drydock-cli 2.2.1__tar.gz → 2.2.2__tar.gz

drydock_cli-2.2.2/CLAUDE.md

@@ -0,0 +1,152 @@
+# 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

{drydock_cli-2.2.1 → drydock_cli-2.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 2.2.1
+Version: 2.2.2
 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/PRD.md

@@ -0,0 +1,81 @@
+# DryDock — Local CLI Coding Agent
+
+**Repository:** https://github.com/fbobe321/drydock
+**PyPI:** https://pypi.org/project/drydock-cli/ (v2.2.1)
+**License:** Apache 2.0 (fork of [mistralai/mistral-vibe](https://github.com/mistralai/mistral-vibe))
+
+## Vision
+
+Best-in-class local coding agent. Build, debug, and ship software using local LLMs on consumer hardware.
+
+## Current Status
+
+- **Model:** devstral-24B-AWQ-4bit on 2x RTX 4060 Ti 16GB
+- **SWE-bench Verified:** 42% file match (target: 68%)
+- **Version:** 2.2.1
+- **Users can:** Build projects from PRDs, fix bugs, review code, refactor
+
+## Features
+
+### Core
+- 24 builtin tools (bash, grep, read_file, write_file, search_replace, etc.)
+- 7 bundled skills (/investigate, /review, /ship, /batch, /simplify, /deep-research, /create-presentation)
+- Multi-agent delegation (explore, diagnostic, planner subagents)
+- Plan→Edit workflow (Mistral's designed pattern)
+- AGENTS.md support (cross-tool standard, auto-created)
+- Textual TUI with wave spinner, message queuing
+
+### Local Model Support
+- `/setup-model` command (vLLM, Ollama, LM Studio, custom)
+- `--local` CLI flag for quick setup
+- Auto-detect model name from server
+
+### Safety
+- Tool permission system (always/ask/never per tool)
+- Loop guidance (never stops, only redirects)
+- API error auto-recovery (retry after 10s)
+
+## Install
+
+```bash
+pip install drydock-cli
+drydock --local http://localhost:8000/v1
+```
+
+## Per-Project Instructions
+
+DryDock loads instructions from these files in the project root:
+- **AGENTS.md** — Cross-tool standard (recommended)
+- **DRYDOCK.md** — DryDock-specific
+- **.drydock/rules/*.md** — Modular rules
+
+Auto-created if none exist. Essential for devstral to use subagents properly.
+
+## Roadmap
+
+### Near-term
+- Reach 68% on SWE-bench Verified (from 42%)
+- Static/dynamic prompt split for prefix caching
+- Layered context compaction (4 fallback layers)
+- Composable prompt sections (task-type specific)
+- Background resource preloading
+
+### Medium-term
+- Real test execution in SWE-bench eval (not just file match)
+- Repo-specific strategies (Django vs sympy vs sphinx)
+- Draft-then-verify pattern for patches
+- Windows Ctrl-C fix
+
+### Long-term
+- Support larger models as hardware improves
+- Plugin marketplace for custom tools/skills
+- Web dashboard for monitoring long runs
+- Multi-model routing (different models for different tasks)
+
+## Architecture Notes
+
+See [CLAUDE.md](CLAUDE.md) for technical details, file locations, constraints, and development workflow.
+
+## Legal
+
+All code is original or forked from Apache 2.0 mistral-vibe. Architecture improvements are standard design patterns implemented from scratch. No proprietary code copied.

{drydock_cli-2.2.1 → drydock_cli-2.2.2}/drydock/cli/textual_ui/app.py

@@ -1706,6 +1706,9 @@ class VibeApp(App):  # noqa: PLR0904
             self.agent_loop.agent_profile
         )
         self._update_profile_widgets(new_profile)
+        # Hide any active loading/spinner widget before switching mode
+        # to prevent the old animation from showing behind the new spinner
+        await self._remove_loading_widget()
         if self._chat_input_container:
             self._chat_input_container.switching_mode = True
 

{drydock_cli-2.2.1 → drydock_cli-2.2.2}/drydock/core/agent_loop.py

@@ -762,7 +762,25 @@ class AgentLoop:
                     text_without_action += 1
                     # Always continue — don't let the agent exit without editing
                     should_break_loop = False
-                    if text_without_action == 1:
+
+                    # Detect model claiming task is done without having edited anything
+                    last_text_lower = (last_message.content or "").lower()
+                    task_complete_phrases = [
+                        "task completed", "task complete", "task is complete",
+                        "task is done", "task has been completed",
+                        "changes have been applied", "fix has been applied",
+                        "the fix is complete", "issue is resolved",
+                    ]
+                    claims_complete = any(p in last_text_lower for p in task_complete_phrases)
+
+                    if claims_complete:
+                        nudge_text = (
+                            "You said the task is complete but NO code was actually changed. "
+                            "Your search_replace call either failed or was never made. "
+                            "Use read_file to read the target file, then use search_replace "
+                            "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
@@ -792,9 +810,11 @@ class AgentLoop:
                             "Pick the most likely file and function, and make a minimal edit."
                         )
                     else:
-                        # Give up after 5 futile nudges — the model can't/won't edit
-                        should_break_loop = True
-                        nudge_text = None
+                        # Keep nudging — never give up
+                        nudge_text = (
+                            "You MUST call a tool. Use grep to find the file, "
+                            "read_file to read it, then search_replace to fix."
+                        )
                     if nudge_text:
                         # Use _inject_system_note (appends to last tool/user message)
                         # User messages cause 'tool after user' ordering violations

