drydock-cli 1.3.6__tar.gz → 1.3.9__tar.gz

{drydock_cli-1.3.6 → drydock_cli-1.3.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 1.3.6
+Version: 1.3.9
 Summary: Drydock — nautical CLI coding agent
 Project-URL: Homepage, https://github.com/fbobe321/drydock
 Project-URL: Repository, https://github.com/fbobe321/drydock

{drydock_cli-1.3.6 → drydock_cli-1.3.9}/drydock/core/agent_loop.py

@@ -617,25 +617,22 @@ class AgentLoop:
                     )
 
                 # Bash abuse detection: model uses bash instead of proper tools
-                # This is the #1 cause of no-patch failures (88% of cases)
-                if not has_made_edit and bash_count >= 5:
-                    if bash_count == 5:
+                if not has_made_edit and bash_count >= 3:
+                    if bash_count == 3:
                         self._inject_system_note(
-                            "You have run 5 bash commands without making any edit. "
-                            "STOP using bash for reading/searching. Use the proper tools: "
-                            "read_file (not cat), grep (not bash grep), search_replace (not sed). "
-                            "Call search_replace NOW to fix the bug."
+                            "You have run 3 bash commands without creating or editing any files. "
+                            "STOP running bash. Start CREATING files with write_file or "
+                            "EDITING files with search_replace. bash does not create code."
                         )
-                    elif bash_count == 8:
+                    elif bash_count == 6:
                         self._inject_system_note(
-                            "WARNING: 8 bash commands, ZERO edits. You are wasting context. "
-                            "Bash does NOT fix bugs — search_replace does. "
-                            "Use search_replace on your NEXT response or the session ends."
+                            "WARNING: 6 bash commands, ZERO file edits. You are stuck in a loop. "
+                            "Your NEXT action MUST be write_file to create a file or "
+                            "search_replace to edit one. Do NOT run any more bash commands."
                         )
-                    elif bash_count >= 12:
-                        # Force stop — model is clearly stuck in bash loop
+                    elif bash_count >= 10:
                         yield AssistantEvent(
-                            content="\n\n[STOPPED: 12+ bash commands without any edit. Use search_replace.]\n",
+                            content="\n\n[STOPPED: 10+ bash commands without creating any files. Use write_file.]\n",
                             stopped_by_middleware=True,
                         )
                         return
@@ -935,22 +932,16 @@ class AgentLoop:
             f"{tool_call.tool_name}:{args_str}".encode()
         ).hexdigest()
 
-        # Determine threshold based on tool type
-        if tool_call.tool_name in ("grep", "read_file", "glob", "tool_search", "lsp"):
-            threshold = 4  # Read-only tools get more room
-        elif tool_call.tool_name == "bash":
-            # Check if it's a read-only bash command
-            cmd = args_str.lower()
-            read_only = any(k in cmd for k in ["ls ", "pwd", "cat ", "head ", "tail ",
-                           "git status", "git log", "git diff", "echo ", "which ", "whoami"])
-            threshold = 4 if read_only else 3
-        elif tool_call.tool_name in ("search_replace", "write_file"):
-            threshold = 2  # Write tools — strict
-        else:
-            threshold = 3
-
         count, last_result = self._tool_call_history.get(sig, (0, ""))
-        if count >= threshold:
+        is_failed = last_result.startswith("FAILED:") if last_result else False
+
+        # ONLY block commands that FAILED — successful commands are allowed to repeat.
+        # The model legitimately runs ls, git status, etc. multiple times during work.
+        # Loop detection (_check_tool_call_repetition) handles pattern-based loops.
+        if not (is_failed and count >= 2):
+            return None
+
+        if count >= 2:
             attempted = self._get_attempted_summary()
             msg = (
                 f"CIRCUIT BREAKER: You already ran `{tool_call.tool_name}` with these "
@@ -993,7 +984,7 @@ class AgentLoop:
         if blocked := self._circuit_breaker_check(tool_call):
             self._consecutive_circuit_breaker_fires += 1
 
-            if self._consecutive_circuit_breaker_fires >= 3:
+            if self._consecutive_circuit_breaker_fires >= 5:
                 # Model is ignoring the circuit breaker — force stop
                 force_msg = (
                     f"FORCED STOP: You ignored the circuit breaker {self._consecutive_circuit_breaker_fires} times. "
@@ -1190,7 +1181,7 @@ class AgentLoop:
             yield event
         for tool_call in resolved.tool_calls:
             # Stop processing more tool calls if circuit breaker force-stopped
-            if self._consecutive_circuit_breaker_fires >= 3:
+            if self._consecutive_circuit_breaker_fires >= 5:
                 break
 
             yield ToolCallEvent(

{drydock_cli-1.3.6 → drydock_cli-1.3.9}/drydock/core/prompts/cli.md

@@ -209,11 +209,13 @@ Bug Fixing in Open-Source Repos
 NEVER edit test files. The bug is in the library source code, not the tests.
 
