drydock-cli 2.6.27__tar.gz → 2.6.29__tar.gz

{drydock_cli-2.6.27 → drydock_cli-2.6.29}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.6.27
+Version: 2.6.29
 Summary: Drydock — nautical CLI coding agent
 Project-URL: Homepage, https://github.com/fbobe321/drydock
 Project-URL: Repository, https://github.com/fbobe321/drydock

{drydock_cli-2.6.27 → drydock_cli-2.6.29}/drydock/core/agent_loop.py

@@ -653,6 +653,36 @@ class AgentLoop:
             return
         last_message = self.messages[-1]
 
+        # Detect thinking-only stall: model produced no visible content
+        # AND no tool calls.  This happens when Gemma 4 leaks thinking
+        # tokens that get stripped, leaving an empty assistant message.
+        # If the previous message was a tool result (model was mid-work),
+        # inject a nudge to keep it going instead of silently ending.
+        if (
+            not last_message.content
+            and not last_message.tool_calls
+            and len(self.messages) >= 3
+        ):
+            prev_role = self.messages[-2].role if len(self.messages) >= 2 else None
+            if prev_role == Role.tool:
+                # Track consecutive empty turns to avoid infinite nudge loops
+                empty_turns = getattr(self, "_consecutive_empty_turns", 0) + 1
+                self._consecutive_empty_turns = empty_turns
+                if empty_turns <= 2:
+                    logger.info("Thinking-only stall detected (attempt %d) — nudging model to continue", empty_turns)
+                    # Remove the empty assistant message and inject a nudge
+                    self.messages.pop()
+                    self._inject_system_note(
+                        "Continue working. Use a tool (read_file, write_file, "
+                        "search_replace, bash) or say what you want to do next."
+                    )
+                    return  # will re-enter _perform_llm_turn on the next loop iteration
+                else:
+                    logger.info("Thinking-only stall persists after %d nudges — letting session end", empty_turns)
+        else:
+            # Reset counter on any productive turn
+            self._consecutive_empty_turns = 0
+
         # Detect repetitive text generation (Gemma 4 sometimes loops text within one response)
         if last_message.content and len(last_message.content) > 200:
             text = last_message.content
@@ -1219,7 +1249,15 @@ class AgentLoop:
                     f"the CODE is. Use read_file to inspect the source, then write_file or "
                     f"search_replace to fix the bug. Do not re-run the same command again."
                 )
-            if stuck_tool in ("write_file", "search_replace"):
+            if stuck_tool == "search_replace":
+                return (
+                    f"STOP: You are calling search_replace with the same arguments "
+                    f"repeatedly. If the tool said 'ALREADY APPLIED', the edit is DONE. "
+                    f"If it said 'not found', you need to read_file FIRST to see the "
+                    f"actual file contents, then use the EXACT text from the file. "
+                    f"Do NOT retry the same search_replace again."
+                )
+            if stuck_tool == "write_file":
                 return (
                     f"You have written the same file multiple times in a row. "
                     f"Move to the NEXT file in your plan. If no next file, run the code "
@@ -1405,6 +1443,12 @@ class AgentLoop:
             start_time = time.perf_counter()
             # Use temperature override if set (loop detection bumps it)
             temp = active_model.temperature
+            # Cap max_tokens for Gemma 4 to prevent infinite thinking loops.
+            # Without a cap, the model can generate 100K+ thinking tokens
+            # (filling the 131K context), hanging for 20+ minutes per turn.
+            effective_max_tokens = max_tokens
+            if effective_max_tokens is None and "gemma" in active_model.name.lower():
+                effective_max_tokens = 8192
             result = await self.backend.complete(
                 model=active_model,
                 messages=self.messages,
@@ -1412,7 +1456,7 @@ class AgentLoop:
                 tools=available_tools,
                 tool_choice=tool_choice,
                 extra_headers=self._get_extra_headers(provider),
-                max_tokens=max_tokens,
+                max_tokens=effective_max_tokens,
                 metadata=self.entrypoint_metadata.model_dump()
                 if self.entrypoint_metadata
                 else None,
@@ -1742,6 +1786,28 @@ class AgentLoop:
 
     def _check_tool_call_repetition(self) -> str | None:
         """Check if recent tool calls are repeating. Returns 'FORCE_STOP', 'WARNING', or None."""
+
+        # Early check: search_replace with the same file + old_string twice
+        # in a row.  This is the #1 user-pain loop — the model retries an
+        # edit that already succeeded or that keeps failing with the same
+        # "not found" error.  Nudge after just 2 identical attempts.
+        recent_sr: list[str] = []
+        for msg in reversed(self.messages[-20:]):
+            if msg.role == Role.assistant and msg.tool_calls:
+                for tc in msg.tool_calls:
+                    if tc.function and tc.function.name == "search_replace":
+                        try:
+                            args = json.loads(tc.function.arguments or "{}")
+                            # Build a key from file_path + old_string (content block)
+                            key = f"{args.get('file_path', '')}:{args.get('content', '')}"
+                            recent_sr.append(key)
+                        except (json.JSONDecodeError, AttributeError):
+                            pass
+                if len(recent_sr) >= 4:
+                    break
+        if len(recent_sr) >= 2 and recent_sr[0] == recent_sr[1]:
+            return "WARNING|search_replace"
+
         sigs: list[str] = []
         tool_names: list[str] = []
         lookback = 0