{drydock_cli-2.2.1 → drydock_cli-2.2.2}/drydock/core/config/harness_files/_harness_manager.py

@@ -69,11 +69,15 @@ class HarnessFilesManager:
         project_skills = Path.cwd() / ".drydock" / "skills"
         if project_skills.is_dir():
             dirs.append(project_skills)
-        # User skills in ~/.drydock/skills/
+        # User skills in ~/.drydock/skills/ (and ~/.vibe/skills/ for backward compat)
         if "user" in self.sources:
             d = GLOBAL_SKILLS_DIR.path
             if d.is_dir():
                 dirs.append(d)
+            # Always check legacy ~/.vibe/skills/ in case migration didn't complete
+            legacy_skills = Path.home() / ".vibe" / "skills"
+            if legacy_skills.is_dir() and legacy_skills != d:
+                dirs.append(legacy_skills)
         # Plugin skills
         try:
             from drydock.core.plugins import get_plugin_skill_dirs

{drydock_cli-2.2.1 → drydock_cli-2.2.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "drydock-cli"
-version = "2.2.1"
+version = "2.2.2"
 description = "Drydock — nautical CLI coding agent"
 readme = "README.md"
 requires-python = ">=3.12"

drydock_cli-2.2.2/scripts/monitor_swebench.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+# Monitor SWE-bench runs, analyze early results, report via Telegram
+# Cron: */10 * * * * /data3/drydock/scripts/monitor_swebench.sh
+
+set -uo pipefail
+export PATH="/home/bobef/miniconda3/bin:$PATH"
+PYTHON="/home/bobef/miniconda3/bin/python3"
+DRYDOCK="/data3/drydock"
+SWEBENCH="/data3/swe_bench_runs"
+LOG="$SWEBENCH/logs/monitor_swebench.log"
+
+log() { echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*" >> "$LOG"; }
+
+# Find the latest running log
+LATEST_LOG=$(ls -t $SWEBENCH/logs/v2*.log 2>/dev/null | head -1)
+if [ -z "$LATEST_LOG" ]; then
+    exit 0
+fi
+
+# Count completed tasks
+DONE=$(grep -c "Files match:" "$LATEST_LOG" 2>/dev/null || echo 0)
+MATCHES=$(grep -c "Files match: True" "$LATEST_LOG" 2>/dev/null || echo 0)
+PATCHES=$(grep "Patch:" "$LATEST_LOG" 2>/dev/null | grep -v "0L\|1L" | wc -l)
+TOTAL=$(grep -c "^\[" "$LATEST_LOG" 2>/dev/null | head -1 || echo 0)
+
+if [ "$DONE" -eq 0 ]; then
+    exit 0
+fi
+
+RATE=$((MATCHES * 100 / DONE))
+
+log "Progress: $DONE done, $MATCHES matches ($RATE%), $PATCHES patches"
+
+# Send Telegram update every 10 tasks
+LAST_REPORTED=$(cat "$SWEBENCH/logs/.last_reported" 2>/dev/null || echo 0)
+if [ "$DONE" -ge $((LAST_REPORTED + 10)) ]; then
+    $PYTHON "$DRYDOCK/scripts/notify_release.py" "swebench" "SWE-bench progress: $DONE tasks, $MATCHES file matches ($RATE%), $PATCHES patches. Log: $(basename $LATEST_LOG)"
+    echo "$DONE" > "$SWEBENCH/logs/.last_reported"
+    log "Sent Telegram update"
+fi

drydock_cli-2.2.1/CLAUDE.md

@@ -1,61 +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
-- **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
-
-## 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 vibe.cli.entrypoint import main; main()" --agent auto-approve -p "your prompt"
-
-# Run tests (note: pytest from swe-bench worktrees can pollute; use -p no:conftest if needed)
-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/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 (two-phase workflow, SWE-bench rules)
-- `drydock/core/types.py` — `MessageList` (custom Sequence, use `.reset()` not `=[]`)
-
-## 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 as safety net. All injection should use `_inject_system_note()`.
-- **`os._exit()` in programmatic.py** — necessary because async cleanup hangs. Ensure stdout is flushed before calling.
-- **Tests may fail due to stale SWE-bench worktrees** poisoning pytest. Use `python3 -c "import ast; ..."` for syntax checks instead.
-
-## SWE-bench Infrastructure
-
-Benchmarking scripts live at `/data3/swe_bench_runs/` (not in the drydock repo):
-- `harness.py` — Runs drydock on SWE-bench tasks, captures patches
-- `continuous_bench.sh` — Continuous improvement loop (cron @reboot + every 6h)
-- `deploy_to_github.sh` → `/data3/drydock/scripts/` — Daily push to GitHub
-
-## When Compacting
-
-Always preserve: the list of files modified in this session, any failing test output, and the current task context from the PRD.
-
-## 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