threadkeeper 0.9.0__tar.gz → 0.9.1__tar.gz

{threadkeeper-0.9.0 → threadkeeper-0.9.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: threadkeeper
-Version: 0.9.0
+Version: 0.9.1
 Summary: Multi-agent shared brain across Claude Code/Desktop, Codex, Gemini, Copilot, VS Code. Cross-session memory, self-improving skill loops, inter-agent signaling — one local MCP server.
 Author: thread-keeper contributors
 License: MIT
@@ -365,7 +365,10 @@ Hard limits: max 2 new skills per pass, `[PROTECTED]` (pinned +
 foreground-authored) skills off-limits. Closes the gap between
 heuristic harvest and SKILL.md materialization — previously pending
 candidates accumulated indefinitely waiting for an agent to call
-`accept_candidate()` manually.
+`accept_candidate()` manually. The loop is machine-wide single-flight:
+while one reviewer child is running, other foreground servers/ticks report
+`candidate_review_running` instead of spawning another child for the same
+queue.
 
 #### 5. Autonomous Curator
 

{threadkeeper-0.9.0 → threadkeeper-0.9.1}/README.md

@@ -324,7 +324,10 @@ Hard limits: max 2 new skills per pass, `[PROTECTED]` (pinned +
 foreground-authored) skills off-limits. Closes the gap between
 heuristic harvest and SKILL.md materialization — previously pending
 candidates accumulated indefinitely waiting for an agent to call
-`accept_candidate()` manually.
+`accept_candidate()` manually. The loop is machine-wide single-flight:
+while one reviewer child is running, other foreground servers/ticks report
+`candidate_review_running` instead of spawning another child for the same
+queue.
 
 #### 5. Autonomous Curator
 

{threadkeeper-0.9.0 → threadkeeper-0.9.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "threadkeeper"
-version = "0.9.0"
+version = "0.9.1"
 description = "Multi-agent shared brain across Claude Code/Desktop, Codex, Gemini, Copilot, VS Code. Cross-session memory, self-improving skill loops, inter-agent signaling — one local MCP server."
 requires-python = ">=3.11"
 authors = [{ name = "thread-keeper contributors" }]

{threadkeeper-0.9.0 → threadkeeper-0.9.1}/tests/test_candidate_reviewer.py

@@ -7,6 +7,7 @@ fork.
 
 from __future__ import annotations
 
+import os
 import sys
 import time
 from pathlib import Path
@@ -205,6 +206,41 @@ def test_run_spawns_when_threshold_met(tmp_path, monkeypatch):
     assert "Bash" not in allowed
 
 
+def test_single_flight_when_reviewer_child_running(tmp_path, monkeypatch):
+    """Candidate review consumes one global queue; don't spawn duplicates."""
+    pkg = _bootstrap(tmp_path, monkeypatch, min_n="3")
+    conn = pkg["db"].get_db()
+    for i in range(4):
+        _seed_pending(conn, "verbatim", f"candidate {i}", age_s=60 + i)
+    conn.execute(
+        "INSERT INTO tasks "
+        "(id, pid, parent_cid, spawned_cid, cwd, prompt, started_at) "
+        "VALUES ('tk_running_review', ?, 'p', 'c', '/x', ?, ?)",
+        (
+            os.getpid(),
+            "You are a CANDIDATE REVIEWER for thread-keeper's extract queue.",
+            int(time.time()) - 30,
+        ),
+    )
+    conn.commit()
+
+    import threadkeeper.tools.spawn as spawn_mod
+
+    def fail_spawn(**kwargs):  # pragma: no cover - should not be called
+        raise AssertionError("spawn should not run while reviewer is active")
+
+    monkeypatch.setattr(spawn_mod, "spawn", fail_spawn)
+
+    out = pkg["candidate_reviewer"].run_review_pass(force=True)
+
+    assert out == "candidate_review_running n=1 (single-flight)"
+    row = conn.execute(
+        "SELECT summary FROM events WHERE kind='candidate_review_pass' "
+        "ORDER BY id DESC LIMIT 1"
+    ).fetchone()
+    assert "candidate_review_running n=1" in row["summary"]
+
+
 # ──────────────────────────────────────────────────────────────────────
 # Daemon lifecycle
 # ──────────────────────────────────────────────────────────────────────