{drydock_cli-2.6.27 → drydock_cli-2.6.29}/drydock/core/llm/format.py

@@ -76,9 +76,26 @@ class APIToolFormatHandler:
         return "auto"
 
     def process_api_response_message(self, message: Any) -> LLMMessage:
+        # Strip ALL Gemma 4 leaked channel tokens from content.
+        # The model emits various formats:
+        #   <|channel>thought...<channel|>  (thinking tokens)
+        #   <|channel>call:tool_name{...}<tool_call|>  (malformed tool calls)
+        #   <|channel><|channel>list_mcp_resources{...}<tool_call|>  (double channel)
+        # Keeping them wastes context and confuses subsequent turns.
+        content = message.content
+        if content and "<|channel>" in content:
+            import re as _re
+            # Strip any <|channel>...<channel|> or <|channel>...<tool_call|> blocks
+            content = _re.sub(
+                r"<\|channel\>.*?(?:<channel\|>|<tool_call\|>)",
+                "",
+                content,
+                flags=_re.DOTALL,
+            ).strip() or None
+
         clean_message = {
             "role": message.role,
-            "content": message.content,
+            "content": content,
             "reasoning_content": getattr(message, "reasoning_content", None),
             "reasoning_signature": getattr(message, "reasoning_signature", None),
         }

{drydock_cli-2.6.27 → drydock_cli-2.6.29}/drydock/core/tools/builtins/search_replace.py

@@ -224,6 +224,38 @@ class SearchReplace(
             if path_match:
                 file_path_str = path_match.group(1)
 
+        # Last resort: if file_path is still missing but we have SEARCH text,
+        # scan project files for a match.  Gemma 4 frequently drops file_path.
+        # Limit scan to avoid hanging on large repos.
+        if not file_path_str and content:
+            blocks = self._parse_search_replace_blocks(content)
+            if blocks:
+                search_text = blocks[0].search.strip()
+                if search_text and len(search_text) >= 10:
+                    project_root = Path.cwd()
+                    candidates: list[Path] = []
+                    files_checked = 0
+                    for ext in ("*.py",):  # Only scan Python files
+                        for f in project_root.rglob(ext):
+                            if files_checked > 200:
+                                break
+                            if "__pycache__" in str(f) or ".git" in str(f):
+                                continue
+                            files_checked += 1
+                            try:
+                                text = f.read_text(encoding="utf-8", errors="replace")
+                                if len(text) > 100_000:
+                                    continue
+                                if search_text in text:
+                                    candidates.append(f)
+                            except Exception:
+                                continue
+                    if len(candidates) == 1:
+                        file_path_str = str(candidates[0])
+                    elif candidates:
+                        candidates.sort(key=lambda p: len(str(p)))
+                        file_path_str = str(candidates[0])
+
         if not file_path_str:
             raise ToolError(
                 "File path is required. Use: search_replace(file_path='path/to/file.py', content='...')"
@@ -333,6 +365,35 @@ class SearchReplace(
 
         for i, (search, replace) in enumerate(blocks, 1):
             if search not in current_content:
+                # Check if this change was ALREADY APPLIED — the replacement
+                # text is in the file but the search text is not.  This is the
+                # #1 cause of edit loops: the model successfully edits a file
+                # then retries the same edit because it doesn't realise it
+                # already worked.
+                if (replace and replace.strip()
+                        and len(replace.strip()) >= 10
+                        and replace.strip() in current_content):
+                    warnings.append(
+                        f"Block {i}: ALREADY APPLIED — the replacement text already "
+                        f"exists in {filepath.name}. The search text is gone because "
+                        f"a previous edit already made this change. "
+                        f"Move on to the next task."
+                    )
+                    # Count as "applied" so we don't error — the file is correct
+                    applied += 1
+                    continue
+                # Also detect deletion that already happened (search text removed,
+                # replace is empty — the line is simply gone).  Only trigger
+                # for meaningful search strings to avoid false positives.
+                if (not replace or not replace.strip()) and len(search.strip()) >= 10:
+                    warnings.append(
+                        f"Block {i}: ALREADY APPLIED — the text you want to remove "
+                        f"is already absent from {filepath.name}. "
+                        f"Move on to the next task."
+                    )
+                    applied += 1
+                    continue
+
                 # Try auto-applying high-confidence fuzzy match (>= 0.95 similarity)
                 auto_apply_threshold = max(fuzzy_threshold, 0.95)
                 best_match = SearchReplace._find_best_fuzzy_match(

