drydock-cli 1.3.9__tar.gz → 1.5.0__tar.gz

{drydock_cli-1.3.9 → drydock_cli-1.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 1.3.9
+Version: 1.5.0
 Summary: Drydock — nautical CLI coding agent
 Project-URL: Homepage, https://github.com/fbobe321/drydock
 Project-URL: Repository, https://github.com/fbobe321/drydock

{drydock_cli-1.3.9 → drydock_cli-1.5.0}/PRD.md

@@ -1,7 +1,7 @@
 # DryDock — Local CLI Coding Agent
 
 **Repository:** https://github.com/fbobe321/drydock
-**PyPI:** https://pypi.org/project/drydock-cli/ (v1.1.5)
+**PyPI:** https://pypi.org/project/drydock-cli/ (v1.5.0)
 **License:** Apache 2.0 (fork of [mistralai/mistral-vibe](https://github.com/mistralai/mistral-vibe))
 **Status:** Active development — continuous improvement running
 
@@ -21,6 +21,7 @@ Every change follows this pipeline:
 Scripts:
 - `scripts/test_smoke.sh` — quick smoke tests (every deploy)
 - `scripts/test_full.sh` — smoke + full regression (nightly)
+- `scripts/test_bank.sh` — 83-test regression bank (10-18 hours, real vLLM)
 - `scripts/deploy_to_github.sh` — smoke tests → sync to GitHub (daily 4 AM)
 - `scripts/publish_to_pypi.sh` — smoke tests → bump → build → PyPI → GitHub
 - `scripts/backup.sh` — rsync to NAS (daily 3 AM)
@@ -248,3 +249,33 @@ Adding Python tools/agents doesn't matter if the prompt doesn't tell the model t
 - Shows exact syntax for delegation
 - Explicit WHEN to delegate rules (3+ files, reviews, bugs, planning)
 - WHEN NOT to delegate (simple fixes, quick questions)
+
+### Phase 11 (Mar 28): Agent Loop Hardening + 83-Test Regression Bank
+
+User reported DryDock was unusable — model looped 12+ times running the same command, search_replace blocked legitimate files, basic tasks couldn't complete.
+
+**Agent loop fixes (5 bugs):**
+- Empty model response: retry 3x with nudge (model sometimes returns nothing)
+- Invalid tool name recovery: model hallucinates `task_agent:`, now gets correct tool list
+- Bash traceback guidance: extracts file:line from crash, directs model to read source
+- Circuit breaker for successful repeats: blocks after 4 identical bash calls (was unlimited)
+- Tighter loop detection: warning at 4 (was 8), force-stop at 8 (was 25)
+
+**System prompt improvements:**
+- Added debugging best practices: RUN → READ → FIX → VERIFY cycle
+- Common bug pattern hints (TypeError, KeyError, IndexError, off-by-one)
+
+**83-test regression bank (real vLLM backend, ~90 min):**
+
+| Suite | Tests | Pass Rate |
+|---|---|---|
+| BUILD (easy/medium/hard) | 25 | 100% |
+| DEBUG (easy/medium/hard) | 20 | 75% |
+| UPDATE (easy/medium/hard) | 15 | 100% |
+| MULTIAGENT | 10 | 100% |
+| TOOLS | 13 | 100% |
+| **Total** | **83** | **94%** |
+
+Run with: `./scripts/test_bank.sh` (full) or `./scripts/test_bank.sh quick` (easy only, ~2h)
+
+**Removed:** False-positive test file guard in search_replace that blocked any file in a directory containing "test" in its name.

