drydock-cli 3.0.123__tar.gz → 3.0.125__tar.gz

{drydock_cli-3.0.123/drydock_cli.egg-info → drydock_cli-3.0.125}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.123
+Version: 3.0.125
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

{drydock_cli-3.0.123 → drydock_cli-3.0.125}/drydock/agent.py

@@ -48,6 +48,7 @@ from drydock.compaction import (
 )
 from drydock.loop_detect import LoopTracker
 from drydock.task_state import TaskState
+from drydock.verification import looks_like_verification
 from drydock.tuning import (
     filter_tool_schemas,
     hallucinated_tool_message,
@@ -153,6 +154,8 @@ def run(
         return cancel is not None and cancel.is_set()
     tool_call_count = 0
     session_has_edited = False
+    ran_verification = False   # a test/check/exec command ran (verification gate)
+    verify_gate_nudges = 0     # bounded "verify before you finish" nudges
     leaked_call_retries = 0
     plan_continue_nudges = 0  # consecutive "you stopped mid-plan" nudges
     empty_response_nudges = 0  # consecutive "you returned nothing" nudges
@@ -332,6 +335,33 @@ def run(
                     ),
                 })
                 continue
+            # VERIFICATION GATE (PRD Epic B): a text-only "done" is not evidence.
+            # If the agent CHANGED files but never ran a test/check/its own code,
+            # don't accept completion — make it verify first. Bounded so it can't
+            # wedge; once it runs any check (ran_verification) the gate is satisfied.
+            if (
+                config.get("verify_gate", True)
+                and allow is None            # main task only — not scoped sub-agents
+                and session_has_edited
+                and not ran_verification
+                and verify_gate_nudges < 2
+            ):
+                verify_gate_nudges += 1
+                state.task.phase = "verify"
+                state.messages.append({
+                    "role": "user",
+                    "content": (
+                        "[SYSTEM] You changed files but have not VERIFIED the work. "
+                        "Before finishing, run a concrete check: the task's own test/"
+                        "eval/build (pytest, test.sh, make, npm test, …) or run the "
+                        "code/script you produced and confirm the output is correct "
+                        "and meets EVERY requirement. Do that now with a tool call; "
+                        "if it fails, fix it and re-check."
+                    ),
+                })
+                continue
+            if session_has_edited and ran_verification:
+                state.task.phase = "complete"
             break
 
         # Execute each tool call
@@ -339,6 +369,10 @@ def run(
             tool_call_count += 1
             if tc["name"] in ("Edit", "Write"):
                 session_has_edited = True
+            elif tc["name"] == "Bash" and looks_like_verification(
+                (tc.get("input") or {}).get("command", "")
+            ):
+                ran_verification = True  # the agent actually ran a test/check/its code
 
             # STOP pressed: don't run the remaining tools, but still record a
             # paired result for each (the assistant message already lists all

{drydock_cli-3.0.123 → drydock_cli-3.0.125}/drydock/config.py

@@ -48,6 +48,10 @@ DEFAULTS: dict[str, object] = {
     # system prompt every turn so they survive context compaction — the model can't
     # drift off the original goal on a long task. Set false to disable.
     "task_anchor": True,
+    # Verification gate (PRD Epic B): if the model claims "done" after changing
+    # files but never ran a test/check/its own code, nudge it to verify first
+    # instead of accepting completion. Bounded. Set false to disable.
+    "verify_gate": True,
     "theme": "harbor",
     # Optional SECOND model ("advisor") for a stronger second opinion — e.g. a
     # Gemini OpenAI-compatible endpoint on another box. Empty = disabled. The

