Skip to content

armgit / Wake

Public Imported from GitHub
0

All your AI agent sessions in one place — browse, search, resume. Rust + GPUI.

Rust

feat(cursor): show the model and token usage Cursor recorded

Corey Chiuc135b1aed6d0
diff --git a/CHANGELOG.md b/CHANGELOG.md
index de5e97c..f78f9fc 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,7 @@
- New: Craft Agents sessions are indexed from its workspaces — titles, models, tool calls and token counts, with archived sessions, sub-tasks and branches, whether the session ran on a Claude connection or on ChatGPT, Copilot or Google. The copy that a Claude connection also leaves in Claude Code's history is hidden while the Craft session exists, so each conversation is listed once. Workspaces kept outside `~/.craft-agent/workspaces` can be added in Settings → Session locations. Read-only, and no Open In (#44, thanks @qiajue)
- New: Devin sessions are indexed from its local database (`~/.local/share/devin/cli/sessions.db`, read-only) — titles, models, thinking, tool calls and token counts included; the visible transcript follows each session's main chain, so retries edited out of the conversation and hidden helper sessions stay out. Open In resumes a session with `devin --resume` in its project directory (#46, thanks @BUKOWSKIREAL for the report and the implementation in #48)
+- Update: Cursor sessions now show the model they used, and their token usage where Cursor recorded it — sessions from late 2025 to early 2026; newer Cursor versions no longer keep token counts on disk. Existing indexes are re-read once after upgrading
## [0.8.1] — 2026-09-23
diff --git a/README.md b/README.md
index 37f05e6..f2cc90e 100644
--- a/README.md
+++ b/README.md
@@ -35,7 +35,7 @@ Your agent history is scattered across `~/.claude`, `~/.codex`, and a dozen othe
| Codex CLI | `~/.codex/sessions` + `state_5.sqlite` (read-only) | ✅ | ✅ |
| Qoder CLI | `~/.qoder/projects/*/*.jsonl` (`QODER_CONFIG_DIR` is respected) | ✅ | — |
| Copilot CLI | `~/.copilot/session-store.db` | — | — |
-| Cursor | `~/.cursor/projects/**/agent-transcripts` (agent transcripts) + `~/Library/Application Support/Cursor/User/globalStorage/state.vscdb` (IDE Chat/Composer history, read-only) | — | — |
+| Cursor | `~/.cursor/projects/**/agent-transcripts` (agent transcripts) + `~/Library/Application Support/Cursor/User/globalStorage/state.vscdb` (IDE Chat/Composer history, read-only) | ✅ | — |
| OpenCode | `~/.local/share/opencode/opencode.db` | ✅ | — |
| OpenCode 2 (`opencode2`) | `~/.local/share/opencode/{opencode.db,opencode-next.db}` (`session_v2` or `session` + `session_message`); both paths are scanned | ✅ | — |
| Kiro | `~/.kiro/sessions/cli` | ✅ | — |
diff --git a/crates/wake-core/src/adapters/cursor.rs b/crates/wake-core/src/adapters/cursor.rs
index d2d4a09..46940ad 100644
--- a/crates/wake-core/src/adapters/cursor.rs
+++ b/crates/wake-core/src/adapters/cursor.rs
@@ -9,8 +9,9 @@ use std::path::{Path, PathBuf};
/// Cursor CLI:`~/.cursor/projects/<slug>/agent-transcripts/<uuid>/<uuid>.jsonl` 明文。
/// 行结构 {role, message:{content:[{type:text|tool_use}]}} + {type:"turn_ended"}。
-/// user 正文包在 <timestamp>/<user_query> 壳里;transcript 不含 cwd,
-/// 优先从同一会话的 IDE 元数据读取路径,缺失时从 slug 目录名还原。
+/// user 正文包在 <timestamp>/<user_query> 壳里;transcript 不含 cwd,也不记模型——
+/// 两样都优先从同一会话的 IDE 元数据借(`cursor_ide::composer_facts`),路径缺失时从
+/// slug 目录名还原。纯 CLI 会话在 IDE 库里没有记录,模型就空着;token 转录这一路不给。
///
/// **同一家的另一个源**:IDE 面板(Chat/Composer)的会话正文在
/// `globalStorage/state.vscdb`,由 cursor_ide.rs 读。某些 Cursor 版本的 IDE
@@ -36,12 +37,53 @@ impl CursorAdapter {
}
}
- fn project_path(&self, r: &SessionFileRef) -> String {
- let metadata = self
- .metadata_db
+ /// 同一个会话在 IDE 库里的 composer(只有默认实例有 IDE 库;自定义根也用于远程缓存,
+ /// 不能混用本机的)。新版 Cursor 给 IDE 会话也写转录、转录胜出,项目与模型得从这份借
+ fn composer_facts(&self, r: &SessionFileRef) -> cursor_ide::ComposerFacts {
+ self.metadata_db
.as_deref()
- .and_then(|db| cursor_ide::project_metadata(db, std::iter::once(r.native_id.as_str())));
- resolve_project_path(r, metadata.as_ref().and_then(|m| m.get(&r.native_id)))
+ .and_then(|db| cursor_ide::composer_facts(db, std::iter::once(r.native_id.as_str())))
+ .and_then(|mut facts| facts.remove(&r.native_id))
+ .unwrap_or_default()
+ }
