drydock-cli 3.0.23__tar.gz → 3.0.25__tar.gz

{drydock_cli-3.0.23/drydock_cli.egg-info → drydock_cli-3.0.25}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.23
+Version: 3.0.25
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

{drydock_cli-3.0.23 → drydock_cli-3.0.25}/drydock/__init__.py

@@ -7,4 +7,4 @@ try:
     __version__ = _version("drydock-cli")
 except PackageNotFoundError:
     # Running from source before install (e.g. PYTHONPATH dev runs).
-    __version__ = "3.0.23"
+    __version__ = "3.0.25"

{drydock_cli-3.0.23 → drydock_cli-3.0.25}/drydock/agent.py

@@ -114,6 +114,7 @@ def run(
     session_has_edited = False
     leaked_call_retries = 0
     plan_continue_nudges = 0  # consecutive "you stopped mid-plan" nudges
+    empty_response_nudges = 0  # consecutive "you returned nothing" nudges
     run_iteration = 0  # stream calls within THIS run() (resets per user message)
     loop_tracker = LoopTracker()
 
@@ -226,6 +227,22 @@ def run(
                     ),
                 })
                 continue
+            # A completely EMPTY response (no text, no tool call, no leaked call)
+            # is a non-answer, not a deliberate "done" — a weak model sometimes
+            # returns nothing on a hard task and the turn dead-ends silently.
+            # Nudge ONCE to produce output. Narrow (only empty text) and capped,
+            # so unlike a blanket auto-continue it can't loop on real answers.
+            if not (assistant_turn.text or "").strip() and empty_response_nudges < 1:
+                empty_response_nudges += 1
+                state.messages.append({
+                    "role": "user",
+                    "content": (
+                        "[SYSTEM] Your last response was empty. Produce your "
+                        "result now: either call a tool to do the work, or reply "
+                        "with text. Do not return an empty message."
+                    ),
+                })
+                continue
             break
 
         # Execute each tool call
@@ -282,6 +299,7 @@ def run(
         # a long, productive plan can run as far as it needs (the cap only
         # bounds back-to-back stalls).
         plan_continue_nudges = 0
+        empty_response_nudges = 0
 
         # Nudge: if past 15 tool calls without any edits, inject gentle guidance
         if tool_call_count == 15 and not session_has_edited and config.get("force_first_tool"):

{drydock_cli-3.0.23 → drydock_cli-3.0.25}/drydock/providers.py

@@ -86,6 +86,33 @@ class AssistantTurn:
     had_leaked_call: bool = False  # model emitted a tool call as text, not a call
 
 
+# ── Tool-call argument parsing ────────────────────────────────────────────
+
+def _parse_tool_args(raw: str) -> dict:
+    """Parse a model's tool-call arguments tolerantly.
+
+    strict=False is the fix that matters: a Bash command built with a heredoc
+    (`cat <<EOF ...`) carries LITERAL newlines, which strict JSON rejects with
+    "Invalid control character". The call then degraded to {"_raw": ...} with no
+    usable args, Bash failed, and the model's corrections re-wrapped into nested
+    {"_raw": "{\"_raw\": ...}"} — a fatal loop. strict=False accepts control
+    chars inside strings (recovering the common multiline case); we also unwrap
+    a few layers of nested {"_raw": ...} the model may have echoed back.
+    """
+    if not raw:
+        return {}
+    for _ in range(4):  # peel nested _raw layers, then parse
+        try:
+            v = json.loads(raw, strict=False)
+        except (json.JSONDecodeError, TypeError):
+            return {"_raw": raw}
+        if isinstance(v, dict) and set(v) == {"_raw"} and isinstance(v["_raw"], str):
+            raw = v["_raw"]
+            continue
+        return v if isinstance(v, dict) else {"_raw": raw}
+    return {"_raw": raw}
+
+
 # ── Message format conversion ─────────────────────────────────────────────
 
 def messages_to_openai(messages: list, system: str) -> list:
@@ -250,10 +277,7 @@ def stream(
     tool_calls = []
     for idx in sorted(tool_buf):
         v = tool_buf[idx]
-        try:
-            inp = json.loads(v["args"]) if v["args"] else {}
-        except json.JSONDecodeError:
-            inp = {"_raw": v["args"]}
+        inp = _parse_tool_args(v["args"])
         tool_calls.append({
             "id": v["id"] or f"call_{idx}",
             "name": v["name"],
@@ -283,11 +307,7 @@ def _complete_nonstreaming(
 
     tool_calls = []
     for i, tc in enumerate(getattr(msg, "tool_calls", None) or []):
-        raw_args = tc.function.arguments or ""
-        try:
-            inp = json.loads(raw_args) if raw_args else {}
-        except json.JSONDecodeError:
-            inp = {"_raw": raw_args}
+        inp = _parse_tool_args(tc.function.arguments or "")
         tool_calls.append({
             "id": tc.id or f"call_{i}",
             "name": tc.function.name,

{drydock_cli-3.0.23 → drydock_cli-3.0.25}/drydock/tools/__init__.py

@@ -428,7 +428,16 @@ def tool_edit(params: dict, config: dict) -> str:
 
 
 def tool_bash(params: dict, config: dict) -> str:
-    cmd = params["command"]
+    cmd = params.get("command")
+    if not cmd:
+        if "_raw" in params:
+            return (
+                "Error: your tool arguments were not valid JSON so the command "
+                "could not be read. Resend as ONE JSON object "
+                '{"command": "..."} with the whole command as a single string '
+                "(escape any newline as \\n; avoid raw line breaks in the JSON)."
+            )
+        return "Error: Bash needs a non-empty 'command'."
     timeout = params.get("timeout", 30)
     reason = bash_safety.dangerous_command(cmd)
     if reason is not None:

