tests: pytest-runnable + finish Smoke→Test scrub

.github/workflows/ci.yml

@@ -30,24 +30,5 @@ jobs:
       - name: Ruff
         run: python -m ruff check pyagent tests
 
-      - name: Smoke tests
-        run: |
-          set +e
-          fail=0
-          failed=""
-          for t in tests/test_*.py; do
-            name=$(basename "$t" .py)
-            echo "::group::$name"
-            python -m "tests.$name"
-            rc=$?
-            echo "::endgroup::"
-            if [ $rc -ne 0 ]; then
-              fail=$((fail+1))
-              failed="$failed $name"
-            fi
-          done
-          if [ $fail -gt 0 ]; then
-            echo "FAILED ($fail tests):$failed"
-            exit 1
-          fi
-          echo "All tests passed"
+      - name: Tests
+        run: python -m pytest tests/

README.md

@@ -85,18 +85,20 @@ Local Ollama needs no key — just the daemon running and a model pulled.
 
 ## Tests
 
-Smoke tests live under `tests/` as standalone scripts (no pytest):
-
 ```bash
 pip install -e '.[dev]'
-pre-commit install                       # one-time: run black + ruff on every commit
-python -m tests.test_token_meter        # run one
-for f in tests/test_*.py; do python -m tests.$(basename "$f" .py); done
+pre-commit install      # one-time: run black + ruff on every commit
+pytest tests/           # run everything
+pytest tests/test_token_meter.py        # run one file
 ```
 
+Each `tests/test_*.py` also runs standalone via
+`python -m tests.test_token_meter`, but pytest is the default
+runner and what CI uses.
+
 Three recall sub-tests in `test_plugins.py` are gated on
 `PYAGENT_HEAVY_TESTS=1` because they download a ~130MB embedding
-model. CI runs the same loop on every push and pull request.
+model.
 
 ## License
 

tests/test_agent_label.py

@@ -1,4 +1,4 @@
-"""Smoke for the CLI's _agent_label rendering.
+"""Test of the CLI's _agent_label rendering.
 
 The label has to survive rich's markup parser and produce
 `[<agent_id>] ` in cyan around the brackets. A previous version used
@@ -51,3 +51,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_arg_scrubbing.py

@@ -1,4 +1,4 @@
-"""Smoke for tool-call arg scrubbing.
+"""Test of tool-call arg scrubbing.
 
 Large string args in tool_call dicts get replaced with a short
 marker after the tool runs, so a write_file with 50KB of content

tests/test_ask_parent.py

@@ -1,4 +1,4 @@
-"""Smoke for mid-task subagent ↔ parent conversation (issue #47).
+"""Test of mid-task subagent ↔ parent conversation (issue #47).
 
 Drives both sides of `ask_parent` / `reply_to_subagent` against
 real `_ChildState` IO threads without spawning subagent
@@ -325,3 +325,8 @@ def asker(question: str) -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_async_subagent.py

@@ -1,4 +1,4 @@
-"""Smoke for async subagent dispatch.
+"""Test of async subagent dispatch.
 
 Covers:
   1. `_drain_pending_async()` — appends queued replies as user-role
@@ -183,3 +183,8 @@ def runner():
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_attachment_lru.py

@@ -1,4 +1,4 @@
-"""Smoke for the per-session attachments LRU cap (issue #86).
+"""Test of the per-session attachments LRU cap (issue #86).
 
 Locks these behaviors:
 
@@ -287,3 +287,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_auto_venv.py

@@ -1,4 +1,4 @@
-"""Smoke for venv discovery / creation + the `python_env` tool.
+"""Test of venv discovery / creation + the `python_env` tool.
 
 Locks:
   1. `discover` finds nothing in an empty workspace.
@@ -242,3 +242,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_background_exec.py

@@ -301,3 +301,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_bench_defaults.py

@@ -1,4 +1,4 @@
-"""Smoke for `pyagent-bench`'s default-budget-by-model table.
+"""Test of `pyagent-bench`'s default-budget-by-model table.
 
 Locks the per-model defaults so a future model rename / pricing-table
 update doesn't silently make Opus runs halt at the Sonnet budget (or
@@ -53,3 +53,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_call_tool.py

@@ -1,4 +1,4 @@
-"""Smoke tests for `PluginAPI.call_tool` — cross-plugin composition.
+"""Tests for `PluginAPI.call_tool` — cross-plugin composition.
 
