topic_history error: {}", e)))?;
|
||||
|
||||
tracing::info!(
|
||||
chat_id = %chat_id,
|
||||
topic_id = %topic_id,
|
||||
compressed_msg_count = compressed.len(),
|
||||
"Two-segment compression committed"
|
||||
);
|
||||
|
||||
session_guard.reload_chat_history(&chat_id)?;
|
||||
session_guard.reload_topic_history(&chat_id, &topic_id)?;
|
||||
|
||||
Ok(())
|
||||
}
|
||||
|
||||
@ -21,6 +21,7 @@
|
||||
- 复杂任务先收敛重点，简单任务直接给结果。
|
||||
- 避免不必要的重复、客套和冗长说明。
|
||||
- 调用工具的时候需要不仅仅回复工具的json，也简短说明你调用工具要完成什么工作
|
||||
- 具备良好的整理意识，同一类的工作输出放到同一个输出文件夹，而不是输出在根目录
|
||||
|
||||
## 回复规则
|
||||
|
||||
@ -32,10 +33,9 @@
|
||||
|
||||
- 回答应以帮助用户完成当前目标为中心。
|
||||
- 在信息不足时先补关键前提，在信息充分时直接执行。
|
||||
- 调用工具的时候必须同时用简短的话告诉用户你调用工具是做什么
|
||||
- 无需担心创建子智能体过多的问题，请按用户或者skill的要求创建对应数量的子智能体，这样可以隔离上下文，更好完成工作
|
||||
- 思考的时候建议用中文思考
|
||||
- 涉及到时间的都用get_time工具获取，避免时间不准确
|
||||
|
||||
## 用户附件
|
||||
|
||||
用户发过来了一些附件，先判断文件后缀名能不能直接读取，如果不能直接read的，比如xlsx,就要通过代码等其他方式读取里面的内容
|
||||
|
||||
@ -64,6 +64,8 @@ pub(crate) struct MessageExecutionRequest<'a> {
|
||||
pub(crate) content: &'a str,
|
||||
pub(crate) media: Vec<MediaItem>,
|
||||
pub(crate) live_emitter: Option<Arc<dyn EmittedMessageHandler>>,
|
||||
/// 消息接收时捕获的 topic_id，全程显式传递避免从共享状态重复读取竞态
|
||||
pub(crate) topic_id: Option<String>,
|
||||
}
|
||||
|
||||
pub(crate) struct ScheduledExecutionRequest<'a> {
|
||||
@ -89,13 +91,19 @@ impl AgentExecutionService {
|
||||
session: &mut Session,
|
||||
request: FinalizeAgentResultRequest<'_>,
|
||||
) -> Result<FinalizedAgentResult, AgentError> {
|
||||
// 检查是否是最新的用户回合
|
||||
let is_current_turn =
|
||||
session.matches_current_user_turn(request.chat_id, request.user_message);
|
||||
// 判断是否是最新的用户回合
|
||||
// 直接比较 current_topic(chat_id) 与 original_topic_id
|
||||
// 这比"检查内存历史最新消息"更可靠，且天然处理"切走又切回"的 case
|
||||
let is_current_turn = match request.original_topic_id.as_deref() {
|
||||
Some(orig_tid) => session.current_topic(request.chat_id).as_deref() == Some(orig_tid),
|
||||
None => true, // 无 topic 时总是视为当前回合
|
||||
};
|
||||
|
||||
if !is_current_turn {
|
||||
let (latest_user_id, latest_user_preview, compression_in_flight, history_len) =
|
||||
session.stale_result_diagnostics(request.chat_id);
|
||||
session.stale_result_diagnostics(
|
||||
request.original_topic_id.as_deref().unwrap_or(request.chat_id),
|
||||
);
|
||||
tracing::info!(
|
||||
channel = %request.channel_name,
|
||||
chat_id = %request.chat_id,
|
||||
@ -111,29 +119,25 @@ impl AgentExecutionService {
|
||||
}
|
||||
|
||||
// 确定保存消息的话题 ID
|
||||
// 如果是最新回合，使用当前话题；否则使用原始话题
|
||||
let target_topic_id = if is_current_turn {
|
||||
session.current_topic(request.chat_id)
|
||||
} else {
|
||||
request.original_topic_id.as_deref()
|
||||
};
|
||||
// 始终使用执行开始时捕获的 original_topic_id，避免从共享状态重复读取竞态
|
||||
let target_topic_id = request.original_topic_id.as_deref();
|
||||
|
||||
// 将结果消息保存到确定的话题
|
||||
if let Some(topic_id) = target_topic_id {
|
||||
if is_current_turn {
|
||||
// 如果是最新回合，使用 append_persisted_messages 保存到数据库并更新内存历史
|
||||
// 话题未切换（current_topic == original_topic_id），安全更新内存历史
|
||||
if let Err(err) = session.append_persisted_messages(
|
||||
request.chat_id,
|
||||
topic_id,
|
||||
request.result.emitted_messages.clone(),
|
||||
) {
|
||||
tracing::error!(
|
||||
error = %err,
|
||||
chat_id = %request.chat_id,
|
||||
topic_id = %topic_id,
|
||||
"Failed to append messages to session history"
|
||||
);
|
||||
}
|
||||
} else {
|
||||
// 如果用户已切换话题，只保存到原始话题（不更新内存历史）
|
||||
// 话题已切换，只写 DB 不更新内存（避免污染新话题的历史）
|
||||
if let Err(err) = session.append_messages_to_topic(
|
||||
request.chat_id,
|
||||
topic_id,
|
||||
@ -147,7 +151,8 @@ impl AgentExecutionService {
|
||||
}
|
||||
}
|
||||
} else if is_current_turn {
|
||||
// 如果没有话题，直接更新内存历史（append_persisted_messages 会处理持久化）
|
||||
// 没有话题：直接更新内存历史（append_persisted_messages 会处理持久化）
|
||||
// 无 topic 场景用 chat_id 作为 topic_histories 的回退 key
|
||||
if let Err(err) = session.append_persisted_messages(
|
||||
request.chat_id,
|
||||
request.result.emitted_messages.clone(),
|
||||
@ -200,11 +205,33 @@ impl AgentExecutionService {
|
||||
&self,
|
||||
request: MessageExecutionRequest<'_>,
|
||||
) -> Result<Vec<OutboundMessage>, AgentError> {
|
||||
// 获取该 topic 的串行锁（通过短暂获取 session 锁）
|
||||
// 同一 topic 的消息处理必须串行执行，防止并发 loop 操作同一历史的不同快照
|
||||
// 不同 topic 之间互不阻塞，支持多话题并发执行
|
||||
let serial_lock = {
|
||||
let mut session_guard = request.session.lock().await;
|
||||
let lock_key = request.topic_id.as_deref().unwrap_or(request.chat_id);
|
||||
session_guard.topic_serial_lock(lock_key)
|
||||
};
|
||||
|
||||
// 等待该 topic 的前一条消息处理完成（含压缩）
|
||||
// await 串行锁时不持有 session 锁，其他 topic 的消息可以正常处理
|
||||
let _serial_guard = serial_lock.lock().await;
|
||||
|
||||
let (history, agent, user_message, user_message_count, original_topic_id) = {
|
||||
let mut session_guard = request.session.lock().await;
|
||||
|
||||
session_guard.ensure_persistent_session(request.chat_id)?;
|
||||
session_guard.ensure_chat_loaded(request.chat_id)?;
|
||||
|
||||
// 优先使用消息接收时捕获的 topic_id，消除 #1 与 #2 之间的竞态
|
||||
let original_topic_id = match &request.topic_id {
|
||||
Some(tid) => Some(tid.clone()),
|
||||
None => session_guard
|
||||
.current_topic(request.chat_id)
|
||||
.map(|s| s.to_string()),
|
||||
};
|
||||
|
||||
session_guard.ensure_chat_loaded(request.chat_id, original_topic_id.as_deref())?;
|
||||
|
||||
session_guard.ensure_agent_prompt_before_user_message(request.chat_id)?;
|
||||
|
||||
@ -219,29 +246,29 @@ impl AgentExecutionService {
|
||||
}
|
||||
let enriched_content =
|
||||
enrich_user_content_with_media_refs(request.content, &media_refs)?;
|
||||
enrich_user_content_with_media_refs(request.content, &media_refs)?;
|
||||
enrich_user_content_with_media_refs(request.content, &media_refs)?;
|
||||
|
||||
// 先计算 user_message_count（在添加新消息之前）
|
||||
let history_before = session_guard.get_or_create_history(request.chat_id).clone();
|
||||
// 无 topic 时用 chat_id 作为 topic_histories 的回退 key
|
||||
let history_key = original_topic_id.as_deref().unwrap_or(request.chat_id);
|
||||
let history_before = session_guard.get_or_create_history(history_key).clone();
|
||||
let user_message_count = history_before.iter().filter(|m| m.role == "user").count();
|
||||
|
||||
// 在添加用户消息前，记录当前话题 ID
|
||||
let original_topic_id = session_guard
|
||||
.current_topic(request.chat_id)
|
||||
.map(|s| s.to_string());
|
||||
|
||||
let user_message = session_guard.create_user_message(&enriched_content, media_refs);
|
||||
session_guard.append_persisted_message(request.chat_id, user_message.clone())?;
|
||||
session_guard.append_persisted_message(
|
||||
request.chat_id,
|
||||
original_topic_id.as_deref(),
|
||||
user_message.clone(),
|
||||
)?;
|
||||
|
||||
// 再获取包含新消息的完整历史记录
|
||||
let history = session_guard.get_or_create_history(request.chat_id).clone();
|
||||
let history = session_guard.get_or_create_history(history_key).clone();
|
||||
session_guard.record_skill_offer(request.chat_id)?;
|
||||
|
||||
let mut agent = session_guard.create_agent(
|
||||
request.chat_id,
|
||||
Some(request.sender_id),
|
||||
Some(&user_message.id),
|
||||
original_topic_id.as_deref(),
|
||||
)?;
|
||||
if let Some(handler) = request.live_emitter.clone() {
|
||||
agent = agent.with_emitted_message_handler(handler);
|
||||
@ -282,27 +309,46 @@ impl AgentExecutionService {
|
||||
&self,
|
||||
request: ScheduledExecutionRequest<'_>,
|
||||
) -> Result<Vec<OutboundMessage>, AgentError> {
|
||||
// 获取该 topic 的串行锁（与普通消息路径共享，保证串行执行）
|
||||
// 定时任务由调度器触发，无用户消息竞态；在锁前一次性捕获 topic_id，
|
||||
// 锁后复用同一值作为 original_topic_id，保证锁键与写入目标一致。
|
||||
let (serial_lock, lock_time_topic_id) = {
|
||||
let mut session_guard = request.session.lock().await;
|
||||
let tid = session_guard
|
||||
.current_topic(request.chat_id)
|
||||
.map(|s| s.to_string());
|
||||
let lock_key = tid.as_deref().unwrap_or(request.chat_id);
|
||||
(session_guard.topic_serial_lock(lock_key), tid)
|
||||
};
|
||||
|
||||
// 等待该 topic 的前一条消息处理完成（含压缩）
|
||||
let _serial_guard = serial_lock.lock().await;
|
||||
|
||||
let (history, mut agent, user_message, user_message_count, original_topic_id, store, session_id) = {
|
||||
let mut session_guard = request.session.lock().await;
|
||||
|
||||
session_guard.ensure_persistent_session(request.chat_id)?;
|
||||
|
||||
// 复用锁前捕获的 topic_id，保证锁键与写入目标一致
|
||||
let original_topic_id = lock_time_topic_id.clone();
|
||||
|
||||
// 如果 fresh_session 为 true，清理历史（内存 + 数据库）
|
||||
if request.fresh_session {
|
||||
session_guard.clear_chat_history(request.chat_id)?;
|
||||
session_guard.clear_chat_history(request.chat_id, original_topic_id.as_deref())?;
|
||||
tracing::info!(
|
||||
chat_id = %request.chat_id,
|
||||
"Fresh session enabled, history cleared"
|
||||
);
|
||||
}
|
||||
|
||||
session_guard.ensure_chat_loaded(request.chat_id)?;
|
||||
session_guard.ensure_chat_loaded(request.chat_id, original_topic_id.as_deref())?;
|
||||
session_guard.ensure_agent_prompt_before_user_message(request.chat_id)?;
|
||||
|
||||
let scheduled_system_prompt =
|
||||
compose_scheduled_task_system_prompt(request.system_prompt);
|
||||
session_guard.append_persisted_message(
|
||||
request.chat_id,
|
||||
original_topic_id.as_deref(),
|
||||
ChatMessage::system_with_context(
|
||||
&scheduled_system_prompt,
|
||||
Some(SYSTEM_CONTEXT_SCHEDULED_PROMPT.to_string()),
|
||||
@ -310,19 +356,19 @@ impl AgentExecutionService {
|
||||
)?;
|
||||
|
||||
// 先计算 user_message_count（在添加新消息之前）
|
||||
let history_before = session_guard.get_or_create_history(request.chat_id).clone();
|
||||
let history_key = original_topic_id.as_deref().unwrap_or(request.chat_id);
|
||||
let history_before = session_guard.get_or_create_history(history_key).clone();
|
||||
let user_message_count = history_before.iter().filter(|m| m.role == "user").count();
|
||||
|
||||
// 在添加用户消息前，记录当前话题 ID
|
||||
let original_topic_id = session_guard
|
||||
.current_topic(request.chat_id)
|
||||
.map(|s| s.to_string());
|
||||
|
||||
let user_message = session_guard.create_user_message(request.prompt, Vec::new());
|
||||
session_guard.append_persisted_message(request.chat_id, user_message.clone())?;
|
||||
session_guard.append_persisted_message(
|
||||
request.chat_id,
|
||||
original_topic_id.as_deref(),
|
||||
user_message.clone(),
|
||||
)?;
|
||||
|
||||
// 再获取包含新消息的完整历史记录
|
||||
let history = session_guard.get_or_create_history(request.chat_id).clone();
|
||||
let history = session_guard.get_or_create_history(history_key).clone();
|
||||
session_guard.record_skill_offer(request.chat_id)?;
|
||||
|
||||
let agent = session_guard.create_agent_with_provider_config(
|
||||
@ -331,6 +377,7 @@ impl AgentExecutionService {
|
||||
Some(request.sender_id),
|
||||
Some(&user_message.id),
|
||||
request.provider_config.clone(),
|
||||
original_topic_id.as_deref(),
|
||||
)?;
|
||||
|
||||
// 获取 store 和 session_id，用于构造消息持久化 handler
|
||||
@ -373,7 +420,7 @@ impl AgentExecutionService {
|
||||
metadata: request.metadata,
|
||||
suppress_live_tool_calls: false,
|
||||
execution_kind: "scheduled_task",
|
||||
original_topic_id,
|
||||
original_topic_id: original_topic_id.clone(),
|
||||
},
|
||||
)
|
||||
.await?;
|
||||
@ -381,7 +428,8 @@ impl AgentExecutionService {
|
||||
// 清理内存历史，释放内存（数据库历史保留）
|
||||
{
|
||||
let mut session_guard = request.session.lock().await;
|
||||
session_guard.remove_history(request.chat_id);
|
||||
let history_key = original_topic_id.as_deref().unwrap_or(request.chat_id);
|
||||
session_guard.remove_history(history_key);
|
||||
tracing::info!(
|
||||
chat_id = %request.chat_id,
|
||||
"Scheduled task completed, memory history released"
|
||||
@ -399,6 +447,7 @@ impl AgentExecutionService {
|
||||
let channel_name = request.channel_name.to_string();
|
||||
let chat_id = request.chat_id.to_string();
|
||||
let execution_kind = request.execution_kind.to_string();
|
||||
let topic_id = request.original_topic_id.clone();
|
||||
|
||||
let finalized_result = {
|
||||
let mut session_guard = session.lock().await;
|
||||
@ -406,8 +455,13 @@ impl AgentExecutionService {
|
||||
};
|
||||
|
||||
if finalized_result.should_schedule_compaction {
|
||||
if let Err(error) =
|
||||
schedule_background_history_compaction(session.clone(), chat_id.clone()).await
|
||||
let compaction_topic_id = topic_id.unwrap_or_else(|| chat_id.clone());
|
||||
if let Err(error) = schedule_background_history_compaction(
|
||||
session.clone(),
|
||||
chat_id.clone(),
|
||||
compaction_topic_id,
|
||||
)
|
||||
.await
|
||||
{
|
||||
tracing::warn!(
|
||||
channel = %channel_name,
|
||||
@ -481,4 +535,62 @@ mod tests {
|
||||
assert!(!should_display_message_to_user(false, &message));
|
||||
assert!(should_display_message_to_user(true, &message));
|
||||
}
|
||||
|
||||
/// 对抗性测试：同一 topic 的串行锁被持有时，第二次获取应阻塞
|
||||
#[tokio::test]
|
||||
async fn test_topic_serial_lock_blocks_concurrent_access() {
|
||||
