drydock-cli 2.2.2__tar.gz → 2.2.4__tar.gz

drydock_cli-2.2.4/CLAUDE.md

@@ -0,0 +1,282 @@
+# Drydock Development Guide
+
+## Project
+
+Drydock is a local CLI coding agent (fork of mistral-vibe, Apache 2.0).
+- **Repo:** https://github.com/fbobe321/drydock
+- **PyPI:** https://pypi.org/project/drydock-cli/ (v2.2.1 for v2, v3.0.0 for v3)
+- **Goal:** Best-in-class SWE-bench Verified pass rate with local LLMs
+- **Hardware:** 2x RTX 4060 Ti 16GB, Gemma 4 26B MoE (A4B) via vLLM Docker at localhost:8000
+- **Server:** remus (Ubuntu 22.04, user: bobef)
+- **Active model:** Gemma 4 26B-A4B-it-AWQ-4bit (replaced devstral-24B — 3-4x faster, better results)
+- **Active codebase:** DryDock v3 at `/data3/drydock-v3/` (clean rewrite)
+
+## Build & Test
+
+### DryDock v3 (active)
+
+```bash
+# Install (minimal deps: only openai>=1.0)
+cd /data3/drydock-v3 && pip install -e .
+
+# Run drydock CLI (defaults to gemma4 model)
+python -m drydock
+python -m drydock --model gemma4
+
+# Run programmatically (headless, for benchmarks)
+PYTHONPATH=/data3/drydock-v3 python3 -c "import sys; sys.path.insert(0, '/data3/drydock-v3'); from drydock.cli import main; main()"
+
+# Start Gemma 4 model server
+bash /data3/Models/start_gemma4.sh
+
+# Check model is running
+curl http://localhost:8000/v1/models
+
+# Syntax check modified files
+python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
+```
+
+### DryDock v2 (legacy)
+
+```bash
+# Install
+uv sync
+
+# Run drydock CLI
+uv run drydock
+
+# Run programmatically (headless, for benchmarks)
+python3 -c "import sys; sys.path.insert(0, '.'); from drydock.cli.entrypoint import main; main()" --agent auto-approve -p "your prompt"
+
+# Run tests
+uv run pytest tests/ -x -q --timeout=30
+
+# Syntax check modified files
+python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
+```
+
+## Key Architecture
+
+### DryDock v3 (active — clean rewrite at `/data3/drydock-v3/`)
+
+v3 is a from-scratch rewrite using nano-claude-code as foundation. 4 core files, ~750 lines total. Provider-agnostic, works with any OpenAI-compatible endpoint.
+
+- `drydock/agent.py` — Core agent loop (161 lines). Multi-turn tool calling, event-driven. **Most changes go here.**
+- `drydock/providers.py` — LLM abstraction (209 lines). Provider registry (vLLM, Ollama, LM Studio, OpenAI, Anthropic). Streams responses, filters Gemma 4 thinking tokens.
+- `drydock/tool_registry.py` — Tool plugin system (46 lines). Register/execute tools by name.
+- `drydock/tools/__init__.py` — 6 built-in tools (258 lines): Read, Write, Edit, Bash, Glob, Grep
+- `drydock/cli.py` — CLI entry point (213 lines). Default model: gemma4.
+- `drydock/compaction.py` — Two-tier context management (119 lines). Normal compaction + emergency compaction on 400 errors.
+
+### DryDock v2 (legacy at `/data3/drydock/`)
+
+- `drydock/core/agent_loop.py` — Main agent loop. Loop detection, message ordering, tool execution.