 Covers:
   - Plugin B's tool calls plugin A's tool through `api.call_tool`.

tests/test_checklist.py

@@ -1,4 +1,4 @@
-"""Smoke for the agent-managed checklist (issue #38).
+"""Test of the agent-managed checklist (issue #38).
 
 Locks the contract for `pyagent.checklist.Checklist`, the three
 tool factories that bind to it, the `checklist` event flowing into
@@ -307,3 +307,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_claude_code_cli.py

@@ -1,4 +1,4 @@
-"""Smoke tests for the claude-code-cli plugin's early-return error paths.
+"""Tests for the claude-code-cli plugin's early-return error paths.
 
 All cases exercised here return before subprocess is invoked, so the
 tests don't require the `claude` binary to be present on PATH. The

tests/test_cli_render.py

@@ -437,3 +437,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_code_mapper.py

@@ -1069,3 +1069,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_context_window.py

@@ -374,3 +374,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_controlling_hooks.py

@@ -1,4 +1,4 @@
-"""Smoke tests for v2 controlling hooks (issue #66).
+"""Tests for v2 controlling hooks (issue #66).
 
 Covers every acceptance-criteria bullet:
 

tests/test_ctrlc.py

@@ -1,4 +1,4 @@
-"""Smoke for clean Ctrl+C behavior.
+"""Test of clean Ctrl+C behavior.
 
 Subprocess-launches the full `pyagent` CLI in its own process group
 (simulating an interactive terminal), waits for the agent to become
@@ -93,3 +93,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_doc_tools.py

@@ -710,3 +710,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_edit_file.py

@@ -1,4 +1,4 @@
-"""Smoke for edit_file.
+"""Test of edit_file.
 
 Exercises (in-process):
   1. Single-match replace, with correct 1-indexed line number in
@@ -134,3 +134,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_glob.py

@@ -1,4 +1,4 @@
-"""Smoke for glob.
+"""Test of glob.
 
 Exercises (in-process):
   1. Single-pattern recursive match returns sorted relative paths.
@@ -150,3 +150,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_grep.py

@@ -1,4 +1,4 @@
-"""Smoke for grep.
+"""Test of grep.
 
 Exercises (in-process):
   1. Default behavior: bare pattern/path returns `path:lineno:line`,
@@ -217,3 +217,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_hn_search.py

@@ -409,3 +409,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_html_tools.py

@@ -240,3 +240,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_kill_active.py

@@ -72,3 +72,8 @@ def runner() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_library_usage.py

@@ -1,4 +1,4 @@
-"""Smoke for the pyagent top-level library surface.
+"""Test of the pyagent top-level library surface.
 
 Locks the public re-exports + auto_client behavior. These are the
 names the README/library-usage.md teach; if they go missing, the
@@ -158,3 +158,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_list_models.py

@@ -334,3 +334,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_memory_category_drift.py

@@ -1,4 +1,4 @@
-"""Smoke for the Tier-2 memory category-drift improvements.
+"""Test of the Tier-2 memory category-drift improvements.
 
 Locks four behaviors:
 
@@ -256,3 +256,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_memory_recall_improvements.py

@@ -1,4 +1,4 @@
-"""Smoke for the Tier-1 memory recall improvements.
+"""Test of the Tier-1 memory recall improvements.
 
 Locks three behaviors:
 
@@ -247,3 +247,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_notify.py

@@ -1,4 +1,4 @@
-"""Smoke for the subagent ↔ parent notification protocol (issue #64).
+"""Test of the subagent ↔ parent notification protocol (issue #64).
 
 Drives both directions of the new `subagent_note` / `parent_note`
 pipe events against real `_ChildState` IO threads without spawning
@@ -341,3 +341,8 @@ def main() -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()

tests/test_notify_surface.py

@@ -1,4 +1,4 @@
-"""Smoke for the parent-side notification surface (issue #65).
+"""Test of the parent-side notification surface (issue #65).
 
 Drives `tell_subagent` and `peek_subagent` against real
 `_ChildState` IO threads with fake subagent pipes — the same
@@ -496,3 +496,8 @@ def _capture_unread(count: int, by_sev: dict[str, int]) -> None:
 
 if __name__ == "__main__":
     main()
+
+
+def test_main() -> None:
+    """Entry point for pytest; runs the standalone main()."""
+    main()