pocketcoder-a1 0.2.5

Autonomous Coding Agent — The Autonomous Gnome

PocketCoder-A1

Autonomous Coding Agent with Web Dashboard

Write tasks before bed, press Start, wake up to results. A1 works autonomously — runs code, verifies output, takes the next task. Real-time dashboard shows everything.

7086 lines of Python. Zero frameworks. 3 providers. Verification that catches a lying agent.

---

Install

pip install pocketcoder-a1

Also need Claude Code CLI:

npm i -g @anthropic-ai/claude-code

Requirements: Python 3.10+, Node.js 18+

Full step-by-step guide: GETTING_STARTED.md

---

Quick Start

# 1. Create venv
python3 -m venv .venv
source .venv/bin/activate

# 2. Install
pip install pocketcoder-a1

# 3. Init on your project
cd /path/to/your-project
pca init .

# 4. Add tasks
pca task add "Add /api/health endpoint"
pca task add "Write unit tests for auth"

# 5. Launch dashboard
pca ui
# Opens http://localhost:7331

---

Dashboard

8 pages, 24 API endpoints, everything in one Python file.

Live Agent Log

Watch what Claude does in real-time. Every file read, edit, bash command — with colored icons.

Tasks with Drag-and-Drop

Cards with priorities, statuses, success criteria. Drag to reorder.

AI Transform

Paste messy notes in free form. AI breaks them into structured tasks with descriptions and criteria.

Completed

5/5 tasks done. 12 sessions. 23 files modified. All verified.

Light Theme

---

How It Works

You add tasks → Start Agent → Agent works autonomously
                                    ↓
                              Takes next task
                                    ↓
                              Writes code (Claude CLI subprocess)
                                    ↓
                              Verifies result ← "Don't trust, verify"
                              (pytest, py_compile, files on disk)
                                    ↓
                         PASS → next task    FAIL → retry (max 5)
                                    ↓
                              All done → COMPLETED

Verification: "Don't Trust, Verify"

Agent says "COMPLETED"? We don't believe it. Three-tier check:

Tier	Type	Checks
1	BLOCKING	py_compile, pytest, files exist, success criteria
2	WARNING	ruff, build, git diff
3	ANTI-LOOP	baseline comparison, max 5 retries, force accept

Caught the agent lying about completion with failing tests. This system prevents that.

---

Providers

Provider	Command	Requires
claude-max	pca start	Claude Max subscription
claude-api	pca start --provider claude-api	Anthropic API key
ollama	pca start --provider ollama	Local Ollama server

Configure in dashboard: Settings page.

---

CLI Commands

Command	What it does
pca init .	Initialize project
pca task add "..."	Add task
pca tasks	Show all tasks
pca start	Start autonomous agent
pca start --task task_001	Work on single task
pca status	Current status
pca validate	Run validation checks
pca ui	Launch dashboard (:7331)
pca config	View settings
pca config set provider ollama	Change provider
pca log	Session history
pca test	Run E2E vision tests

---

Dashboard Pages

Page	URL	What
Dashboard	/	6 metric cards, Start/Stop, live log, Quick Add
Tasks	/tasks	List + drag-and-drop + bulk add + AI Transform
Task Detail	/task/{id}	Full view + logs + Start/Stop/Delete
Sessions	/sessions	Session history with metrics
Activity Log	/log	Timeline of actions
Settings	/settings	Provider, API key, parameters
Commits	/commits	Git history
Transform	/transform	Free text → structured tasks via AI

6 Metric Cards

Card	Example
Tasks	2/5 done + progress bar
Session	#3 + WORKING badge
Tokens	12.4K in / 3.2K out
Cost	$0.08
Duration	48s (live timer)
Files	3 modified

---

Project Structure

a1/
├── loop.py        (1219)  # Brain: subprocess → stream-json → verify
├── dashboard.py   (3491)  # Web UI: 8 pages, 24 API, inline HTML/CSS/JS
├── validator.py    (361)  # Eyes: syntax, tests, lint, build, criteria
├── cli.py          (387)  # CLI: pca init/task/start/ui/test/config
├── config.py       (126)  # Settings: CLI > env > config.json > defaults
├── tasks.py        (236)  # Tasks: CRUD, priority, reorder, criteria
├── checkpoint.py   (173)  # State: session, status, metrics, decisions
└── tester/        (1087)  # Vision QA: Playwright + Claude Vision

---

API

Full REST API — 24 endpoints. Monitor from scripts:

# Current status
curl http://localhost:7331/api/status | python -m json.tool

# Agent log
curl "http://localhost:7331/api/log?since=0" | python -m json.tool

# Add task
curl -X POST http://localhost:7331/add-task -d "task=Fix+the+bug&description=details"

---

Links

• Habr article (RU) — PocketCoder v1
• Telegram
• BVMax

License

MIT