Merge pull request #47 from Digital-Threads/feat/session-end-clear-catchup

feat: SessionEnd(clear) catch-up so /clear keeps the last segment (0.26.0)

Author: Shahinyanm

CHANGELOG.md

@@ -7,6 +7,19 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.26.0] - 2026-06-14
+
+### Added
+- **`/clear` no longer drops the last segment.** A new `SessionEnd` hook (wired
+  by `install-hooks --auto-capture`) runs the same transcript catch-up as `Stop`
+  when the session ends with reason `clear` — the last chance to capture the
+  final segment before `/clear` orphans the transcript. Gated to `clear` so it
+  doesn't re-process what `Stop` already handled on other exits.
+- **`recall --json`** — `task-journal recall "<context>" --json` emits the
+  cross-project memory hits as a JSON array (task_id, project_hash, event_type,
+  text, score) for machine consumers like the Loom host; empty/missing memory
+  yields `[]`.
+
 ## [0.25.1] - 2026-06-14
 
 ### Fixed

Cargo.lock

@@ -7,7 +7,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.25.1"
+version = "0.26.0"
 edition = "2021"
 rust-version = "1.88"
 license = "MIT"

Cargo.toml

@@ -23,7 +23,7 @@ default = ["embed"]
 embed = ["tj-core/embed"]
 
 [dependencies]
-tj-core = { package = "task-journal-core", version = "0.25.1", path = "../tj-core", default-features = false }
+tj-core = { package = "task-journal-core", version = "0.26.0", path = "../tj-core", default-features = false }
 anyhow = { workspace = true }
 clap = { workspace = true }
 tracing = { workspace = true }

crates/tj-cli/Cargo.toml

@@ -1809,7 +1809,7 @@ fn main() -> Result<()> {
                             { "type": "command", "command": cmd },
                         ]}]),
                     );
-                    for ev in ["PostToolUse", "Stop", "PreCompact"] {
+                    for ev in ["PostToolUse", "Stop", "PreCompact", "SessionEnd"] {
                         obj.insert(
                             ev.to_string(),
                             serde_json::json!([{ "matcher": "", "hooks": [{ "type": "command", "command": cmd }] }]),
@@ -2493,6 +2493,21 @@ runs in the background and won't block you; it only fills gaps and never closes
                 return Ok(());
             }
 
+            // SessionEnd with reason "clear": /clear discards the conversation
+            // and the transcript orphans, so this is the LAST chance to capture
+            // the final segment. Extracted to its own function so its locals do
+            // NOT bloat `main`'s already-huge stack frame — inlining it here
+            // overflowed the 1 MiB Windows main-thread stack on every command.
+            if kind == "SessionEnd" {
+                return run_session_end_catchup(
+                    &payload,
+                    &events_path,
+                    &project_hash,
+                    &backend,
+                    live_session_id.as_deref(),
+                );
+            }
+
             // Drain any pending entries first (Task 10 fills the real-classifier branch).
             drain_pending(
                 &events_path,
@@ -3924,6 +3939,58 @@ fn run_nudge() -> anyhow::Result<()> {
 }
 