+
+ fn build_meta(&self, r: &SessionFileRef, p: &CursorParse) -> SessionMeta {
+ let facts = self.composer_facts(r);
+ let cwd = resolve_project_path(r, facts.project.as_ref());
+ let title = title_from_messages(&p.messages).unwrap_or_else(|| UNTITLED.to_string());
+ SessionMeta {
+ key: format!("cursor:{}", r.native_id),
+ host: String::new(),
+ id: r.native_id.clone(),
+ agent: AgentId::Cursor,
+ title,
+ project_path: cwd.clone(),
+ project_name: project_name_of(&cwd),
+ file_path: r.file_path.clone(),
+ created_at: if p.created_at > 0 {
+ p.created_at
+ } else {
+ r.mtime_ms
+ },
+ updated_at: if p.updated_at > 0 {
+ p.updated_at
+ } else {
+ r.mtime_ms
+ },
+ message_count: p
+ .messages
+ .iter()
+ .filter(|m| m.kind == MessageKind::Text)
+ .count() as i64,
+ size_bytes: r.size,
+ git_branch: None,
+ model: facts.model,
+ tokens_used: None,
+ archived: false,
+ source: None,
+ favorite: false,
+ pinned: false,
+ }
}
}
@@ -354,43 +396,6 @@ fn resolve_project_path(
decode_slug(&slug)
}
-fn build_meta(r: &SessionFileRef, p: &CursorParse, cwd: String) -> SessionMeta {
- let title = title_from_messages(&p.messages).unwrap_or_else(|| UNTITLED.to_string());
- SessionMeta {
- key: format!("cursor:{}", r.native_id),
- host: String::new(),
- id: r.native_id.clone(),
- agent: AgentId::Cursor,
- title,
- project_path: cwd.clone(),
- project_name: project_name_of(&cwd),
- file_path: r.file_path.clone(),
- created_at: if p.created_at > 0 {
- p.created_at
- } else {
- r.mtime_ms
- },
- updated_at: if p.updated_at > 0 {
- p.updated_at
- } else {
- r.mtime_ms
- },
- message_count: p
- .messages
- .iter()
- .filter(|m| m.kind == MessageKind::Text)
- .count() as i64,
- size_bytes: r.size,
- git_branch: None,
- model: None,
- tokens_used: None,
- archived: false,
- source: None,
- favorite: false,
- pinned: false,
- }
-}
-
/// 这个 transcript 只有回合标记、没有任何对话行吗?
///
/// IDE 面板里跑的会话在这里只留 `{"type":"turn_ended"}`,正文全在
@@ -510,7 +515,7 @@ impl AgentAdapter for CursorAdapter {
fn parse_session(&self, r: &SessionFileRef) -> Result<ParsedSession> {
let parsed = parse_cursor_jsonl(Path::new(&r.file_path), false)?;
- let meta = build_meta(r, &parsed, self.project_path(r));
+ let meta = self.build_meta(r, &parsed);
Ok(ParsedSession::derive(
meta,
&parsed.messages,
@@ -538,7 +543,7 @@ impl AgentAdapter for CursorAdapter {
}
}
Ok(ParsedTranscript {
- meta: build_meta(r, &parsed, self.project_path(r)),
+ meta: self.build_meta(r, &parsed),
mainline: parsed.messages,
sidechains,
unknown_line_count: parsed.unknown_lines,
@@ -594,23 +599,29 @@ impl AgentAdapter for CursorAdapter {
}
}
- fn project_path_updates(
+ /// IDE 库里的 composer 可能比转录的最后一次写入还晚落盘:项目与模型每轮扫描都按它
+ /// 刷一遍,转录没变也跟上(与解析时借的是同一份,不会来回翻)
+ fn sidecar_updates(
&self,
refs: &[SessionFileRef],
- ) -> std::collections::HashMap<String, String> {
+ ) -> std::collections::HashMap<String, SidecarMeta> {
if refs.is_empty() {
return Default::default();
}
- let metadata = self.metadata_db.as_deref().and_then(|db| {
- cursor_ide::project_metadata(db, refs.iter().map(|r| r.native_id.as_str()))
+ let facts = self.metadata_db.as_deref().and_then(|db| {
+ cursor_ide::composer_facts(db, refs.iter().map(|r| r.native_id.as_str()))
});
- let Some(metadata) = metadata else {
+ let Some(facts) = facts else {
return Default::default();
};
refs.iter()
.filter_map(|r| {
- let project = metadata.get(&r.native_id)?;
- Some((r.file_path.clone(), resolve_project_path(r, Some(project))))
+ let f = facts.get(&r.native_id)?;
+ let update = SidecarMeta {
+ project: f.project.as_ref().map(|p| resolve_project_path(r, Some(p))),
+ model: f.model.clone(),
+ };
+ Some((r.file_path.clone(), update))
})
.collect()
}
@@ -948,7 +959,7 @@ mod tests {
let (dir, adapter, r) =
metadata_fixture(r#"{"workspaceIdentifier":{"uri":{"fsPath":"/wakefx/My Project"}}}"#);
let parsed = adapter.parse_session(&r).unwrap();
- let updates = adapter.project_path_updates(std::slice::from_ref(&r));
+ let updates = adapter.sidecar_updates(std::slice::from_ref(&r));
for case in ["unchanged", "newer", "resized", "other-copy", "child"] {
let index = dir.path().join(format!("{case}.db"));
let store = Store::open(&index).unwrap();
@@ -974,7 +985,7 @@ mod tests {
.unwrap();
}
let changed = store
