drydock-cli 2.8.2

Local-first CLI coding agent — tested with Gemma 4 26B via vLLM

DryDock

     ____             ____             _    
    |  _ \ _ __ _   _|  _ \  ___   ___| | __
    | | | | '__| | | | | | |/ _ \ / __| |/ /
    | |_| | |  | |_| | |_| | (_) | (__|   < 
    |____/|_|   \__, |____/ \___/ \___|_|\_\
                |___/                       

Local-first CLI coding agent. Chart your course. Execute with precision.

DryDock is a TUI coding assistant designed to work with local LLMs. It provides a conversational interface to your codebase — explore, modify, build, and test projects through natural language and a powerful set of tools.

[!IMPORTANT] DryDock is tested and optimized for Gemma 4 26B-A4B (26B MoE, 4B active parameters). Recommended serving stack: llama.cpp with --jinja (the chat-template fix that prevents the tool-call loops Gemma 4 hits under other backends). vLLM is also documented below as a higher-throughput alternative for batch/eval workloads. Other models and providers are supported (Mistral, OpenAI, Anthropic, Ollama) but are not as thoroughly tested. If you use a different model, expect to tune prompts and tool settings.

Tested Hardware + Model

Component	Spec
GPUs	2× NVIDIA RTX 4060 Ti 16GB
Model (llama.cpp, recommended)	unsloth/gemma-4-26B-A4B-it-GGUF — UD-Q3_K_M (12.7GB) or UD-Q4_K_M (16.9GB)
Model (vLLM, alternative)	casperhansen/gemma-4-26b-a4b-it-AWQ-4bit
Performance	~15–17 tok/s decode (llama.cpp Q3), ~70 tok/s decode (vLLM AWQ)
Active params	4B per token (MoE architecture — fast inference)

Recommended path: llama.cpp + Unsloth GGUF

Why this is the recommended setup: Gemma 4's tool-calling format requires precise chat-template handling. Without --jinja, tool results get injected without the right turn markers and the model loops or returns empty assistant messages — the exact 400 Bad Request loop fixed in v2.7.39 (GH #14). With --jinja, the GGUF's bundled chat template handles tool turns natively and the loops disappear.

# 1. Download Unsloth's GGUF (Q3_K_M is the article-recommended quant;
#    UD-Q4_K_M is a higher-quality alternative if you have ~17GB VRAM)
huggingface-cli download unsloth/gemma-4-26B-A4B-it-GGUF \
    --include "gemma-4-26B-A4B-it-UD-Q3_K_M.gguf" \
    --local-dir /path/to/models

# 2. Build llama.cpp with CUDA (or use the Docker image
#    ghcr.io/ggml-org/llama.cpp:server-cuda)
git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=OFF
cmake --build build --config Release -j8 --target llama-server

# 3. Start the server with the article recipe
./build/bin/llama-server \
    -m /path/to/models/gemma-4-26B-A4B-it-UD-Q3_K_M.gguf \
    --host 0.0.0.0 --port 8000 \
    -ngl 99 -c 32768 -np 1 \
    --jinja \
    -ctk q8_0 -ctv q8_0 \
    --alias gemma4

Critical flags:

• --jinja — the loop-fix. Required for tool-using workflows. Without it, Gemma 4 enters infinite retry loops on multi-turn tool sessions.
• -ngl 99 — offload all layers to GPU
• -c 32768 — 32K context (fits in 16GB VRAM with q8 KV cache)
• -ctk q8_0 -ctv q8_0 — quantize KV cache for longer contexts
• -np 1 — single slot (concurrent requests serialize)
• --alias gemma4 — what the API reports as the model field

Drydock config (~/.drydock/config.toml):

active_model = "gemma4"

[[providers]]
name = "local"
api_base = "http://localhost:8000/v1"
api_key_env_var = ""
backend = "generic"

