drydock-cli 2.2.7__tar.gz → 2.2.9__tar.gz

{drydock_cli-2.2.7 → drydock_cli-2.2.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.2.7
+Version: 2.2.9
 Summary: Drydock — nautical CLI coding agent
 Project-URL: Homepage, https://github.com/fbobe321/drydock
 Project-URL: Repository, https://github.com/fbobe321/drydock

{drydock_cli-2.2.7 → drydock_cli-2.2.9}/drydock/core/agent_loop.py

@@ -497,6 +497,8 @@ class AgentLoop:
             files_explored: list[str] = []  # Track files read for context state
             files_modified: list[str] = []  # Track files edited for blast radius
             context_summary_injected = False
+            self._temp_override: float | None = None  # Temperature bump on loop detection
+            self._failed_approaches: list[str] = []  # Persistent failed approach tracker
             while not should_break_loop:
                 # Loop protection: prevent infinite tool-call loops
                 tool_turns += 1
@@ -861,57 +863,31 @@ class AgentLoop:
                     if nudge_text:
                         self._inject_system_note(nudge_text)
 
-                # Loop guidance — prune duplicates and nudge, NEVER stop
+                # Loop guidance — retrospection, NEVER stop
                 if not should_break_loop:
                     rep = self._check_tool_call_repetition()
                     if rep == "FORCE_STOP" or (rep and rep.startswith("WARNING")):
-                        stuck_tool = ""
-                        if rep and "|" in rep:
-                            stuck_tool = rep.split("|", 1)[1]
-
                         # Prune duplicate tool calls to give model fresh context
                         self._prune_repeated_tool_calls()
 
-                        # Gentle nudge — just suggest a different approach
-                        nudge = "Try a different approach."
-                        if stuck_tool in ("read_file",):
-                            nudge = "You already read this. Try search_replace or a different file."
-                        elif stuck_tool in ("grep",):
-                            nudge = "Try a different search term or read a file you found."
-                        elif stuck_tool in ("bash", "run_command"):
-                            nudge = "If the command failed, read the error and fix the code."
-                        self._inject_system_note(nudge)
-
-                        # Extract tool-specific nudge
-                        nudge = ""
-                        if stuck_tool in ("read_file",):
-                            nudge = " You have been READING the same code repeatedly. You already know what it says. Make your EDIT now using search_replace or write_file."
-                        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 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 = (
-                                "BLOCKED: Your tool call was not executed because you have been repeating the same action. "
-                                "You already have all the information you need from previous tool results. "
-                                "Your ONLY valid next action is: use search_replace to make your code fix. "
-                                "Do NOT read, grep, or bash. Use search_replace NOW."
-                            )
-                            self._inject_system_note(warning_text, replace_last_tool=True)
-                        else:
-                            warning_text = (
-                                "WARNING: You are repeating the same tool call with the same arguments. "
-                                "This is not making progress. Try a DIFFERENT approach, use a different tool, "
-                                "or tell the user what's blocking you." + nudge
-                            )
-                            self._inject_system_note(warning_text)
+                        # Temperature bump: force model to explore different paths
+                        self._temp_override = 0.5
+
+                        # Track failed approach for accumulator
+                        self._record_failed_approach()
+
+                        # RETROSPECTION: Summarize recent actions and let the model
+                        # decide its own next step — no hardcoded nudges
+                        retro = self._build_retrospection()
+                        if retro:
+                            # Include failed approaches if any
+                            if self._failed_approaches:
+                                approaches = "\n".join(f"  - {a}" for a in self._failed_approaches[-5:])
+                                retro += f"\n\nApproaches already tried (don't repeat these):\n{approaches}"
+                            self._inject_system_note(retro)
+                    else:
+                        # Loop broken — reset temperature to normal
+                        self._temp_override = None
 