+- `drydock/core/programmatic.py` — Headless API entry point used by SWE-bench harness
+- `drydock/core/build_orchestrator.py` — Multi-phase build pipeline (currently disabled, model drives builds)
+- `drydock/core/tools/builtins/bash.py` — Shell tool with allowlist/denylist, conda support
+- `drydock/core/tools/builtins/search_replace.py` — File editing tool with fuzzy matching
+- `drydock/core/prompts/cli.md` — System prompt (workflow, delegation, tool rules)
+- `drydock/core/prompts/builder.md` — Minimal prompt for file builder subagents
+- `drydock/core/system_prompt.py` — Builds system prompt, loads AGENTS.md/DRYDOCK.md
+- `drydock/core/types.py` — `MessageList` (custom Sequence, use `.reset()` not `=[]`)
+- `drydock/core/middleware.py` — Tiered context warnings, auto-compaction
+- `drydock/core/agents/models.py` — Agent profiles (explore, diagnostic, planner, builder)
+- `drydock/core/skills/manager.py` — Skill discovery and loading
+- `drydock/cli/textual_ui/app.py` — TUI application
+- `drydock/cli/commands.py` — Slash commands (/help, /setup-model, /consult, etc.)
+
+## Critical Constraints
+
+### v3 Constraints
+- **Gemma 4 leaks thinking tokens.** The model emits `<|channel>thought<channel|>` tokens that must be filtered out. `providers.py` strips these via regex before displaying output.
+- **Two-tier compaction.** Normal compaction truncates old tool results. Emergency compaction triggers on 400 errors (context overflow) with aggressive truncation.
+- **Provider-agnostic design.** No model-specific hacks in the agent loop. All model quirks handled in `providers.py`.
+
+### v2 Constraints (still apply to v2 codebase)
+- **MessageList is not a plain list.** Never `self.messages = [...]`. Use `self.messages.reset([...])`.
+- **vLLM/Mistral rejects `user` after `tool` messages.** `_sanitize_message_ordering()` runs before every LLM call. All injection must use `_inject_system_note()`.
+- **`os._exit()` in programmatic.py** — necessary because async cleanup hangs. Ensure stdout is flushed before calling.
+- **AGENTS.md is CRITICAL for devstral.** Without per-project instructions, the model loops on ls/bash. Auto-created by `_ensure_agents_md()`.
+
+### Shared Constraints
+- **"Restate the goal" blocks tool calling.** Never ask the model to output text before its first tool call — it pre-empts the tool-calling mechanism.
+- **Additive-only harness changes.** Every control flow modification (circuit breaker, orchestrator, nudges) caused regressions. Just inject better context before the agent loop.
+- **Circuit breaker is disabled.** It blocked valid retries. Loop detection prunes duplicates and nudges instead. Never stops the session.
+- **Loop detection never stops.** Only prunes repeated tool calls and injects guidance. The only hard stop is MAX_TOOL_TURNS (200).
+
+## Gemma 4 Tool Calling (v3 — active)
+
+- **Model:** Gemma-4-26B-A4B-it-AWQ-4bit (26B MoE, only 4B active params per token)
+- **Serving:** vLLM Docker image `vllm/vllm-openai:gemma4` with `--tool-call-parser gemma4 --enable-auto-tool-choice`
+- **Start script:** `/data3/Models/start_gemma4.sh` (Docker: 2 GPU tensor parallel, 131K context, fp8 KV cache)
+- **Performance:** 3-4x faster than devstral-24B, 0% timeouts on SWE-bench, 70% file match (vs 60% devstral)
+- **Thinking tokens:** Model leaks `<|channel>thought<channel|>` — filtered in `providers.py` via regex
+- **Tool choice:** `tool_choice="auto"` (default), model natively decides tool vs text
+- **No AGENTS.md dependency:** Gemma 4 does not loop without AGENTS.md (unlike devstral)
+- **Docker management:** `docker stop gemma4 && docker rm gemma4` to restart
+
+## Mistral/devstral Tool Calling (v2 — legacy)
+
+- `tool_choice="auto"` (default): model decides whether to call a tool or respond with text
+- Never ask for text before first tool call — blocks tool calling
+- `tool_choice="required"` forces a tool call (useful for first-turn delegation)
+- AGENTS.md anchors model behavior — without it, devstral loops
+- Plan→Edit workflow: use plan mode first, then switch to accept-edits
+
+## File Locations
+
+```
+/data3/drydock-v3/                 ← DryDock v3 (active, clean rewrite)
+├── drydock/                       ← Python package (~750 lines total)
+│   ├── agent.py                   ← Core agent loop (THE most important file)
+│   ├── providers.py               ← LLM abstraction (vLLM, Ollama, etc.)