let lock = std::sync::Arc::new(tokio::sync::Mutex::new(()));
|
||||
let _guard1 = lock.lock().await;
|
||||
|
||||
// 第二次获取应阻塞，1ms 超时验证
|
||||
let result = tokio::time::timeout(
|
||||
std::time::Duration::from_millis(1),
|
||||
lock.lock(),
|
||||
)
|
||||
.await;
|
||||
|
||||
assert!(result.is_err(), "第二次获取同一锁应阻塞");
|
||||
}
|
||||
|
||||
/// 对抗性测试：不同 topic 的串行锁互不影响，可同时获取
|
||||
#[tokio::test]
|
||||
async fn test_different_topic_locks_independent() {
|
||||
let lock_a = std::sync::Arc::new(tokio::sync::Mutex::new(()));
|
||||
let lock_b = std::sync::Arc::new(tokio::sync::Mutex::new(()));
|
||||
|
||||
let _guard_a = lock_a.lock().await;
|
||||
|
||||
// 不同锁应立即可获取
|
||||
let result = tokio::time::timeout(
|
||||
std::time::Duration::from_millis(100),
|
||||
lock_b.lock(),
|
||||
)
|
||||
.await;
|
||||
|
||||
assert!(result.is_ok(), "不同 topic 的锁应互不影响");
|
||||
}
|
||||
|
||||
/// 对抗性测试：错误返回路径锁被正确释放（RAII 保证）
|
||||
#[tokio::test]
|
||||
async fn test_serial_lock_released_on_error() {
|
||||
let lock = std::sync::Arc::new(tokio::sync::Mutex::new(()));
|
||||
|
||||
// 模拟 prepare_and_execute_message 的错误路径：
|
||||
// 获取锁 → 返回错误 → 锁应通过 RAII 释放
|
||||
{
|
||||
let _serial_guard = lock.lock().await;
|
||||
// 模拟错误返回（`?` 或 `Err` 分支）
|
||||
let _result: Result<(), AgentError> = Err(AgentError::Other("simulated".to_string()));
|
||||
// _serial_guard 在此块结束时 Drop，释放锁
|
||||
}
|
||||
|
||||
// 锁应已释放，可再次获取
|
||||
let result = tokio::time::timeout(
|
||||
std::time::Duration::from_millis(100),
|
||||
lock.lock(),
|
||||
)
|
||||
.await;
|
||||
|
||||
assert!(result.is_ok(), "错误返回后锁应已释放");
|
||||
}
|
||||
}
|
||||
|
||||
@ -269,7 +269,7 @@ impl InboundProcessor {
|
||||
if let Some(ref topic_id) = current_topic {
|
||||
let cancel_rx = self.cancel_manager.register(topic_id).await;
|
||||
self.session_manager
|
||||
.set_agent_cancel_token(&channel, &chat_id, cancel_rx)
|
||||
.set_agent_cancel_token(&channel, &chat_id, Some(topic_id.as_str()), cancel_rx)
|
||||
.await;
|
||||
}
|
||||
|
||||
@ -282,6 +282,7 @@ impl InboundProcessor {
|
||||
&inbound.content,
|
||||
inbound.media,
|
||||
Some(live_emitter),
|
||||
current_topic.as_deref(),
|
||||
)
|
||||
.await
|
||||
{
|
||||
|
||||
@ -201,7 +201,6 @@ pub(crate) fn build_session_manager_with_sender(
|
||||
let runtime_config = SubAgentRuntimeConfig {
|
||||
default_allowed_tools: task_config.allowed_tools.iter().cloned().collect(),
|
||||
default_max_execution_secs: task_config.max_execution_secs,
|
||||
explore_max_execution_secs: task_config.explore_max_execution_secs,
|
||||
ttl_hours: task_config.ttl_hours,
|
||||
skills_index: skills.system_index_prompt(),
|
||||
max_nesting_depth: task_config.max_nesting_depth,
|
||||
|
||||
@ -323,14 +323,16 @@ impl Session {
|
||||
/// 存入待使用的取消信号接收端。
|
||||
///
|
||||
/// 在 Agent 执行前由处理器调用，Agent 构建时（create_agent）自动消费。
|
||||
/// 每个 chat_id 同时只允许一个 pending token；新 token 会替换旧 token。
|
||||
/// 优先按 topic_id 键化（不同 topic 的 token 互不覆盖）；
|
||||
/// 无 topic 时回退到 chat_id。
|
||||
pub fn set_cancel_receiver(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
topic_id: Option<&str>,
|
||||
receiver: tokio::sync::watch::Receiver<()>,
|
||||
) {
|
||||
self.pending_cancel_tokens
|
||||
.insert(chat_id.to_string(), receiver);
|
||||
let key = topic_id.unwrap_or(chat_id).to_string();
|
||||
self.pending_cancel_tokens.insert(key, receiver);
|
||||
}
|
||||
|
||||
/// 获取当前话题 ID（指定 chat）
|
||||
@ -338,27 +340,20 @@ impl Session {
|
||||
self.history.chat_topic(chat_id)
|
||||
}
|
||||
|
||||
/// 获取历史所对应的话题 ID（指定 chat）
|
||||
pub fn history_topic(&self, chat_id: &str) -> Option<&str> {
|
||||
self.history.history_topic(chat_id)
|
||||
}
|
||||
|
||||
/// 切换话题 - 清除当前历史并加载新话题的历史
|
||||
/// 切换话题 - 设置当前 topic 并加载新话题的历史到内存
|
||||
/// 不同 topic 的历史在 topic_histories 中独立存储，切换不互斥。
|
||||
pub fn switch_topic(&mut self, chat_id: &str, topic_id: &str) -> Result<(), AgentError> {
|
||||
// 清除当前历史
|
||||
self.history.remove_history(chat_id);
|
||||
|
||||
// 先设置当前话题（set_history 需要这个）
|
||||
// 设置当前 topic（UI 状态）
|
||||
self.history.set_chat_topic(chat_id, topic_id.to_string());
|
||||
|
||||
// 加载新话题的历史（按 session_id 过滤，排除子智能体消息）
|
||||
// 加载新 topic 的历史到内存（按 topic_id 键化）
|
||||
let session_id = self.persistent_session_id(chat_id);
|
||||
let messages = self
|
||||
.store
|
||||
.load_messages_for_topic(topic_id, Some(&session_id))
|
||||
.map_err(|e| AgentError::Other(format!("load topic messages error: {}", e)))?;
|
||||
|
||||
self.history.set_history(chat_id, messages);
|
||||
self.history.set_history(topic_id, messages);
|
||||
|
||||
tracing::info!(
|
||||
topic_id = %topic_id,
|
||||
@ -372,32 +367,14 @@ impl Session {
|
||||
self.history.ensure_persistent_session(chat_id)
|
||||
}
|
||||
|
||||
pub fn ensure_chat_loaded(&mut self, chat_id: &str) -> Result<(), AgentError> {
|
||||
// 检查历史是否存在且对应正确的话题
|
||||
// 先获取 topic 信息并转换为 owned String，避免借用冲突
|
||||
let current_topic: Option<String> = self.history.chat_topic(chat_id).map(|s| s.to_string());
|
||||
let stored_topic = self.history.history_topic(chat_id);
|
||||
|
||||
if self.chat_history_exists(chat_id) {
|
||||
// 如果历史已存在，但话题不匹配，需要重新加载
|
||||
if current_topic.as_deref() != stored_topic {
|
||||
tracing::info!(
|
||||
chat_id = %chat_id,
|
||||
current_topic = ?current_topic,
|
||||
stored_topic = ?stored_topic,
|
||||
"Topic changed, reloading history"
|
||||
);
|
||||
self.reload_chat_history(chat_id)?;
|
||||
}
|
||||
return Ok(());
|
||||
}
|
||||
|
||||
// 历史不存在，按 topic 加载（如果设置了 topic）
|
||||
self.history.ensure_chat_loaded(chat_id, current_topic.as_deref())
|
||||
}
|
||||
|
||||
fn chat_history_exists(&self, chat_id: &str) -> bool {
|
||||
self.history.get_history(chat_id).is_some()
|
||||
/// 确保指定 topic 的历史已加载到内存。
|
||||
/// 按 topic_id 键化查找，已存在则直接返回，否则从 DB 加载。
|
||||
pub fn ensure_chat_loaded(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
topic_id: Option<&str>,
|
||||
) -> Result<(), AgentError> {
|
||||
self.history.ensure_chat_loaded(chat_id, topic_id)
|
||||
}
|
||||
|
||||
pub fn ensure_agent_prompt_before_user_message(
|
||||
@ -408,43 +385,68 @@ impl Session {
|
||||
.ensure_agent_prompt_before_user_message(chat_id)
|
||||
}
|
||||
|
||||
/// 获取或创建指定 chat_id 的会话历史
|
||||
pub fn get_or_create_history(&mut self, chat_id: &str) -> &mut Vec<ChatMessage> {
|
||||
self.history.get_or_create_history(chat_id)
|
||||
/// 获取或创建指定 topic_id 的会话历史
|
||||
pub fn get_or_create_history(&mut self, topic_id: &str) -> &mut Vec<ChatMessage> {
|
||||
self.history.get_or_create_history(topic_id)
|
||||
}
|
||||
|
||||
/// 获取指定 chat_id 的会话历史（不创建）
|
||||
pub fn get_history(&self, chat_id: &str) -> Option<&Vec<ChatMessage>> {
|
||||
self.history.get_history(chat_id)
|
||||
/// 获取指定 topic_id 的会话历史（不创建）
|
||||
pub fn get_history(&self, topic_id: &str) -> Option<&Vec<ChatMessage>> {
|
||||
self.history.get_history(topic_id)
|
||||
}
|
||||
|
||||
/// 使用完整消息追加到历史
|
||||
pub fn add_message(&mut self, chat_id: &str, message: ChatMessage) {
|
||||
self.history.add_message(chat_id, message);
|
||||
/// 使用完整消息追加到指定 topic 的历史
|
||||
pub fn add_message(&mut self, topic_id: &str, message: ChatMessage) {
|
||||
self.history.add_message(topic_id, message);
|
||||
}
|
||||
|
||||
pub fn remove_history(&mut self, chat_id: &str) {
|
||||
self.history.remove_history(chat_id);
|
||||
pub fn remove_history(&mut self, topic_id: &str) {
|
||||
self.history.remove_history(topic_id);
|
||||
}
|
||||
|
||||
pub fn clear_chat_history(&mut self, chat_id: &str) -> Result<(), AgentError> {
|
||||
self.history.clear_chat_history(chat_id)
|
||||
pub fn clear_chat_history(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
topic_id: Option<&str>,
|
||||
) -> Result<(), AgentError> {
|
||||
self.history.clear_chat_history(chat_id, topic_id)
|
||||
}
|
||||
|
||||
/// 将消息写入内存与持久化层（使用当前 topic）
|
||||
/// 将消息写入内存与持久化层。
|
||||
/// 优先使用显式传入的 topic_id；未传入时回退到当前 chat 的活跃 topic。
|
||||
/// 只有当写入的 topic 匹配当前活跃 topic 时才更新内存历史，
|
||||
/// 避免旧 topic 的消息污染已切换到的新 topic 的内存历史。
|
||||
pub fn append_persisted_message(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
explicit_topic_id: Option<&str>,
|
||||
message: ChatMessage,
|
||||
) -> Result<(), AgentError> {
|
||||
let session_id = self.persistent_session_id(chat_id);
|
||||
let topic_id = self.history.chat_topic(chat_id).map(|s| s.to_string());
|
||||
let topic_id = explicit_topic_id
|
||||
.map(|s| s.to_string())
|
||||
.or_else(|| self.history.chat_topic(chat_id).map(|s| s.to_string()));
|
||||
self.store
|
||||
.append_message_with_topic(&session_id, topic_id.as_deref(), &message)
|
||||
.map_err(|err| {
|
||||
AgentError::Other(format!("append message persistence error: {}", err))
|
||||
})?;
|
||||
self.add_message(chat_id, message);
|
||||
|
||||
// 只有当写入的 topic 匹配当前活跃 topic 时才更新内存历史。
|
||||
// 当用户已切换到新 topic 时，旧 topic 的排队消息不应污染新 topic 的内存历史。
|
||||
let current_chat_topic = self.history.chat_topic(chat_id);
|
||||
if topic_id.as_deref() == current_chat_topic {
|
||||
if let Some(ref tid) = topic_id {
|
||||
self.add_message(tid, message);
|
||||
}
|
||||
} else {
|
||||
tracing::info!(
|
||||
chat_id = %chat_id,
|
||||
write_topic_id = ?topic_id,
|
||||
current_topic_id = ?current_chat_topic,
|
||||
"Skipping memory history update: message belongs to a different topic"
|
||||
);
|
||||
}
|
||||
|
||||
// 更新 topic 的最后活跃时间
|
||||
if let Some(ref topic_id) = topic_id {
|
||||
@ -458,13 +460,13 @@ impl Session {
|
||||
|
||||
pub fn append_persisted_messages<I>(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
topic_id: &str,
|
||||
messages: I,
|
||||
) -> Result<(), AgentError>
|
||||
where
|
||||
I: IntoIterator<Item = ChatMessage>,
|
||||
{
|
||||
self.history.append_persisted_messages(chat_id, messages)
|
||||
self.history.append_persisted_messages(topic_id, messages)
|
||||
}
|
||||
|
||||
/// 将消息保存到指定话题（直接写入数据库，不更新内存历史）
|
||||
@ -486,32 +488,32 @@ impl Session {
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
fn latest_user_message_id(&self, chat_id: &str) -> Option<&str> {
|
||||
self.latest_user_message(chat_id)
|
||||
fn latest_user_message_id(&self, topic_id: &str) -> Option<&str> {
|
||||
self.latest_user_message(topic_id)
|
||||
.map(|message| message.id.as_str())
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
fn latest_user_message(&self, chat_id: &str) -> Option<&ChatMessage> {
|
||||
self.history.latest_user_message(chat_id)
|
||||
fn latest_user_message(&self, topic_id: &str) -> Option<&ChatMessage> {
|
||||
self.history.latest_user_message(topic_id)
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
fn is_latest_user_message(&self, chat_id: &str, message_id: &str) -> bool {
|
||||
self.latest_user_message_id(chat_id)
|
||||
fn is_latest_user_message(&self, topic_id: &str, message_id: &str) -> bool {
|
||||
self.latest_user_message_id(topic_id)
|
||||
.map(|current_id| current_id == message_id)
|
||||
.unwrap_or(false)
|
||||
}
|
||||
|
||||
pub(crate) fn matches_current_user_turn(&self, chat_id: &str, message: &ChatMessage) -> bool {
|
||||
self.history.matches_current_user_turn(chat_id, message)
|
||||
pub(crate) fn matches_current_user_turn(&self, topic_id: &str, message: &ChatMessage) -> bool {
|
||||
self.history.matches_current_user_turn(topic_id, message)
|
||||
}
|
||||
|
||||
pub(crate) fn stale_result_diagnostics(
|
||||
&self,
|
||||
chat_id: &str,
|
||||
topic_id: &str,
|
||||
) -> (Option<&str>, Option<String>, bool, usize) {
|
||||
self.history.stale_result_diagnostics(chat_id)
|
||||
self.history.stale_result_diagnostics(topic_id)
|
||||
}
|
||||
|
||||
/// 清除所有历史
|
||||
@ -533,20 +535,20 @@ impl Session {
|
||||
&self.compressor
|
||||
}
|
||||
|
||||
pub(crate) fn reload_chat_history(&mut self, chat_id: &str) -> Result<(), AgentError> {
|
||||
// 如果当前有 topic，加载该 topic 的消息（按 session_id 过滤，排除子智能体消息）
|
||||
if let Some(topic_id) = self.history.chat_topic(chat_id) {
|
||||
let session_id = self.persistent_session_id(chat_id);
|
||||
let messages = self
|
||||
.store
|
||||
.load_messages_for_topic(topic_id, Some(&session_id))
|
||||
.map_err(|e| AgentError::Other(format!("load topic messages error: {}", e)))?;
|
||||
self.history.set_history(chat_id, messages);
|
||||
} else {
|
||||
// 否则加载 session 的所有消息
|
||||
self.history.reload_chat_history(chat_id)?;
|
||||
}
|
||||
Ok(())
|
||||
/// 获取该 topic 的串行化锁。
|
||||
/// 同一 topic 的消息处理（agent loop + 压缩）共享此锁，保证串行执行；
|
||||
/// 不同 topic 之间互不阻塞。
|
||||
pub(crate) fn topic_serial_lock(&mut self, topic_id: &str) -> Arc<tokio::sync::Mutex<()>> {
|
||||
self.history.topic_serial_lock(topic_id)
|
||||
}
|
||||
|
||||
/// 按 topic_id 从 DB 重新加载历史到内存
|
||||
pub(crate) fn reload_topic_history(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
topic_id: &str,
|
||||
) -> Result<(), AgentError> {
|
||||
self.history.reload_topic_history(chat_id, topic_id)
|
||||
}
|
||||
|
||||
pub(crate) fn store(&self) -> Arc<dyn ConversationRepository> {