{threadkeeper-0.9.0 → threadkeeper-0.9.1}/threadkeeper/candidate_reviewer.py

@@ -35,6 +35,7 @@ Hardstops:
 
 from __future__ import annotations
 
+from contextlib import contextmanager
 import logging
 import sqlite3
 import threading
@@ -43,6 +44,7 @@ import time
 from .config import (
     CANDIDATE_REVIEW_INTERVAL_S,
     CANDIDATE_REVIEW_MIN,
+    DB_PATH,
 )
 from .db import get_db
 from . import identity
@@ -51,6 +53,7 @@ logger = logging.getLogger(__name__)
 
 _started = False
 
+CANDIDATE_REVIEW_PROMPT_PREFIX = "You are a CANDIDATE REVIEWER"
 
 CANDIDATE_REVIEW_PROMPT = """\
 You are a CANDIDATE REVIEWER for thread-keeper's extract queue.
@@ -247,6 +250,67 @@ def _collect_pending(conn: sqlite3.Connection) -> tuple[str, int]:
     return ("\n".join(parts), len(rows))
 
 
+def _running_reviewer_children(conn: sqlite3.Connection) -> list[str]:
+    """Running candidate-review task ids, reaping dead rows.
+
+    Candidate review consumes one global queue. Two reviewers launched from
+    different foreground MCP servers will read the same candidates and burn
+    memory doing duplicate work, so this loop is machine-wide single-flight.
+    """
+    from .helpers import alive
+    try:
+        rows = conn.execute(
+            "SELECT id, pid FROM tasks WHERE ended_at IS NULL "
+            "AND prompt LIKE ?",
+            (CANDIDATE_REVIEW_PROMPT_PREFIX + "%",),
+        ).fetchall()
+    except sqlite3.OperationalError:
+        return []
+    now = int(time.time())
+    running: list[str] = []
+    touched = False
+    for r in rows:
+        pid = int(r["pid"] or 0)
+        if pid > 0 and not alive(pid):
+            conn.execute(
+                "UPDATE tasks SET ended_at=? WHERE id=? AND ended_at IS NULL",
+                (now, r["id"]),
+            )
+            touched = True
+            continue
+        running.append(r["id"])
+    if touched:
+        conn.commit()
+    return running
+
+
+@contextmanager
+def _review_spawn_lock():
+    """Cross-process guard for check-running-then-spawn.
+
+    The tasks-table running check is not atomic across foreground MCP
+    processes. A short file lock prevents two daemon ticks from both observing
+    no reviewer and spawning duplicate children for the same pending queue.
+    """
+    try:
+        import fcntl
+    except ImportError:  # pragma: no cover - thread-keeper runs on Unix CLIs.
+        yield True
+        return
+    lock_path = DB_PATH.parent / "candidate-reviewer.lock"
+    lock_path.parent.mkdir(parents=True, exist_ok=True)
+    with lock_path.open("w") as lock:
+        try:
+            fcntl.flock(lock.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
+        except BlockingIOError:
+            yield False
+            return
+        try:
+            yield True
+        finally:
+            fcntl.flock(lock.fileno(), fcntl.LOCK_UN)
+
+
 # ──────────────────────────────────────────────────────────────────────
 # Synchronous pass + daemon loop
 # ──────────────────────────────────────────────────────────────────────
@@ -263,53 +327,63 @@ def run_review_pass(force: bool = False) -> str:
     """
     if CANDIDATE_REVIEW_INTERVAL_S <= 0 and not force:
         return "disabled"
-    conn = get_db()
-    inventory, n_pending = _collect_pending(conn)
-    now = int(time.time())
-    if n_pending < CANDIDATE_REVIEW_MIN:
-        _record_review_pass(
-            conn, now,
-            f"below_threshold pending={n_pending} "
-            f"min={CANDIDATE_REVIEW_MIN}",
+    with _review_spawn_lock() as locked:
+        if not locked:
+            return "candidate_review_running n=1 (single-flight lock)"
+
+        conn = get_db()
+        now = int(time.time())
+        running = _running_reviewer_children(conn)
+        if running:
+            out = f"candidate_review_running n={len(running)} (single-flight)"
+            _record_review_pass(conn, now, out)
+            return out
+
+        inventory, n_pending = _collect_pending(conn)
+        if n_pending < CANDIDATE_REVIEW_MIN:
+            _record_review_pass(
+                conn, now,
+                f"below_threshold pending={n_pending} "
+                f"min={CANDIDATE_REVIEW_MIN}",
+            )
+            return f"below_threshold n={n_pending}"
+
+        active_skills = _active_skills_dump(conn)
+        full_prompt = (
+            CANDIDATE_REVIEW_PROMPT
+            + active_skills
+            + inventory
         )
-        return f"below_threshold n={n_pending}"
 