[[models]]
name = "gemma4"
provider = "local"
alias = "gemma4"
temperature = 1.0           # MUST be 1.0 with --jinja — lower temps reinforce loops
context_window = 32768       # Match `-c 32768` from llama-server. Drydock
                             # auto-clamps auto_compact_threshold to
                             # context_window − 4096 so we never blow past
                             # the server's max input.
auto_compact_threshold = 28000

# Article-recommended sampling (passed through extra_sampling to llama-server).
# Drydock auto-bakes these on first launch when llama.cpp is detected at
# 127.0.0.1:8080 / :8000, but you can override here.
[models.extra_params]
top_k = 40
top_p = 0.95
frequency_penalty = 1.1
max_tokens = 2048

Alternative: vLLM (higher throughput, no --jinja equivalent)

vLLM has its own --tool-call-parser gemma4 path that works for most workflows, but has been observed to enter tool-call loops on long multi-turn sessions (GH #14, fixed at the drydock side in v2.7.39 by filtering empty assistant messages before re-call). Use vLLM when you need higher decode throughput (~70 tok/s vs llama.cpp's ~15–17) for batch eval or non-interactive workloads where loop-fix matters less.

huggingface-cli download casperhansen/gemma-4-26b-a4b-it-AWQ-4bit \
    --local-dir /path/to/models/Gemma-4-26B-A4B-it-AWQ-4bit

docker run -d \
    --gpus all \
    --name gemma4 \
    -p 8000:8000 \
    -v /path/to/models:/models \
    --ipc=host \
    vllm/vllm-openai:gemma4 \
    --model /models/Gemma-4-26B-A4B-it-AWQ-4bit \
    --quantization compressed-tensors \
    --tensor-parallel-size 2 \
    --max-model-len 131072 \
    --max-num-seqs 2 \
    --gpu-memory-utilization 0.95 \
    --kv-cache-dtype fp8 \
    --served-model-name gemma4 \
    --trust-remote-code \
    --tool-call-parser gemma4 \
    --enable-auto-tool-choice \
    --attention-backend TRITON_ATTN

Key flags:

• --tensor-parallel-size 2 — split across 2 GPUs
• --kv-cache-dtype fp8 — reduce KV cache memory for longer contexts
• --tool-call-parser gemma4 + --enable-auto-tool-choice — required for Gemma 4 tool calling under vLLM
• --max-num-seqs 2 — limit concurrent requests (prevents OOM on 16GB GPUs)

Verify either backend is running:

curl http://localhost:8000/v1/models

For vLLM, drydock config is the same as the llama.cpp block above, except temperature = 0.2 is fine — the --jinja requirement only applies to llama.cpp.

Install

pip install drydock-cli

Or with uv:

uv tool install drydock-cli

[!TIP] New install hitting weird behavior? See DEPLOYMENT.md for the exact known-working vLLM launch flags, ~/.drydock/config.toml, env vars, and a diagnostic checklist. Most "DryDock doesn't work" issues trace back to missing vLLM flags (--tool-call-parser gemma4, --enable-auto-tool-choice) or temperature/thinking config drift.

Quick Start

cd your-project/
drydock

First run creates a config at ~/.drydock/config.toml and prompts for your provider setup.

> Review the PRD and build the package

Features

• TUI Interface: Full terminal UI with streaming output, tool approval, and session management.
• Adaptive Thinking: Automatically adjusts reasoning depth per turn — full thinking for planning, fast mode for file writes.
• Powerful Toolset: Read, write, and patch files. Execute shell commands. Search code with grep. Delegate to subagents.
• Project-Aware: Scans project structure, loads AGENTS.md / DRYDOCK.md for context.
• Subagent Delegation: Large tasks can be delegated to builder/planner/explorer subagents with isolated context.
• Loop Detection: Advisory-only detection that nudges the model away from repetitive actions without blocking.
• Conda/Pip Support: Auto-approves pip install, conda install, pytest, and other dev commands.
• Bundled Skills: Ships with skills like create-presentation for PowerPoint generation.
• MCP Support: Connect Model Context Protocol servers for extended capabilities.
• Safety First: Tool execution approval with --dangerously-skip-permissions for full auto-approve.