|
||||
@ -572,6 +574,7 @@ impl Session {
|
||||
chat_id: &str,
|
||||
sender_id: Option<&str>,
|
||||
message_id: Option<&str>,
|
||||
explicit_topic_id: Option<&str>,
|
||||
) -> Result<AgentLoop, AgentError> {
|
||||
self.create_agent_with_provider_config(
|
||||
chat_id,
|
||||
@ -579,6 +582,7 @@ impl Session {
|
||||
sender_id,
|
||||
message_id,
|
||||
self.provider_config.clone(),
|
||||
explicit_topic_id,
|
||||
)
|
||||
}
|
||||
|
||||
@ -589,10 +593,21 @@ impl Session {
|
||||
sender_id: Option<&str>,
|
||||
message_id: Option<&str>,
|
||||
provider_config: LLMProviderConfig,
|
||||
explicit_topic_id: Option<&str>,
|
||||
) -> Result<AgentLoop, AgentError> {
|
||||
// 优先使用显式传入的 topic_id；回退到当前 chat 的活跃 topic
|
||||
let topic_id = explicit_topic_id
|
||||
.map(|s| s.to_string())
|
||||
.or_else(|| self.current_topic(session_chat_id).map(|s| s.to_string()));
|
||||
|
||||
// 消费 pending 的取消信号接收端（如果存在）
|
||||
let cancel_token = self.pending_cancel_tokens.remove(session_chat_id);
|
||||
let topic_id = self.current_topic(session_chat_id).map(|s| s.to_string());
|
||||
// 优先按 topic_id 查找；无 topic 时回退 chat_id
|
||||
let cancel_token = match &topic_id {
|
||||
Some(tid) => self.pending_cancel_tokens.remove(tid)
|
||||
.or_else(|| self.pending_cancel_tokens.remove(session_chat_id)),
|
||||
None => self.pending_cancel_tokens.remove(session_chat_id),
|
||||
};
|
||||
|
||||
self.agent_factory.create(AgentBuildRequest {
|
||||
channel_name: &self.channel_name,
|
||||
session_chat_id,
|
||||
@ -812,10 +827,11 @@ impl SessionManager {
|
||||
&self,
|
||||
channel_name: &str,
|
||||
chat_id: &str,
|
||||
topic_id: Option<&str>,
|
||||
token: tokio::sync::watch::Receiver<()>,
|
||||
) {
|
||||
if let Some(session) = self.get(channel_name).await {
|
||||
session.lock().await.set_cancel_receiver(chat_id, token);
|
||||
session.lock().await.set_cancel_receiver(chat_id, topic_id, token);
|
||||
}
|
||||
}
|
||||
|
||||
@ -837,6 +853,7 @@ impl SessionManager {
|
||||
content: &str,
|
||||
media: Vec<crate::bus::MediaItem>,
|
||||
live_emitter: Option<Arc<dyn EmittedMessageHandler>>,
|
||||
topic_id: Option<&str>,
|
||||
) -> Result<Vec<OutboundMessage>, AgentError> {
|
||||
self.messages
|
||||
.handle_message(
|
||||
@ -846,6 +863,7 @@ impl SessionManager {
|
||||
content,
|
||||
media,
|
||||
live_emitter,
|
||||
topic_id,
|
||||
)
|
||||
.await
|
||||
}
|
||||
@ -953,7 +971,7 @@ mod tests {
|
||||
user_tx,
|
||||
tools,
|
||||
skills,
|
||||
store,
|
||||
store.clone(),
|
||||
100,
|
||||
Arc::new(SubagentRuntime::from_config(Default::default())),
|
||||
)
|
||||
@ -961,19 +979,22 @@ mod tests {
|
||||
.unwrap();
|
||||
|
||||
session.ensure_persistent_session("chat-1").unwrap();
|
||||
session.ensure_chat_loaded("chat-1").unwrap();
|
||||
let session_id = session.persistent_session_id("chat-1");
|
||||
let topic = store.create_topic(&session_id, "test topic", None).unwrap();
|
||||
let topic_id = topic.id.clone();
|
||||
session.switch_topic("chat-1", &topic_id).unwrap();
|
||||
|
||||
let first = session.create_user_message("first", Vec::new());
|
||||
let first_id = first.id.clone();
|
||||
session.append_persisted_message("chat-1", first).unwrap();
|
||||
assert!(session.is_latest_user_message("chat-1", &first_id));
|
||||
session.append_persisted_message("chat-1", Some(&topic_id), first).unwrap();
|
||||
assert!(session.is_latest_user_message(&topic_id, &first_id));
|
||||
|
||||
let second = session.create_user_message("second", Vec::new());
|
||||
let second_id = second.id.clone();
|
||||
session.append_persisted_message("chat-1", second).unwrap();
|
||||
session.append_persisted_message("chat-1", Some(&topic_id), second).unwrap();
|
||||
|
||||
assert!(!session.is_latest_user_message("chat-1", &first_id));
|
||||
assert!(session.is_latest_user_message("chat-1", &second_id));
|
||||
assert!(!session.is_latest_user_message(&topic_id, &first_id));
|
||||
assert!(session.is_latest_user_message(&topic_id, &second_id));
|
||||
}
|
||||
|
||||
#[tokio::test]
|
||||
@ -1010,46 +1031,37 @@ mod tests {
|
||||
.unwrap();
|
||||
|
||||
session.ensure_persistent_session("chat-1").unwrap();
|
||||
session.ensure_chat_loaded("chat-1").unwrap();
|
||||
let session_id = session.persistent_session_id("chat-1");
|
||||
let topic = store.create_topic(&session_id, "test topic", None).unwrap();
|
||||
let topic_id = topic.id.clone();
|
||||
session.switch_topic("chat-1", &topic_id).unwrap();
|
||||
|
||||
let first = session.create_user_message("first", Vec::new());
|
||||
let first_id = first.id.clone();
|
||||
session.append_persisted_message("chat-1", first).unwrap();
|
||||
session.append_persisted_message("chat-1", Some(&topic_id), first).unwrap();
|
||||
session
|
||||
.append_persisted_message("chat-1", ChatMessage::assistant("answer-1"))
|
||||
.append_persisted_message("chat-1", Some(&topic_id), ChatMessage::assistant("answer-1"))
|
||||
.unwrap();
|
||||
|
||||
let second = session.create_user_message("second", Vec::new());
|
||||
session
|
||||
.append_persisted_message("chat-1", second.clone())
|
||||
.append_persisted_message("chat-1", Some(&topic_id), second.clone())
|
||||
.unwrap();
|
||||
session
|
||||
.append_persisted_message("chat-1", ChatMessage::assistant("answer-2"))
|
||||
.append_persisted_message("chat-1", Some(&topic_id), ChatMessage::assistant("answer-2"))
|
||||
.unwrap();
|
||||
|
||||
let session_id = session.persistent_session_id("chat-1");
|
||||
let snapshot_end_seq = store
|
||||
.get_session(&session_id)
|
||||
.unwrap()
|
||||
.unwrap()
|
||||
.message_count;
|
||||
let preserved_messages = session.get_history("chat-1").unwrap().clone();
|
||||
let preserved_messages = session.get_history(&topic_id).unwrap().clone();
|
||||
|
||||
store
|
||||
.compact_active_history(
|
||||
&session_id,
|
||||
snapshot_end_seq,
|
||||
&[],
|
||||
&ChatMessage::system("[Compressed History]\n\nsummary"),
|
||||
&preserved_messages,
|
||||
)
|
||||
.replace_topic_history(&session_id, &topic_id, &preserved_messages)
|
||||
.unwrap();
|
||||
|
||||
session.reload_chat_history("chat-1").unwrap();
|
||||
session.reload_topic_history("chat-1", &topic_id).unwrap();
|
||||
|
||||
assert!(!session.is_latest_user_message("chat-1", &first_id));
|
||||
assert!(!session.is_latest_user_message("chat-1", &second.id));
|
||||
assert!(session.matches_current_user_turn("chat-1", &second));
|
||||
assert!(!session.is_latest_user_message(&topic_id, &first_id));
|
||||
assert!(session.is_latest_user_message(&topic_id, &second.id));
|
||||
assert!(session.matches_current_user_turn(&topic_id, &second));
|
||||
}
|
||||
|
||||
async fn start_mock_openai_server() -> String {
|
||||
@ -1209,7 +1221,7 @@ mod tests {
|
||||
.unwrap();
|
||||
|
||||
let outbound = session_manager
|
||||
.handle_message("test-channel", "user-1", "chat-1", "hello", Vec::new(), None)
|
||||
.handle_message("test-channel", "user-1", "chat-1", "hello", Vec::new(), None, None)
|
||||
.await
|
||||
.unwrap();
|
||||
|
||||
@ -2037,7 +2049,7 @@ mod tests {
|
||||
.unwrap();
|
||||
|
||||
session.ensure_persistent_session("chat-1").unwrap();
|
||||
session.ensure_chat_loaded("chat-1").unwrap();
|
||||
session.switch_topic("chat-1", "chat-1").unwrap();
|
||||
|
||||
let history = session.get_history("chat-1").unwrap();
|
||||
// 新设计：系统提示词不再持久化到历史记录，而是每次请求时动态注入
|
||||
@ -2078,11 +2090,14 @@ mod tests {
|
||||
.unwrap();
|
||||
|
||||
session.ensure_persistent_session("chat-1").unwrap();
|
||||
session.ensure_chat_loaded("chat-1").unwrap();
|
||||
let session_id = session.persistent_session_id("chat-1");
|
||||
let topic = store.create_topic(&session_id, "test topic", None).unwrap();
|
||||
let topic_id = topic.id.clone();
|
||||
session.switch_topic("chat-1", &topic_id).unwrap();
|
||||
|
||||
for turn in 0..100 {
|
||||
session
|
||||
.append_persisted_message("chat-1", ChatMessage::user(format!("user-{turn}")))
|
||||
.append_persisted_message("chat-1", Some(&topic_id), ChatMessage::user(format!("user-{turn}")))
|
||||
.unwrap();
|
||||
}
|
||||
|
||||
@ -2091,7 +2106,7 @@ mod tests {
|
||||
.unwrap();
|
||||
|
||||
// 新设计：系统提示词不再持久化到历史记录
|
||||
let history = session.get_history("chat-1").unwrap();
|
||||
let history = session.get_history(&topic_id).unwrap();
|
||||
let user_messages = history
|
||||
.iter()
|
||||
.filter(|message| message.role == "user")
|
||||
@ -2110,7 +2125,7 @@ mod tests {
|
||||
session
|
||||
.ensure_agent_prompt_before_user_message("chat-1")
|
||||
.unwrap();
|
||||
let history = session.get_history("chat-1").unwrap();
|
||||
let history = session.get_history(&topic_id).unwrap();
|
||||
let user_messages = history
|
||||
.iter()
|
||||
.filter(|message| message.role == "user")
|
||||
@ -2152,11 +2167,14 @@ mod tests {
|
||||
.unwrap();
|
||||
|
||||
session.ensure_persistent_session("chat-1").unwrap();
|
||||
session.ensure_chat_loaded("chat-1").unwrap();
|
||||
let session_id = session.persistent_session_id("chat-1");
|
||||
let topic = store.create_topic(&session_id, "test topic", None).unwrap();
|
||||
let topic_id = topic.id.clone();
|
||||
session.switch_topic("chat-1", &topic_id).unwrap();
|
||||
|
||||
for turn in 0..100 {
|
||||
session
|
||||
.append_persisted_message("chat-1", ChatMessage::user(format!("user-{turn}")))
|
||||
.append_persisted_message("chat-1", Some(&topic_id), ChatMessage::user(format!("user-{turn}")))
|
||||
.unwrap();
|
||||
}
|
||||
|
||||
@ -2165,7 +2183,7 @@ mod tests {
|
||||
.unwrap();
|
||||
|
||||
// 新设计：系统提示词不再持久化到历史记录
|
||||
let history = session.get_history("chat-1").unwrap();
|
||||
let history = session.get_history(&topic_id).unwrap();
|
||||
let user_messages = history
|
||||
.iter()
|
||||
.filter(|message| message.role == "user")
|
||||
|
||||
@ -17,10 +17,18 @@ fn preview_text(content: &str, max_chars: usize) -> String {
|
||||
|
||||
pub(crate) struct SessionHistory {
|
||||
channel_name: String,
|
||||
chat_histories: HashMap<String, Vec<ChatMessage>>,
|
||||
chat_topic_ids: HashMap<String, String>, // 每个 chat 的当前 topic
|
||||
history_topic_ids: HashMap<String, String>, // 每个 chat 的历史所对应的话题
|
||||
/// 按 topic_id 键化的内存历史缓存。
|
||||
/// 不同 topic 的历史独立存储，互不干扰，支持多话题并发执行。
|
||||
topic_histories: HashMap<String, Vec<ChatMessage>>,
|
||||
/// UI 状态：每个 chat 当前活跃的 topic（按 chat_id 键）。
|
||||
chat_topic_ids: HashMap<String, String>,
|
||||
/// 正在压缩中的 topic_id 集合
|
||||
compression_in_flight: HashSet<String>,
|
||||
/// 按 topic_id 的串行化锁。
|
||||
/// 同一 topic 的消息处理（agent loop + 压缩）必须串行执行，
|
||||
/// 防止并发 loop 操作同一历史的不同快照产生交错序列。
|
||||
/// 不同 topic 之间互不阻塞，支持多话题并发执行。
|
||||
topic_serial_locks: HashMap<String, Arc<tokio::sync::Mutex<()>>>,
|
||||
conversations: Arc<dyn ConversationRepository>,
|
||||
skill_events: Arc<dyn SkillEventRepository>,
|
||||
}
|
||||
@ -33,15 +41,25 @@ impl SessionHistory {
|
||||
) -> Self {
|
||||
Self {
|
||||
channel_name: channel_name.into(),
|
||||
chat_histories: HashMap::new(),
|
||||
topic_histories: HashMap::new(),
|
||||
chat_topic_ids: HashMap::new(),
|
||||
history_topic_ids: HashMap::new(),
|
||||
compression_in_flight: HashSet::new(),
|
||||
topic_serial_locks: HashMap::new(),
|
||||
conversations,
|
||||
skill_events,
|
||||
}
|
||||
}
|
||||
|
||||
/// 获取或创建该 topic 的串行化锁。
|
||||
/// 同一 topic 的所有消息处理共享同一个锁，保证串行执行；
|
||||
/// 不同 topic 之间互不阻塞，支持多话题并发执行。
|
||||
pub(crate) fn topic_serial_lock(&mut self, topic_id: &str) -> Arc<tokio::sync::Mutex<()>> {
|
||||
self.topic_serial_locks
|
||||
.entry(topic_id.to_string())
|
||||
.or_insert_with(|| Arc::new(tokio::sync::Mutex::new(())))
|
||||
.clone()
|
||||
}
|
||||
|
||||
pub(crate) fn persistent_session_id(&self, chat_id: &str) -> String {
|
||||
persistent_session_id(&self.channel_name, chat_id)
|
||||
}
|
||||
@ -55,38 +73,36 @@ impl SessionHistory {
|
||||
.map_err(|err| AgentError::Other(format!("session persistence error: {}", err)))
|
||||
}
|
||||
|
||||
/// 确保指定 topic 的历史已加载到内存。
|
||||
/// 按 topic_id 键化查找，如果已存在则直接返回，否则从 DB 加载。
|
||||
pub(crate) fn ensure_chat_loaded(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
topic_id: Option<&str>,
|
||||
) -> Result<(), AgentError> {
|
||||
if self.chat_histories.contains_key(chat_id) {
|
||||
let Some(tid) = topic_id else {
|
||||
return Ok(());
|
||||
};
|
||||
if self.topic_histories.contains_key(tid) {
|
||||
return Ok(());
|
||||
}
|
||||
|
||||
// 如果提供了 topic_id，按 topic 加载；否则按 session 加载
|
||||
let mut history = if let Some(tid) = topic_id {
|
||||
let sid = self.persistent_session_id(chat_id);
|
||||
self.conversations
|
||||
.load_messages_for_topic(tid, Some(&sid))
|
||||
.map_err(|err| AgentError::Other(format!("session history load error: {}", err)))?
|
||||
} else {
|
||||
self.conversations
|
||||
.load_messages(&self.persistent_session_id(chat_id))
|
||||
.map_err(|err| AgentError::Other(format!("session history load error: {}", err)))?