- .update_project_paths(std::slice::from_ref(&r), &updates)
+ .update_sidecar_meta(std::slice::from_ref(&r), &updates)
.unwrap();
assert_eq!(changed, case == "unchanged", "{case}");
let after = store.get_session(&saved.key).unwrap().unwrap();
@@ -991,6 +1002,38 @@ mod tests {
}
}
+ #[test]
+ fn sidecar_refresh_fills_the_model_without_clearing_it() {
+ use crate::db::Store;
+
+ // composer 比转录的最后一次写入还晚落盘:解析那一刻借不到模型,之后每轮扫描的
+ // 侧档刷新补上;侧档不知道模型时(Auto 档、读不出来)不清掉库里已有的
+ let (dir, adapter, r) =
+ metadata_fixture(r#"{"modelConfig":{"modelName":"claude-4.5-sonnet"}}"#);
+ let parsed = adapter.parse_session(&r).unwrap();
+ assert_eq!(parsed.meta.model.as_deref(), Some("claude-4.5-sonnet"));
+ let store = Store::open(&dir.path().join("index.db")).unwrap();
+ let mut saved = parsed.meta.clone();
+ saved.model = None;
+ store
+ .write_session(&saved, r.mtime_ms, &parsed.units)
+ .unwrap();
+ let model = || store.get_session(&saved.key).unwrap().unwrap().model;
+
+ let updates = adapter.sidecar_updates(std::slice::from_ref(&r));
+ assert!(store
+ .update_sidecar_meta(std::slice::from_ref(&r), &updates)
+ .unwrap());
+ assert_eq!(model().as_deref(), Some("claude-4.5-sonnet"));
+
+ let unknown =
+ std::collections::HashMap::from([(r.file_path.clone(), SidecarMeta::default())]);
+ assert!(!store
+ .update_sidecar_meta(std::slice::from_ref(&r), &unknown)
+ .unwrap());
+ assert_eq!(model().as_deref(), Some("claude-4.5-sonnet"));
+ }
+
#[test]
fn custom_root_does_not_use_default_project_metadata() {
let (_dir, adapter, r) =
@@ -1004,9 +1047,7 @@ mod tests {
custom.parse_transcript(&r).unwrap().meta.project_path,
"/wakefx/My/Project"
);
- assert!(custom
- .project_path_updates(std::slice::from_ref(&r))
- .is_empty());
+ assert!(custom.sidecar_updates(std::slice::from_ref(&r)).is_empty());
}
#[test]
diff --git a/crates/wake-core/src/adapters/cursor_ide.rs b/crates/wake-core/src/adapters/cursor_ide.rs
index ac848df..1537446 100644
--- a/crates/wake-core/src/adapters/cursor_ide.rs
+++ b/crates/wake-core/src/adapters/cursor_ide.rs
@@ -14,7 +14,7 @@ use std::path::{Path, PathBuf};
/// (`{"type":"turn_ended"}`)写进那里、正文全留在本库——所以 IDE 会话在只读
/// JSONL 的旧实现里是空壳。两源同 native_id 的重叠交 scanner 的副本裁决,
/// 本源以 `dedup_rank` 排在 CLI 源之后:转录带正文时 CLI 那份胜出,
-/// 并可从本库补充项目路径;本源那份正文留作解析失败的回退;
+/// 并可从本库补充项目路径与模型 / token;本源那份正文留作解析失败的回退;
/// 只有转录缺失或只剩空壳的会话才由本源胜出。不按 mtime 定胜负——两边写盘
/// 先后不固定,同一会话会在项目之间跳(2026-09-15 实测)。
///
@@ -82,28 +82,9 @@ pub(super) struct ProjectMetadata {
pub repositories: Vec<String>,
}
-/// Read only the requested composers, sharing one connection across a scan.
-/// Keep workspace and repository paths distinct: a repo may contain the actual
-/// workspace, or be one of several repos inside it.
-pub(super) fn project_metadata<'a>(
- db: &Path,
- ids: impl Iterator<Item = &'a str>,
-) -> Option<HashMap<String, ProjectMetadata>> {
- let ro = open_sqlite_ro(db, "cursor-project")?;
- let mut stmt = ro
- .conn
- .prepare("SELECT CAST(value AS TEXT) FROM cursorDiskKV WHERE key = ?1")
- .ok()?;
- let mut paths = HashMap::new();
- for id in ids {
- let Ok(raw) = stmt.query_row([format!("{COMPOSER_PREFIX}{id}")], |r| {
- r.get::<_, String>(0)
- }) else {
- continue;
- };
- let Ok(data) = serde_json::from_str::<Value>(&raw) else {
- continue;
- };
+impl ProjectMetadata {
+ /// composer 记的工作区与 Git 仓库;两样都没有给 None
+ fn from_data(data: &Value) -> Option<Self> {
let workspace = data
.pointer("/workspaceIdentifier/uri/fsPath")
.and_then(Value::as_str)
@@ -118,17 +99,140 @@ pub(super) fn project_metadata<'a>(
.filter(|s| !s.is_empty())
.map(str::to_string)
.collect::<Vec<_>>();
- if workspace.is_some() || !repositories.is_empty() {
- paths.insert(
- id.to_string(),
- ProjectMetadata {
- workspace,
- repositories,
- },
- );
+ (workspace.is_some() || !repositories.is_empty()).then_some(Self {
+ workspace,
+ repositories,
+ })
+ }
+}
+
+/// 转录那一路向本库借的东西(转录不记 cwd、模型与 token)。**只读 composer 那一行**:气泡动辄
+/// 几十上百 MB(本机最大一条 15k 条气泡、98 MB),而转录胜出的都是新版 Cursor 的会话,气泡里
+/// 本来就没有逐条模型与 token——逐条那一级只在 `CursorIdeAdapter::parse` 里用,那边气泡已经读进来了
+#[derive(Default)]
+pub(super) struct ComposerFacts {
+ /// Keep workspace and repository paths distinct: a repo may contain the
+ /// actual workspace, or be one of several repos inside it.