+│   ├── tool_registry.py           ← Tool plugin system
+│   ├── tools/__init__.py          ← 6 built-in tools (Read, Write, Edit, Bash, Glob, Grep)
+│   ├── cli.py                     ← CLI entry point
+│   ├── compaction.py              ← Two-tier context management
+│   ├── memory/                    ← Memory system
+│   └── skills/                    ← Skill framework
+├── scripts/
+│   └── monitor.sh                 ← Process monitor (restarts Gemma 4 if needed)
+├── logs/                          ← Runtime logs
+├── pyproject.toml                 ← v3.0.0, depends only on openai>=1.0
+└── AGENTS.md                      ← Project instructions
+```
+
+```
+/data3/drydock/                    ← DryDock v2 (legacy, original codebase)
+├── drydock/                       ← Python package
+│   ├── core/
+│   │   ├── agent_loop.py          ← v2 agent loop
+│   │   ├── prompts/cli.md         ← System prompt
+│   │   ├── build_orchestrator.py  ← Multi-phase builder (disabled)
+│   │   ├── tools/builtins/        ← 24 builtin tools
+│   │   ├── agents/                ← Subagent definitions
+│   │   └── skills/                ← Skill framework
+│   ├── cli/textual_ui/app.py     ← TUI application
+│   └── skills/                    ← 7 bundled skills
+├── tests/
+│   ├── test_smoke.py              ← 20 fast tests
+│   ├── test_loop_detection.py     ← 38 loop tests
+│   ├── test_bank_prd.py           ← 15 PRD-driven tests
+│   └── testbank_helpers.py        ← Shared test infrastructure
+├── scripts/
+│   ├── deploy_to_github.sh        ← Daily push (4 AM)
+│   ├── publish_to_pypi.sh         ← Tests → build → PyPI → tag
+│   ├── notify_release.py          ← Telegram notifications
+│   └── backup.sh                  ← rsync to NAS (3 AM)
+└── test_bank_results/             ← Test logs
+```
+
+```
+/data3/swe_bench_runs/             ← SWE-bench infrastructure (separate)
+├── harness.py                     ← Runs drydock on SWE-bench tasks (supports v2 + v3 backends)
+├── continuous_bench.sh            ← Continuous improvement loop (v2/devstral)
+├── continuous_gemma4.sh           ← Continuous improvement loop (v3/Gemma 4, 20-task batches)
+└── logs/                          ← Batch results
+```
+
+```
+/data3/Models/                     ← Model files and startup scripts
+├── Gemma-4-26B-A4B-it-AWQ-4bit/  ← Gemma 4 model weights
+└── start_gemma4.sh                ← Docker startup (vLLM + Gemma 4)
+```
+
+## Configuration Management
+
+**All scripts use explicit Python path:** `/home/bobef/miniconda3/bin/python3`
+**Cron PATH doesn't inherit shell.** Every script exports PATH explicitly.