{drydock_cli-1.3.9 → drydock_cli-1.5.0}/drydock/core/agent_loop.py

@@ -124,8 +124,8 @@ class ToolDecision(BaseModel):
 
 MAX_TOOL_TURNS = 200  # Bug fixes rarely need more than 50 turns; 200 is generous ceiling
 MAX_API_ERRORS = 5
-REPEAT_WARNING_THRESHOLD = 8  # Same exact call 8+ times before warning
-REPEAT_FORCE_STOP_THRESHOLD = 25  # Same exact call 25+ times before force-stop
+REPEAT_WARNING_THRESHOLD = 4  # Same exact call 4+ times before warning
+REPEAT_FORCE_STOP_THRESHOLD = 8  # Same exact call 8+ times before force-stop
 
 logger = logging.getLogger(__name__)
 
@@ -192,6 +192,7 @@ class AgentLoop:
         # Key: hash(tool_name + args), Value: (count, last_result_snippet)
         self._tool_call_history: dict[str, tuple[int, str]] = {}
         self._consecutive_circuit_breaker_fires: int = 0
+        self._empty_responses: int = 0
 
         system_prompt = get_universal_system_prompt(
             self.tool_manager, self.config, self.skill_manager, self.agent_manager
@@ -534,9 +535,19 @@ class AgentLoop:
                             stopped_by_middleware=True,
                         )
                         return
-                    # Inject error notice — ALWAYS append to last tool result
-                    # vLLM/Mistral rejects user messages after tool messages
-                    error_text = f"API error occurred: {e}. Please continue with your task."
+                    # Check if the error is about invalid function/tool name
+                    error_str = str(e)
+                    if "Function name" in error_str or "function" in error_str.lower() and "must be" in error_str.lower():
+                        # Model hallucinated a tool name — give it the correct list
+                        available = ", ".join(sorted(self.tool_manager.available_tools.keys())[:15])
+                        error_text = (
+                            f"ERROR: You tried to call a tool that does not exist. "
+                            f"Available tools: {available}. "
+                            f"Use one of these exact tool names. "
+                            f"For subagent delegation, use 'task'. For file search, use 'grep'."
+                        )
+                    else:
+                        error_text = f"API error occurred: {e}. Please continue with your task."
                     self._inject_system_note(error_text)
                     continue
 
@@ -670,10 +681,26 @@ class AgentLoop:
 
                 should_break_loop = last_message.role != Role.tool
 
+                # Handle empty response (no content AND no tools) — retry
+                if (should_break_loop and last_message.role == Role.assistant
+                        and not last_message.content and not last_message.tool_calls):
+                    empty_response_count = getattr(self, '_empty_responses', 0) + 1
+                    self._empty_responses = empty_response_count
+                    if empty_response_count <= 3:
+                        should_break_loop = False
+                        self._inject_system_note(
+                            "Your response was empty. You MUST call a tool. "
+                            "Start by reading the relevant file with read_file, "
+                            "or run the failing command with bash to see the error."
+                        )
+                        logger.warning("Empty model response — nudging (attempt %d)", empty_response_count)
+                        continue
+                    # After 3 empty responses, let it exit
+
                 # If model gives text without tool calls and hasn't edited anything,
                 # nudge it to make an edit instead of just describing what to do.
                 # Never let the agent exit without at least attempting an edit.
-                if should_break_loop and not has_made_edit and tool_turns >= 2:
+                if should_break_loop and not has_made_edit and tool_turns >= 1:
                     text_without_action += 1
                     # Always continue — don't let the agent exit without editing
                     should_break_loop = False
@@ -742,8 +769,8 @@ class AgentLoop:
                     if rep == "FORCE_STOP":
                         repeat_warnings += 1
                         logger.warning("Detected infinite loop: same tool call %d+ times (warning %d)", REPEAT_FORCE_STOP_THRESHOLD, repeat_warnings)
-                        if repeat_warnings >= 12:
-                            # True last resort — only stop after many redirects
+                        if repeat_warnings >= 5:
+                            # Stop after 5 redirects — model is not recovering
                             yield AssistantEvent(
                                 content="\n\n[Stopping: exhausted all retry attempts.]\n",
                                 stopped_by_middleware=True,
@@ -773,7 +800,7 @@ class AgentLoop:
                         else:
                             repeat_warnings += 1
                         logger.warning("Detected repeated tool calls (warning %.1f, investigation=%s)", repeat_warnings, is_investigation_warning)
-                        if repeat_warnings >= 15:
+                        if repeat_warnings >= 6:
                             yield AssistantEvent(
                                 content="\n\n[Stopping: too many repeated actions despite warnings.]\n",
                                 stopped_by_middleware=True,
@@ -792,7 +819,13 @@ class AgentLoop:
                         elif stuck_tool in ("grep",):
                             nudge = " You have been SEARCHING repeatedly. Pick the most relevant file from your search results and READ it, then make your edit."
                         elif stuck_tool in ("bash", "run_command"):
-                            nudge = " You have been running bash commands repeatedly without making progress. STOP testing and make your code edit with search_replace now. If your fix doesn't work, try a DIFFERENT approach."
+                            nudge = (
+                                " You have been running the SAME bash command repeatedly. "
+                                "Running the same command again will give the same result. "
+                                "If it succeeded but didn't do what you expected, READ the source code "
+                                "to understand WHY, then FIX the code with search_replace. "
+                                "If it failed, try a DIFFERENT command or approach."
+                            )
                         if repeat_warnings >= 4:
                             # Escalated: replace tool result with strong directive
                             warning_text = (
@@ -935,10 +968,19 @@ class AgentLoop:
         count, last_result = self._tool_call_history.get(sig, (0, ""))
         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):
+        # Block failed commands after 2 repeats.
+        # Block SUCCESSFUL commands after 4 repeats — the model should not
+        # run the exact same command with the exact same args 5+ times.
+        # Read-only checks (ls, pwd, git status) get a higher threshold.
+        tool_name = tool_call.tool_name
+        is_readonly = tool_name in ("grep", "read_file", "glob", "ls")
+        success_threshold = 6 if is_readonly else 4
+
+        if is_failed and count >= 2:
+            pass  # will be blocked below
+        elif not is_failed and count >= success_threshold:
+            pass  # will be blocked below
+        else:
             return None
 