{drydock_cli-3.0.23 → drydock_cli-3.0.25/drydock_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.23
+Version: 3.0.25
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

{drydock_cli-3.0.23 → drydock_cli-3.0.25}/drydock_cli.egg-info/SOURCES.txt

@@ -34,6 +34,7 @@ tests/test_cli_agents.py
 tests/test_compaction.py
 tests/test_config.py
 tests/test_detect.py
+tests/test_empty_response.py
 tests/test_guards_and_tools.py
 tests/test_hallucinated_tools.py
 tests/test_leaked_tool_call.py
@@ -43,6 +44,7 @@ tests/test_providers_unreachable.py
 tests/test_stop.py
 tests/test_subagent.py
 tests/test_todo.py
+tests/test_tool_arg_parsing.py
 tests/test_tools_undo.py
 tests/test_tui.py
 tests/test_tuning.py

{drydock_cli-3.0.23 → drydock_cli-3.0.25}/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 = "3.0.23"
+version = "3.0.25"
 description = "Drydock — a local, provider-agnostic terminal coding agent for local LLMs"
 readme = "README.md"
 requires-python = ">=3.12"

drydock_cli-3.0.25/tests/test_empty_response.py

@@ -0,0 +1,42 @@
+"""A completely empty assistant response (no text, no tool call, no leaked
+call) should be nudged ONCE to produce output rather than dead-ending the user
+with a silent stop — but it must not loop on real (non-empty) answers."""
+from __future__ import annotations
+
+import drydock.agent as agent_mod
+from drydock.agent import AgentState, run
+from drydock.providers import AssistantTurn
+
+
+def test_empty_response_gets_one_nudge(monkeypatch):
+    seq = [
+        AssistantTurn("", [], 1, 1),          # empty → nudge
+        AssistantTurn("here is the answer", [], 1, 1),  # then a real reply
+    ]
+    monkeypatch.setattr(agent_mod, "stream", lambda **kw: iter([seq.pop(0)]))
+    st = AgentState()
+    list(run("do it", st, {"model": "m"}, "sys"))
+    nudges = [m for m in st.messages
+              if m["role"] == "user" and "response was empty" in m["content"]]
+    assert len(nudges) == 1
+    assert not seq  # both turns consumed
+
+
+def test_empty_nudge_capped_at_one(monkeypatch):
+    # Persistently empty → nudged once, then the loop gives control back.
+    seq = [AssistantTurn("", [], 1, 1) for _ in range(6)]
+    monkeypatch.setattr(agent_mod, "stream", lambda **kw: iter([seq.pop(0)]))
+    st = AgentState()
+    list(run("do it", st, {"model": "m"}, "sys"))
+    nudges = [m for m in st.messages
+              if m["role"] == "user" and "response was empty" in m["content"]]
+    assert len(nudges) == 1  # not a spin
+
+
+def test_nonempty_text_is_not_nudged(monkeypatch):
+    seq = [AssistantTurn("all done", [], 1, 1)]
+    monkeypatch.setattr(agent_mod, "stream", lambda **kw: iter([seq.pop(0)]))
+    st = AgentState()
+    list(run("hi", st, {"model": "m"}, "sys"))
+    assert not [m for m in st.messages
+                if m["role"] == "user" and "response was empty" in m["content"]]

drydock_cli-3.0.25/tests/test_tool_arg_parsing.py

@@ -0,0 +1,45 @@
+"""Tolerant parsing of a model's tool-call arguments.
+
+Gemma builds Bash commands with heredocs (`cat <<EOF ...`) whose arguments
+carry LITERAL newlines. Strict json.loads rejects those ("Invalid control
+character"), the call degraded to {"_raw": ...} with no `command`, Bash failed,
+and the model's retries nested {"_raw": "{\\"_raw\\": ...}"} into a fatal loop
+(seen live on crack-7z-hash). _parse_tool_args fixes both."""
+from __future__ import annotations
+
+import json
+
+from drydock.providers import _parse_tool_args
+from drydock.tools import tool_bash
+
+
+def test_heredoc_with_literal_newlines_recovers_command():
+    raw = '{"command": "cat <<EOF > p.txt\npassword\n123456\nEOF"}'
+    out = _parse_tool_args(raw)
+    assert out.get("command", "").startswith("cat <<EOF > p.txt")
+    assert "\n" in out["command"]  # newlines preserved
+
+
+def test_nested_raw_is_unwrapped():
+    triple = json.dumps({"_raw": json.dumps({"_raw": json.dumps({"command": "ls -la"})})})
+    assert _parse_tool_args(triple) == {"command": "ls -la"}
+
+
+def test_normal_and_empty_and_broken():
+    assert _parse_tool_args('{"command":"echo hi"}') == {"command": "echo hi"}
+    assert _parse_tool_args("") == {}
+    assert _parse_tool_args("{not json") == {"_raw": "{not json"}
+
+
+def test_bash_gives_recovery_hint_on_unparsed_args():
+    out = tool_bash({"_raw": "garbage"}, {})
+    assert "not valid JSON" in out and '{"command"' in out
+
+
+def test_bash_runs_recovered_heredoc(tmp_path):
+    # End-to-end: a heredoc command recovered by _parse_tool_args actually runs.
+    raw = ('{"command": "cat <<EOF > note.txt\nhello\nworld\nEOF"}')
+    params = _parse_tool_args(raw)
+    out = tool_bash(params, {"cwd": str(tmp_path)})
+    assert "Error" not in out
+    assert (tmp_path / "note.txt").read_text() == "hello\nworld\n"