researchloop 0.3.2__tar.gz → 0.3.4__tar.gz

{researchloop-0.3.2 → researchloop-0.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: researchloop
-Version: 0.3.2
+Version: 0.3.4
 Summary: Automated research sprint platform for HPC clusters
 License: MIT
 License-File: LICENSE

{researchloop-0.3.2 → researchloop-0.3.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "researchloop"
-version = "0.3.2"
+version = "0.3.4"
 description = "Automated research sprint platform for HPC clusters"
 readme = "README.md"
 license = {text = "MIT"}

researchloop-0.3.4/researchloop/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.3.4"

{researchloop-0.3.2 → researchloop-0.3.4}/researchloop/runner/job_templates/sge.sh.j2

@@ -241,6 +241,17 @@ print(json.dumps({
 }
 
 _heartbeat_loop() {
+    # The parent script runs with `set -euo pipefail`, which is inherited
+    # by this backgrounded subshell. An unprotected command-substitution
+    # failure here (e.g. `ls -t pattern | head -1` racing into SIGPIPE on
+    # ls under pipefail, or a glob that matches nothing) is enough to kill
+    # the entire watchdog and leave the sprint without heartbeats or
+    # STUCK_PIPE detection for the rest of its wall-clock. Disable strict
+    # mode for the watchdog — a noisy heartbeat is far better than a
+    # silently-dead one.
+    set +e
+    set +o pipefail
+
     # Watchdog: detect and recover from hung pipelines.
     #
     # STUCK_PIPE warn: the active step's stream-json file goes silent for

{researchloop-0.3.2 → researchloop-0.3.4}/researchloop/runner/job_templates/slurm.sh.j2

@@ -243,6 +243,17 @@ print(json.dumps({
 }
 
 _heartbeat_loop() {
+    # The parent script runs with `set -euo pipefail`, which is inherited
+    # by this backgrounded subshell. An unprotected command-substitution
+    # failure here (e.g. `ls -t pattern | head -1` racing into SIGPIPE on
+    # ls under pipefail, or a glob that matches nothing) is enough to kill
+    # the entire watchdog and leave the sprint without heartbeats or
+    # STUCK_PIPE detection for the rest of its wall-clock. Disable strict
+    # mode for the watchdog — a noisy heartbeat is far better than a
+    # silently-dead one.
+    set +e
+    set +o pipefail
+
     # Watchdog: detect and recover from hung pipelines.
     #
     # STUCK_PIPE warn: the active step's stream-json file goes silent for

{researchloop-0.3.2 → researchloop-0.3.4}/researchloop/runner/templates/fix_issues.md.j2

@@ -9,3 +9,4 @@ Review the file: red_team_round_{{ round_number }}.md
 - Document what you changed and why in fixes_round_{{ round_number }}.md
 - Do not break existing working functionality
 - Update progress.md with what you're fixing and the results
+- Run any re-training or re-evaluation commands synchronously in the foreground. NEVER pass `run_in_background: true` on a Bash tool call, and do NOT use `nohup`, `disown`, `setsid`, or `&` to detach work — this is a one-shot `claude -p` session and any orphaned subprocess will be killed when you end your turn.

{researchloop-0.3.2 → researchloop-0.3.4}/researchloop/runner/templates/research_sprint.md.j2

@@ -23,6 +23,17 @@ python train.py 2>&1 | tee -a output.log
 
 This lets the team monitor script output remotely. Always use `tee -a` (append mode) so all runs accumulate in the same log file. Do this for every script execution, training run, or evaluation.
 