|
||||
};
|
||||
let sid = self.persistent_session_id(chat_id);
|
||||
let mut history = self
|
||||
.conversations
|
||||
.load_messages_for_topic(tid, Some(&sid))
|
||||
.map_err(|err| AgentError::Other(format!("session history load error: {}", err)))?;
|
||||
|
||||
// 清理 DB 加载的历史中可能存在的不完整 tool_call 序列
|
||||
let removed = crate::bus::message::sanitize_incomplete_tool_call_sequences(&mut history);
|
||||
if removed > 0 {
|
||||
tracing::warn!(
|
||||
chat_id = %chat_id,
|
||||
topic_id = %tid,
|
||||
removed_count = removed,
|
||||
"Sanitized incomplete tool_call sequences on history load"
|
||||
);
|
||||
}
|
||||
|
||||
self.chat_histories.insert(chat_id.to_string(), history);
|
||||
self.topic_histories.insert(tid.to_string(), history);
|
||||
Ok(())
|
||||
}
|
||||
|
||||
@ -98,58 +114,56 @@ impl SessionHistory {
|
||||
Ok(())
|
||||
}
|
||||
|
||||
pub(crate) fn get_or_create_history(&mut self, chat_id: &str) -> &mut Vec<ChatMessage> {
|
||||
self.chat_histories.entry(chat_id.to_string()).or_default()
|
||||
pub(crate) fn get_or_create_history(&mut self, topic_id: &str) -> &mut Vec<ChatMessage> {
|
||||
self.topic_histories.entry(topic_id.to_string()).or_default()
|
||||
}
|
||||
|
||||
pub(crate) fn get_history(&self, chat_id: &str) -> Option<&Vec<ChatMessage>> {
|
||||
self.chat_histories.get(chat_id)
|
||||
pub(crate) fn get_history(&self, topic_id: &str) -> Option<&Vec<ChatMessage>> {
|
||||
self.topic_histories.get(topic_id)
|
||||
}
|
||||
|
||||
pub(crate) fn set_history(&mut self, chat_id: &str, history: Vec<ChatMessage>) {
|
||||
self.chat_histories.insert(chat_id.to_string(), history);
|
||||
// 记录历史对应的话题（当前设置的话题）
|
||||
if let Some(topic_id) = self.chat_topic_ids.get(chat_id) {
|
||||
self.history_topic_ids.insert(chat_id.to_string(), topic_id.clone());
|
||||
}
|
||||
pub(crate) fn set_history(&mut self, topic_id: &str, history: Vec<ChatMessage>) {
|
||||
self.topic_histories.insert(topic_id.to_string(), history);
|
||||
}
|
||||
|
||||
/// 获取指定 chat 的历史所对应的话题
|
||||
pub(crate) fn history_topic(&self, chat_id: &str) -> Option<&str> {
|
||||
self.history_topic_ids.get(chat_id).map(|s| s.as_str())
|
||||
}
|
||||
|
||||
/// 设置指定 chat 的当前 topic
|
||||
/// 设置指定 chat 的当前 topic（UI 状态）
|
||||
pub(crate) fn set_chat_topic(&mut self, chat_id: &str, topic_id: String) {
|
||||
self.chat_topic_ids.insert(chat_id.to_string(), topic_id);
|
||||
}
|
||||
|
||||
/// 获取指定 chat 的当前 topic
|
||||
/// 获取指定 chat 的当前 topic（UI 状态）
|
||||
pub(crate) fn chat_topic(&self, chat_id: &str) -> Option<&str> {
|
||||
self.chat_topic_ids.get(chat_id).map(|s| s.as_str())
|
||||
}
|
||||
|
||||
/// 清除指定 chat 的 topic
|
||||
/// 清除指定 chat 的 topic（UI 状态）
|
||||
pub(crate) fn clear_chat_topic(&mut self, chat_id: &str) {
|
||||
self.chat_topic_ids.remove(chat_id);
|
||||
}
|
||||
|
||||
pub(crate) fn add_message(&mut self, chat_id: &str, message: ChatMessage) {
|
||||
self.get_or_create_history(chat_id).push(message);
|
||||
pub(crate) fn add_message(&mut self, topic_id: &str, message: ChatMessage) {
|
||||
self.get_or_create_history(topic_id).push(message);
|
||||
}
|
||||
|
||||
pub(crate) fn remove_history(&mut self, chat_id: &str) {
|
||||
self.chat_histories.remove(chat_id);
|
||||
self.compression_in_flight.remove(chat_id);
|
||||
self.history_topic_ids.remove(chat_id);
|
||||
pub(crate) fn remove_history(&mut self, topic_id: &str) {
|
||||
self.topic_histories.remove(topic_id);
|
||||
self.compression_in_flight.remove(topic_id);
|
||||
}
|
||||
|
||||
pub(crate) fn clear_chat_history(&mut self, chat_id: &str) -> Result<(), AgentError> {
|
||||
if let Some(history) = self.chat_histories.get_mut(chat_id) {
|
||||
let len = history.len();
|
||||
history.clear();
|
||||
#[cfg(debug_assertions)]
|
||||
tracing::debug!(chat_id = %chat_id, previous_len = len, "Chat history cleared");
|
||||
/// 清空指定 chat/topic 的内存历史和 DB 消息。
|
||||
/// 内存按 topic_id 清，DB 按 session_id 清（保留原行为以兼容无 topic 场景）。
|
||||
pub(crate) fn clear_chat_history(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
topic_id: Option<&str>,
|
||||
) -> Result<(), AgentError> {
|
||||
if let Some(tid) = topic_id {
|
||||
if let Some(history) = self.topic_histories.get_mut(tid) {
|
||||
let len = history.len();
|
||||
history.clear();
|
||||
#[cfg(debug_assertions)]
|
||||
tracing::debug!(topic_id = %tid, previous_len = len, "Topic history cleared");
|
||||
}
|
||||
}
|
||||
|
||||
self.conversations
|
||||
@ -159,7 +173,7 @@ impl SessionHistory {
|
||||
|
||||
pub(crate) fn append_persisted_messages<I>(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
topic_id: &str,
|
||||
messages: I,
|
||||
) -> Result<(), AgentError>
|
||||
where
|
||||
@ -170,13 +184,11 @@ impl SessionHistory {
|
||||
return Ok(());
|
||||
}
|
||||
|
||||
// 在追加新消息前，先清理内存历史中的不完整 tool_call 序列
|
||||
// 这防止脏数据（如取消时产生的孤立 assistant(tool_calls)）在内存历史中累积
|
||||
if let Some(history) = self.chat_histories.get_mut(chat_id) {
|
||||
if let Some(history) = self.topic_histories.get_mut(topic_id) {
|
||||
let removed = crate::bus::message::sanitize_incomplete_tool_call_sequences(history);
|
||||
if removed > 0 {
|
||||
tracing::warn!(
|
||||
chat_id = %chat_id,
|
||||
topic_id = %topic_id,
|
||||
removed_count = removed,
|
||||
"Sanitized in-memory history before appending persisted messages"
|
||||
);
|
||||
@ -184,7 +196,7 @@ impl SessionHistory {
|
||||
}
|
||||
|
||||
for message in messages {
|
||||
self.add_message(chat_id, message);
|
||||
self.add_message(topic_id, message);
|
||||
}
|
||||
Ok(())
|
||||
}
|
||||
@ -203,13 +215,13 @@ impl SessionHistory {
|
||||
Ok(())
|
||||
}
|
||||
|
||||
pub(crate) fn latest_user_message(&self, chat_id: &str) -> Option<&ChatMessage> {
|
||||
self.get_history(chat_id)
|
||||
pub(crate) fn latest_user_message(&self, topic_id: &str) -> Option<&ChatMessage> {
|
||||
self.get_history(topic_id)
|
||||
.and_then(|history| history.iter().rev().find(|message| message.role == "user"))
|
||||
}
|
||||
|
||||
pub(crate) fn matches_current_user_turn(&self, chat_id: &str, message: &ChatMessage) -> bool {
|
||||
self.latest_user_message(chat_id)
|
||||
pub(crate) fn matches_current_user_turn(&self, topic_id: &str, message: &ChatMessage) -> bool {
|
||||
self.latest_user_message(topic_id)
|
||||
.map(|current| {
|
||||
current.id == message.id
|
||||
|| (current.content == message.content
|
||||
@ -221,14 +233,14 @@ impl SessionHistory {
|
||||
|
||||
pub(crate) fn stale_result_diagnostics(
|
||||
&self,
|
||||
chat_id: &str,
|
||||
topic_id: &str,
|
||||
) -> (Option<&str>, Option<String>, bool, usize) {
|
||||
let latest_user = self.latest_user_message(chat_id);
|
||||
let latest_user = self.latest_user_message(topic_id);
|
||||
let latest_user_id = latest_user.map(|message| message.id.as_str());
|
||||
let latest_user_preview = latest_user.map(|message| preview_text(&message.content, 80));
|
||||
let compression_in_flight = self.compression_in_flight.contains(chat_id);
|
||||
let compression_in_flight = self.compression_in_flight.contains(topic_id);
|
||||
let history_len = self
|
||||
.get_history(chat_id)
|
||||
.get_history(topic_id)
|
||||
.map(|history| history.len())
|
||||
.unwrap_or(0);
|
||||
|
||||
@ -240,31 +252,29 @@ impl SessionHistory {
|
||||
)
|
||||
}
|
||||
|
||||
/// 清空所有内存历史（主要用于测试全局重置）。
|
||||
/// 不遍历清 DB，生产环境如需清 DB 应由调用方显式调用。
|
||||
pub(crate) fn clear_all_history(&mut self) -> Result<(), AgentError> {
|
||||
let chat_ids: Vec<String> = self.chat_histories.keys().cloned().collect();
|
||||
let total: usize = self.chat_histories.values().map(|h| h.len()).sum();
|
||||
self.chat_histories.clear();
|
||||
let total: usize = self.topic_histories.values().map(|h| h.len()).sum();
|
||||
self.topic_histories.clear();
|
||||
self.compression_in_flight.clear();
|
||||
#[cfg(debug_assertions)]
|
||||
tracing::debug!(previous_total = total, "All chat histories cleared");
|
||||
|
||||
for chat_id in chat_ids {
|
||||
self.conversations
|
||||
.clear_messages(&self.persistent_session_id(&chat_id))
|
||||
.map_err(|err| {
|
||||
AgentError::Other(format!("clear history persistence error: {}", err))
|
||||
})?;
|
||||
}
|
||||
|
||||
tracing::debug!(previous_total = total, "All topic histories cleared");
|
||||
Ok(())
|
||||
}
|
||||
|
||||
pub(crate) fn reload_chat_history(&mut self, chat_id: &str) -> Result<(), AgentError> {
|
||||
/// 按 topic_id 从 DB 重新加载历史到内存
|
||||
pub(crate) fn reload_topic_history(
|
||||
&mut self,
|
||||
chat_id: &str,
|
||||
topic_id: &str,
|
||||
) -> Result<(), AgentError> {
|
||||
let sid = self.persistent_session_id(chat_id);
|
||||
let history = self
|
||||
.conversations
|
||||
.load_messages(&self.persistent_session_id(chat_id))
|
||||
.load_messages_for_topic(topic_id, Some(&sid))
|
||||
.map_err(|err| AgentError::Other(format!("session history reload error: {}", err)))?;
|
||||
self.chat_histories.insert(chat_id.to_string(), history);
|
||||
self.topic_histories.insert(topic_id.to_string(), history);
|
||||
Ok(())
|
||||
}
|
||||
|
||||
|
||||
@ -28,6 +28,7 @@ impl SessionMessageService {
|
||||
content: &str,
|
||||
media: Vec<MediaItem>,
|
||||
live_emitter: Option<Arc<dyn EmittedMessageHandler>>,
|
||||
topic_id: Option<&str>,
|
||||
) -> Result<Vec<OutboundMessage>, AgentError> {
|
||||
#[cfg(debug_assertions)]
|
||||
{
|
||||
@ -54,6 +55,7 @@ impl SessionMessageService {
|
||||
content,
|
||||
media,
|
||||
live_emitter,
|
||||
topic_id: topic_id.map(|s| s.to_string()),
|
||||
})
|
||||
.await?;
|
||||
|
||||
|
||||
@ -36,152 +36,49 @@ impl SystemPromptProvider for ToolPromptProvider {
|
||||
/// memory_search / memory_manage 工具使用说明
|
||||
const MEMORY_TOOLS_INSTRUCTIONS: &str = r#"# 记忆工具
|
||||
|
||||
## 记忆检索
|
||||
|
||||
在绝大多数请求开始时，都应先使用长期记忆检索工具 memory_search 来召回相关记忆，再决定如何回答或是否需要写入记忆。先检索通常能帮助识别用户长期偏好、稳定事实、历史决策、持续任务和上下文约束。
|
||||
|
||||
### 默认流程
|
||||
- 先使用长期记忆检索工具 memory_search，优先调用 memory_search(action='search')。
|
||||
- 只有在你已经明确知道 namespace 和 key 时，才改用 get。
|
||||
- 只有在需要浏览最近几条记忆时，才用 list。
|
||||
- 即使用户没有明确提到「记忆」或「偏好」，也应该先搜记忆，不要因为你自认为已经能直接回答就省略检索。
|
||||
|
||||
### 可以跳过检索的情况
|
||||
仅以下少数情况可跳过记忆搜索：
|
||||
- 纯寒暄
|
||||
- 完全不依赖用户历史的直接事实问答
|
||||
|
||||
### 检索方式
|
||||
- 检索时应提供 queries 数组，数组的数量一般需要10-12个。
|
||||
## 检索关键词建议
|
||||
- 提供 queries 数组，数量一般 10-12 个
|
||||
- 同时放入中文关键词、英文单词
|
||||
- 越靠近最新会话，生成关键词的比例或者权重应该更高
|
||||
- 例如：queries=['email', '邮件', 'folder',"preference"]
|
||||
- 越靠近最新会话，关键词比例或权重应更高
|
||||
|
||||
## 记忆写入
|
||||
|
||||
### 命名空间分类
|
||||
记忆必须使用以下命名空间之一：
|
||||
- `user` - 用户记忆：用户长期偏好、身份背景和历史协作信息
|
||||
- `semantic` - 语义记忆：结构化或非结构化知识内容
|
||||
- `episodic` - 情景记忆：历史对话、任务执行过程及关键事件
|
||||
- `skill` - 技能记忆：技能定义、工作流、工具调用策略及最佳实践
|
||||
- `environment` - 环境记忆：外部系统状态、运行环境配置和实时资源信息
|
||||
- `reflection` - 反思记忆：成功经验、失败原因和优化建议
|
||||
- `other` - 其他记忆：不属于以上分类的其他内容
|
||||
|
||||
### 写入规则
|
||||
- 写入或修改记忆时使用 memory_manage。
|
||||
- 遇到未来仍有用的信息时写入记忆：用户长期偏好、稳定事实、用户对你的纠正、持续任务或项目上下文、明确决策等。
|
||||
|
||||
### 【重要注意！】以下场景视为高价值加分，必须记录记忆
|
||||
- 用户多次跟你交互去优化输出
|
||||
## 写入规则
|
||||
遇到以下高价值信息时，使用 memory_manage 写入记忆：
|
||||
- 用户多次与你交互优化输出
|
||||
- 用户对你的纠正
|
||||
- 确定的事实，路径/地址/网址等
|
||||
- 用户独特的表达，缩写/非常规的表达
|
||||
- 因为你的错误，你道歉了
|
||||
- 用户说默认xxx的消息
|
||||
- 入口信息，比如链接、应用包名等
|
||||
- 确定的事实（路径/地址/网址等）
|
||||
- 用户独特的表达、缩写/非常规表达
|
||||
- 入口信息（链接、应用包名等）
|
||||
- 用户说"默认 xxx"的消息
|
||||
|
||||
### 注意
|
||||
- 如果你决定不再调用工具，则反思一下是否使用 memory_manage 保存记忆"#;
|
||||
## 注意
|
||||
- 如果决定不再调用工具，反思一下是否需要用 memory_manage 保存记忆"#;
|
||||
|
||||
/// skill_activate / skill_manage 工具使用说明
|
||||
const SKILL_TOOLS_INSTRUCTIONS: &str = r#"# 技能工具
|
||||
|
||||
## 技能存储路径
|
||||
- 项目级: `{project-root}/.picobot/skills/{skill-name}/SKILL.md`
|
||||
- 用户级: `~/.picobot/skills/{skill-name}/SKILL.md`
|
||||
|
||||
## 创建/修改技能
|
||||
- 必须使用 `skill_manage` 工具的 `create` 或 `update` action
|
||||
- 不要使用 `write` 工具直接写入技能文件
|
||||
- `skill_manage` 会自动创建正确的目录结构
|
||||
|
||||
## 使用技能
|
||||
## 使用规则
|
||||
- 技能名称不是工具名称，不能直接调用
|
||||
- 必须先调用 `skill_activate` 工具激活技能，再按指令执行
|
||||
- 一次只能激活一个技能，激活后会返回该技能的完整说明
|
||||
- 必须先调用 `skill_activate` 激活技能，再按返回的说明执行
|
||||
- 一次只能激活一个技能
|
||||
- 创建/修改技能必须用 `skill_manage` 的 `create` / `update` action，不要用 `write` 工具直接写技能文件
|
||||
|
||||
## 何时使用技能
|
||||
当满足以下条件时，应该使用技能：
|
||||
- 当前任务与某个技能的描述相匹配
|
||||
## 何时使用
|
||||
- 当前任务与某个技能的描述匹配
|
||||
- 需要执行特定领域的专业化工作流
|
||||
- 任务涉及多步骤操作，且有现成技能可用
|
||||
|
||||
## 如何使用技能
|
||||
1. **查看可用技能**: 浏览系统提示词中的 <available_skills> 列表
|
||||
2. **匹配任务**: 判断是否有技能的描述与当前任务匹配
|
||||
3. **激活技能**: 调用 `skill_activate` 工具，传入技能名称（name 参数）
|
||||
4. **执行指令**: 根据 skill_activate 返回的详细说明执行任务"#;
|
||||
- 任务涉及多步骤操作且有现成技能可用"#;
|
||||
|
||||
/// todo_write / todo_read 工具使用说明
|
||||
const TODO_WRITE_INSTRUCTIONS: &str = r#"# TodoWrite 工具
|
||||
|
||||
你可以使用 `todo_write` 工具在对话中维护结构化的任务列表。
|
||||
|
||||
## 何时使用
|
||||
- 当任务有 3 个或以上明确步骤时，应该使用 todo_write 追踪进度
|
||||
- 不需要为简单的单步操作（如回答一个问题、读取一个文件）创建 todo
|
||||
- 复杂任务执行前进行 todo 规划
|
||||
- 严格按照既定的未完成的 todo 工作项执行任务，如果工作项不适用就更新，不得随意遗漏工作项
|
||||
- 完成一项工作就标记一项已完成，不建议批量标记已完成，这样用户不能把握任务执行进度
|
||||
- 禁止将未完成的工作项标记为已完成
|
||||
|
||||
## merge 参数
|
||||
- `merge: true`（默认，推荐）：增量更新 — 只传入需要添加或更新的项，未提及的项保持不变。**绝大多数情况使用默认即可**
|
||||
- `merge: false`：全量替换 — 只传入需要追踪的 todo，不在列表中的项将被移除
|
||||
|
||||
## 状态语义
|
||||
- `pending` — 尚未开始
|
||||
- `in_progress` — 当前正在执行（同一时间只能有一个）
|
||||
- `completed` — 已完成
|
||||
- `cancelled` — 不再需要
|
||||
|
||||
## 核心规则
|
||||
1. 同一时间只能有一个任务处于 `in_progress` 状态
|
||||
2. 必须先完成当前 `in_progress` 的任务，再开始下一个
|
||||
3. `completed` 和 `cancelled` 的项可以重新激活（改回 `in_progress` 或 `pending`），用于任务返工或恢复
|
||||
4. `in_progress` 不能退回 `pending`，应直接标记为 `completed` 或 `cancelled`
|
||||
5. 不要先标记 completed 再去实际执行 — 先完成工作，再标记
|
||||
6. `content` 字段保持简洁、可执行
|
||||
7. **每个任务都必须传 `id`**。新任务由你生成一个短随机字符串作为 id（如 `"r9Tg8Kq2"`），更新任务时使用相同的 id。id 可以从之前 todo_write 返回的 `current_todos` 中获取
|
||||
|
||||
## 使用范例
|
||||
|
||||
创建新任务（生成随机 id）：
|
||||
```json
|
||||
{"merge": true, "todos": [{"id": "aB3kLm9x", "content": "修复登录 bug", "status": "in_progress"}]}
|
||||
```
|
||||
|
||||
追加新任务：
|
||||
```json
|
||||
{"merge": true, "todos": [{"id": "pQ7nWy2z", "content": "补充测试", "status": "pending"}]}
|
||||
```
|
||||
|
||||
更新已有任务（使用创建时的 id）：
|