+ pub project: Option<ProjectMetadata>,
+ pub model: Option<String>,
+}
+
+/// 批量读,一次连接、一条缓存的语句(转录解析读一个,每轮扫描的侧档刷新读全部)。库打不开
+/// 给 None;库里没有的 id(纯 CLI 会话)不出现在结果里
+pub(super) fn composer_facts<'a>(
+ db: &Path,
+ ids: impl Iterator<Item = &'a str>,
+) -> Option<HashMap<String, ComposerFacts>> {
+ let ro = open_sqlite_ro(db, "cursor-project")?;
+ Some(
+ ids.filter_map(|id| {
+ let data = read_composer(&ro.conn, id)?;
+ let facts = ComposerFacts {
+ project: ProjectMetadata::from_data(&data),
+ model: composer_model(&data),
+ };
+ Some((id.to_string(), facts))
+ })
+ .collect(),
+ )
+}
+
+/// IDE 会话用过的模型与 token——Cursor 在本地记下的那部分
+#[derive(Default)]
+struct Usage {
+ model: Option<String>,
+ tokens: Option<i64>,
+}
+
+/// 模型:对话里最后一条带 `modelInfo.modelName` 的气泡(逐次请求实际用的模型,2025-10 ~
+/// 2026-01 的版本写在用户气泡上)> `composer_model`。token:各气泡 `tokenCount` 的输入加输出
+/// 按调用累加——只有 2025-09 ~ 2026-01 的版本记过,之后的 Cursor 本地不记用量,给 None
+fn usage_from(data: &Value, bubbles: &HashMap<String, Value>) -> Usage {
+ let mut model = None;
+ let mut tokens = 0i64;
+ for (_, bubble) in in_order(data, bubbles) {
+ if let Some(name) =
+ optional_string(bubble.pointer("/modelInfo/modelName")).filter(|m| is_model_name(m))
+ {
+ model = Some(name);
}
+ let count = |pointer: &str| bubble.pointer(pointer).and_then(Value::as_i64).unwrap_or(0);
+ tokens += count("/tokenCount/inputTokens") + count("/tokenCount/outputTokens");
}
- Some(paths)
+ Usage {
+ model: model.or_else(|| composer_model(data)),
+ tokens: (tokens > 0).then_some(tokens),
+ }
+}
+
+/// 会话级的模型:`modelConfig.modelName`(当前选中的)> `usageData` 里请求数最多的模型
+/// (`{模型: {costInCents, amount}}`,实际计过费的)
+fn composer_model(data: &Value) -> Option<String> {
+ optional_string(data.pointer("/modelConfig/modelName"))
+ .filter(|m| is_model_name(m))
+ .or_else(|| {
+ data.get("usageData")?
+ .as_object()?
+ .iter()
+ .filter(|(name, _)| is_model_name(name))
+ .max_by_key(|(_, used)| used.get("amount").and_then(Value::as_i64).unwrap_or(0))
+ .map(|(name, _)| name.clone())
+ })
+}
+
+/// "default" 是 Auto 档的占位,不是模型名
+fn is_model_name(name: &str) -> bool {
+ !name.trim().is_empty() && name != "default"
+}
+
+/// 一个 composer 的元数据行(`composerData:<id>`);语句按连接缓存,批量读时只编译一次
+fn read_composer(conn: &rusqlite::Connection, id: &str) -> Option<Value> {
+ let raw: String = conn
+ .prepare_cached("SELECT CAST(value AS TEXT) FROM cursorDiskKV WHERE key = ?1")
+ .ok()?
+ .query_row([format!("{COMPOSER_PREFIX}{id}")], |x| x.get(0))
+ .ok()?;
+ serde_json::from_str(&raw).ok()
+}
+
+/// 一个 composer 的全部气泡正文(`bubbleId:<id>:<气泡>`),按气泡 id 索引。一次范围扫描全取
+/// 回来(逐条点查是 N 次往返);上界用同前缀接 U+FFFF:UUID 只含 [0-9a-f-],不会越界到别的
+/// 会话。Cursor 清理过的气泡会留下 value 为 NULL 的行(本机 2.5 GB 库里 983 行、波及 107 个
+/// 会话),按"已被清理"跳过——当成错误会让整个会话解析失败
+fn read_bubbles(conn: &rusqlite::Connection, id: &str) -> rusqlite::Result<HashMap<String, Value>> {
+ let prefix = format!("{BUBBLE_PREFIX}{id}:");
+ let mut stmt = conn.prepare(
+ "SELECT substr(key, ?2), CAST(value AS TEXT) FROM cursorDiskKV
+ WHERE key >= ?1 AND key < ?3",
+ )?;
+ let upper = format!("{prefix}\u{FFFF}");
+ let cut = prefix.len() as i64 + 1;
+ let mut bubbles = HashMap::new();
+ let mut found = stmt.query(rusqlite::params![&prefix, cut, &upper])?;
+ while let Some(row) = found.next()? {
+ let bubble_id: String = row.get(0)?;
+ let Some(body) = row.get::<_, Option<String>>(1)? else {
+ continue;
+ };
+ if let Ok(v) = serde_json::from_str::<Value>(&body) {
+ bubbles.insert(bubble_id, v);
+ }
+ }
+ Ok(bubbles)
+}
+
+/// 按 `fullConversationHeadersOnly` 的顺序配出(顺序表项, 气泡正文)。**顺序只认这张表**——
+/// KV 表按 key 字典序,而 bubbleId 是随机 UUID,照 key 序读会把对话打乱。顺序表里有、KV 里
+/// 没有的(气泡行被 Cursor 清理过,老会话常见)跳过
+fn in_order<'a>(
+ data: &'a Value,
+ bubbles: &'a HashMap<String, Value>,
+) -> impl Iterator<Item = (&'a Value, &'a Value)> {
+ data.get("fullConversationHeadersOnly")
+ .and_then(Value::as_array)
+ .into_iter()
+ .flatten()
+ .filter_map(|head| Some((head, bubbles.get(head.get("bubbleId")?.as_str()?)?)))