drydock_cli-3.0.125/drydock/verification.py

@@ -0,0 +1,34 @@
+"""Verification gating (Agent-Buildout PRD, Epic B / verification):
+
+  "A text-only model response is not evidence that a task is complete."
+
+The agent loop uses this to decide, when the model claims "done", whether it
+actually VERIFIED — i.e. ran a test, a linter/type-check, a build, or executed the
+code it produced. If it changed files but ran nothing, the loop nudges it to verify
+instead of accepting completion. Deliberately small (a command classifier + the loop
+gate) so it drops into the current loop without a rewrite. Logic original to Drydock.
+"""
+from __future__ import annotations
+
+import re
+
+# Explicit test / lint / type-check / build commands.
+_VERIFY_KEYWORDS = re.compile(
+    r"\b(pytest|py\.test|unittest|nosetests?|tox|test\.sh|run_tests|"
+    r"make(\s+\w+)?|npm\s+(test|run)|yarn\s+(test|run)|pnpm\s+(test|run)|jest|mocha|vitest|"
+    r"cargo\s+(test|check|build|run|clippy)|go\s+(test|build|run|vet)|"
+    r"ruff|flake8|pylint|mypy|pyright|eslint|tsc|prettier|"
+    r"ctest|cmake|gradle|mvn|rspec|phpunit|dotnet\s+test|bats)\b",
+    re.I,
+)
+# Executing something — running the produced script/binary counts as checking it.
+_EXEC = re.compile(r"(?:^|[\s;&|(])(?:python3?|node|deno|bun|bash|sh|ruby|perl|"
+                   r"java|javac|gcc|g\+\+|clang|\./|\.\\)")
+
+
+def looks_like_verification(cmd) -> bool:
+    """True if a shell command constitutes verifying work — running tests/checks/
+    builds, or executing the produced code. Best-effort; never raises."""
+    if not cmd or not isinstance(cmd, str):
+        return False
+    return bool(_VERIFY_KEYWORDS.search(cmd) or _EXEC.search(cmd))

{drydock_cli-3.0.123 → drydock_cli-3.0.125/drydock_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.123
+Version: 3.0.125
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

{drydock_cli-3.0.123 → drydock_cli-3.0.125}/drydock_cli.egg-info/SOURCES.txt

@@ -29,6 +29,7 @@ drydock/suggest.py
 drydock/task_state.py
 drydock/tool_registry.py
 drydock/tuning.py
+drydock/verification.py
 drydock/web.py
 drydock/builtin_skills/__init__.py
 drydock/builtin_skills/ml-data.md
@@ -126,6 +127,7 @@ tests/test_tool_arg_parsing.py
 tests/test_tools_undo.py
 tests/test_tui.py
 tests/test_tuning.py
+tests/test_verification_gate.py
 tests/test_viewimage.py
 tests/test_vision_input.py
 tests/test_web_tools.py

{drydock_cli-3.0.123 → drydock_cli-3.0.125}/pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "setuptools.build_meta"
 # PyPI distribution name is drydock-cli (the established install name, continued
 # from the retired v2 fork); the import package + CLI command stay `drydock`.
 name = "drydock-cli"
-version = "3.0.123"
+version = "3.0.125"
 description = "Drydock — a local, provider-agnostic terminal coding agent for local LLMs"
 readme = "README.md"
 requires-python = ">=3.11"

{drydock_cli-3.0.123 → drydock_cli-3.0.125}/tests/test_failure_loop.py

@@ -38,7 +38,7 @@ def test_changing_args_do_not_trip_valve(monkeypatch):
     monkeypatch.setattr(agent_mod, "stream", lambda **kw: iter([seq.pop(0)]))
     st = AgentState()
     import tempfile
-    list(run("go", st, {"model": "m", "cwd": tempfile.mkdtemp()}, "sys"))
+    list(run("go", st, {"model": "m", "cwd": tempfile.mkdtemp(), "verify_gate": False}, "sys"))
     assert any(m.get("role") == "assistant" and m.get("content") == "done"
                for m in st.messages)
 
@@ -57,7 +57,7 @@ def test_identical_SUCCESS_streak_ends_turn(monkeypatch):
 
     monkeypatch.setattr(agent_mod, "stream", counting)
     st = AgentState()
-    list(run("go", st, {"model": "m", "cwd": tempfile.mkdtemp()}, "sys"))
+    list(run("go", st, {"model": "m", "cwd": tempfile.mkdtemp(), "verify_gate": False}, "sys"))
     assert calls["n"] <= 12  # identical success → stopped at the cap, not 200
 