||||
```json
|
||||
{"merge": true, "todos": [{"id": "aB3kLm9x", "content": "修复登录 bug", "status": "completed"}]}
|
||||
```
|
||||
|
||||
同时更新多项：
|
||||
```json
|
||||
{"merge": true, "todos": [
|
||||
{"id": "aB3kLm9x", "content": "修复登录 bug", "status": "completed"},
|
||||
{"id": "pQ7nWy2z", "content": "补充测试", "status": "in_progress"}
|
||||
]}
|
||||
```
|
||||
## 行为规则
|
||||
- 严格按照未完成的 todo 工作项执行任务，不适用就更新，不得随意遗漏
|
||||
- 完成一项就标记一项已完成，不建议批量标记（用户无法把握进度）
|
||||
- 禁止将未完成的工作项标记为已完成（先完成工作，再标记）
|
||||
- `in_progress` 不能退回 `pending`，应直接标记为 `completed` 或 `cancelled`
|
||||
- `completed` / `cancelled` 的项可重新激活用于返工或恢复
|
||||
|
||||
## 查询当前列表
|
||||
|
||||
使用 `todo_read` 工具查看当前任务列表，无需任何参数：
|
||||
```json
|
||||
{}
|
||||
```
|
||||
|
||||
在以下场景应主动调用 `todo_read`：
|
||||
- 对话开始时，检查是否有未完成的任务
|
||||
- 不确定当前任务状态时，先查询再操作
|
||||
- 完成一个任务后，查看剩余任务"#;
|
||||
- 对话开始时、不确定状态时、完成一项后，应主动调用 `todo_read` 查看任务列表"#;
|
||||
|
||||
/// shell / bash 工具使用说明
|
||||
const SHELL_TOOLS_INSTRUCTIONS: &str = r#"# Shell 交互终端
|
||||
|
||||
@ -8,6 +8,7 @@ pub mod command;
|
||||
pub mod config;
|
||||
pub mod domain;
|
||||
pub mod experts;
|
||||
pub mod frontmatter;
|
||||
pub mod gateway;
|
||||
pub mod logging;
|
||||
pub mod mcp;
|
||||
|
||||
@ -88,6 +88,9 @@ pub struct TodoItemSummary {
|
||||
pub id: String,
|
||||
pub content: String,
|
||||
pub status: String,
|
||||
pub priority: String,
|
||||
pub created_at: i64,
|
||||
pub updated_at: i64,
|
||||
pub created_by_message_id: Option<String>,
|
||||
}
|
||||
|
||||
|
||||
@ -394,6 +394,17 @@ impl OpenAIProvider {
|
||||
let status = resp.status();
|
||||
if !status.is_success() {
|
||||
let text = resp.text().await.unwrap_or_default();
|
||||
let sequence = format_message_sequence(&body);
|
||||
tracing::error!(
|
||||
provider = %self.name,
|
||||
model = %self.model_id,
|
||||
url = %url,
|
||||
status = %status,
|
||||
response_len = text.len(),
|
||||
response_body = %text,
|
||||
sequence = ?sequence,
|
||||
"OpenAI-compatible streaming API request failed"
|
||||
);
|
||||
return Err(format!("API error {}: {}", status, text).into());
|
||||
}
|
||||
|
||||
@ -672,6 +683,82 @@ impl OpenAIProvider {
|
||||
}
|
||||
}
|
||||
|
||||
// Forward-order check: verify tool messages IMMEDIATELY follow the
|
||||
// assistant(tool_calls). If any non-tool message appears between the
|
||||
// assistant and its tool results, the API rejects with
|
||||
// "insufficient tool messages following tool_calls message".
|
||||
//
|
||||
// The reverse scan above only checks existence (tool result appears
|
||||
// somewhere after assistant), NOT immediacy. This forward pass catches:
|
||||
// [assistant(tool_calls=[A]), user, tool(A)]
|
||||
// ^ reverse scan sees tool(A) after assistant → "resolved"
|
||||
// but API requires tool(A) to be IMMEDIATELY after assistant
|
||||
{
|
||||
let mut pending_tool_ids: std::collections::HashSet<&str> = std::collections::HashSet::new();
|
||||
let mut pending_assistant_idx: Option<usize> = None;
|
||||
|
||||
for (i, m) in request.messages.iter().enumerate() {
|
||||
// If we have pending tool_ids and encounter a non-tool message,
|
||||
// the assistant's tool results were NOT immediately following.
|
||||
if !pending_tool_ids.is_empty() && m.role != "tool" {
|
||||
if let Some(idx) = pending_assistant_idx {
|
||||
skip_assistant_indices.insert(idx);
|
||||
tracing::warn!(
|
||||
message_index = idx,
|
||||
interrupted_by_index = i,
|
||||
interrupted_by_role = %m.role,
|
||||
pending_tool_call_count = pending_tool_ids.len(),
|
||||
"build_request_body: assistant tool_calls not immediately \
|
||||
followed by tool results — stripping tool_calls"
|
||||
);
|
||||
// Remove this assistant's tool_call_ids from with_parent
|
||||
// so orphaned tool messages are dropped during serialization
|
||||
if let Some(calls) = &request.messages[idx].tool_calls {
|
||||
for tc in calls.iter() {
|
||||
with_parent.remove(tc.id.as_str());
|
||||
}
|
||||
}
|
||||
}
|
||||
pending_tool_ids.clear();
|
||||
pending_assistant_idx = None;
|
||||
}
|
||||
|
||||
if m.role == "assistant" {
|
||||
if let Some(ref calls) = m.tool_calls {
|
||||
if !calls.is_empty() && !skip_assistant_indices.contains(&i) {
|
||||
pending_tool_ids = calls.iter().map(|tc| tc.id.as_str()).collect();
|
||||
pending_assistant_idx = Some(i);
|
||||
}
|
||||
}
|
||||
} else if m.role == "tool" {
|
||||
if let Some(ref tc_id) = m.tool_call_id {
|
||||
pending_tool_ids.remove(tc_id.as_str());
|
||||
if pending_tool_ids.is_empty() {
|
||||
pending_assistant_idx = None;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Handle trailing assistant with unresolved immediate tool results
|
||||
if !pending_tool_ids.is_empty() {
|
||||
if let Some(idx) = pending_assistant_idx {
|
||||
skip_assistant_indices.insert(idx);
|
||||
tracing::warn!(
|
||||
message_index = idx,
|
||||
pending_tool_call_count = pending_tool_ids.len(),
|
||||
"build_request_body: trailing assistant tool_calls without \
|
||||
immediately following tool results — stripping tool_calls"
|
||||
);
|
||||
if let Some(calls) = &request.messages[idx].tool_calls {
|
||||
for tc in calls.iter() {
|
||||
with_parent.remove(tc.id.as_str());
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// valid_tool_call_parent_ids = with_parent (assistant tool_call_ids
|
||||
// whose parent assistant has ALL results after it)
|
||||
let valid_tool_call_parent_ids = &with_parent;
|
||||
@ -799,6 +886,38 @@ impl OpenAIProvider {
|
||||
}
|
||||
}
|
||||
|
||||
/// Builds a compact, human-readable summary of the message sequence in `body`
|
||||
/// for diagnostic logging. Only emitted on API errors (e.g. 400 responses) to
|
||||
/// avoid flooding logs on every request — see callers in `chat` and
|
||||
/// `chat_streaming_internal`.
|
||||
fn format_message_sequence(body: &Value) -> Vec<String> {
|
||||
body["messages"].as_array()
|
||||
.map(|msgs| msgs.iter().enumerate().map(|(i, m)| {
|
||||
let role = m.get("role").and_then(|r| r.as_str()).unwrap_or("?");
|
||||
match role {
|
||||
"assistant" => {
|
||||
let tc_count = m.get("tool_calls")
|
||||
.and_then(|t| t.as_array())
|
||||
.map(|a| a.len())
|
||||
.unwrap_or(0);
|
||||
if tc_count > 0 {
|
||||
format!("[{}] assistant(tool_calls={})", i, tc_count)
|
||||
} else {
|
||||
format!("[{}] assistant", i)
|
||||
}
|
||||
}
|
||||
"tool" => {
|
||||
let tcid = m.get("tool_call_id")
|
||||
.and_then(|t| t.as_str())
|
||||
.unwrap_or("??");
|
||||
format!("[{}] tool(id={})", i, tcid)
|
||||
}
|
||||
_ => format!("[{}] {}", i, role),
|
||||
}
|
||||
}).collect())
|
||||
.unwrap_or_default()
|
||||
}
|
||||
|
||||
#[derive(Deserialize)]
|
||||
struct OpenAIResponse {
|
||||
id: String,
|
||||
@ -945,6 +1064,7 @@ impl LLMProvider for OpenAIProvider {
|
||||
|
||||
// Debug: Log LLM response (only in debug builds)
|
||||
if !status.is_success() {
|
||||
let sequence = format_message_sequence(&body);
|
||||
tracing::error!(
|
||||
provider = %self.name,
|
||||
model = %self.model_id,
|
||||
@ -952,6 +1072,7 @@ impl LLMProvider for OpenAIProvider {
|
||||
status = %status,
|
||||
response_len = text.len(),
|
||||
response_body = %text,
|
||||
sequence = ?sequence,
|
||||
"OpenAI-compatible API request failed"
|
||||
);
|
||||
return Err(format!("API error {}: {}", status, text).into());
|
||||
@ -1496,4 +1617,131 @@ mod tests {
|
||||
// custom_param 应该保留
|
||||
assert_eq!(body["custom_param"], Value::String("value".to_string()));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_build_request_body_strips_tool_calls_when_not_immediately_followed() {
|
||||
// [assistant(tool_calls=[A]), user, tool(A)] → should strip tool_calls
|
||||
// The tool result exists but a user message interrupts between assistant
|
||||
// and tool result. The API would reject this with
|
||||
// "insufficient tool messages following tool_calls message".
|
||||
let provider = OpenAIProvider::new(
|
||||
"test".to_string(),
|
||||
"key".to_string(),
|
||||
"https://example.com/v1".to_string(),
|
||||
HashMap::new(),
|
||||
120,
|
||||
"gpt-test".to_string(),
|
||||
None,
|
||||
None,
|
||||
HashMap::new(),
|
||||
);
|
||||
|
||||
let request = ChatCompletionRequest {
|
||||
messages: vec![
|
||||
Message {
|
||||
role: "assistant".to_string(),
|
||||
content: vec![ContentBlock::text("calling tool")],
|
||||
reasoning_content: None,
|
||||
tool_call_id: None,
|
||||
name: None,
|
||||
tool_calls: Some(vec![ToolCall {
|
||||
id: "call_A".to_string(),
|
||||
name: "search".to_string(),
|
||||
arguments: json!({}),
|
||||
}]),
|
||||
},
|
||||
Message {
|
||||
role: "user".to_string(),
|
||||
content: vec![ContentBlock::text("interrupting message")],
|
||||
reasoning_content: None,
|
||||
tool_call_id: None,
|
||||
name: None,
|
||||
tool_calls: None,
|
||||
},
|
||||
Message {
|
||||
role: "tool".to_string(),
|
||||
content: vec![ContentBlock::text("result")],
|
||||
reasoning_content: None,
|
||||
tool_call_id: Some("call_A".to_string()),
|
||||
name: Some("search".to_string()),
|
||||
tool_calls: None,
|
||||
},
|
||||
],
|
||||
temperature: None,
|
||||
max_tokens: None,
|
||||
tools: None,
|
||||
};
|
||||
|
||||
let body = provider.build_request_body(&request);
|
||||
let messages = body["messages"].as_array().unwrap();
|
||||
|
||||
// Assistant should NOT have tool_calls (stripped because not immediately followed)
|
||||
assert!(
|
||||
messages[0].get("tool_calls").is_none(),
|
||||
"tool_calls should be stripped when tool results are not immediately following"
|
||||
);
|
||||
// Tool message should be dropped (orphaned after stripping)
|
||||
assert_eq!(
|
||||
messages.len(),
|
||||
2,
|
||||
"tool message should be dropped as orphaned, got {} messages",
|
||||
messages.len()
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_build_request_body_preserves_tool_calls_when_immediately_followed() {
|
||||
// [assistant(tool_calls=[A]), tool(A)] → should keep tool_calls (valid sequence)
|
||||
let provider = OpenAIProvider::new(
|
||||
"test".to_string(),
|
||||
"key".to_string(),
|
||||
"https://example.com/v1".to_string(),
|
||||
HashMap::new(),
|
||||
120,
|
||||
"gpt-test".to_string(),
|
||||
None,
|
||||
None,
|
||||
HashMap::new(),
|
||||
);
|
||||
|
||||
let request = ChatCompletionRequest {
|
||||
messages: vec![
|
||||
Message {
|
||||
role: "assistant".to_string(),
|
||||
content: vec![ContentBlock::text("calling tool")],
|
||||
reasoning_content: None,
|
||||
tool_call_id: None,
|
||||
name: None,
|
||||
tool_calls: Some(vec![ToolCall {
|
||||
id: "call_A".to_string(),
|
||||
name: "search".to_string(),
|
||||
arguments: json!({}),
|
||||
}]),
|
||||
},
|
||||
Message {
|
||||
role: "tool".to_string(),
|
||||
content: vec![ContentBlock::text("result")],
|
||||
reasoning_content: None,
|
||||
tool_call_id: Some("call_A".to_string()),
|
||||
name: Some("search".to_string()),
|
||||
tool_calls: None,
|
||||
},
|
||||
],
|
||||
temperature: None,
|
||||
max_tokens: None,
|
||||
tools: None,
|
||||
};
|
||||
|
||||
let body = provider.build_request_body(&request);
|
||||
let messages = body["messages"].as_array().unwrap();
|
||||
|
||||
// Assistant should keep tool_calls (valid immediate sequence)
|
||||
let tool_calls = messages[0].get("tool_calls")
|
||||
.and_then(|t| t.as_array())
|
||||
.expect("tool_calls should be preserved when immediately followed");
|
||||
assert_eq!(tool_calls.len(), 1);
|
||||
// Tool message should be present
|
||||
assert_eq!(messages.len(), 2);
|
||||
assert_eq!(messages[1]["role"], "tool");
|
||||
}
|
||||
}
|
||||
|
||||
@ -867,11 +867,14 @@ struct SkillFrontmatter {
|
||||
fn parse_skill_file(path: &Path, source: SkillSource) -> Result<Skill, String> {
|
||||
let content = fs::read_to_string(path).map_err(|e| format!("failed to read file: {}", e))?;
|
||||
|
||||
let (frontmatter_raw, body) =
|
||||
split_frontmatter(&content).ok_or_else(|| "missing YAML frontmatter block".to_string())?;
|
||||
|
||||
let frontmatter: SkillFrontmatter = serde_yaml::from_str(frontmatter_raw)
|
||||
.map_err(|e| format!("invalid YAML frontmatter: {}", e))?;
|
||||
let (frontmatter, body) = match crate::frontmatter::parse::<SkillFrontmatter>(&content) {
|
||||
Ok(v) => v,
|
||||
Err(err) => {
|
||||
let bytes = content.len();
|
||||
let crlf = content.contains('\r');
|
||||
return Err(format!("{} (bytes={}, crlf={})", err, bytes, crlf));
|
||||
}
|
||||
};
|
||||
|
||||
let description = frontmatter.description.trim();
|
||||
if description.is_empty() {
|
||||
@ -895,15 +898,6 @@ fn parse_skill_file(path: &Path, source: SkillSource) -> Result<Skill, String> {
|
||||
})
|
||||
}
|
||||
|
||||
fn split_frontmatter(content: &str) -> Option<(&str, &str)> {
|
||||
let rest = content.strip_prefix("---\n")?;
|
||||
let marker = "\n---\n";
|
||||
let idx = rest.find(marker)?;
|
||||
let frontmatter = &rest[..idx];
|
||||
let body = &rest[idx + marker.len()..];
|
||||
Some((frontmatter, body))
|
||||
}
|
||||
|
||||
// 使用 platform 模块提供的 xml_escape 和 path_to_uri 函数
|
||||
|
||||
// SkillPromptProvider 实现
|
||||
@ -999,11 +993,17 @@ mod tests {
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_split_frontmatter() {
|
||||
let input = "---\ndescription: demo\n---\nhello";
|
||||
let (fm, body) = split_frontmatter(input).unwrap();
|
||||
assert!(fm.contains("description"));
|
||||
assert_eq!(body, "hello");
|
||||
fn test_parse_skill_file_handles_crlf_endings() {
|
||||
let dir = tempfile::tempdir().unwrap();
|
||||
let skill_dir = dir.path().join("demo");
|
||||
fs::create_dir_all(&skill_dir).unwrap();
|
||||
let skill_md = skill_dir.join("SKILL.md");
|
||||
fs::write(&skill_md, "---\r\ndescription: demo skill\r\n---\r\nStep A\r\nStep B").unwrap();
|
||||
|
||||
let skill = parse_skill_file(&skill_md, SkillSource::Project).unwrap();
|
||||
assert_eq!(skill.name, "demo");
|
||||
assert_eq!(skill.description, "demo skill");
|
||||
assert_eq!(skill.body, "Step A\nStep B");
|
||||
}
|
||||
|
||||
#[test]
|
||||
|
||||
@ -875,6 +875,74 @@ impl SessionStore {
|
||||
Ok(())
|
||||
}
|
||||
|
||||
/// Replace the entire history for a specific topic.