                 if user_cancelled:
                     return
@@ -1291,11 +1267,31 @@ class AgentLoop:
                         )
                 except (json.JSONDecodeError, AttributeError):
                     pass
+                # AUTO-READ: Automatically read the target file so the model has
+                # the actual content — don't just tell it to read, DO it for it
+                auto_read_content = ""
+                try:
+                    sr_args = json.loads(tool_call.raw_arguments or "{}")
+                    sr_path = sr_args.get("file_path", sr_args.get("path", ""))
+                    if sr_path and Path(sr_path).exists():
+                        with open(sr_path, "r", encoding="utf-8", errors="replace") as f:
+                            lines = f.readlines()
+                        # Show first 50 lines or the whole file if small
+                        preview_lines = lines[:50]
+                        numbered = [f"{i+1}\t{line.rstrip()}" for i, line in enumerate(preview_lines)]
+                        auto_read_content = (
+                            f"\n\nAUTO-READ of {sr_path} (first {len(preview_lines)} lines):\n"
+                            + "\n".join(numbered)
+                        )
+                        if len(lines) > 50:
+                            auto_read_content += f"\n[... {len(lines) - 50} more lines]"
+                except Exception:
+                    pass
                 error_msg += (
                     "\n\n[RECOVERY: Your search text didn't match the file contents. "
-                    "Use read_file with offset/limit to re-read the exact area you want to change, "
-                    "then retry search_replace with the EXACT text from read_file output."
+                    "The actual file content is shown below — use EXACT text from it for your next edit."
                     f"{sr_file_hint}]"
+                    f"{auto_read_content}"
                 )
             elif tool_call.tool_name == "search_replace" and "multiple" in str(exc).lower():
                 error_msg += (
@@ -1429,10 +1425,12 @@ class AgentLoop:
 
         try:
             start_time = time.perf_counter()
+            # Use temperature override if set (loop detection bumps it)
+            temp = getattr(self, '_temp_override', None) or active_model.temperature
             result = await self.backend.complete(
                 model=active_model,
                 messages=self.messages,
-                temperature=active_model.temperature,
+                temperature=temp,
                 tools=available_tools,
                 tool_choice=tool_choice,
                 extra_headers=self._get_extra_headers(provider),
@@ -1477,10 +1475,12 @@ class AgentLoop:
             start_time = time.perf_counter()
             usage = LLMUsage()
             chunk_agg = LLMChunk(message=LLMMessage(role=Role.assistant))
+            # Use temperature override if set (loop detection bumps it)
+            temp = getattr(self, '_temp_override', None) or active_model.temperature
             async for chunk in self.backend.complete_streaming(
                 model=active_model,
                 messages=self.messages,
-                temperature=active_model.temperature,
+                temperature=temp,
                 tools=available_tools,
                 tool_choice=tool_choice,
                 extra_headers=self._get_extra_headers(provider),
@@ -1584,6 +1584,96 @@ class AgentLoop:
                     feedback=feedback,
                 )
 