+**User's DryDock install:** `/home/bobef/miniforge3/envs/drydock/bin/python3` (Python 3.14)
+**Dev/test Python:** `/home/bobef/miniconda3/bin/python3` (Python 3.12)
+**Config:** `~/.drydock/config.toml`
+**Tokens:** `~/.config/drydock/github_token`, `~/.config/drydock/pypi_token`
+**Git remotes:** `drydock` → github.com/fbobe321/drydock, `origin` → github.com/mistralai/mistral-vibe
+
+## Deployment Process
+
+1. Code → modify files
+2. Syntax check → `python3 -c "import ast; ..."`
+3. Tests → `pytest tests/test_smoke.py tests/test_loop_detection.py -q`
+4. Commit → descriptive message
+5. Publish → `./scripts/publish_to_pypi.sh` (tests → build → PyPI → tag → integration test → GitHub → Telegram)
+6. Install on user's env → `/home/bobef/miniforge3/envs/drydock/bin/pip install --force-reinstall --no-deps --no-cache-dir drydock-cli==X.Y.Z`
+
+## SWE-bench
+
+### Current Results (Gemma 4 + DryDock v3)
+- **70% file match** (vs 60% devstral v2), 0% timeouts, 8x faster per task
+- Harness at `/data3/swe_bench_runs/harness.py` supports both v2 (devstral) and v3 (gemma4) backends
+- Continuous bench loop: `/data3/swe_bench_runs/continuous_gemma4.sh` (20-task batches with Telegram updates)
+- AGENTS.md auto-created per task with bug-fix workflow
+- Environment bootstrapping (Meta-Harness technique) gathers repo structure before first LLM call
+
+### Previous Results (devstral + DryDock v2)
+- 42% file match, 62% real patches (50-task eval)
+- Target was 68% (devstral-small-2 published score)
+- Continuous bench loop: `/data3/swe_bench_runs/continuous_bench.sh`
+
+## 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)
+- ✅ DryDock v3 clean rewrite (provider-agnostic, ~750 lines)
+- ✅ Gemma 4 MoE model (3-4x faster, 70% file match)
+- ✅ Two-tier context compaction (normal + emergency on 400 errors)
+- ✅ Docker-based model serving (vLLM + Gemma 4)
+- ✅ Multi-backend SWE-bench harness (v2/devstral + v3/gemma4)
+- ✅ Thinking token filtering (Gemma 4 `<|channel>` leak)
+
+**Legal note:** All patterns are standard design concepts implemented from scratch. No proprietary code copied.
+
+## Key Learnings
+
+1. AGENTS.md is essential — devstral loops without it (Gemma 4 does not have this problem)
+2. "Restate the goal" blocked tool calling for months — one line
+3. Additive-only changes work best — control flow mods regress
+4. The harness matters as much as the model (Meta-Harness paper)
+5. Environment bootstrapping saves 2-4 exploratory turns
+6. Plan→Edit workflow is Mistral's design, don't fight it
+7. Circuit breaker was net negative — removed entirely
+8. Loop detection should guide, never stop
+9. Test the installed package, not source (different Python envs)
+10. Never wait for full test runs — fix issues from early results
+11. Clean rewrites beat incremental fixes — v3 (~750 lines) outperforms v2 (~5000+ lines)
+12. Model choice matters more than agent complexity — Gemma 4 MoE gives 3-4x speed + better accuracy than devstral
+13. Gemma 4 leaks thinking tokens (`<|channel>thought<channel|>`) — must filter in provider layer
+14. Two-tier compaction (normal + emergency on 400) is more robust than single-tier
+15. Provider-agnostic design enables fast model switching — v3 works with any OpenAI-compatible endpoint
+16. Docker-based model serving (vLLM) is more reliable than bare-metal — easier restarts, GPU management
+
+## When Compacting
+
+Always preserve: the list of files modified in this session, any failing test output, current SWE-bench results, and the user's latest instructions.
+
+## Style
+
+- Python 3.12+, type hints, pathlib over os.path
+- Match existing patterns in the codebase
+- No unnecessary abstractions — keep fixes minimal
+- Always run syntax check after editing .py files
+- Don't add features beyond what was asked

{drydock_cli-2.2.2 → drydock_cli-2.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.2.2
+Version: 2.2.4
 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-2.2.4}/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-2.2.4}/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-2.2.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.2.2"
+version = "2.2.4"
 description = "Drydock — nautical CLI coding agent"
 readme = "README.md"
 requires-python = ">=3.12"

drydock_cli-2.2.2/CLAUDE.md

@@ -1,152 +0,0 @@
-# Drydock Development Guide
-
-## Project
-
-Drydock is a local CLI coding agent (fork of mistral-vibe, Apache 2.0).