}
impl CursorIdeAdapter {
@@ -196,7 +300,13 @@ impl CursorIdeAdapter {
})
}
- fn build_meta(&self, r: &SessionFileRef, row: &IdeRow, message_count: i64) -> SessionMeta {
+ fn build_meta(
+ &self,
+ r: &SessionFileRef,
+ row: &IdeRow,
+ message_count: i64,
+ usage: Usage,
+ ) -> SessionMeta {
let title = clean_title_candidate(&row.name);
SessionMeta {
key: format!("cursor:{}", row.id),
@@ -224,8 +334,8 @@ impl CursorIdeAdapter {
message_count,
size_bytes: r.size,
git_branch: None,
- model: None,
- tokens_used: None,
+ model: usage.model,
+ tokens_used: usage.tokens,
archived: false,
source: None,
favorite: false,
@@ -234,62 +344,16 @@ impl CursorIdeAdapter {
}
/// 单会话解析:一次连接,先读 composerData 拿气泡顺序,再按
- /// `bubbleId:<cid>:` 前缀范围扫出正文。**顺序只认
- /// `fullConversationHeadersOnly`**——KV 表按 key 字典序,而 bubbleId 是
- /// 随机 UUID,照 key 序读会把对话打乱。
+ /// `bubbleId:<cid>:` 前缀范围扫出正文,按 `in_order` 排好
fn parse(&self, r: &SessionFileRef) -> Result<(SessionMeta, Vec<TranscriptMessage>)> {
let ro = open_sqlite_ro(&self.db, "cursor-ide")
.ok_or_else(|| anyhow!("cannot open cursor IDE store"))?;
- let raw: String = ro
- .conn
- .query_row(
- "SELECT CAST(value AS TEXT) FROM cursorDiskKV WHERE key = ?1",
- [format!("{COMPOSER_PREFIX}{}", r.native_id)],
- |x| x.get(0),
- )
- .map_err(|_| anyhow!("cursor composer {} not in store", r.native_id))?;
- let data: Value = serde_json::from_str(&raw)?;
-
- // 气泡正文单独存 KV,一次范围扫描全取回来(逐条点查是 N 次往返)。
- // 上界用同前缀接 U+FFFF:UUID 只含 [0-9a-f-],不会越界到别的会话
- let prefix = format!("{BUBBLE_PREFIX}{}:", r.native_id);
- let mut bubbles: HashMap<String, Value> = HashMap::new();
- {
- let mut stmt = ro.conn.prepare(
- "SELECT substr(key, ?2), CAST(value AS TEXT) FROM cursorDiskKV
- WHERE key >= ?1 AND key < ?3",
- )?;
- let upper = format!("{prefix}\u{FFFF}");
- let cut = prefix.len() as i64 + 1;
- let mut found = stmt.query(rusqlite::params![&prefix, cut, &upper])?;
- while let Some(row) = found.next()? {
- let id: String = row.get(0)?;
- // Cursor 清理过的气泡会留下 value 为 NULL 的行(本机 2.5 GB 库里
- // 983 行、波及 107 个会话),按"已被清理"跳过——当成错误会让
- // 整个会话解析失败
- let Some(body) = row.get::<_, Option<String>>(1)? else {
- continue;
- };
- if let Ok(v) = serde_json::from_str::<Value>(&body) {
- bubbles.insert(id, v);
- }
- }
- }
+ let data = read_composer(&ro.conn, &r.native_id)
+ .ok_or_else(|| anyhow!("cursor composer {} not in store", r.native_id))?;
+ let bubbles = read_bubbles(&ro.conn, &r.native_id)?;
- let order = data
- .get("fullConversationHeadersOnly")
- .and_then(|v| v.as_array())
- .cloned()
- .unwrap_or_default();
let mut messages: Vec<TranscriptMessage> = Vec::new();
- for head in &order {
- let Some(bid) = head.get("bubbleId").and_then(|v| v.as_str()) else {
- continue;
- };
- // 顺序表里有、KV 里没有:气泡行已被 Cursor 清理(老会话常见)
- let Some(bubble) = bubbles.get(bid) else {
- continue;
- };
+ for (head, bubble) in in_order(&data, &bubbles) {
// 角色以顺序表为准、气泡自带的 type 兜底(两处写同一枚举)
let kind = head
.get("type")
@@ -306,12 +370,10 @@ impl CursorIdeAdapter {
.iter()
.filter(|m| m.kind == MessageKind::Text)
.count() as i64;
- // 枚举快照优先(与列表页同源);库在扫描与打开详情之间变过时就地重建
- let row = self
- .rows()
- .and_then(|rows| rows.into_iter().find(|x| x.id == r.native_id))
- .unwrap_or_else(|| IdeRow::from_data(&r.native_id, &data, order.len() as i64));
- let mut meta = self.build_meta(r, &row, count);
+ // 行信息从已经读到的 composer 现算(与 `rows` 的 SQL 同一算法,列表与详情对得上),
+ // 不去翻枚举快照:Cursor 正在写库时快照随库戳失效,每条会话解析都会触发一次整库重列
+ let row = IdeRow::from_data(&r.native_id, &data);
+ let mut meta = self.build_meta(r, &row, count, usage_from(&data, &bubbles));
// composer 没起名时(新建会话、或 Cursor 还没生成摘要)回退首条用户消息
if meta.title == UNTITLED {
if let Some(t) = title_from_messages(&messages) {
@@ -493,25 +555,30 @@ struct IdeRow {
}
impl IdeRow {
- /// 枚举快照里没有这条时的兜底(库在扫描与打开详情之间变过)
- fn from_data(id: &str, data: &Value, bubble_count: i64) -> Self {
- let created = data.get("createdAt").and_then(|v| v.as_i64()).unwrap_or(0);
- let updated = data
- .get("lastUpdatedAt")
- .and_then(|v| v.as_i64())
+ /// 从 composer 的 JSON 现算,与 `rows` 的 SQL 逐列同一算法:改了一边另一边照改,