{drydock_cli-2.6.27 → drydock_cli-2.6.29}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.6.27"
+version = "2.6.29"
 description = "Drydock — nautical CLI coding agent"
 readme = "README.md"
 requires-python = ">=3.12"

{drydock_cli-2.6.27 → drydock_cli-2.6.29}/scripts/shakedown.py

@@ -400,6 +400,27 @@ class DrydockDriver:
         time.sleep(0.2)
         self.child.send("\r")  # submit
 
+    def dismiss_trust_dialog(self) -> bool:
+        """Detect and dismiss the 'Trust this folder?' dialog.
+
+        This dialog can appear AFTER spawn() — e.g., when drydock lazily
+        resolves the project folder.  If we don't catch it, the prompt
+        gets eaten by the dialog and the session never starts.
+        """
+        if self.child is None:
+            return False
+        try:
+            recent = self.child.before or ""
+        except Exception:
+            return False
+        if "Trust this folder" in recent or "trust this folder" in recent.lower():
+            self.child.send("\x1b[D")  # Left arrow (No → Yes)
+            time.sleep(0.2)
+            self.child.send("\r")      # Enter
+            time.sleep(2)
+            return True
+        return False
+
     def dismiss_permission_prompts(self) -> bool:
         """Detect and auto-approve any tool permission dialog.
 
@@ -528,6 +549,13 @@ def run_test(cwd: Path, prompt: str, pkg: str) -> int:
                 print(f"\n[*] TUI exited at {int(elapsed)}s")
                 break
 
+            # Trust dialog check — can appear AFTER spawn, eating the prompt
+            if driver.dismiss_trust_dialog():
+                print(f"  [{int(elapsed):4d}s] Trust dialog dismissed — re-sending prompt")
+                time.sleep(1)
+                driver.type_message(prompt)
+                time.sleep(2)
+
             # Permission dialog check — fast, just peeks at buffer
             if driver.dismiss_permission_prompts():
                 print(f"  [{int(elapsed):4d}s] permission dialog auto-approved")
@@ -588,8 +616,30 @@ def run_test(cwd: Path, prompt: str, pkg: str) -> int:
                     print(f"\n[*] Model finished (text response, no tool call)")
                     break
 
-            # True dead silence: no new MESSAGES of any kind
+            # True dead silence: no new MESSAGES of any kind.
+            # But first: if the package already works, the model just
+            # stalled after finishing — that's a PASS, not a FAIL.
+            # Gemma 4 often goes silent for 5+ min on a final thinking
+            # turn after successfully writing all files.
             if detector.silent_seconds() > DEAD_SILENCE_SECONDS:
+                # Early-exit check: does the package already work?
+                if pkg_dir.exists():
+                    try:
+                        check = subprocess.run(
+                            ["python3", "-m", pkg, "--help"],
+                            cwd=str(cwd),
+                            capture_output=True,
+                            text=True,
+                            timeout=10,
+                        )
+                        if check.returncode == 0 and len(check.stdout.strip()) > 0:
+                            print(
+                                f"\n[*] Model silent {int(detector.silent_seconds())}s "
+                                f"but {pkg} --help works — treating as done"
+                            )
+                            break  # exit cleanly — no fail_reason added
+                    except Exception:
+                        pass
                 fail_reasons.append(
                     f"dead silence for {int(detector.silent_seconds())}s "
                     f"(no new messages of any kind)"

{drydock_cli-2.6.27 → drydock_cli-2.6.29}/scripts/shakedown_regression.sh

@@ -79,8 +79,13 @@ for proj in "${projs[@]}"; do
     dir="$PROJECTS_DIR/$proj"
     done_n=$((done_n + 1))
 
-    # Derive package name: strip leading "<num>_" from dir
-    pkg=$(echo "$proj" | sed 's/^[0-9]*_//')
+    # Derive package name from the PRD's "## Package Name" section if it
+    # exists, otherwise fall back to stripping the leading number from the
+    # directory name. Many PRDs use a different package name than the dir
+    # (e.g., dir=08_todo_list but PRD says `todo_manager`).
+    pkg=$(grep -A1 "## Package Name" "$dir/PRD.md" 2>/dev/null \
+        | tail -1 | grep -oP '`\K[a-z0-9_]+(?=`)' | head -1)
+    [ -z "$pkg" ] && pkg=$(echo "$proj" | sed 's/^[0-9]*_//')
 
     # Snapshot PRD if no master exists yet
     if [ -f "$dir/PRD.md" ] && [ ! -f "$dir/PRD.master.md" ]; then

{drydock_cli-2.6.27 → drydock_cli-2.6.29}/uv.lock

@@ -326,7 +326,7 @@ wheels = [
 
 [[package]]
 name = "drydock-cli"
-version = "2.4.7"
+version = "2.6.28"
 source = { editable = "." }
 dependencies = [
     { name = "agent-client-protocol" },

drydock_cli-2.6.27/.auto_release.lock

@@ -1 +0,0 @@
-306344