         if count >= 2:
@@ -1161,6 +1203,19 @@ class AgentLoop:
                     "Add more surrounding context lines to old_str to make it unique.]"
                 )
 
+            # RECOVERY: After bash failure with traceback, extract file/line and nudge
+            if tool_call.tool_name in ("bash", "run_command") and "Traceback" in str(exc):
+                import re
+                # Extract last File "path", line N from traceback
+                tb_matches = re.findall(r'File "([^"]+)", line (\d+)', str(exc))
+                if tb_matches:
+                    tb_file, tb_line = tb_matches[-1]
+                    if not tb_file.startswith("/home") and not "site-packages" in tb_file:
+                        error_msg += (
+                            f"\n\n[NEXT STEP: Read the source at {tb_file} around line {tb_line} "
+                            f"with read_file to understand the bug, then fix it with search_replace.]"
+                        )
+
             yield ToolResultEvent(
                 tool_name=tool_call.tool_name,
                 tool_class=tool_call.tool_class,
@@ -1485,11 +1540,11 @@ class AgentLoop:
         # Lower thresholds catch loops where the model uses the same tool
         # with slightly different args (e.g., grep with different paths).
         if last_tool in ("bash", "run_command"):
-            same_tool_limit = 8
+            same_tool_limit = 5
         elif last_tool in ("grep", "read_file"):
-            same_tool_limit = 10  # investigation tools need some room
+            same_tool_limit = 7  # investigation tools need some room
         else:
-            same_tool_limit = 6
+            same_tool_limit = 5
         if (
             len(tool_names) >= same_tool_limit
             and all(n == tool_names[-1] for n in tool_names[-same_tool_limit:])

{drydock_cli-1.3.9 → drydock_cli-1.5.0}/drydock/core/prompts/cli.md

@@ -223,6 +223,21 @@ Error Handling:
 - After ANY error, your next action must be DIFFERENT from what just failed.
 - Never run the same failing command twice — the result won't change.
 