-- **Repo:** https://github.com/fbobe321/drydock
-- **PyPI:** https://pypi.org/project/drydock-cli/ (v2.2.1)
-- **Goal:** Best-in-class SWE-bench Verified pass rate with local LLMs
-- **Hardware:** 2x RTX 4060 Ti 16GB, devstral-24B-AWQ-4bit via vLLM at localhost:8000
-- **Server:** remus (Ubuntu 22.04, user: bobef)
-
-## Build & Test
-
-```bash
-# Install
-uv sync
-
-# Run drydock CLI
-uv run drydock
-
-# Run programmatically (headless, for benchmarks)
-python3 -c "import sys; sys.path.insert(0, '.'); from drydock.cli.entrypoint import main; main()" --agent auto-approve -p "your prompt"
-
-# Run tests
-uv run pytest tests/ -x -q --timeout=30
-
-# Syntax check modified files
-python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
-```
-
-## Key Architecture
-
-- `drydock/core/agent_loop.py` — Main agent loop. Loop detection, message ordering, tool execution. **Most changes go here.**
-- `drydock/core/programmatic.py` — Headless API entry point used by SWE-bench harness
-- `drydock/core/build_orchestrator.py` — Multi-phase build pipeline (currently disabled, model drives builds)
-- `drydock/core/tools/builtins/bash.py` — Shell tool with allowlist/denylist, conda support
-- `drydock/core/tools/builtins/search_replace.py` — File editing tool with fuzzy matching
-- `drydock/core/prompts/cli.md` — System prompt (workflow, delegation, tool rules)
-- `drydock/core/prompts/builder.md` — Minimal prompt for file builder subagents
-- `drydock/core/system_prompt.py` — Builds system prompt, loads AGENTS.md/DRYDOCK.md
-- `drydock/core/types.py` — `MessageList` (custom Sequence, use `.reset()` not `=[]`)
-- `drydock/core/middleware.py` — Tiered context warnings, auto-compaction
-- `drydock/core/agents/models.py` — Agent profiles (explore, diagnostic, planner, builder)
-- `drydock/core/skills/manager.py` — Skill discovery and loading
-- `drydock/cli/textual_ui/app.py` — TUI application
-- `drydock/cli/commands.py` — Slash commands (/help, /setup-model, /consult, etc.)
-
-## Critical Constraints
-
-- **MessageList is not a plain list.** Never `self.messages = [...]`. Use `self.messages.reset([...])`.
-- **vLLM/Mistral rejects `user` after `tool` messages.** `_sanitize_message_ordering()` runs before every LLM call. All injection must use `_inject_system_note()`.
-- **`os._exit()` in programmatic.py** — necessary because async cleanup hangs. Ensure stdout is flushed before calling.
-- **AGENTS.md is CRITICAL for devstral.** Without per-project instructions, the model loops on ls/bash. Auto-created by `_ensure_agents_md()`.
-- **"Restate the goal" blocks tool calling.** Never ask the model to output text before its first tool call — it pre-empts the tool-calling mechanism.
-- **Additive-only harness changes.** Every control flow modification (circuit breaker, orchestrator, nudges) caused regressions. Just inject better context before the agent loop.
-- **Circuit breaker is disabled.** It blocked valid retries. Loop detection prunes duplicates and nudges instead. Never stops the session.
-- **Loop detection never stops.** Only prunes repeated tool calls and injects guidance. The only hard stop is MAX_TOOL_TURNS (200).
