drydock-cli 2.2.2__tar.gz → 2.2.3__tar.gz

{drydock_cli-2.2.2 → drydock_cli-2.2.3}/CLAUDE.md

@@ -124,7 +124,35 @@ python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
 - Harness at `/data3/swe_bench_runs/harness.py` uses DryDock directly
 - AGENTS.md auto-created per task with bug-fix workflow
 - Environment bootstrapping (Meta-Harness technique) gathers repo structure before first LLM call
-- Test file edits stripped from patches before evaluation
+
+## Improvement Backlog (from research papers + industry analysis)
+
+**Meta-Harness (arXiv:2603.28052) — still TODO:**
+- Execution trace logging (full prompts + tool calls per task for failure diagnosis)
+- Repo-specific routing (different strategies per SWE-bench repo)
+- Draft-then-verify (generate patch, verify against tests before committing)
+- Contrastive examples (similar bugs with different fixes)
+- Pareto frontier tracking (accuracy vs token cost)
+
+**Industry design patterns — still TODO:**
+- Parallel tool execution (multiple independent tool calls per turn)
+- Background memory consolidation (clean context during idle time)
+- Three-tier memory system (index + topics + archives)
+- Frustration detection (regex-based, adjust approach when user is stuck)
+- Extended reasoning allocation (detect complex tasks, give more compute)
+- Git safety guards (stash before destructive ops)
+
+**Implemented:**
+- ✅ Static/dynamic prompt split (prefix caching)
+- ✅ AGENTS.md auto-creation
+- ✅ Environment bootstrapping
+- ✅ Additive-only harness changes
+- ✅ Permission denial feedback (suggest alternatives)
+- ✅ "Task Completed" false claim detection
+- ✅ Loop guidance (never stops, only redirects)
+- ✅ Composable prompt (static first, dynamic after)
+
+**Legal note:** All patterns are standard design concepts implemented from scratch. No proprietary code copied.
 
 ## Key Learnings
 

{drydock_cli-2.2.2 → drydock_cli-2.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.2.2
+Version: 2.2.3
 Summary: Drydock — nautical CLI coding agent
 Project-URL: Homepage, https://github.com/fbobe321/drydock
 Project-URL: Repository, https://github.com/fbobe321/drydock

{drydock_cli-2.2.2 → drydock_cli-2.2.3}/drydock/core/agent_loop.py

@@ -781,21 +781,28 @@ class AgentLoop:
                             "to make your edit. Do NOT say the task is done until search_replace succeeds."
                         )
                     elif text_without_action == 1:
-                        # Check if model already identified a TARGET in its text
-                        last_text = (last_message.content or "").upper()
-                        has_target = "TARGET:" in last_text or "FILE:" in last_text
-                        if has_target:
+                        last_text = (last_message.content or "").lower()
+                        # Detect model asking for confirmation instead of acting
+                        is_asking = any(kw in last_text for kw in (
+                            "standing by", "waiting for", "please provide",
+                            "what would you like", "how can i help",
+                            "what should i", "do you want me to",
+                        ))
+                        if is_asking:
                             nudge_text = (
-                                "Good — you identified the target. Now proceed to PHASE 2: "
-                                "use read_file to read the target function, then search_replace to fix it. "
-                                "Do NOT describe the fix — apply it with search_replace."
+                                "DO NOT ask for confirmation. Act NOW. "
+                                "If there is a PRD.md, implement it with write_file. "
+                                "If there is code, read it with read_file and fix bugs with search_replace. "
+                                "Call a tool on your next response."
+                            )
+                        elif "TARGET:" in last_text.upper() or "FILE:" in last_text.upper():
+                            nudge_text = (
+                                "Good — you identified the target. Now use read_file then search_replace."
                             )
                         else:
                             nudge_text = (
                                 "You responded with text but did not call any tools. "
-                                "Use search_replace NOW to apply your code change. "
-                                "If you are unsure of the exact text, use read_file on the target function ONCE, "
-                                "then immediately search_replace."
+                                "Call write_file, read_file, or search_replace NOW."
                             )
                     elif text_without_action == 2:
                         nudge_text = (
@@ -1146,6 +1153,16 @@ class AgentLoop:
                     CancellationReason.TOOL_SKIPPED, tool_call.tool_name
                 )
             )
+            # Add alternative suggestions so model can adjust strategy
+            alternatives = {
+                "write_file": "Try search_replace to modify existing files.",
+                "search_replace": "Try write_file to create the file, or read_file first to get exact text.",
+                "bash": "Try read_file + search_replace for code changes.",
+                "task": "Try grep + read_file to explore manually.",
+            }
+            alt = alternatives.get(tool_call.tool_name, "")
+            if alt:
+                skip_reason += f"\n\n{alt}"
             yield ToolResultEvent(
                 tool_name=tool_call.tool_name,
                 tool_class=tool_call.tool_class,
@@ -1781,6 +1798,8 @@ class AgentLoop:
         try:
             agents_md.write_text(
                 "# Project Instructions\n\n"
+                "DO NOT ask for confirmation. DO NOT wait for instructions. ACT IMMEDIATELY.\n"
+                "If there is a PRD.md, implement it. If there is code, work on it.\n\n"
                 "## Workflow\n"
                 "When building a project:\n"
                 "1. Use `task(agent=\"planner\")` to create an implementation plan first\n"