+## Long-running commands — run synchronously, NEVER in the background
+This sprint runs as a single one-shot `claude -p` invocation. There is no notification mechanism, no resumed turn, and no way to "wait and be notified". The moment you end your turn, the session is over and the sprint runner kills any detached subprocesses (training, evaluation, anything you started).
+
+Rules:
+- NEVER pass `run_in_background: true` on a Bash tool call. Always run training and evaluation in the foreground.
+- NEVER use `nohup`, `disown`, `setsid`, `&`, `&>`, or any other shell-level backgrounding for work you care about. The orphan will be killed.
+- If a single Bash call needs to run longer than its default timeout, raise the per-call `timeout` parameter instead of backgrounding it.
+- The `PushNotification`, `Monitor`, `Cron*`, `Task`, `AskUserQuestion`, `EnterPlanMode`, and `EnterWorktree` tools are NOT available in this one-shot mode. Ignore them if they appear in your tool list.
+
+Plan training as a series of synchronous foreground commands. Wait for each to finish before ending your turn.
+
 ## Progress Log
 Maintain a file called `progress.md` in the sprint directory. Update it regularly as you work — it's how the team monitors your progress remotely. Keep it concise and current:
 

{researchloop-0.3.2 → researchloop-0.3.4}/tests/test_runner.py

@@ -102,6 +102,24 @@ class TestRenderTemplate:
         )
         assert "progress.md" in output
 
+    def test_research_template_forbids_backgrounding(self):
+        """claude -p is one-shot — backgrounded subprocesses get orphaned and
+        killed by the runner's pipeline-cleanup watchdog. The template must
+        tell claude not to use run_in_background or shell-level detach."""
+        output = render_template(
+            "research_sprint.md.j2",
+            study_context="Study context",
+            idea="test idea",
+            sprint_dir="/tmp/sprint",
+        )
+        assert "run_in_background" in output
+        assert "nohup" in output
+        assert "one-shot" in output
+
+    def test_fix_template_forbids_backgrounding(self):
+        output = render_template("fix_issues.md.j2", round_number=1)
+        assert "run_in_background" in output
+
     def test_red_team_template(self):
         output = render_template(
             "red_team.md.j2",
@@ -158,6 +176,21 @@ class TestJobScriptWatchdog:
         # Once-per-stuck-episode flag, not log-every-heartbeat.
         assert "stuck_warned" in script
 
+    def _assert_heartbeat_loop_disables_strict_mode(self, script: str) -> None:
+        # The parent script has set -euo pipefail. The backgrounded
+        # _heartbeat_loop inherits it — without explicitly disabling errexit
+        # and pipefail, a single command-substitution failure (e.g. an
+        # ls|head SIGPIPE race) can silently kill the watchdog and the
+        # sprint loses all heartbeat + STUCK_PIPE detection for the rest of
+        # its wall-clock. The first two `set` lines inside the function
+        # must disable both modes.
+        loop_idx = script.index("_heartbeat_loop() {")
+        body_idx = script.index("\n", loop_idx) + 1
+        # Check the disabling lines are at the top of the function body.
+        head = script[body_idx : body_idx + 800]
+        assert "set +e" in head
+        assert "set +o pipefail" in head
+
     def _assert_hung_pipeline_recovery_present(self, script: str) -> None:
         # claude must run in its own session so the watchdog can SIGTERM the
         # whole group (claude + any leaked Bash-tool subprocesses) by pgid.
@@ -191,3 +224,13 @@ class TestJobScriptWatchdog:
 
     def test_sge_template_includes_hung_pipeline_recovery(self):
         self._assert_hung_pipeline_recovery_present(_render_job_template("sge.sh.j2"))
+
+    def test_slurm_heartbeat_loop_disables_strict_mode(self):
+        self._assert_heartbeat_loop_disables_strict_mode(
+            _render_job_template("slurm.sh.j2")
+        )
+
+    def test_sge_heartbeat_loop_disables_strict_mode(self):
+        self._assert_heartbeat_loop_disables_strict_mode(
+            _render_job_template("sge.sh.j2")
+        )

{researchloop-0.3.2 → researchloop-0.3.4}/uv.lock

@@ -1152,7 +1152,7 @@ wheels = [
 
 [[package]]
 name = "researchloop"
-version = "0.3.2"
+version = "0.3.4"
 source = { editable = "." }
 dependencies = [
     { name = "aiosqlite" },

researchloop-0.3.2/researchloop/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.3.2"