-
-## Mistral/devstral Tool Calling
-
-- `tool_choice="auto"` (default): model decides whether to call a tool or respond with text
-- Never ask for text before first tool call — blocks tool calling
-- `tool_choice="required"` forces a tool call (useful for first-turn delegation)
-- AGENTS.md anchors model behavior — without it, devstral loops
-- Plan→Edit workflow: use plan mode first, then switch to accept-edits
-
-## File Locations
-
-```
-/data3/drydock/                    ← Main repo (git)
-├── drydock/                       ← Python package
-│   ├── core/
-│   │   ├── agent_loop.py          ← THE most important file
-│   │   ├── prompts/cli.md         ← System prompt
-│   │   ├── build_orchestrator.py  ← Multi-phase builder (disabled)
-│   │   ├── tools/builtins/        ← 24 builtin tools
-│   │   ├── agents/                ← Subagent definitions
-│   │   └── skills/                ← Skill framework
-│   ├── cli/textual_ui/app.py     ← TUI application
-│   └── skills/                    ← 7 bundled skills
-├── tests/
-│   ├── test_smoke.py              ← 20 fast tests
-│   ├── test_loop_detection.py     ← 38 loop tests
-│   ├── test_bank_prd.py           ← 15 PRD-driven tests
-│   └── testbank_helpers.py        ← Shared test infrastructure
-├── scripts/
-│   ├── deploy_to_github.sh        ← Daily push (4 AM)
-│   ├── publish_to_pypi.sh         ← Tests → build → PyPI → tag
-│   ├── notify_release.py          ← Telegram notifications
-│   └── backup.sh                  ← rsync to NAS (3 AM)
-└── test_bank_results/             ← Test logs
-```
-
-```
-/data3/swe_bench_runs/             ← SWE-bench infrastructure (separate)
-├── harness.py                     ← Runs drydock on SWE-bench tasks
-├── continuous_bench.sh            ← Continuous improvement loop
-└── logs/                          ← Batch results
-```
-
-## Configuration Management
-
-**All scripts use explicit Python path:** `/home/bobef/miniconda3/bin/python3`
-**Cron PATH doesn't inherit shell.** Every script exports PATH explicitly.
-**User's DryDock install:** `/home/bobef/miniforge3/envs/drydock/bin/python3` (Python 3.14)
-**Dev/test Python:** `/home/bobef/miniconda3/bin/python3` (Python 3.12)
-**Config:** `~/.drydock/config.toml`
-**Tokens:** `~/.config/drydock/github_token`, `~/.config/drydock/pypi_token`
-**Git remotes:** `drydock` → github.com/fbobe321/drydock, `origin` → github.com/mistralai/mistral-vibe
-
-## Deployment Process
-
-1. Code → modify files
-2. Syntax check → `python3 -c "import ast; ..."`
-3. Tests → `pytest tests/test_smoke.py tests/test_loop_detection.py -q`
-4. Commit → descriptive message
-5. Publish → `./scripts/publish_to_pypi.sh` (tests → build → PyPI → tag → integration test → GitHub → Telegram)
-6. Install on user's env → `/home/bobef/miniforge3/envs/drydock/bin/pip install --force-reinstall --no-deps --no-cache-dir drydock-cli==X.Y.Z`
-
-## SWE-bench
-
-- Current: 42% file match, 62% real patches (50-task eval)
-- Target: 68% (devstral-small-2 published score)
-- 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
-
-## Key Learnings
-
-1. AGENTS.md is essential — devstral loops without it
-2. "Restate the goal" blocked tool calling for months — one line
-3. Additive-only changes work best — control flow mods regress
-4. The harness matters as much as the model (Meta-Harness paper)
-5. Environment bootstrapping saves 2-4 exploratory turns
-6. Plan→Edit workflow is Mistral's design, don't fight it
-7. Circuit breaker was net negative — removed entirely
-8. Loop detection should guide, never stop
-9. Test the installed package, not source (different Python envs)
-10. Never wait for full test runs — fix issues from early results
-
-## When Compacting
-
-Always preserve: the list of files modified in this session, any failing test output, current SWE-bench results, and the user's latest instructions.
-
-## Style
-
-- Python 3.12+, type hints, pathlib over os.path
-- Match existing patterns in the codebase
-- No unnecessary abstractions — keep fixes minimal
-- Always run syntax check after editing .py files
-- Don't add features beyond what was asked