+Debugging Best Practices:
+When fixing bugs, follow this cycle: RUN → READ → FIX → VERIFY.
+1. RUN the failing command to see the exact error (traceback, assertion, etc.)
+2. READ the source file at the line mentioned in the traceback
+3. FIX the bug with search_replace — common patterns:
+   - TypeError (str + int): add type conversion (int(), float(), str())
+   - KeyError: add .get() with default, or check key existence
+   - IndexError: fix loop bounds or add bounds check
+   - ImportError: fix the import path or module name
+   - Off-by-one: check if index starts at 0 or 1
+   - Missing return value: check if function returns in all paths
+4. VERIFY by running the command again — if it still fails, read the NEW error
+
+CRITICAL: After fixing a bug, ALWAYS run the command again to verify. Never assume your fix worked.
+
 Context Budget:
 Your context window is your most important resource. Every grep result and file read consumes it.
 Performance degrades as context fills. Budget your investigation:

{drydock_cli-1.3.9 → drydock_cli-1.5.0}/drydock/core/tools/builtins/search_replace.py

@@ -119,16 +119,6 @@ class SearchReplace(
     ) -> AsyncGenerator[ToolStreamEvent | SearchReplaceResult, None]:
         file_path, search_replace_blocks = self._prepare_and_validate_args(args)
 
-        # BLOCK editing test files — the bug is in source code, not tests
-        file_str = str(file_path)
-        if ("/tests/" in file_str or "/test_" in file_str or
-                file_str.endswith("_test.py") or "/testing/" in file_str):
-            raise ToolError(
-                f"BLOCKED: You are trying to edit a test file ({file_path.name}). "
-                f"The bug is in LIBRARY SOURCE code, not tests. "
-                f"Use grep to find the corresponding source file and edit that instead."
-            )
-
         # Injection guard: scan replacement content for suspicious patterns
         from drydock.core.tools.injection_guard import check_content_for_injection
         if warning := check_content_for_injection(args.content, args.file_path):

{drydock_cli-1.3.9 → drydock_cli-1.5.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "1.3.9"
+version = "1.5.0"
 description = "Drydock — nautical CLI coding agent"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-1.3.9 → drydock_cli-1.5.0}/scripts/deploy_to_github.sh

@@ -37,7 +37,8 @@ AUTH_URL="https://${GITHUB_TOKEN}@github.com/fbobe321/drydock.git"
 
 log "Running regression tests..."
 cd "$DRYDOCK_SRC"
-if ! python3 -m pip install -q "pytest>=9.0" 2>/dev/null; python3 -m pytest tests/test_drydock_regression.py tests/test_drydock_tasks.py tests/test_loop_detection.py tests/test_agent_tasks.py tests/test_integration.py tests/test_user_issues.py tests/test_real_issues.py \
+python3 -m pip install -q "pytest>=9.0" 2>/dev/null
+if ! python3 -m pytest tests/test_drydock_regression.py tests/test_drydock_tasks.py tests/test_loop_detection.py tests/test_agent_tasks.py tests/test_integration.py tests/test_user_issues.py tests/test_real_issues.py \
     -p no:xdist -p no:cov --override-ini="addopts=" -q 2>&1; then
     log "TESTS FAILED — deploy aborted. Fix the tests before deploying."
     exit 1

{drydock_cli-1.3.9 → drydock_cli-1.5.0}/scripts/publish_to_pypi.sh

@@ -25,7 +25,8 @@ PYPI_TOKEN=$(cat "$TOKEN_FILE")
 
 # --- Run tests ---
 log "Running regression tests..."
-if ! $PYTHON -m pip install -q "pytest>=9.0" 2>/dev/null; $PYTHON -m pytest tests/test_drydock_regression.py tests/test_drydock_tasks.py tests/test_loop_detection.py tests/test_agent_tasks.py tests/test_integration.py tests/test_user_issues.py tests/test_real_issues.py \
+$PYTHON -m pip install -q "pytest>=9.0" 2>/dev/null
+if ! $PYTHON -m pytest tests/test_drydock_regression.py tests/test_drydock_tasks.py tests/test_loop_detection.py tests/test_agent_tasks.py tests/test_integration.py tests/test_user_issues.py tests/test_real_issues.py \
     -p no:xdist -p no:cov --override-ini="addopts=" -q 2>&1; then
     log "TESTS FAILED — publish aborted."
     exit 1