|
||||
///
|
||||
/// Deletes only messages belonging to the given topic_id (preserving
|
||||
/// other topics' messages), then inserts the new messages with topic_id
|
||||
/// set correctly. Used by the compressor when it has produced a
|
||||
/// complete, validated message list for a single topic.
|
||||
///
|
||||
/// Seq numbers continue from the current session-wide max (not reset to
|
||||
/// 1) so we don't collide with other topics' messages. Gaps in seq
|
||||
/// (from the deleted old messages) are harmless — per-topic loading
|
||||
/// orders by seq and gaps don't affect ordering.
|
||||
pub fn replace_topic_history(
|
||||
&self,
|
||||
session_id: &str,
|
||||
topic_id: &str,
|
||||
messages: &[ChatMessage],
|
||||
) -> Result<(), StorageError> {
|
||||
let conn = self.pool.get()?;
|
||||
let tx = conn.unchecked_transaction()?;
|
||||
let now = current_timestamp();
|
||||
|
||||
// Delete only messages belonging to this topic — other topics'
|
||||
// messages are preserved (the pre-existing `replace_active_history`
|
||||
// clobbered the entire session, which broke multi-topic isolation).
|
||||
tx.execute(
|
||||
"DELETE FROM messages WHERE session_id = ?1 AND topic_id = ?2",
|
||||
params![session_id, topic_id],
|
||||
)?;
|
||||
|
||||
// Continue seq from the session-wide max so we don't violate
|
||||
// UNIQUE(session_id, seq). Other topics' messages keep their seqs.
|
||||
let start_seq: i64 = tx.query_row(
|
||||
"SELECT COALESCE(MAX(seq), 0) + 1 FROM messages WHERE session_id = ?1",
|
||||
params![session_id],
|
||||
|row| row.get(0),
|
||||
)?;
|
||||
|
||||
for (i, message) in messages.iter().enumerate() {
|
||||
let seq = start_seq + i as i64;
|
||||
insert_message_with_topic_seq(&tx, session_id, topic_id, seq, message)?;
|
||||
}
|
||||
|
||||
// Update this topic's message_count and timestamps.
|
||||
tx.execute(
|
||||
"UPDATE topics SET message_count = ?2, last_active_at = ?3, updated_at = ?3 WHERE id = ?1",
|
||||
params![topic_id, messages.len() as i64, now],
|
||||
)?;
|
||||
|
||||
// Recompute session-wide counts from the messages table so they stay
|
||||
// consistent after a partial replacement (we only touched one topic,
|
||||
// so we can't just set the session count to `messages.len()`).
|
||||
let (total_count, user_turn_count): (i64, i64) = tx.query_row(
|
||||
"SELECT COUNT(*), COALESCE(SUM(CASE WHEN role = 'user' THEN 1 ELSE 0 END), 0) \
|
||||
FROM messages WHERE session_id = ?1",
|
||||
params![session_id],
|
||||
|row| Ok((row.get(0)?, row.get(1)?)),
|
||||
)?;
|
||||
tx.execute(
|
||||
"UPDATE sessions SET message_count = ?2, user_turn_count = ?3, \
|
||||
updated_at = ?4, last_active_at = ?4, archived_at = NULL \
|
||||
WHERE id = ?1 AND deleted_at IS NULL",
|
||||
params![session_id, total_count, user_turn_count, now],
|
||||
)?;
|
||||
|
||||
tx.commit()?;
|
||||
Ok(())
|
||||
}
|
||||
|
||||
pub fn mark_agent_prompt_reinjected(&self, session_id: &str) -> Result<(), StorageError> {
|
||||
let now = current_timestamp();
|
||||
let conn = self.pool.get()?;
|
||||
@ -1667,6 +1735,51 @@ fn insert_message_with_seq(
|
||||
Ok(())
|
||||
}
|
||||
|
||||
/// Insert a message with an explicit `topic_id` and `seq`.
|
||||
///
|
||||
/// Used by `replace_topic_history` to insert compressed messages while
|
||||
/// preserving topic association (the plain `insert_message_with_seq` would
|
||||
/// set topic_id to NULL).
|
||||
fn insert_message_with_topic_seq(
|
||||
conn: &rusqlite::Transaction<'_>,
|
||||
session_id: &str,
|
||||
topic_id: &str,
|
||||
seq: i64,
|
||||
message: &ChatMessage,
|
||||
) -> Result<(), StorageError> {
|
||||
let media_refs_json = serde_json::to_string(&message.media_refs)?;
|
||||
let tool_calls_json = message
|
||||
.tool_calls
|
||||
.as_ref()
|
||||
.map(serde_json::to_string)
|
||||
.transpose()?;
|
||||
conn.execute(
|
||||
"
|
||||
INSERT INTO messages (
|
||||
id, session_id, topic_id, seq, role, content,
|
||||
system_context, reasoning_content, media_refs_json, tool_call_id, tool_name, tool_calls_json, tool_duration_ms, created_at
|
||||
) VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8, ?9, ?10, ?11, ?12, ?13, ?14)
|
||||
",
|
||||
params![
|
||||
message.id,
|
||||
session_id,
|
||||
topic_id,
|
||||
seq,
|
||||
message.role,
|
||||
message.content,
|
||||
message.system_context,
|
||||
message.reasoning_content,
|
||||
media_refs_json,
|
||||
message.tool_call_id,
|
||||
message.tool_name,
|
||||
tool_calls_json,
|
||||
message.tool_duration_ms.map(|v| v as i64),
|
||||
message.timestamp,
|
||||
],
|
||||
)?;
|
||||
Ok(())
|
||||
}
|
||||
|
||||
fn clone_message_for_compaction(message: &ChatMessage, timestamp: i64) -> ChatMessage {
|
||||
ChatMessage {
|
||||
id: uuid::Uuid::new_v4().to_string(),
|
||||
|
||||
@ -63,6 +63,17 @@ pub trait ConversationRepository: Send + Sync + 'static {
|
||||
session_id: &str,
|
||||
messages: &[ChatMessage],
|
||||
) -> Result<(), StorageError>;
|
||||
|
||||
/// Replace the entire history for a specific topic.
|
||||
/// Deletes only messages belonging to the given topic_id, then inserts
|
||||
/// the new messages with topic_id set correctly. Used by compressor when
|
||||
/// it has produced a complete, validated message list for a single topic.
|
||||
fn replace_topic_history(
|
||||
&self,
|
||||
session_id: &str,
|
||||
topic_id: &str,
|
||||
messages: &[ChatMessage],
|
||||
) -> Result<(), StorageError>;
|
||||
}
|
||||
|
||||
pub trait PromptInjectionRepository: Send + Sync + 'static {
|
||||
@ -252,6 +263,15 @@ impl ConversationRepository for super::SessionStore {
|
||||
) -> Result<(), StorageError> {
|
||||
super::SessionStore::replace_active_history(self, session_id, messages)
|
||||
}
|
||||
|
||||
fn replace_topic_history(
|
||||
&self,
|
||||
session_id: &str,
|
||||
topic_id: &str,
|
||||
messages: &[ChatMessage],
|
||||
) -> Result<(), StorageError> {
|
||||
super::SessionStore::replace_topic_history(self, session_id, topic_id, messages)
|
||||
}
|
||||
}
|
||||
|
||||
impl PromptInjectionRepository for super::SessionStore {
|
||||
|
||||
@ -33,11 +33,31 @@ impl FileEditTool {
|
||||
|
||||
// Check directory restriction
|
||||
if let Some(ref allowed) = self.allowed_dir {
|
||||
let allowed_path = Path::new(allowed);
|
||||
if !resolved.starts_with(allowed_path) {
|
||||
// canonicalize both paths to resolve symlinks and prevent path traversal
|
||||
// via symlinks inside allowed_dir pointing outside.
|
||||
// For edit tool the target file may not exist yet; fall back to
|
||||
// canonicalizing the parent directory.
|
||||
let canonical_allowed = std::fs::canonicalize(allowed)
|
||||
.map_err(|e| format!("Failed to canonicalize allowed dir '{}': {}", allowed, e))?;
|
||||
let canonical_resolved = match std::fs::canonicalize(&resolved) {
|
||||
Ok(c) => c,
|
||||
Err(_) => {
|
||||
// File doesn't exist yet; canonicalize parent directory
|
||||
let parent = resolved.parent().ok_or_else(|| {
|
||||
format!("Path '{}' has no parent directory", path)
|
||||
})?;
|
||||
let canonical_parent = std::fs::canonicalize(parent).map_err(|e| {
|
||||
format!("Failed to canonicalize parent directory of '{}': {}", path, e)
|
||||
})?;
|
||||
canonical_parent.join(resolved.file_name().ok_or_else(|| {
|
||||
format!("Path '{}' has no file name component", path)
|
||||
})?)
|
||||
}
|
||||
};
|
||||
if !canonical_resolved.starts_with(&canonical_allowed) {
|
||||
return Err(format!(
|
||||
"Path '{}' is outside allowed directory '{}'",
|
||||
path, allowed
|
||||
"Path '{}' (resolves to '{}') is outside allowed directory '{}'",
|
||||
path, canonical_resolved.display(), canonical_allowed.display()
|
||||
));
|
||||
}
|
||||
}
|
||||
|
||||
@ -37,11 +37,18 @@ impl FileReadTool {
|
||||
|
||||
// Check directory restriction
|
||||
if let Some(ref allowed) = self.allowed_dir {
|
||||
let allowed_path = Path::new(allowed);
|
||||
if !resolved.starts_with(allowed_path) {
|
||||
// canonicalize both paths to resolve symlinks and prevent path traversal
|
||||
// via symlinks inside allowed_dir pointing outside
|
||||
let canonical_allowed = std::fs::canonicalize(allowed)
|
||||
.map_err(|e| format!("Failed to canonicalize allowed dir '{}': {}", allowed, e))?;
|
||||
// For read tool, file must exist; canonicalize will fail for non-existent paths
|
||||
// which is acceptable (returns error)
|
||||
let canonical_resolved = std::fs::canonicalize(&resolved)
|
||||
.map_err(|e| format!("Failed to canonicalize path '{}': {}", path, e))?;
|
||||
if !canonical_resolved.starts_with(&canonical_allowed) {
|
||||
return Err(format!(
|
||||
"Path '{}' is outside allowed directory '{}'",
|
||||
path, allowed
|
||||
"Path '{}' (resolves to '{}') is outside allowed directory '{}'",
|
||||
path, canonical_resolved.display(), canonical_allowed.display()
|
||||
));
|
||||
}
|
||||
}
|
||||
|
||||
@ -32,11 +32,31 @@ impl FileWriteTool {
|
||||
|
||||
// Check directory restriction
|
||||
if let Some(ref allowed) = self.allowed_dir {
|
||||
let allowed_path = Path::new(allowed);
|
||||
if !resolved.starts_with(allowed_path) {
|
||||
// canonicalize both paths to resolve symlinks and prevent path traversal
|
||||
// via symlinks inside allowed_dir pointing outside.
|
||||
// For write tool the target file may not exist yet; fall back to
|
||||
// canonicalizing the parent directory.
|
||||
let canonical_allowed = std::fs::canonicalize(allowed)
|
||||
.map_err(|e| format!("Failed to canonicalize allowed dir '{}': {}", allowed, e))?;
|
||||
let canonical_resolved = match std::fs::canonicalize(&resolved) {
|
||||
Ok(c) => c,
|
||||
Err(_) => {
|
||||
// File doesn't exist yet; canonicalize parent directory
|
||||
let parent = resolved.parent().ok_or_else(|| {
|
||||
format!("Path '{}' has no parent directory", path)
|
||||
})?;
|
||||
let canonical_parent = std::fs::canonicalize(parent).map_err(|e| {
|
||||
format!("Failed to canonicalize parent directory of '{}': {}", path, e)
|
||||
})?;
|
||||
canonical_parent.join(resolved.file_name().ok_or_else(|| {
|
||||
format!("Path '{}' has no file name component", path)
|
||||
})?)