+ /// 列表(快照)与详情(现算)才对得上
+ fn from_data(id: &str, data: &Value) -> Self {
+ let int = |key: &str| data.get(key).and_then(Value::as_i64).unwrap_or(0);
+ let heads = data
+ .get("fullConversationHeadersOnly")
+ .and_then(Value::as_array);
+ let last_bubble = heads
+ .and_then(|h| h.last())
+ .and_then(|h| h.get("createdAt"))
+ .and_then(Value::as_str)
+ .map(iso_ms)
.unwrap_or(0);
- let cwd = project_path_from_data(data).unwrap_or_default();
Self {
id: id.to_string(),
name: data
.get("name")
- .and_then(|v| v.as_str())
+ .and_then(Value::as_str)
.unwrap_or_default()
.to_string(),
- created_ms: created,
- updated_ms: pick_updated(updated, 0, created),
- bubble_count,
- cwd: cwd.to_string(),
+ created_ms: int("createdAt"),
+ updated_ms: pick_updated(int("lastUpdatedAt"), last_bubble, int("createdAt")),
+ bubble_count: heads.map_or(0, |h| h.len() as i64),
+ cwd: project_path_from_data(data).unwrap_or_default().to_string(),
}
}
}
@@ -552,7 +619,7 @@ impl AgentAdapter for CursorIdeAdapter {
if let Some(row) = by_id.get(r.native_id.as_str()) {
out.insert(
r.file_path.clone(),
- self.build_meta(r, row, row.bubble_count),
+ self.build_meta(r, row, row.bubble_count, Usage::default()),
);
}
}
diff --git a/crates/wake-core/src/adapters/mod.rs b/crates/wake-core/src/adapters/mod.rs
index e6e5135..b8be2a0 100644
--- a/crates/wake-core/src/adapters/mod.rs
+++ b/crates/wake-core/src/adapters/mod.rs
@@ -67,13 +67,14 @@ pub trait AgentAdapter: Send + Sync {
) -> Option<std::collections::HashMap<String, SessionMeta>> {
None
}
- /// Refresh project metadata from sidecars even when the transcript did not
- /// change. Keys are source file paths; only existing, unchanged winning
+ /// Refresh metadata kept in sidecars (project path, model) even when the
+ /// transcript did not change: a sidecar can be written after the transcript's
+ /// final write. Keys are source file paths; only existing, unchanged winning
/// copies are updated, without replacing their body or source timestamps.
- fn project_path_updates(
+ fn sidecar_updates(
&self,
_refs: &[SessionFileRef],
- ) -> std::collections::HashMap<String, String> {
+ ) -> std::collections::HashMap<String, SidecarMeta> {
Default::default()
}
/// quick 与 parsed 的合并策略:默认 parsed 为准、quick 补缺。
diff --git a/crates/wake-core/src/adapters/remote.rs b/crates/wake-core/src/adapters/remote.rs
index 51cb7f1..4d1e5c3 100644
--- a/crates/wake-core/src/adapters/remote.rs
+++ b/crates/wake-core/src/adapters/remote.rs
@@ -111,11 +111,11 @@ impl AgentAdapter for RemoteAdapter {
self.inner.merge_quick_meta(parsed, quick)
}
- fn project_path_updates(
+ fn sidecar_updates(
&self,
refs: &[SessionFileRef],
- ) -> std::collections::HashMap<String, String> {
- self.inner.project_path_updates(refs)
+ ) -> std::collections::HashMap<String, SidecarMeta> {
+ self.inner.sidecar_updates(refs)
}
fn parse_session(&self, r: &SessionFileRef) -> Result<ParsedSession> {
diff --git a/crates/wake-core/src/db.rs b/crates/wake-core/src/db.rs
index 129d625..d015250 100644
--- a/crates/wake-core/src/db.rs
+++ b/crates/wake-core/src/db.rs
@@ -205,8 +205,10 @@ CREATE TABLE IF NOT EXISTS claimed_sessions (
/// 记下 agent 查 Wake 的每次调用,落 wake_lookups 表,老库靠这轮重解析回填),
/// "7" = 三条分支合到 main 时对齐(2026-09-21/22):Codex spawn_agent 子线程折叠、
/// wake_lookups、记忆层一起回填——开发库可能戳着 5 或 6,换代判据是精确不等,
-/// 两个都跳过。
-pub const FTS_FORMAT: &str = "7";
+/// 两个都跳过;
+/// "8" = Cursor 会话补上模型与 token(IDE 库的 modelInfo / modelConfig / usageData /
+/// tokenCount;转录胜出的会话向同一个 composer 借),老行靠这轮重解析回填。
+pub const FTS_FORMAT: &str = "8";
fn open_conn(path: &Path) -> Result<Connection> {
if let Some(dir) = path.parent() {
@@ -803,28 +805,39 @@ impl Store {
/// Sidecar-only updates must not replace transcript contents or steal a
/// session from another copy. Changed files go through normal parsing;
/// children inherit their project through replace_parent_links instead.
- pub(crate) fn update_project_paths(
+ /// A field the sidecar does not know (`None` / empty) is left as it is.