 
@@ -73,7 +73,7 @@ def test_polling_with_changing_result_not_tripped(monkeypatch):
     seq.append(AssistantTurn("done", [], 1, 1))
     monkeypatch.setattr(agent_mod, "stream", lambda **kw: iter([seq.pop(0)]))
     st = AgentState()
-    list(run("go", st, {"model": "m", "cwd": tempfile.mkdtemp()}, "sys"))
+    list(run("go", st, {"model": "m", "cwd": tempfile.mkdtemp(), "verify_gate": False}, "sys"))
     assert any(m.get("content") == "done" for m in st.messages)
 
 

drydock_cli-3.0.125/tests/test_verification_gate.py

@@ -0,0 +1,71 @@
+"""Verification gating (Agent-Buildout PRD Epic B): a text-only 'done' after editing
+files is not accepted until the agent runs a check. Model can't self-declare done."""
+from __future__ import annotations
+
+from drydock import agent
+from drydock.providers import AssistantTurn
+from drydock.verification import looks_like_verification as v
+
+
+def test_classifier():
+    for cmd in ("pytest -q", "python -m pytest tests/", "bash test.sh", "make test",
+                "npm test", "cargo test", "ruff check .", "mypy src", "python solve.py",
+                "./run_check", "node index.js"):
+        assert v(cmd), cmd
+    for cmd in ("ls -la", "cat file.txt", "grep foo bar", "echo hi", "git status", ""):
+        assert not v(cmd), cmd
+    assert not v(None) and not v(123)
+
+
+def _drive(cfg, turns):
+    """turns: list of AssistantTurn to yield in order; returns the messages seen."""
+    calls = {"i": 0}
+    def fake_stream(model, system, messages, tool_schemas, config):
+        i = calls["i"]; calls["i"] += 1
+        yield turns[min(i, len(turns) - 1)]
+    orig = agent.stream; agent.stream = fake_stream
+    try:
+        st = agent.AgentState()
+        list(agent.run("do it", st, cfg, "SYS"))
+        return st, calls["i"]
+    finally:
+        agent.stream = orig
+
+
+CFG = {"model": "gemma4", "max_tokens": 8192, "context_limit": 65536, "verify_gate": True}
+
+
+def test_gate_fires_after_edit_without_verification():
+    # turn 1: Write a file (no tool result needed — execute() runs the real tool);
+    # turn 2: text-only "done". Gate should nudge -> a 3rd call happens.
+    write = AssistantTurn("", [{"id": "1", "name": "Write",
+                                "input": {"file_path": "/tmp/dd_vg.txt", "content": "x"}}], 5, 5)
+    done = AssistantTurn("All done!", [], 5, 5)
+    st, n_calls = _drive(CFG, [write, done, done])
+    # the gate injected a verify nudge → the loop asked the model again (>2 calls)
+    assert n_calls >= 3
+    assert any("VERIFIED" in m.get("content", "") for m in st.messages if m.get("role") == "user")
+
+
+def test_no_gate_when_verification_ran():
+    write = AssistantTurn("", [{"id": "1", "name": "Write",
+                                "input": {"file_path": "/tmp/dd_vg2.txt", "content": "x"}}], 5, 5)
+    check = AssistantTurn("", [{"id": "2", "name": "Bash", "input": {"command": "python -c 'print(1)'"}}], 5, 5)
+    done = AssistantTurn("Done, verified.", [], 5, 5)
+    st, _ = _drive(CFG, [write, check, done])
+    assert not any("VERIFIED" in m.get("content", "") for m in st.messages if m.get("role") == "user")
+    assert st.task.phase == "complete"
+
+
+def test_no_gate_when_no_edits():
+    # pure Q&A: no files changed → nothing to verify → accept immediately
+    st, _ = _drive(CFG, [AssistantTurn("The answer is 42.", [], 5, 5)])
+    assert not any("VERIFIED" in m.get("content", "") for m in st.messages if m.get("role") == "user")
+
+
+def test_disabled():
+    write = AssistantTurn("", [{"id": "1", "name": "Write",
+                                "input": {"file_path": "/tmp/dd_vg3.txt", "content": "x"}}], 5, 5)
+    done = AssistantTurn("done", [], 5, 5)
+    st, _ = _drive({**CFG, "verify_gate": False}, [write, done])
+    assert not any("VERIFIED" in m.get("content", "") for m in st.messages if m.get("role") == "user")