Built-in Agents

• default: Standard agent that requires approval for tool executions.
• plan: Read-only agent for exploration and planning.
• accept-edits: Auto-approves file edits only.
• auto-approve: Auto-approves all tool executions.

drydock --agent plan

Gemma 4 Optimizations

DryDock includes several optimizations specifically tuned for Gemma 4:

• Simplified prompt (gemma4.md): 20-line system prompt instead of 125 lines. Complex prompts cause Gemma 4 to plan instead of act.
• Non-streaming mode: Streaming breaks Gemma 4 tool call JSON parsing. DryDock automatically disables streaming for Gemma 4.
• Thinking token filtering: Gemma 4 leaks <|channel>thought<channel|> tokens into text output. DryDock strips these before storing in context.
• Adaptive thinking: Full thinking for planning (turn 1) and error recovery. Thinking OFF for routine file writes — eliminates 30-120s hangs between files.
• search_replace resilience: Auto-detects already-applied edits, infers missing file paths, fuzzy-matches whitespace differences.
• Reduced tool set: Disables tools that confuse Gemma 4 (ask_user_question, task_create, etc.).

Usage

Interactive Mode

drydock                        # Start interactive session
drydock "Fix the login bug"    # Start with a prompt
drydock --continue             # Resume last session
drydock --resume abc123        # Resume specific session

Keyboard shortcuts:

• Ctrl+C — Cancel current operation (double-tap to quit)
• Shift+Tab — Toggle auto-approve mode
• Ctrl+O — Toggle tool output
• Ctrl+G — Open external editor
• @ — File path autocompletion
• !command — Run shell command directly

Programmatic Mode

drydock --prompt "Analyze the codebase" --max-turns 5 --output json
drydock --dangerously-skip-permissions -p "Fix all lint errors"

Configuration

DryDock is configured via config.toml. It looks first in ./.drydock/config.toml, then ~/.drydock/config.toml.

API Key

drydock --setup                              # Interactive setup
export MISTRAL_API_KEY="your_key"            # Or set env var

Keys are saved to ~/.drydock/.env.

Consultant Model

Set a smarter model for the /consult command:

consultant_model = "gemini-2.5-pro"

The consultant provides read-only advice — it never calls tools. Use /consult <question> to ask it.

Custom Agents

Create agent configs in ~/.drydock/agents/:

# ~/.drydock/agents/redteam.toml
active_model = "devstral-2"
system_prompt_id = "redteam"
disabled_tools = ["search_replace", "write_file"]

Skills

DryDock discovers skills from:

1. Custom paths in config.toml via skill_paths
2. Project .drydock/skills/ or .agents/skills/
3. Global ~/.drydock/skills/
4. Bundled skills (shipped with the package)

MCP Servers

[[mcp_servers]]
name = "fetch_server"
transport = "stdio"
command = "uvx"
args = ["mcp-server-fetch"]

Testing

DryDock uses a shakedown harness (scripts/shakedown.py) that drives the real TUI via pexpect and judges on user-perceptible criteria — not tool-call counts.

# Single project test
python3 scripts/shakedown.py \
    --cwd /path/to/project \
    --prompt "review the PRD and build the package" \
    --pkg package_name

# Interactive back-and-forth test
python3 scripts/shakedown_interactive.py \
    --cwd /path/to/project \
    --pkg package_name

# Full regression suite (370 PRDs)
bash scripts/shakedown_suite.sh

Pass criteria: no write loops, no ignored interrupts, no search_replace cascades, package executes, session finishes within time budget.

Slash Commands

Type /help in the input for available commands. Create custom slash commands via the skills system.

Session Management

drydock --continue              # Continue last session
drydock --resume abc123         # Resume specific session
drydock --workdir /path/to/dir  # Set working directory

License

Copyright 2025 Mistral AI (original work) Copyright 2026 DryDock contributors (modifications)

Licensed under the Apache License, Version 2.0. See LICENSE for details.

DryDock is a fork of mistralai/mistral-vibe (Apache 2.0). See NOTICE for attribution.