drydock-cli 2.2.3__tar.gz → 2.2.4__tar.gz

{drydock_cli-2.2.3 → drydock_cli-2.2.4}/CLAUDE.md

@@ -4,13 +4,40 @@
 
 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)
+- **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, devstral-24B-AWQ-4bit via vLLM at localhost:8000
+- **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
@@ -30,7 +57,20 @@ 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 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
@@ -47,16 +87,35 @@ python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
 
 ## 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).
 
-## Mistral/devstral Tool Calling
+## 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
@@ -67,10 +126,28 @@ python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
 ## File Locations
 
 ```
-/data3/drydock/                    ← Main repo (git)
+/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          ← THE most important file
+│   │   ├── agent_loop.py          ← v2 agent loop
 │   │   ├── prompts/cli.md         ← System prompt
 │   │   ├── build_orchestrator.py  ← Multi-phase builder (disabled)
 │   │   ├── tools/builtins/        ← 24 builtin tools
@@ -93,11 +170,18 @@ python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
 
 ```
 /data3/swe_bench_runs/             ← SWE-bench infrastructure (separate)
-├── harness.py                     ← Runs drydock on SWE-bench tasks
-├── continuous_bench.sh            ← Continuous improvement loop
+├── 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`
@@ -119,12 +203,18 @@ python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
 
 ## 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
+### 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:**
@@ -151,12 +241,18 @@ python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
 - ✅ "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
+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)
@@ -166,6 +262,12 @@ python3 -c "import ast; ast.parse(open('path/to/file.py').read())"
 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
 

{drydock_cli-2.2.3 → drydock_cli-2.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.2.3
+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-2.2.3 → drydock_cli-2.2.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.2.3"
+version = "2.2.4"
 description = "Drydock — nautical CLI coding agent"
 readme = "README.md"
 requires-python = ">=3.12"