 /// Proactive recall injector (opt-in hook). Reads the UserPromptSubmit payload
+/// SessionEnd(reason=clear) catch-up: enqueue transcript chunks newer than the
+/// active task's last event, then spawn the classify-worker. Kept OUT of `main`
+/// so its locals don't grow `main`'s already-huge stack frame — inlining it
+/// overflowed the 1 MiB Windows main-thread stack on every command.
+fn run_session_end_catchup(
+    payload: &serde_json::Value,
+    events_path: &std::path::Path,
+    project_hash: &str,
+    backend: &str,
+    live_session_id: Option<&str>,
+) -> anyhow::Result<()> {
+    let reason = payload.get("reason").and_then(|x| x.as_str()).unwrap_or("");
+    if reason != "clear" || !events_path.exists() {
+        return Ok(());
+    }
+    let state_path = tj_core::paths::state_dir()?.join(format!("{project_hash}.sqlite"));
+    let conn = tj_core::db::open(&state_path)?;
+    tj_core::db::ingest_new_events(&conn, events_path, project_hash)?;
+    let Some(tc) = recent_task_contexts(&conn, 1)?.into_iter().next() else {
+        return Ok(());
+    };
+    let last_event_ts: Option<String> = conn
+        .query_row(
+            "SELECT timestamp FROM events_index WHERE task_id=?1 ORDER BY timestamp DESC LIMIT 1",
+            rusqlite::params![&tc.task_id],
+            |r| r.get::<_, String>(0),
+        )
+        .ok();
+    let transcript_path = payload
+        .get("transcript_path")
+        .and_then(|x| x.as_str())
+        .map(std::path::PathBuf::from);
+    if let Some(tp) = transcript_path.as_ref() {
+        if tp.exists() {
+            let enq = enqueue_transcript_chunks_since_last_event(
+                tp,
+                events_path,
+                project_hash,
+                backend,
+                last_event_ts.as_deref(),
+                "SessionEndChunk",
+                live_session_id,
+            )
+            .unwrap_or(0);
+            if enq > 0 && std::env::var("TJ_DISABLE_CLASSIFY_SPAWN").is_err() {
+                let _ = spawn_classify_worker(backend);
+            }
+        }
+    }
+    Ok(())
+}
+
 /// from stdin, keyword-searches the global index for relevant prior
 /// decisions/rejections/constraints across all projects, and emits a budgeted
 /// `additionalContext` block. Never blocks the prompt: any miss, empty result,

crates/tj-cli/src/main.rs

@@ -884,11 +884,38 @@ fn install_hooks_auto_capture_wires_all_events() {
         "PostToolUse",
         "Stop",
         "PreCompact",
+        "SessionEnd",
     ] {
         assert!(content.contains(ev), "--auto-capture must wire {ev}");
     }
 }
 
+#[test]
+fn session_end_hook_is_clean_noop_without_journal() {
+    // SessionEnd(clear) with no journal yet must exit cleanly (it's the
+    // last-chance catch-up; nothing to catch when there's no project journal).
+    let dir = assert_fs::TempDir::new().unwrap();
+    let proj = assert_fs::TempDir::new().unwrap();
+    for reason in ["clear", "logout"] {
+        let payload = serde_json::json!({
+            "hook_event_name": "SessionEnd",
+            "reason": reason,
+            "session_id": "s-end",
+            "transcript_path": "/nonexistent/x.jsonl",
+            "cwd": proj.path().to_string_lossy(),
+        })
+        .to_string();
+        Command::cargo_bin("task-journal")
+            .unwrap()
+            .current_dir(proj.path())
+            .env("XDG_DATA_HOME", dir.path())
+            .args(["ingest-hook", "--backend", "hybrid"])
+            .write_stdin(payload)
+            .assert()
+            .success();
+    }
+}
+
 #[test]
 fn install_hooks_merges_and_preserves_third_party_hooks() {
     let dir = assert_fs::TempDir::new().unwrap();

crates/tj-cli/tests/cli.rs

@@ -17,7 +17,7 @@ path = "src/main.rs"
 
 [dependencies]
 # Lean: the MCP server doesn't embed yet, so it skips the model2vec backend.
-tj-core = { package = "task-journal-core", version = "0.25.1", path = "../tj-core", default-features = false }
+tj-core = { package = "task-journal-core", version = "0.26.0", path = "../tj-core", default-features = false }
 anyhow = { workspace = true }
 tokio = { workspace = true }
 tracing = { workspace = true }

crates/tj-mcp/Cargo.toml

@@ -1,6 +1,6 @@
 {
   "name": "task-journal",
-  "version": "0.25.1",
+  "version": "0.26.0",
   "description": "Append-only journal of AI-coding task reasoning chains: hypotheses, decisions, rejections, evidence. Renders compact resume packs so an agent can pick up a 2-week-old task with full context.",
   "author": {
     "name": "Mher Shahinyan"