+ pub(crate) fn update_sidecar_meta(
&self,
refs: &[SessionFileRef],
- paths: &HashMap<String, String>,
+ updates: &HashMap<String, SidecarMeta>,
) -> Result<bool> {
- if paths.is_empty() {
+ if updates.is_empty() {
return Ok(false);
}
let mut conn = self.write.lock().unwrap();
let tx = conn.transaction()?;
let mut changed = false;
for r in refs {
- let Some(project) = paths.get(&r.file_path).filter(|p| !p.is_empty()) else {
+ let Some(update) = updates.get(&r.file_path) else {
continue;
};
- let name = crate::adapters::parse_utils::project_name_of(project);
- changed |= tx.execute(
- "UPDATE sessions SET project_path = ?1, project_name = ?2
- WHERE file_path = ?3 AND file_mtime = ?4 AND file_size = ?5
- AND parent_key = '' AND (project_path <> ?1 OR project_name <> ?2)",
- params![project, name, r.file_path, r.mtime_ms, r.size],
- )? > 0;
+ if let Some(project) = update.project.as_deref().filter(|p| !p.is_empty()) {
+ let name = crate::adapters::parse_utils::project_name_of(project);
+ changed |= tx.execute(
+ "UPDATE sessions SET project_path = ?1, project_name = ?2
+ WHERE file_path = ?3 AND file_mtime = ?4 AND file_size = ?5
+ AND parent_key = '' AND (project_path <> ?1 OR project_name <> ?2)",
+ params![project, name, r.file_path, r.mtime_ms, r.size],
+ )? > 0;
+ }
+ if let Some(model) = update.model.as_deref().filter(|m| !m.is_empty()) {
+ changed |= tx.execute(
+ "UPDATE sessions SET model = ?1
+ WHERE file_path = ?2 AND file_mtime = ?3 AND file_size = ?4
+ AND model IS NOT ?1",
+ params![model, r.file_path, r.mtime_ms, r.size],
+ )? > 0;
+ }
}
tx.commit()?;
Ok(changed)
diff --git a/crates/wake-core/src/mcp/tools.rs b/crates/wake-core/src/mcp/tools.rs
index adc799f..fc09546 100644
--- a/crates/wake-core/src/mcp/tools.rs
+++ b/crates/wake-core/src/mcp/tools.rs
@@ -34,7 +34,9 @@ struct CachedTranscript {
path: String,
mtime: i64,
size: i64,
+ // 侧档(Cursor 的 IDE 库)在转录不变时也会改项目与模型:两样都进键,索引一刷就重解析
indexed_project: String,
+ indexed_model: Option<String>,
transcript: Arc<ParsedTranscript>,
}
@@ -59,6 +61,7 @@ impl TranscriptCache {
&& cached.mtime == stamp
&& cached.size == r.size
&& cached.indexed_project == meta.project_path
+ && cached.indexed_model == meta.model
{
return Ok(cached.transcript.clone());
}
@@ -69,6 +72,7 @@ impl TranscriptCache {
mtime: stamp,
size: r.size,
indexed_project: meta.project_path.clone(),
+ indexed_model: meta.model.clone(),
transcript: t.clone(),
});
Ok(t)
diff --git a/crates/wake-core/src/models.rs b/crates/wake-core/src/models.rs
index be3d79c..b8e9ea0 100644
--- a/crates/wake-core/src/models.rs
+++ b/crates/wake-core/src/models.rs
@@ -408,6 +408,14 @@ impl SessionFileRef {
}
}
+/// 转录没变时也能从侧档刷新的元数据(`AgentAdapter::sidecar_updates`)。`None` = 这份侧档
+/// 不知道,不是"清空"
+#[derive(Debug, Clone, Default, PartialEq)]
+pub struct SidecarMeta {
+ pub project: Option<String>,
+ pub model: Option<String>,
+}
+
#[derive(Debug, Clone)]
pub struct ParsedSession {
pub meta: SessionMeta,
diff --git a/crates/wake-core/src/scanner.rs b/crates/wake-core/src/scanner.rs
index 1bdbcfe..3593405 100644
--- a/crates/wake-core/src/scanner.rs
+++ b/crates/wake-core/src/scanner.rs
@@ -444,7 +444,7 @@ fn run_scan_inner(
// Sidecar metadata can arrive after the transcript's final write.
// Refresh it independently of the mtime/size gate and notify the UI.
- if store.update_project_paths(&refs, &adapter.project_path_updates(&refs))? {
+ if store.update_sidecar_meta(&refs, &adapter.sidecar_updates(&refs))? {
events.on_sessions_changed();
}
diff --git a/crates/wake-core/tests/adapter_contracts.rs b/crates/wake-core/tests/adapter_contracts.rs
index 3a88e43..e6ccc1d 100644
--- a/crates/wake-core/tests/adapter_contracts.rs
+++ b/crates/wake-core/tests/adapter_contracts.rs
@@ -1267,6 +1267,13 @@ fn cursor_ide_parse_contract() {
assert_eq!(s.meta.updated_at, 1786300600000);
assert!(s.meta.file_path.ends_with("#cide-0001"));
assert_eq!(s.unknown_line_count, 0);
+ // 模型:逐条请求记的实发模型(用户气泡的 modelInfo)压过会话级的 modelConfig;
+ // token 是各气泡 tokenCount 的输入加输出按调用累加。详情页与 MCP 读的 parse_transcript
+ // 同一口径
+ for meta in [&s.meta, &t.meta] {
+ assert_eq!(meta.model.as_deref(), Some("claude-4.5-sonnet"));
+ assert_eq!(meta.tokens_used, Some(1200 + 80 + 1500 + 120));
+ }
// 空壳流式气泡(bb)、已被清理的气泡(cc)与 value 为 NULL 的气泡行(dd)
// 都不产出消息,也不让整段解析失败
@@ -1341,6 +1348,10 @@ fn cursor_ide_parse_contract() {