|
||||
}
|
||||
};
|
||||
if !canonical_resolved.starts_with(&canonical_allowed) {
|
||||
return Err(format!(
|
||||
"Path '{}' is outside allowed directory '{}'",
|
||||
path, allowed
|
||||
"Path '{}' (resolves to '{}') is outside allowed directory '{}'",
|
||||
path, canonical_resolved.display(), canonical_allowed.display()
|
||||
));
|
||||
}
|
||||
}
|
||||
|
||||
@ -311,6 +311,7 @@ impl Tool for HttpRequestTool {
|
||||
|
||||
let client = match reqwest::Client::builder()
|
||||
.timeout(Duration::from_secs(self.timeout_secs))
|
||||
.redirect(reqwest::redirect::Policy::none())
|
||||
.build()
|
||||
{
|
||||
Ok(c) => c,
|
||||
|
||||
@ -87,6 +87,21 @@ impl ToolRegistry {
|
||||
*new_registry.tools.write().expect("ToolRegistry lock poisoned") = filtered;
|
||||
new_registry
|
||||
}
|
||||
|
||||
/// 创建一个仅包含指定工具的新 registry 副本（白名单）。
|
||||
/// include 中不存在于当前 registry 的名称会被静默跳过（取交集语义）。
|
||||
pub fn only(&self, include: &[&str]) -> Self {
|
||||
let include_set: std::collections::HashSet<&str> = include.iter().copied().collect();
|
||||
let tools = self.tools.read().expect("ToolRegistry lock poisoned");
|
||||
let filtered: HashMap<String, Arc<dyn ToolTrait>> = tools
|
||||
.iter()
|
||||
.filter(|(name, _)| include_set.contains(name.as_str()))
|
||||
.map(|(k, v)| (k.clone(), v.clone()))
|
||||
.collect();
|
||||
let new_registry = ToolRegistry::new();
|
||||
*new_registry.tools.write().expect("ToolRegistry lock poisoned") = filtered;
|
||||
new_registry
|
||||
}
|
||||
}
|
||||
|
||||
impl Default for ToolRegistry {
|
||||
@ -94,3 +109,82 @@ impl Default for ToolRegistry {
|
||||
Self::new()
|
||||
}
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
use crate::tools::traits::ToolResult;
|
||||
use async_trait::async_trait;
|
||||
|
||||
/// 仅用于测试的占位工具，按构造名注册
|
||||
struct FakeTool {
|
||||
tool_name: String,
|
||||
}
|
||||
|
||||
#[async_trait]
|
||||
impl ToolTrait for FakeTool {
|
||||
fn name(&self) -> &str {
|
||||
&self.tool_name
|
||||
}
|
||||
fn description(&self) -> &str {
|
||||
"fake"
|
||||
}
|
||||
fn parameters_schema(&self) -> serde_json::Value {
|
||||
serde_json::json!({})
|
||||
}
|
||||
async fn execute(&self, _args: serde_json::Value) -> anyhow::Result<ToolResult> {
|
||||
Ok(ToolResult {
|
||||
success: true,
|
||||
output: String::new(),
|
||||
error: None,
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
fn registry_with(names: &[&str]) -> ToolRegistry {
|
||||
let reg = ToolRegistry::new();
|
||||
for n in names {
|
||||
reg.register(FakeTool {
|
||||
tool_name: n.to_string(),
|
||||
});
|
||||
}
|
||||
reg
|
||||
}
|
||||
|
||||
fn sorted_names(reg: &ToolRegistry) -> Vec<String> {
|
||||
let mut v = reg.tool_names();
|
||||
v.sort();
|
||||
v
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn only_keeps_listed_tools() {
|
||||
let reg = registry_with(&["read", "edit", "write", "bash"]);
|
||||
let filtered = reg.only(&["read", "bash"]);
|
||||
assert_eq!(sorted_names(&filtered), vec!["bash", "read"]);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn only_silently_skips_missing_names() {
|
||||
let reg = registry_with(&["read", "edit"]);
|
||||
let filtered = reg.only(&["read", "nonexistent", "glob"]);
|
||||
assert_eq!(sorted_names(&filtered), vec!["read"]);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn only_with_empty_include_returns_empty() {
|
||||
let reg = registry_with(&["read", "edit"]);
|
||||
let filtered = reg.only(&[]);
|
||||
assert!(filtered.tool_names().is_empty());
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn only_does_not_mutate_source() {
|
||||
let reg = registry_with(&["read", "edit", "write"]);
|
||||
let _ = reg.only(&["read"]);
|
||||
// 源 registry 不受影响
|
||||
let mut v = reg.tool_names();
|
||||
v.sort();
|
||||
assert_eq!(v, vec!["edit", "read", "write"]);
|
||||
}
|
||||
}
|
||||
|
||||
@ -102,6 +102,7 @@ mod tests {
|
||||
prompt_template: "任务: {{description}}\n指令: {{prompt}}".to_string(),
|
||||
body: None,
|
||||
allowed_tools: None,
|
||||
denied_tools: None,
|
||||
max_execution_secs: None,
|
||||
source: SubagentSource::Builtin,
|
||||
path: None,
|
||||
|
||||
@ -29,8 +29,6 @@ pub struct SubAgentRuntimeConfig {
|
||||
pub default_allowed_tools: HashSet<String>,
|
||||
/// 默认最大执行时间（秒）
|
||||
pub default_max_execution_secs: u64,
|
||||
/// Explore 类型的最大执行时间（秒）
|
||||
pub explore_max_execution_secs: u64,
|
||||
/// 任务 TTL（小时）
|
||||
pub ttl_hours: u64,
|
||||
/// 技能索引（可选，预生成的技能列表字符串）
|
||||
@ -57,7 +55,6 @@ impl Default for SubAgentRuntimeConfig {
|
||||
"send_session_message".to_string(), // 用于进度通知
|
||||
]),
|
||||
default_max_execution_secs: 3600, // 60分钟
|
||||
explore_max_execution_secs: 3600, // 60分钟
|
||||
ttl_hours: 24,
|
||||
skills_index: None,
|
||||
max_nesting_depth: 1,
|
||||
@ -394,23 +391,80 @@ impl DefaultSubAgentRuntime {
|
||||
.unwrap_or(self.config.default_max_execution_secs)
|
||||
}
|
||||
|
||||
/// 根据 def 与嵌套深度构建子代理工具集。
|
||||
/// 过滤顺序：base → allowed_tools 白名单 → denied_tools 黑名单 + depth 达到上限移除 task。
|
||||
/// - `allowed_tools` 为 Some 时取交集（白名单），None 表示不限制。
|
||||
/// - `denied_tools` 为 Some 时扣除（黑名单），在白名单之后应用。
|
||||
/// - 当 child_depth >= max_nesting_depth 时移除 task 工具（防无限嵌套的安全兜底，
|
||||
/// 不可被 def 覆盖）。默认 max_nesting_depth=2，即孙代理（depth=2）无法再创建子代理。
|
||||
fn build_subagent_tools_registry(
|
||||
&self,
|
||||
def: Option<&SubagentDef>,
|
||||
child_depth: u32,
|
||||
) -> Arc<ToolRegistry> {
|
||||
let depth_deny_task = child_depth >= self.config.max_nesting_depth;
|
||||
let allowed: Option<&Vec<String>> = def.and_then(|d| d.allowed_tools.as_ref());
|
||||
let denied_tools: Option<&Vec<String>> = def.and_then(|d| d.denied_tools.as_ref());
|
||||
|
||||
// 快速路径：无白名单、无黑名单、无需 depth 兜底 → 直接复用 Arc（避免拷贝）
|
||||
if allowed.is_none() && denied_tools.is_none() && !depth_deny_task {
|
||||
return self.subagent_tools.clone();
|
||||
}
|
||||
|
||||
Arc::new(Self::filter_tool_registry(
|
||||
&self.subagent_tools,
|
||||
allowed,
|
||||
denied_tools,
|
||||
depth_deny_task,
|
||||
))
|
||||
}
|
||||
|
||||
/// 纯函数：在 base 之上应用白名单/黑名单/depth 规则。
|
||||
/// 抽取为关联函数便于单元测试（无需构造整个 DefaultSubAgentRuntime）。
|
||||
fn filter_tool_registry(
|
||||
base: &ToolRegistry,
|
||||
allowed: Option<&Vec<String>>,
|
||||
denied_tools: Option<&Vec<String>>,
|
||||
depth_deny_task: bool,
|
||||
) -> ToolRegistry {
|
||||
// 1. 应用白名单（若存在），否则取得 owned 副本以便后续黑名单过滤
|
||||
let tools: ToolRegistry = match allowed {
|
||||
Some(list) => {
|
||||
let refs: Vec<&str> = list.iter().map(|s| s.as_str()).collect();
|
||||
base.only(&refs)
|
||||
}
|
||||
None => base.without(&[]),
|
||||
};
|
||||
|
||||
// 2. 合并黑名单（depth 规则 + denied_tools）
|
||||
let mut denied: Vec<&str> = Vec::new();
|
||||
if depth_deny_task {
|
||||
denied.push(TaskTool::TOOL_NAME);
|
||||
}
|
||||
if let Some(dt) = denied_tools {
|
||||
denied.extend(dt.iter().map(|s| s.as_str()));
|
||||
}
|
||||
|
||||
if denied.is_empty() {
|
||||
tools
|
||||
} else {
|
||||
tools.without(&denied)
|
||||
}
|
||||
}
|
||||
|
||||
/// 创建子代理实例
|
||||
fn create_subagent(
|
||||
&self,
|
||||
session: &TaskSession,
|
||||
system_prompt: String,
|
||||
def: Option<&SubagentDef>,
|
||||
parent_nesting_depth: u32,
|
||||
parent_task_id: Option<String>,
|
||||
) -> Result<AgentLoop, TaskError> {
|
||||
let prompt_provider = Arc::new(StaticSystemPromptProvider::new(system_prompt));
|
||||
|
||||
// 孙智能体（depth >= 2）不注册 task 工具，防止无限嵌套
|
||||
let child_depth = parent_nesting_depth + 1;
|
||||
let tools = if child_depth >= 2 {
|
||||
Arc::new(self.subagent_tools.without(&[TaskTool::TOOL_NAME]))
|
||||
} else {
|
||||
self.subagent_tools.clone()
|
||||
};
|
||||
let tools = self.build_subagent_tools_registry(def, child_depth);
|
||||
|
||||
AgentLoop::with_tools_and_system_prompt_provider(
|
||||
AgentRuntimeConfig::from(self.provider_config.clone()),
|
||||
@ -481,11 +535,7 @@ impl DefaultSubAgentRuntime {
|
||||
};
|
||||
|
||||
// 设置超时
|
||||
let max_secs = if session.subagent_type == "explore" {
|
||||
self.config.explore_max_execution_secs
|
||||
} else {
|
||||
self.effective_max_execution_secs(def)
|
||||
};
|
||||
let max_secs = self.effective_max_execution_secs(def);
|
||||
let timeout_duration = Duration::from_secs(max_secs);
|
||||
|
||||
let result = tokio::time::timeout(
|
||||
@ -663,7 +713,7 @@ impl SubAgentRuntime for DefaultSubAgentRuntime {
|
||||
);
|
||||
|
||||
// 7. 创建子代理
|
||||
let agent = self.create_subagent(&session, system_prompt, parent_context.nesting_depth, parent_context.task_id.clone())?;
|
||||
let agent = self.create_subagent(&session, system_prompt, Some(&def), parent_context.nesting_depth, parent_context.task_id.clone())?;
|
||||
|
||||
// 8. 执行任务
|
||||
let result = self
|
||||
@ -747,8 +797,16 @@ impl SubAgentRuntime for DefaultSubAgentRuntime {
|
||||
&additional_prompt,
|
||||
);
|
||||
|
||||
// 4.1 重新解析 def 以应用工具过滤。
|
||||
// 安全要求：def 被删除/禁用时必须失败恢复，而不是降级为完整工具集——
|
||||
// 否则一个受限子代理（如 allowed_tools: [read]）在 def 失踪后会获得全部工具，
|
||||
// 构成权限提升。与 spawn 保持一致：def 不可用即拒绝执行。
|
||||
let def = self
|
||||
.find_subagent_def(&session.subagent_type)
|
||||
.map_err(TaskError::InvalidArguments)?;
|
||||
|
||||
// 5. 创建子代理
|
||||
let agent = self.create_subagent(&session, system_prompt, parent_context.nesting_depth, parent_context.task_id.clone())?;
|
||||
let agent = self.create_subagent(&session, system_prompt, Some(&def), parent_context.nesting_depth, parent_context.task_id.clone())?;
|
||||
|
||||
// 6. 使用历史继续执行
|
||||
let result = self
|
||||
@ -809,7 +867,6 @@ impl SubagentCatalog {
|
||||
pub fn new() -> Self {
|
||||
let mut catalog = Self::default();
|
||||
catalog.register(SubagentDef::builtin_general());
|
||||
catalog.register(SubagentDef::builtin_explore());
|
||||
catalog
|
||||
}
|
||||
|
||||
@ -826,7 +883,6 @@ impl SubagentCatalog {
|
||||
// 先内置作为基础
|
||||
let mut merged: std::collections::HashMap<String, SubagentDef> = std::collections::HashMap::new();
|
||||
merged.insert("general".to_string(), SubagentDef::builtin_general());
|
||||
merged.insert("explore".to_string(), SubagentDef::builtin_explore());
|
||||
|
||||
tracing::debug!(cwd = %cwd.display(), "Discovering subagents from cwd");
|
||||
|
||||
@ -960,6 +1016,12 @@ pub struct SubagentWithStatus {
|
||||
pub source: String,
|
||||
/// Which scopes have this subagent disabled. Empty means enabled.
|
||||
pub disabled_in_scopes: Vec<String>,
|
||||
/// 工具白名单（None 表示不过滤）
|
||||
#[serde(default, skip_serializing_if = "Option::is_none")]
|
||||
pub allowed_tools: Option<Vec<String>>,
|
||||
/// 工具黑名单（None 表示不过滤）
|
||||
#[serde(default, skip_serializing_if = "Option::is_none")]
|
||||
pub denied_tools: Option<Vec<String>>,
|
||||
}
|
||||
|
||||
#[derive(Debug, Clone)]
|
||||
@ -1105,6 +1167,8 @@ impl SubagentRuntime {
|
||||
description: def.description.clone(),
|
||||
source: def.source.as_str().to_string(),
|
||||
disabled_in_scopes: scopes.iter().map(|s| s.as_str().to_string()).collect(),
|
||||
allowed_tools: def.allowed_tools.clone(),
|
||||
denied_tools: def.denied_tools.clone(),
|
||||
}
|
||||
})
|
||||
.collect();
|
||||
@ -1341,6 +1405,8 @@ struct SubagentFrontmatter {
|
||||
#[serde(default)]
|
||||
allowed_tools: Option<Vec<String>>,
|
||||
#[serde(default)]
|
||||
denied_tools: Option<Vec<String>>,
|
||||
#[serde(default)]
|
||||
max_execution_secs: Option<u64>,
|
||||
}
|
||||
|
||||
@ -1401,11 +1467,14 @@ fn parse_subagent_file(path: &Path, source: SubagentSource) -> Result<SubagentDe
|
||||
let content = fs::read_to_string(path)
|
||||
.map_err(|e| format!("failed to read file: {}", e))?;
|
||||
|
||||
let (frontmatter_raw, body) = split_frontmatter(&content)
|
||||
.ok_or_else(|| "missing YAML frontmatter block".to_string())?;
|
||||
|
||||
let frontmatter: SubagentFrontmatter = serde_yaml::from_str(frontmatter_raw)
|
||||
.map_err(|e| format!("invalid YAML frontmatter: {}", e))?;
|
||||
let (frontmatter, body) = match crate::frontmatter::parse::<SubagentFrontmatter>(&content) {
|
||||
Ok(v) => v,
|
||||
Err(err) => {
|
||||
let bytes = content.len();
|
||||
let crlf = content.contains('\r');
|
||||
return Err(format!("{} (bytes={}, crlf={})", err, bytes, crlf));
|
||||
}
|
||||
};
|
||||
|
||||
if frontmatter.description.trim().is_empty() {
|
||||
return Err("description is required and cannot be empty".to_string());
|
||||
@ -1428,43 +1497,13 @@ fn parse_subagent_file(path: &Path, source: SubagentSource) -> Result<SubagentDe
|
||||
prompt_template,
|
||||
body: if body_content.is_empty() { None } else { Some(body_content) },
|
||||
allowed_tools: frontmatter.allowed_tools,
|
||||
denied_tools: frontmatter.denied_tools,
|
||||
max_execution_secs: frontmatter.max_execution_secs,
|
||||
source,
|
||||
path: Some(path.to_path_buf()),
|
||||
})
|
||||
}
|
||||
|
||||
/// 分割 frontmatter 和 body
|
||||
fn split_frontmatter(content: &str) -> Option<(&str, &str)> {
|
||||
// 跳过开头的 ---
|
||||
let content = content
|
||||
.strip_prefix("---")
|
||||
.or_else(|| content.strip_prefix("---"))?;
|
||||
|
||||
// 跳过 --- 后的换行符和可能的空行
|
||||
let content = content.trim_start_matches('\r').trim_start_matches('\n');
|
||||
|
||||
// 找结束标记（容忍不同的换行符格式和前面的空行）
|
||||
// 尝试多种可能的结束标记格式
|
||||
let end_markers = ["\n---\n", "\n---", "\r\n---\r\n", "\r\n---"];
|
||||
let mut idx = None;
|
||||
let mut marker_len = 0;
|
||||
for marker in end_markers {
|
||||
if let Some(pos) = content.find(marker) {
|
||||
idx = Some(pos);
|
||||
marker_len = marker.len();
|
||||
break;
|
||||
}
|
||||
}
|
||||
let idx = idx?;
|
||||
|
||||
let frontmatter = &content[..idx];
|
||||
let body = &content[idx + marker_len..];
|
||||
let body = body.trim_start_matches('\r').trim_start_matches('\n');
|
||||
|
||||
Some((frontmatter, body))
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
@ -1538,10 +1577,9 @@ mod tests {
|
||||
assert!(change.changed);
|
||||
assert!(!change.available);
|
||||
|
||||
// 禁用后 prompt 不应包含 general（explore 仍可用，所以 prompt 仍为 Some）
|
||||
let prompt = runtime.system_index_prompt_filtered().unwrap();
|
||||
assert!(!prompt.contains("<name>general</name>"));
|
||||
assert!(prompt.contains("<name>explore</name>"));
|
||||
// 禁用后 prompt 不应包含 general（无可用子代理时返回 None）
|
||||
let prompt = runtime.system_index_prompt_filtered();
|
||||
assert!(prompt.map_or(true, |p| !p.contains("<name>general</name>")));
|
||||
}
|
||||
|
||||
#[test]
|
||||
@ -1556,8 +1594,8 @@ mod tests {
|
||||
runtime
|
||||
