feat: ask --json + clean resume packs (0.26.2) (#50)

* feat(cli): ask --json for tooling / Loom host (additive)

Add --json to `ask` (semantic search, current project), mirroring recall --json:
emits a JSON array [{task_id, project_hash, event_type, text, score}]. Default
(no flag) behaviour unchanged. 119 cli tests green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(pack): drop compaction noise + dedupe so the dossier reads clean

The pack's Active decisions / Rejected / Recent events sections were polluted
with machine-generated markers (Claude Code's 'This session is being
continued…' / 'Conversation compacted at…') that the classifier filed as
decisions, plus exact duplicates — so a task's reasoning read as repeated
essays. Add a conservative is_noise() filter for those known prefixes and
de-duplicate by text within each section. Additive: same pack structure, just
cleaner content; existing behavior for noise-free journals is unchanged. 291
tests green (+ golden packs).

* fix(pack): rustfmt + update precompact test for clean-pack behavior; release 0.26.2

- cargo fmt on pack.rs noise tests (CI fmt gate was red)
- precompact_hook_appends_marker_decision_to_open_task now asserts the
  marker is recorded in the append-only journal but filtered OUT of the
  rendered pack (the new is_noise behavior), instead of expecting it in
  the pack
- bump workspace to 0.26.2 + CHANGELOG (ask --json, clean pack)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: Shahinyanm

CHANGELOG.md

@@ -7,6 +7,23 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.26.2] - 2026-06-16
+
+### Added
+- **`ask --json`** — `task-journal ask "<query>" --json` emits the semantic
+  search hits as a JSON array (task_id, project_hash, event_type, text, score)
+  for machine consumers like the Loom host; no matches yields `[]`. Mirrors the
+  existing `recall --json`.
+
+### Changed
+- **Resume packs read clean.** Auto-recorded compaction / session-continuation
+  markers (e.g. "This session is being continued…", "Conversation compacted
+  at…") are machine noise the classifier sometimes files as decisions. The pack
+  now drops them from the recent-events, rejected and active-decisions sections
+  and de-duplicates exact repeats, so the dossier reads as crisp reasoning. The
+  append-only journal still records every marker — only the rendered pack hides
+  them.
+
 ## [0.26.1] - 2026-06-14
 
 ### Fixed

Cargo.lock

@@ -7,7 +7,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.26.1"
+version = "0.26.2"
 edition = "2021"
 rust-version = "1.88"
 license = "MIT"

Cargo.toml

@@ -631,6 +631,9 @@ enum Commands {
         /// Maximum number of results.
         #[arg(long, default_value_t = 5)]
         k: usize,
+        /// Emit a JSON array instead of human lines (for tooling / the Loom host).
+        #[arg(long)]
+        json: bool,
     },
     /// Cross-project recall (Pillar B): search EVERY project's decisions,
     /// rejections and constraints for reasoning relevant to the query —
@@ -1277,7 +1280,7 @@ fn real_main() -> Result<()> {
                 embedder.dim()
             );
         }
-        Commands::Ask { query, k } => {
+        Commands::Ask { query, k, json } => {
             let cwd = std::env::current_dir()?;
             let project_hash = tj_core::project_hash::from_path(&cwd)?;
             let events_path = tj_core::paths::events_dir()?.join(format!("{project_hash}.jsonl"));
@@ -1298,7 +1301,21 @@ fn real_main() -> Result<()> {
             let qv = embedder.embed_one(&query)?;
             let hits =
                 tj_core::db::semantic_search(&conn, &project_hash, &qv, embedder.model_id(), k)?;
-            if hits.is_empty() {
+            if json {
+                let arr: Vec<serde_json::Value> = hits
+                    .iter()
+                    .map(|h| {
+                        serde_json::json!({
+                            "task_id": h.task_id,
+                            "project_hash": project_hash,
+                            "event_type": h.event_type,
+                            "text": h.text,
+                            "score": h.score,
+                        })
+                    })
+                    .collect();
+                println!("{}", serde_json::to_string(&arr)?);
+            } else if hits.is_empty() {
                 println!("no matches");
             } else {
                 for h in hits {

crates/tj-cli/src/main.rs

@@ -2660,6 +2660,32 @@ fn precompact_hook_appends_marker_decision_to_open_task() {
         .assert()
         .success();
 