-    active_skills = _active_skills_dump(conn)
-    full_prompt = (
-        CANDIDATE_REVIEW_PROMPT
-        + active_skills
-        + inventory
-    )
+        from .tools.spawn import spawn  # type: ignore
+        try:
+            result = spawn(
+                prompt=full_prompt,
+                visible=False,
+                capture_output=True,
+                permission_mode="auto",
+                role="candidate_reviewer",
+                write_origin="candidate_review",
+                slim=True,
+                extra_allowed_tools=(
+                    "mcp__thread-keeper__skill_manage,"
+                    "mcp__thread-keeper__skill_list,"
+                    "mcp__thread-keeper__accept_candidate,"
+                    "mcp__thread-keeper__reject_candidate,"
+                    "mcp__thread-keeper__lesson_append,"
+                    "mcp__thread-keeper__mark_skill_materialized,"
+                    "Read,Write"
+                ),
+            )
+        except Exception as e:
+            _record_review_pass(conn, now, f"spawn_error: {e}")
+            return f"spawn_error: {e}"
 
-    from .tools.spawn import spawn  # type: ignore
-    try:
-        result = spawn(
-            prompt=full_prompt,
-            visible=False,
-            capture_output=True,
-            permission_mode="auto",
-            role="candidate_reviewer",
-            write_origin="candidate_review",
-            slim=True,
-            extra_allowed_tools=(
-                "mcp__thread-keeper__skill_manage,"
-                "mcp__thread-keeper__skill_list,"
-                "mcp__thread-keeper__accept_candidate,"
-                "mcp__thread-keeper__reject_candidate,"
-                "mcp__thread-keeper__lesson_append,"
-                "mcp__thread-keeper__mark_skill_materialized,"
-                "Read,Write"
-            ),
+        _record_review_pass(
+            conn, now,
+            f"spawned pending={n_pending} :: {str(result)[:140]}",
         )
-    except Exception as e:
-        _record_review_pass(conn, now, f"spawn_error: {e}")
-        return f"spawn_error: {e}"
-
-    _record_review_pass(
-        conn, now,
-        f"spawned pending={n_pending} :: {str(result)[:140]}",
-    )
-    return str(result)
+        return str(result)
 
 
 def _serve_loop() -> None:

{threadkeeper-0.9.0 → threadkeeper-0.9.1}/threadkeeper.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: threadkeeper
-Version: 0.9.0
+Version: 0.9.1
 Summary: Multi-agent shared brain across Claude Code/Desktop, Codex, Gemini, Copilot, VS Code. Cross-session memory, self-improving skill loops, inter-agent signaling — one local MCP server.
 Author: thread-keeper contributors
 License: MIT
@@ -365,7 +365,10 @@ Hard limits: max 2 new skills per pass, `[PROTECTED]` (pinned +
 foreground-authored) skills off-limits. Closes the gap between
 heuristic harvest and SKILL.md materialization — previously pending
 candidates accumulated indefinitely waiting for an agent to call
-`accept_candidate()` manually.
+`accept_candidate()` manually. The loop is machine-wide single-flight:
+while one reviewer child is running, other foreground servers/ticks report
+`candidate_review_running` instead of spawning another child for the same
+queue.
 
 #### 5. Autonomous Curator