.disable_subagent(SubagentScope::Project, "general")
|
||||
.unwrap();
|
||||
let prompt = runtime.system_index_prompt_filtered().unwrap();
|
||||
assert!(!prompt.contains("<name>general</name>"));
|
||||
// 无可用子代理时返回 None
|
||||
assert!(runtime.system_index_prompt_filtered().is_none());
|
||||
|
||||
let change = runtime
|
||||
.enable_subagent(SubagentScope::Project, "general")
|
||||
@ -1586,10 +1624,6 @@ mod tests {
|
||||
assert!(general
|
||||
.disabled_in_scopes
|
||||
.contains(&"project".to_string()));
|
||||
|
||||
// explore 应仍启用
|
||||
let explore = items.iter().find(|i| i.name == "explore").unwrap();
|
||||
assert!(explore.disabled_in_scopes.is_empty());
|
||||
}
|
||||
|
||||
#[test]
|
||||
@ -1605,12 +1639,10 @@ mod tests {
|
||||
.unwrap();
|
||||
|
||||
assert!(runtime.find_available("general").is_none());
|
||||
assert!(runtime.find_available("explore").is_some());
|
||||
|
||||
// available_names 不应包含 general
|
||||
let names = runtime.available_names();
|
||||
assert!(!names.contains(&"general".to_string()));
|
||||
assert!(names.contains(&"explore".to_string()));
|
||||
}
|
||||
|
||||
#[test]
|
||||
@ -1626,4 +1658,225 @@ mod tests {
|
||||
.unwrap_err();
|
||||
assert!(err.contains("not found"));
|
||||
}
|
||||
|
||||
// ===== 工具过滤（allowed_tools / denied_tools）测试 =====
|
||||
|
||||
use crate::tools::traits::{Tool as ToolTrait, ToolResult};
|
||||
|
||||
/// 占位工具，按构造名注册
|
||||
struct FakeTool {
|
||||
tool_name: String,
|
||||
}
|
||||
|
||||
#[async_trait::async_trait]
|
||||
impl ToolTrait for FakeTool {
|
||||
fn name(&self) -> &str {
|
||||
&self.tool_name
|
||||
}
|
||||
fn description(&self) -> &str {
|
||||
"fake"
|
||||
}
|
||||
fn parameters_schema(&self) -> serde_json::Value {
|
||||
serde_json::json!({})
|
||||
}
|
||||
async fn execute(&self, _args: serde_json::Value) -> anyhow::Result<ToolResult> {
|
||||
Ok(ToolResult {
|
||||
success: true,
|
||||
output: String::new(),
|
||||
error: None,
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
fn base_registry() -> ToolRegistry {
|
||||
let reg = ToolRegistry::new();
|
||||
for name in &["read", "edit", "write", "bash", "task"] {
|
||||
reg.register(FakeTool {
|
||||
tool_name: name.to_string(),
|
||||
});
|
||||
}
|
||||
reg
|
||||
}
|
||||
|
||||
fn sorted_names(reg: &ToolRegistry) -> Vec<String> {
|
||||
let mut v = reg.tool_names();
|
||||
v.sort();
|
||||
v
|
||||
}
|
||||
|
||||
/// 把 &str 切片转为 Some(Vec<String>)，便于构造过滤参数
|
||||
fn s(v: &[&str]) -> Option<Vec<String>> {
|
||||
Some(v.iter().map(|x| x.to_string()).collect())
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn filter_no_restriction_returns_all() {
|
||||
let base = base_registry();
|
||||
let reg = DefaultSubAgentRuntime::filter_tool_registry(&base, None, None, false);
|
||||
assert_eq!(
|
||||
sorted_names(®),
|
||||
vec!["bash", "edit", "read", "task", "write"]
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn filter_depth_deny_task_removes_task() {
|
||||
let base = base_registry();
|
||||
let reg = DefaultSubAgentRuntime::filter_tool_registry(&base, None, None, true);
|
||||
assert_eq!(
|
||||
sorted_names(®),
|
||||
vec!["bash", "edit", "read", "write"]
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn filter_whitelist_keeps_only_listed() {
|
||||
let base = base_registry();
|
||||
let allowed = s(&["read", "bash"]);
|
||||
let reg = DefaultSubAgentRuntime::filter_tool_registry(&base, allowed.as_ref(), None, false);
|
||||
assert_eq!(sorted_names(®), vec!["bash", "read"]);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn filter_whitelist_skips_missing_names() {
|
||||
let base = base_registry();
|
||||
// 包含未注册的工具名应被静默跳过
|
||||
let allowed = s(&["read", "nonexistent", "glob"]);
|
||||
let reg = DefaultSubAgentRuntime::filter_tool_registry(&base, allowed.as_ref(), None, false);
|
||||
assert_eq!(sorted_names(®), vec!["read"]);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn filter_blacklist_removes_listed() {
|
||||
let base = base_registry();
|
||||
let denied = s(&["bash", "task"]);
|
||||
let reg = DefaultSubAgentRuntime::filter_tool_registry(&base, None, denied.as_ref(), false);
|
||||
assert_eq!(sorted_names(®), vec!["edit", "read", "write"]);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn filter_whitelist_then_blacklist() {
|
||||
let base = base_registry();
|
||||
let allowed = s(&["read", "bash"]);
|
||||
let denied = s(&["bash"]);
|
||||
let reg = DefaultSubAgentRuntime::filter_tool_registry(
|
||||
&base,
|
||||
allowed.as_ref(),
|
||||
denied.as_ref(),
|
||||
false,
|
||||
);
|
||||
// 白名单留下 read+bash，黑名单再扣除 bash
|
||||
assert_eq!(sorted_names(®), vec!["read"]);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn filter_empty_whitelist_yields_empty() {
|
||||
let base = base_registry();
|
||||
let allowed = s(&[]);
|
||||
let reg = DefaultSubAgentRuntime::filter_tool_registry(&base, allowed.as_ref(), None, false);
|
||||
assert!(reg.tool_names().is_empty());
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn filter_depth_rule_overrides_whitelist_task() {
|
||||
let base = base_registry();
|
||||
// 白名单显式包含 task，但 depth≥2 安全兜底仍应移除它
|
||||
let allowed = s(&["read", "task"]);
|
||||
let reg = DefaultSubAgentRuntime::filter_tool_registry(&base, allowed.as_ref(), None, true);
|
||||
assert_eq!(sorted_names(®), vec!["read"]);
|
||||
}
|
||||
|
||||
// ===== frontmatter 解析（denied_tools）测试 =====
|
||||
|
||||
#[test]
|
||||
fn parse_subagent_file_handles_crlf_endings() {
|
||||
let temp = tempfile::tempdir().unwrap();
|
||||
let path = temp.path().join("SUBAGENT.md");
|
||||
std::fs::write(
|
||||
&path,
|
||||
"---\r\nname: demo\r\ndescription: demo subagent\r\n---\r\nStep A\r\nStep B",
|
||||
)
|
||||
.unwrap();
|
||||
|
||||
let subagent = parse_subagent_file(&path, SubagentSource::Project).unwrap();
|
||||
assert_eq!(subagent.name, "demo");
|
||||
assert_eq!(subagent.description, "demo subagent");
|
||||
assert_eq!(subagent.body.as_deref(), Some("Step A\nStep B"));
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn parse_subagent_file_reads_denied_tools() {
|
||||
let temp = tempfile::tempdir().unwrap();
|
||||
let path = temp.path().join("SUBAGENT.md");
|
||||
std::fs::write(
|
||||
&path,
|
||||
"---\n\
|
||||
name: sandbox\n\
|
||||
description: sandbox agent\n\
|
||||
allowed_tools: [read, todo_write]\n\
|
||||
denied_tools: [bash, task]\n\
|
||||
---\n\
|
||||
body instructions",
|
||||
)
|
||||
.unwrap();
|
||||
|
||||
let def = parse_subagent_file(&path, SubagentSource::Project).unwrap();
|
||||
assert_eq!(def.name, "sandbox");
|
||||
assert_eq!(
|
||||
def.allowed_tools.as_deref(),
|
||||
Some(["read".to_string(), "todo_write".to_string()].as_slice())
|
||||
);
|
||||
assert_eq!(
|
||||
def.denied_tools.as_deref(),
|
||||
Some(["bash".to_string(), "task".to_string()].as_slice())
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn parse_subagent_file_denied_tools_default_none_when_absent() {
|
||||
let temp = tempfile::tempdir().unwrap();
|
||||
let path = temp.path().join("SUBAGENT.md");
|
||||
std::fs::write(
|
||||
&path,
|
||||
"---\nname: basic\ndescription: basic agent\n---\nbody",
|
||||
)
|
||||
.unwrap();
|
||||
|
||||
let def = parse_subagent_file(&path, SubagentSource::User).unwrap();
|
||||
assert!(def.allowed_tools.is_none());
|
||||
assert!(def.denied_tools.is_none());
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn list_with_status_projects_tool_fields() {
|
||||
let temp = tempfile::tempdir().unwrap();
|
||||
let mut catalog = SubagentCatalog::new();
|
||||
catalog.register(SubagentDef {
|
||||
name: "sandbox".to_string(),
|
||||
description: "sandbox agent".to_string(),
|
||||
prompt_template: String::new(),
|
||||
body: None,
|
||||
allowed_tools: Some(vec!["read".to_string(), "todo_write".to_string()]),
|
||||
denied_tools: Some(vec!["bash".to_string()]),
|
||||
max_execution_secs: None,
|
||||
source: SubagentSource::Builtin,
|
||||
path: None,
|
||||
});
|
||||
let runtime = SubagentRuntime::new(
|
||||
SubagentsConfig::default(),
|
||||
Arc::new(catalog),
|
||||
temp.path().to_path_buf(),
|
||||
);
|
||||
|
||||
let items = runtime.list_with_status();
|
||||
let item = items.iter().find(|i| i.name == "sandbox").unwrap();
|
||||
assert_eq!(
|
||||
item.allowed_tools.as_deref(),
|
||||
Some(["read".to_string(), "todo_write".to_string()].as_slice())
|
||||
);
|
||||
assert_eq!(
|
||||
item.denied_tools.as_deref(),
|
||||
Some(["bash".to_string()].as_slice())
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
@ -60,8 +60,10 @@ pub struct SubagentDef {
|
||||
pub prompt_template: String,
|
||||
/// 可选的详细指令（body 部分）
|
||||
pub body: Option<String>,
|
||||
/// 工具白名单（None 表示使用默认）
|
||||
/// 工具白名单（None 表示不过滤，Some 时仅这些工具可用）
|
||||
pub allowed_tools: Option<Vec<String>>,
|
||||
/// 工具黑名单（None 表示不过滤，Some 时这些工具被禁用；在白名单之后应用）
|
||||
pub denied_tools: Option<Vec<String>>,
|
||||
/// 最大执行时间（秒），None 表示使用默认
|
||||
pub max_execution_secs: Option<u64>,
|
||||
/// 来源
|
||||
@ -79,20 +81,7 @@ impl SubagentDef {
|
||||
prompt_template: "你是一个专注的子代理，正在执行一个独立任务。\n\n任务描述: {{description}}\n\n你应该:\n1. 专注于完成任务，不要偏离目标\n2. 使用可用的工具进行必要操作\n3. 完成后给出简洁的总结\n4. 不要尝试创建新的子代理任务\n\n任务追踪:\n你可以使用 `todo_write` 工具追踪子任务进度。规则：同一时间只有一个 in_progress，完成后再标记下一个，3步以上才使用。\n\n注意: 你没有访问主对话历史的权限，这是一个独立的执行上下文。".to_string(),
|
||||
body: None,
|
||||
allowed_tools: None,
|
||||
max_execution_secs: None,
|
||||
source: SubagentSource::Builtin,
|
||||
path: None,
|
||||
}
|
||||
}
|
||||
|
||||
/// 创建内置 explore 子代理定义
|
||||
pub fn builtin_explore() -> Self {
|
||||
Self {
|
||||
name: "explore".to_string(),
|
||||
description: "探索型子代理 - 只读搜索代理".to_string(),
|
||||
prompt_template: "你是一个只读探索代理，用于代码库探索和信息收集。\n\n任务描述: {{description}}\n\n你应该:\n1. 只使用只读工具进行探索\n2. 专注于理解和收集信息\n3. 不要进行任何写操作\n4. 给出简洁的发现总结\n\n注意: 你是一个只读代理，禁止执行任何修改操作。".to_string(),
|
||||
body: None,
|
||||
allowed_tools: None,
|
||||
denied_tools: None,
|
||||
max_execution_secs: None,
|
||||
source: SubagentSource::Builtin,
|
||||
path: None,
|
||||
|
||||
@ -526,7 +526,6 @@ export function ConfigPage({ onClose, onSaveConnection, initialTab }: ConfigPage
|
||||
<SectionCard title="Task 子代理">
|
||||
<div className="flex items-center justify-between"><span className="text-sm text-[var(--text-secondary)]">启用 Task 工具</span><Toggle checked={config.tools.task.enabled} onChange={v => update('tools', { ...config.tools, task: { ...config.tools.task, enabled: v } })} /></div>
|
||||
<Field label="最大执行时间 (秒)"><input type="number" value={config.tools.task.max_execution_secs} onChange={e => update('tools', { ...config.tools, task: { ...config.tools.task, max_execution_secs: +e.target.value } })} className={inputCls} /></Field>
|
||||
<Field label="探索模式最大执行时间 (秒)"><input type="number" value={config.tools.task.explore_max_execution_secs} onChange={e => update('tools', { ...config.tools, task: { ...config.tools.task, explore_max_execution_secs: +e.target.value } })} className={inputCls} /></Field>
|
||||
<Field label="TTL (小时)"><input type="number" value={config.tools.task.ttl_hours} onChange={e => update('tools', { ...config.tools, task: { ...config.tools.task, ttl_hours: +e.target.value } })} className={inputCls} /></Field>
|
||||
</SectionCard>
|
||||
<SectionCard title="允许的工具列表">
|
||||
@ -612,6 +611,31 @@ export function ConfigPage({ onClose, onSaveConnection, initialTab }: ConfigPage
|
||||
}
|
||||
}
|
||||
|
||||
const toolLabel = (key: string): string => {
|
||||
const known = TASK_KNOWN_TOOLS.find(t => t.key === key)
|
||||
return known ? known.label : key
|
||||
}
|
||||
|
||||
const renderToolTags = (
|
||||
label: string,
|
||||
tools: string[] | undefined,
|
||||
tone: 'allow' | 'deny',
|
||||
) => {
|
||||
if (!tools || tools.length === 0) return null
|
||||
const tagCls =
|
||||
tone === 'allow'
|
||||
? 'bg-emerald-500/10 text-emerald-600 dark:text-emerald-400'
|
||||
: 'bg-rose-500/10 text-rose-600 dark:text-rose-400'
|
||||
return (
|
||||
<div className="flex items-center gap-1 flex-wrap mt-1">
|
||||
<span className="text-[10px] text-[var(--text-muted)]">{label}:</span>
|
||||
{tools.map(t => (
|
||||
<span key={t} className={`text-[10px] px-1.5 py-0.5 rounded ${tagCls}`}>{toolLabel(t)}</span>
|
||||
))}
|
||||
</div>
|
||||
)
|
||||
}
|
||||
|
||||
return (
|
||||
<SectionCard title="已发现子代理" subtitle="即时生效">
|
||||
{subagentListLoading && subagents.length === 0 ? (
|
||||
@ -632,6 +656,8 @@ export function ConfigPage({ onClose, onSaveConnection, initialTab }: ConfigPage
|
||||
<span className="text-[10px] px-1.5 py-0.5 rounded bg-[var(--bg-tertiary)] text-[var(--text-muted)] uppercase tracking-wider">{subagent.source}</span>
|
||||
</div>
|
||||
<p className="text-xs text-[var(--text-muted)] truncate mt-0.5">{subagent.description}</p>
|
||||
{renderToolTags('允许', subagent.allowed_tools, 'allow')}
|
||||
{renderToolTags('禁用', subagent.denied_tools, 'deny')}
|
||||
</div>
|
||||
<Toggle checked={isEnabled} onChange={() => handleToggle(subagent.name, isEnabled)} />
|
||||
</div>
|
||||
|
||||
@ -7,7 +7,7 @@ export interface GatewayConfig { host: string; port: number; show_tool_results:
|
||||
export interface TimeConfig { timezone: string }
|
||||
export interface SchedulerConfig { enabled: boolean; tick_resolution_ms: number; worker_queue_capacity: number; misfire_policy: 'skip' | 'catch_up'; jobs?: SchedulerJobConfig[] }
|
||||
export interface SkillsConfig { enabled: boolean; sources: string[]; max_index_chars: number; max_listed_skills: number }
|
||||
export interface TaskConfig { enabled: boolean; max_execution_secs: number; explore_max_execution_secs: number; ttl_hours: number; allowed_tools: string[] }
|
||||
export interface TaskConfig { enabled: boolean; max_execution_secs: number; ttl_hours: number; allowed_tools: string[] }
|
||||
export interface ToolsConfig { disabled: string[]; task: TaskConfig }
|
||||
export interface MemoryMaintenanceConfig { max_merge_ratio: number; min_memories_to_keep: number; max_merge_per_group: number }
|
||||
export interface ImageContextConfig { max_images_in_context: number; max_image_age_rounds: number }
|
||||
@ -45,6 +45,8 @@ export interface SubagentItem {
|
||||
description: string
|
||||
source: string
|
||||
disabled_in_scopes: string[]
|
||||
allowed_tools?: string[]
|
||||
denied_tools?: string[]
|
||||
}
|
||||
|
||||
export interface SubagentListResponse {
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user