 CRITICAL — Tool Usage:
+- To CREATE files: use write_file. NEVER use bash with touch/echo>/cat<<EOF.
 - To EDIT files: use search_replace. NEVER use bash with sed/awk/echo to edit files.
 - To READ files: use read_file. NEVER use bash with cat/head/tail.
 - To SEARCH code: use grep. NEVER use bash with grep/find/rg.
-- bash is ONLY for: running tests, checking git status, installing packages, running scripts.
-- If you have not called search_replace, you have not fixed anything. Bash commands do not fix bugs.
+- To CREATE directories: bash mkdir -p is fine.
+- bash is ONLY for: running tests, git commands, installing packages, mkdir, running scripts.
+- If you have not called write_file or search_replace, you have not created or fixed anything.
 
 Error Handling:
 - If a command FAILS, do NOT repeat it. Read the error message and adapt.

{drydock_cli-1.3.6 → drydock_cli-1.3.9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "1.3.6"
+version = "1.3.9"
 description = "Drydock — nautical CLI coding agent"
 readme = "README.md"
 requires-python = ">=3.12"

drydock_cli-1.3.9/tests/test_real_workflow.py

@@ -0,0 +1,86 @@
+"""Real workflow tests — the EXACT scenarios users hit.
+
+These run against the real vLLM backend with realistic prompts.
+If these pass, DryDock is usable for normal work.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+from pathlib import Path
+
+import httpx
+import pytest
+
+from drydock.core.config.harness_files import init_harness_files_manager
+try:
+    init_harness_files_manager("user", "project")
+except RuntimeError:
+    pass
+
+from drydock.core.agent_loop import AgentLoop
+from drydock.core.agents.models import BuiltinAgentName
+from drydock.core.config import Backend, ModelConfig, ProviderConfig, VibeConfig
+from drydock.core.types import AssistantEvent, ToolCallEvent, ToolResultEvent
+
+
+def _vllm_ok():
+    try:
+        return httpx.get("http://localhost:8000/v1/models", timeout=3).status_code == 200
+    except Exception:
+        return False
+
+pytestmark = pytest.mark.skipif(not _vllm_ok(), reason="vLLM not running")
+
+
+def _agent(tmp_path, max_turns=20):
+    config = VibeConfig(
+        active_model="devstral", auto_approve=True, enable_telemetry=False,
+        include_project_context=False, system_prompt_id="cli",
+        providers=[ProviderConfig(name="local", api_base="http://localhost:8000/v1", api_key_env_var="", backend=Backend.GENERIC)],
+        models=[ModelConfig(name="devstral", provider="local", input_price=0, output_price=0)],
+        session_logging={"enabled": False, "save_dir": str(tmp_path / "logs")},
+    )
+    os.chdir(tmp_path)
+    return AgentLoop(config=config, agent_name=BuiltinAgentName.AUTO_APPROVE, max_turns=max_turns)
+
+
+async def _run(agent, prompt, max_events=150):
+    events = []
+    async for ev in agent.act(prompt):
+        events.append(ev)
+        if len(events) >= max_events:
+            break
+    return events
+
+
+@pytest.mark.asyncio
+async def test_build_project_from_prd(tmp_path):
+    """THE user's exact scenario: read PRD, build project.
+
+    Must NOT trigger circuit breaker or crash.
+    Must create at least some files.
+    """
+    (tmp_path / "PRD.md").write_text(
+        "# CLI Tool\\n\\nBuild a Python CLI tool.\\n"
+        "## Features\\n- Parse input files\\n- Show statistics\\n"
+        "## Structure\\n- myapp/ package\\n- parser.py, cli.py\\n"
+    )
+
+    agent = _agent(tmp_path)
+    events = await _run(agent, "review the PRD and get started")
+
+    errors = [e for e in events if isinstance(e, ToolResultEvent) and e.error]
+    cb_errors = [e for e in errors if "CIRCUIT BREAKER" in str(e.error)]
+    force_stops = [e for e in events if isinstance(e, AssistantEvent) and e.stopped_by_middleware
+                   and "FORCED STOP" in (e.content or "")]
+    ordering = [e for e in errors if "Unexpected role" in str(e.error)]
+
+    assert not force_stops, f"Session force-stopped during normal project setup"
+    assert not ordering, f"Message ordering crash: {ordering[0].error[:100]}"
+    assert len(cb_errors) <= 2, f"Circuit breaker fired {len(cb_errors)} times during normal work"
+
+    # Should have created some files
+    created = list(tmp_path.rglob("*.py"))
+    assert len(created) > 0, "No Python files created"