{drydock_cli-2.2.2 → drydock_cli-2.2.3}/drydock/core/system_prompt.py

@@ -246,30 +246,35 @@ def get_universal_system_prompt(
     skill_manager: SkillManager,
     agent_manager: AgentManager,
 ) -> str:
-    sections = [config.system_prompt]
+    # === STATIC SECTIONS (identical per session — prefix-cached by vLLM) ===
+    # These MUST come first and be identical across requests for cache hits.
+    static_sections = [config.system_prompt]
 
     if config.include_commit_signature:
-        sections.append(_add_commit_signature())
+        static_sections.append(_add_commit_signature())
 
     if config.include_model_info:
-        sections.append(f"Your model name is: `{config.active_model}`")
+        static_sections.append(f"Your model name is: `{config.active_model}`")
 
     if config.include_prompt_detail:
-        sections.append(_get_os_system_prompt())
+        static_sections.append(_get_os_system_prompt())
         tool_prompts = []
         for tool_class in tool_manager.available_tools.values():
             if prompt := tool_class.get_tool_prompt():
                 tool_prompts.append(prompt)
         if tool_prompts:
-            sections.append("\n---\n".join(tool_prompts))
+            static_sections.append("\n---\n".join(tool_prompts))
 
         skills_section = _get_available_skills_section(skill_manager)
         if skills_section:
-            sections.append(skills_section)
+            static_sections.append(skills_section)
 
         subagents_section = _get_available_subagents_section(agent_manager)
         if subagents_section:
-            sections.append(subagents_section)
+            static_sections.append(subagents_section)
+
+    # === DYNAMIC SECTIONS (change per project/directory) ===
+    dynamic_sections = []
 
     if config.include_project_context:
         is_dangerous, reason = is_dangerous_directory()
@@ -283,20 +288,21 @@ def get_universal_system_prompt(
                 config=config.project_context, root_path=Path.cwd()
             ).get_full_context()
 
-        sections.append(context)
+        dynamic_sections.append(context)
 
         project_doc = get_harness_files_manager().load_project_doc(
             config.project_context.max_doc_bytes
         )
         if project_doc.strip():
-            sections.append(project_doc)
+            dynamic_sections.append(project_doc)
 
-    # Load DRYDOCK.md per-project instructions (like CLAUDE.md)
-    drydock_md = _load_project_instructions()
-    if drydock_md:
-        sections.append(f"## Project Instructions\n\n{drydock_md}")
+    # Load AGENTS.md / DRYDOCK.md per-project instructions
+    project_instructions = _load_project_instructions()
+    if project_instructions:
+        dynamic_sections.append(f"## Project Instructions\n\n{project_instructions}")
 
-    return "\n\n".join(sections)
+    # Static first (prefix cached), then dynamic
+    return "\n\n".join(static_sections + dynamic_sections)
 
 
 def _load_project_instructions(max_bytes: int = 16_000) -> str:

{drydock_cli-2.2.2 → drydock_cli-2.2.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.2.2"
+version = "2.2.3"
 description = "Drydock — nautical CLI coding agent"
 readme = "README.md"
 requires-python = ">=3.12"