.expect("cursor ide fallback parse");
assert_eq!(s2.meta.title, "空 name 会话的兜底标题应取这句");
assert_eq!(s2.meta.created_at, 1786310000000);
+ // modelConfig 是 "default"(Auto 档,不是模型名)→ 退到 usageData 里请求最多的模型;
+ // 没有 tokenCount(新版 Cursor 本地不记)→ 空着,不写 0
+ assert_eq!(s2.meta.model.as_deref(), Some("claude-4-sonnet"));
+ assert_eq!(s2.meta.tokens_used, None);
assert_eq!(s2.meta.updated_at, ms("2026-08-09T11:00:20.000Z"));
// 子代理归属来自 composerHeaders.subagentInfo
@@ -1418,6 +1429,20 @@ fn cursor_parse_contract() {
assert_eq!(s.meta.message_count, 3);
assert_eq!(s.unknown_line_count, 1); // session_started;turn_ended 不计
+ // 转录不记模型:向 IDE 库里同一个 composer("CLI twin")的会话级设置借,两条解析路径
+ // 一致。只读 composer 那一行、不读气泡——那份气泡里的 tokenCount 不会被算进来;
+ // 自定义根(也用于远程缓存)不借本机 IDE 库
+ for meta in [&s.meta, &t.meta] {
+ assert_eq!(meta.model.as_deref(), Some("gpt-5.6-sol"));
+ assert_eq!(meta.tokens_used, None);
+ }
+ let custom = CursorAdapter::new()
+ .with_custom_root(fixture("cursor"))
+ .parse_session(&r)
+ .expect("cursor custom-root parse");
+ assert_eq!(custom.meta.model, None);
+ assert_eq!(custom.meta.tokens_used, None);
+
assert_eq!(
roles_kinds(&t.mainline),
vec![
diff --git a/crates/wake-core/tests/common/mod.rs b/crates/wake-core/tests/common/mod.rs
index dc48937..ac1fa02 100644
--- a/crates/wake-core/tests/common/mod.rs
+++ b/crates/wake-core/tests/common/mod.rs
@@ -698,6 +698,8 @@ pub fn build_cursor_ide_db(path: &Path) {
"createdAt":1786300000000,"lastUpdatedAt":1786300600000,
"workspaceIdentifier":{"id":"ws1","uri":{"fsPath":"/Users/tester/Github/wakefx","path":"/Users/tester/Github/wakefx","scheme":"file"}},
"trackedGitRepos":[],
+ "modelConfig":{"modelName":"claude-4.5-opus-high","maxMode":false},
+ "usageData":{"claude-4.5-opus-high":{"costInCents":12,"amount":3}},
"fullConversationHeadersOnly":[
{"bubbleId":"zz-user-1","type":1,"createdAt":"2026-08-09T10:00:05.000Z"},
{"bubbleId":"aa-think-1","type":2,"createdAt":"2026-08-09T10:00:08.000Z"},
@@ -709,10 +711,12 @@ pub fn build_cursor_ide_db(path: &Path) {
{"bubbleId":"kk-final-1","type":2,"createdAt":"2026-08-09T10:00:12.000Z"}]}'),
('bubbleId:cide-0001:zz-user-1',
'{"_v":3,"type":1,"bubbleId":"zz-user-1","createdAt":"2026-08-09T10:00:05.000Z",
+ "modelInfo":{"modelName":"claude-4.5-sonnet"},"tokenCount":{"inputTokens":0,"outputTokens":0},
"text":"Cursor IDE 看看二维码扫描为何闪退,是不是 useEffect() 的问题"}'),
('bubbleId:cide-0001:aa-think-1',
'{"_v":3,"type":2,"bubbleId":"aa-think-1","createdAt":"2026-08-09T10:00:08.000Z",
- "text":"","thinking":{"text":"先查 effect 依赖和清理函数"}}'),
+ "text":"","thinking":{"text":"先查 effect 依赖和清理函数"},
+ "tokenCount":{"inputTokens":1200,"outputTokens":80}}'),
('bubbleId:cide-0001:mm-tool-1',
'{"_v":3,"type":2,"bubbleId":"mm-tool-1","createdAt":"2026-08-09T10:00:09.000Z","text":"",
"toolFormerData":{"toolCallId":"call_ide_1","name":"grep","status":"completed",
@@ -729,11 +733,14 @@ pub fn build_cursor_ide_db(path: &Path) {
('bubbleId:cide-0001:dd-null-1', NULL),
('bubbleId:cide-0001:kk-final-1',
'{"_v":3,"type":2,"bubbleId":"kk-final-1","createdAt":"2026-08-09T10:00:12.000Z",
- "text":"找到泄漏点,已在清理回调里停止扫描。"}'),
+ "text":"找到泄漏点,已在清理回调里停止扫描。",
+ "tokenCount":{"inputTokens":1500,"outputTokens":120}}'),
('composerData:cide-0002',
'{"_v":13,"composerId":"cide-0002","name":"",
"createdAt":1786310000000,
+ "modelConfig":{"modelName":"default"},
+ "usageData":{"gpt-5.2":{"costInCents":3,"amount":2},"claude-4-sonnet":{"costInCents":9,"amount":5}},
"fullConversationHeadersOnly":[
{"bubbleId":"q1","type":1},
{"bubbleId":"q2","type":2,"createdAt":"2026-08-09T11:00:20.000Z"}]}'),
@@ -758,9 +765,11 @@ pub fn build_cursor_ide_db(path: &Path) {
('composerData:33333333-aaaa-bbbb-cccc-000000000003',
'{"_v":18,"composerId":"33333333-aaaa-bbbb-cccc-000000000003","name":"CLI twin",
"createdAt":1786340000000,"lastUpdatedAt":1786340100000,
+ "modelConfig":{"modelName":"gpt-5.6-sol"},
"fullConversationHeadersOnly":[{"bubbleId":"t1","type":1}]}'),
('bubbleId:33333333-aaaa-bbbb-cccc-000000000003:t1',
- '{"_v":3,"type":1,"bubbleId":"t1","createdAt":"2026-08-09T13:00:00.000Z","text":"IDE 库里的同一会话"}'),
+ '{"_v":3,"type":1,"bubbleId":"t1","createdAt":"2026-08-09T13:00:00.000Z","text":"IDE 库里的同一会话",
+ "tokenCount":{"inputTokens":40,"outputTokens":2}}'),
('composerData:44444444-aaaa-bbbb-cccc-000000000004',
'{"_v":18,"composerId":"44444444-aaaa-bbbb-cccc-000000000004","name":"Stub twin",