+    // The hook still appends the marker decision to the append-only journal…
+    let events_glob = dir.path().join("task-journal").join("events");
+    let mut marker_lines = 0;
+    for entry in std::fs::read_dir(&events_glob).unwrap() {
+        let p = entry.unwrap().path();
+        if p.extension().and_then(|e| e.to_str()) == Some("jsonl") {
+            let body = std::fs::read_to_string(&p).unwrap();
+            for line in body.lines() {
+                let v: serde_json::Value = serde_json::from_str(line).unwrap();
+                if v["type"] == "decision"
+                    && v["text"]
+                        .as_str()
+                        .unwrap_or("")
+                        .contains("Conversation compacted at")
+                {
+                    marker_lines += 1;
+                }
+            }
+        }
+    }
+    assert_eq!(
+        marker_lines, 1,
+        "marker decision must be recorded in the journal"
+    );
+
+    // …but the pack filters it out as machine noise so the dossier reads clean.
     Command::cargo_bin("task-journal")
         .unwrap()
         .env("XDG_DATA_HOME", dir.path())
@@ -2668,9 +2694,9 @@ fn precompact_hook_appends_marker_decision_to_open_task() {
         .assert()
         .success()
         .stdout(
-            contains("[decision]")
-                .and(contains("Conversation compacted at"))
-                .and(contains("single reasoning unit")),
+            contains("Conversation compacted at")
+                .not()
+                .and(contains("single reasoning unit").not()),
         );
 }
 

crates/tj-cli/tests/cli.rs

@@ -27,6 +27,20 @@ pub struct PackMetadata {
 
 use anyhow::Context;
 use rusqlite::Connection;
+use std::collections::HashSet;
+
+/// Auto-recorded compaction / session-continuation markers are not real
+/// reasoning — Claude Code injects them when a conversation is compacted or
+/// resumed, and the classifier sometimes files them as decisions. Drop them
+/// from the pack so the reasoning reads clean. Conservative: only well-known
+/// machine-generated prefixes.
+pub(crate) fn is_noise(text: &str) -> bool {
+    let t = text.trim_start();
+    t.starts_with("This session is being continued from a previous conversation")
+        || t.starts_with("Conversation compacted at")
+        || t.starts_with("[Conversation compacted")
+        || t.starts_with("Caveat: The messages below were generated by the user")
+}
 
 fn render_recent_events(conn: &Connection, task_id: &str, limit: usize) -> anyhow::Result<String> {
     let mut out = format!("## Recent events (last {limit})\n");
@@ -44,6 +58,9 @@ fn render_recent_events(conn: &Connection, task_id: &str, limit: usize) -> anyho
     })?;
     for row in rows {
         let (ts, ty, st, txt) = row?;
+        if is_noise(&txt) {
+            continue;
+        }
         let one_line = txt
             .lines()
             .next()
@@ -97,8 +114,12 @@ fn render_rejected(conn: &Connection, task_id: &str) -> anyhow::Result<String> {
         .query_map(rusqlite::params![task_id], |r| r.get::<_, String>(0))?
         .collect::<Result<_, _>>()?;
     let mut count = 0;
+    let mut seen: HashSet<String> = HashSet::new();
     for eid in event_ids {
         let text: String = text_stmt.query_row(rusqlite::params![eid], |r| r.get(0))?;
+        if is_noise(&text) || !seen.insert(text.trim().to_string()) {
+            continue;
+        }
         out.push_str(&format!("- {text}\n"));
         count += 1;
     }
@@ -122,8 +143,14 @@ fn render_active_decisions(conn: &Connection, task_id: &str) -> anyhow::Result<S
         Ok((r.get::<_, String>(0)?, r.get::<_, Option<String>>(1)?))
     })?;
     let mut count = 0;
+    let mut seen: HashSet<String> = HashSet::new();
     for row in rows {
         let (text, alternatives) = row?;
+        // Skip machine noise (compaction markers) and exact duplicates so the
+        // section reads as crisp decisions, not repeated essays.
+        if is_noise(&text) || !seen.insert(text.trim().to_string()) {
+            continue;
+        }
         out.push_str(&format!("- {text}\n"));
         // v0.12.0: structured alternatives render under the decision so the
         // pack shows "considered A/B/C, chose X" without reconstructing it
@@ -451,6 +478,27 @@ mod tests {
         assert_eq!(s, "\"Compact\"");
     }
 
+    #[test]
+    fn is_noise_flags_machine_markers_not_real_reasoning() {
+        assert!(is_noise(
+            "This session is being continued from a previous conversation that ran out of context."
+        ));
+        assert!(is_noise(
+            "Conversation compacted at 2026-06-07T11:47:33Z; preceding events…"
+        ));
+        assert!(is_noise("  [Conversation compacted]"));
+        assert!(is_noise(
+            "Caveat: The messages below were generated by the user while running local commands."
+        ));
+        // real reasoning is kept
+        assert!(!is_noise(
+            "Use axum for the server because it has better middleware."
+        ));
+        assert!(!is_noise(
+            "Rejected: per-stage sessions lose accumulated context."
+        ));
+    }
+
     #[test]
     fn parent_pack_contains_subtasks_section() {
         let d = tempfile::TempDir::new().unwrap();