fix: make launchd service importable

Author: lamenting-hawthorn

README.md

@@ -315,18 +315,21 @@ See `docs/safety.md` for details.
 - memory/skill proposals
 - review/apply lifecycle
 - dataset exports and manifests
-- replay benchmark reports
+- benchmark reports
 - training config generation
 - Hermes setup/status/controller run UX
 - persisted controller run history
+- macOS launchd service plist generation
+- controller-managed dataset auto-export behind explicit policy gates
 
 ### Next product steps
 
-1. Add a real background service runner (`launchd` first, then Linux systemd/cron).
-2. Add dataset readiness judging before any training plan or training runner.
-3. Add evaluator staleness detection when evaluator component hashes change.
-4. Add stronger evidence-trust scoring so exports learn from verified work, not claims.
-5. Add training planner artifacts, still approval-gated and no auto-run.
+0. Run a local real-system smoke test against this machine's Hermes `state.db` in an isolated SkillLoop project. Verify `setup --connect hermes --start`, `status`, `controller history/show`, `service install/status/uninstall`, and generated dataset manifests before loading any persistent service.
+1. Build the P1 dataset readiness judge before any training plan or training runner. It should read dataset manifests and return `ready`, `collect_more_data`, or `blocked` with machine-readable reasons.
+2. Add evaluator staleness detection when evaluator component hashes change, so controller-managed datasets do not depend on outdated scores.
+3. Add stronger evidence-trust scoring so exports learn from verified work, not assistant claims.
+4. Add approval-gated training planner artifacts after readiness/staleness gates exist.
+5. Add Linux service generation (`systemd` unit or cron) after the macOS launchd path has real local use.
 
 ## Development checks
 

docs/architecture.md

@@ -222,8 +222,9 @@ This boundary is deliberate: it keeps the learning layer inspectable, reviewable
 
 ## Roadmap priorities
 
+0. Run a local real-system smoke test against this machine's Hermes `state.db` in an isolated SkillLoop project. Verify controller ticks, status/history/show, launchd plist generation, service uninstall, and dataset manifest generation before loading any persistent service.
 1. Add a dataset readiness judge before any training plan is recommended.
 2. Add evaluator staleness detection when evaluator code/provenance changes.
 3. Add stronger evidence-trust scoring so learning artifacts depend on tool/user evidence rather than assistant claims.
-4. Add Linux service generation (`systemd` unit or cron) after the macOS launchd path has had real local use.
-5. Add approval-gated training plans; keep training execution separate until readiness, cost, evaluation, and promotion gates exist.
+4. Add approval-gated training plans; keep training execution separate until readiness, cost, evaluation, and promotion gates exist.
+5. Add Linux service generation (`systemd` unit or cron) after the macOS launchd path has had real local use.

skillloop/service.py

@@ -20,6 +20,7 @@ class ServiceSpec:
     state_dir: Path
     interval_seconds: int = DEFAULT_INTERVAL_SECONDS
     python_executable: str = sys.executable
+    python_path: str | None = None
 
     @property
     def stdout_path(self) -> Path:
@@ -58,6 +59,7 @@ def build_service_spec(
         state_dir=state,
         interval_seconds=interval_seconds,
         python_executable=python_executable or sys.executable,
+        python_path=str(Path(__file__).resolve().parents[1]),
     )
 
 
@@ -80,6 +82,7 @@ def launchd_plist(spec: ServiceSpec) -> dict[str, Any]:
         "StandardErrorPath": str(spec.stderr_path),
         "EnvironmentVariables": {
             "PYTHONUNBUFFERED": "1",
+            **({"PYTHONPATH": spec.python_path} if spec.python_path else {}),
         },
     }
 
@@ -121,6 +124,7 @@ def write_service_metadata(spec: ServiceSpec, *, kind: str, path: str | Path) ->
             "controller",
             "run",
         ],
+        "python_path": spec.python_path,
     }
     spec.metadata_path.write_text(json.dumps(payload, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
     return spec.metadata_path

tests/test_service.py

@@ -31,6 +31,8 @@ def test_launchd_plist_runs_controller_tick(tmp_path):
         "controller",
         "run",
     ]
+    assert payload["EnvironmentVariables"]["PYTHONUNBUFFERED"] == "1"
+    assert payload["EnvironmentVariables"]["PYTHONPATH"]
 
 
 def test_service_install_status_and_uninstall_cli(tmp_path, capsys):
@@ -57,6 +59,7 @@ def test_service_install_status_and_uninstall_cli(tmp_path, capsys):
     assert metadata["kind"] == "launchd"
     assert metadata["label"] == "com.skillloop.test"
     assert metadata["path"] == str(plist_path)
+    assert metadata["python_path"]
 
     assert main(["--path", str(tmp_path), "service", "status"]) == 0
     assert main(["--path", str(tmp_path), "service", "status", "--json"]) == 0