drydock_cli-1.5.0/scripts/test_bank.sh

@@ -0,0 +1,78 @@
+#!/bin/bash
+# TEST BANK — Full regression suite against real vLLM backend.
+# Requires vLLM at localhost:8000. Takes 10-18 hours.
+#
+# Usage:
+#   ./scripts/test_bank.sh           # Run everything
+#   ./scripts/test_bank.sh build     # Run only build tests
+#   ./scripts/test_bank.sh debug     # Run only debug tests
+#   ./scripts/test_bank.sh update    # Run only update tests
+#   ./scripts/test_bank.sh multi     # Run only multi-agent tests
+#   ./scripts/test_bank.sh tools     # Run only tool integration tests
+#   ./scripts/test_bank.sh quick     # Run only easy tests (~2 hours)
+
+set -euo pipefail
+cd "$(dirname "$0")/.."
+
+pip install -q "pytest>=9.0" pytest-asyncio httpx 2>/dev/null
+
+COMMON_ARGS="-v -s -p no:xdist -p no:cov --override-ini=addopts="
+CATEGORY="${1:-all}"
+TIMESTAMP=$(date +%Y%m%d_%H%M%S)
+LOG_DIR="test_bank_results"
+mkdir -p "$LOG_DIR"
+
+run_suite() {
+    local name="$1"
+    local file="$2"
+    local extra="${3:-}"
+    echo ""
+    echo "================================================================"
+    echo "  $name — $(date)"
+    echo "================================================================"
+    local logfile="$LOG_DIR/${name}_${TIMESTAMP}.log"
+    python3 -m pytest "$file" $COMMON_ARGS $extra 2>&1 | tee "$logfile"
+    echo ""
+    echo "Results saved to: $logfile"
+}
+
+case "$CATEGORY" in
+    build)
+        run_suite "BUILD" "tests/test_bank_build.py"
+        ;;
+    debug)
+        run_suite "DEBUG" "tests/test_bank_debug.py"
+        ;;
+    update)
+        run_suite "UPDATE" "tests/test_bank_update.py"
+        ;;
+    multi)
+        run_suite "MULTI-AGENT" "tests/test_bank_multiagent.py"
+        ;;
+    tools)
+        run_suite "TOOLS" "tests/test_bank_tools.py"
+        ;;
+    quick)
+        echo "=== QUICK TEST BANK (Easy tests only, ~2 hours) ==="
+        run_suite "BUILD-EASY" "tests/test_bank_build.py" "-k TestBuildEasy"
+        run_suite "DEBUG-EASY" "tests/test_bank_debug.py" "-k TestDebugEasy"
+        run_suite "UPDATE-EASY" "tests/test_bank_update.py" "-k TestUpdateEasy"
+        run_suite "TOOLS" "tests/test_bank_tools.py"
+        ;;
+    all)
+        echo "=== FULL TEST BANK — Estimated 10-18 hours ==="
+        echo "=== Started: $(date) ==="
+        run_suite "BUILD" "tests/test_bank_build.py"
+        run_suite "DEBUG" "tests/test_bank_debug.py"
+        run_suite "UPDATE" "tests/test_bank_update.py"
+        run_suite "MULTI-AGENT" "tests/test_bank_multiagent.py"
+        run_suite "TOOLS" "tests/test_bank_tools.py"
+        echo ""
+        echo "=== FULL TEST BANK COMPLETE — $(date) ==="
+        echo "Results in: $LOG_DIR/"
+        ;;
+    *)
+        echo "Usage: $0 {all|build|debug|update|multi|tools|quick}"
+        exit 1
+        ;;
+esac

{drydock_cli-1.3.9 → drydock_cli-1.5.0}/scripts/test_full.sh

@@ -11,3 +11,7 @@ python3 -m pytest tests/test_smoke.py -p no:xdist -p no:cov --override-ini="addo
 echo ""
 echo "=== FULL REGRESSION (real backend) ==="
 python3 -m pytest tests/test_full_regression.py -v -p no:xdist -p no:cov --override-ini="addopts="
+
+echo ""
+echo "=== BUILD PROJECT TESTS (real backend, 30-60 min) ==="
+python3 -m pytest tests/test_build_projects.py -v -s -p no:xdist -p no:cov --override-ini="addopts="