+    def _record_failed_approach(self) -> None:
+        """Record a one-line summary of the most recent failed approach.
+
+        Survives pruning and compaction so the model doesn't retry
+        the same strategy after context cleanup.
+        """
+        if not hasattr(self, '_failed_approaches'):
+            self._failed_approaches = []
+
+        # Extract the last tool call and its result
+        last_tool_name = ""
+        last_tool_args = ""
+        last_result = ""
+        for msg in reversed(self.messages[-6:]):
+            if msg.role == Role.tool and not last_result:
+                result = str(msg.content or "")[:100]
+                if "error" in result.lower() or "not found" in result.lower():
+                    last_result = result[:80]
+            elif msg.role == Role.assistant and msg.tool_calls and not last_tool_name:
+                for tc in msg.tool_calls:
+                    if tc.function:
+                        last_tool_name = tc.function.name or ""
+                        args = tc.function.arguments or ""
+                        # Extract file path if present
+                        try:
+                            parsed = json.loads(args)
+                            last_tool_args = parsed.get("file_path", parsed.get("path", parsed.get("command", "")))[:60]
+                        except (json.JSONDecodeError, TypeError, AttributeError):
+                            last_tool_args = args[:40]
+                        break
+
+        if last_tool_name:
+            summary = f"{last_tool_name}({last_tool_args})"
+            if last_result:
+                summary += f" → {last_result}"
+            # Don't add duplicates
+            if not self._failed_approaches or self._failed_approaches[-1] != summary:
+                self._failed_approaches.append(summary)
+                # Keep only last 10
+                self._failed_approaches = self._failed_approaches[-10:]
+
+    def _build_retrospection(self) -> str | None:
+        """Build a retrospection summary of recent tool calls.
+
+        Instead of hardcoded nudges, show the model what it's been doing
+        and let it decide on a different approach. The model is better at
+        self-correcting when it can see its own pattern of behavior.
+        """
+        # Collect last 8 tool calls with their results (truncated)
+        recent: list[str] = []
+        count = 0
+        for msg in reversed(self.messages):
+            if count >= 8:
+                break
+            if msg.role == Role.assistant and msg.tool_calls:
+                for tc in reversed(msg.tool_calls or []):
+                    if tc.function:
+                        name = tc.function.name or "?"
+                        args_str = tc.function.arguments or ""
+                        # Truncate args for readability
+                        if len(args_str) > 120:
+                            args_str = args_str[:120] + "..."
+                        recent.append(f"  {count+1}. {name}({args_str})")
+                        count += 1
+                        if count >= 8:
+                            break
+            elif msg.role == Role.tool and recent:
+                # Add result summary to last entry
+                result = str(msg.content or "")[:150]
+                if "error" in result.lower() or "Error" in result:
+                    recent[-1] += f" → ERROR"
+                elif "not found" in result.lower():
+                    recent[-1] += f" → NOT FOUND"
+                else:
+                    recent[-1] += f" → ok"
+
+        if len(recent) < 3:
+            return None
+
+        recent.reverse()
+        summary = "\n".join(recent)
+
+        return (
+            f"RETROSPECTION — Your last {len(recent)} tool calls:\n"
+            f"{summary}\n\n"
+            f"You are repeating a pattern that is not making progress. "
+            f"Review the sequence above. What went wrong? What should you do differently? "
+            f"Choose a different approach on your own."
+        )
+
     def _inject_system_note(self, text: str, replace_last_tool: bool = False) -> None:
         """Safely inject a system note into conversation without breaking message ordering.
 

{drydock_cli-2.2.7 → drydock_cli-2.2.9}/drydock/core/tools/builtins/search_replace.py

@@ -23,11 +23,12 @@ from drydock.core.tools.utils import resolve_file_tool_permission
 from drydock.core.types import ToolResultEvent, ToolStreamEvent
 
 SEARCH_REPLACE_BLOCK_RE = re.compile(
-    r"<{5,} SEARCH\r?\n(.*?)\r?\n?={5,}\r?\n(.*?)\r?\n?>{5,} REPLACE", flags=re.DOTALL
+    r"<{5,}\s*(?:SEARCH\s*)+\r?\n(.*?)\r?\n?={5,}\r?\n(.*?)\r?\n?>{5,}\s*(?:REPLACE\s*)+",
+    flags=re.DOTALL,
 )
 
 SEARCH_REPLACE_BLOCK_WITH_FENCE_RE = re.compile(
-    r"```[\s\S]*?\n<{5,} SEARCH\r?\n(.*?)\r?\n?={5,}\r?\n(.*?)\r?\n?>{5,} REPLACE\s*\n```",
+    r"```[\s\S]*?\n<{5,}\s*(?:SEARCH\s*)+\r?\n(.*?)\r?\n?={5,}\r?\n(.*?)\r?\n?>{5,}\s*(?:REPLACE\s*)+\s*\n```",
     flags=re.DOTALL,
 )
 
@@ -213,6 +214,16 @@ class SearchReplace(
                 ">>>>>>> REPLACE"
             )
 
+        # Detect garbled token output (common with some models/parsers)
+        for block in search_replace_blocks:
+            if re.search(r'(.)\1{4,}', block.search) or re.search(r'<{3,}[a-z]', block.search):
+                raise ToolError(
+                    "Your search text appears garbled (repeated/corrupted characters). "
+                    "This happens when regex or special characters confuse the tool call parser. "
+                    "WORKAROUND: Use read_file to get the exact line numbers, then use write_file "
+                    "to rewrite the entire function instead of search_replace on complex regex patterns."
+                )
+
         return file_path, search_replace_blocks
 
     async def _read_file(self, file_path: Path) -> str:

{drydock_cli-2.2.7 → drydock_cli-2.2.9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.2.7"
+version = "2.2.9"
 description = "Drydock — nautical CLI coding agent"
 readme = "README.md"
 requires-python = ">=3.12"