main
Web UI (tinyclaw web): - Flask + SSE real-time progress streaming - Queue view: add/delete tasks, status colours, error display - Live log panel with coloured output - History tab with unified diffs (Rosé Pine theme) - Model picker dropdown - Auto-refresh every 5s History (history.py): - JSONL-based, records before/after content + unified diff - Wired into driver for EDIT and CREATE actions - Failed tasks also recorded QA benchmark (tinyclaw qa <models...>): - 5 fixed tasks: divide/greet/is_even/repeat/safe_get - Scores syntax validity + instruction following - Comparison table: pass%, syntax%, latency, cost QA results (3 models): - gemma-3-12b-it: 80% pass, 100% syntax, ~944ms ← winner - gemma-3-4b-it: 60% pass, 100% syntax, ~917ms - qwen2.5-coder-7b: 0% pass, 0% syntax (max_tokens issue — hits 512 limit mid-function)
TinyClaw
Orchestrator that makes 1-4B parameter language models capable of real project work.
How it works
Tiny models (1-4B) can't plan, can't manage state, and can't reliably use tools. But they can edit a single function when given clear instructions.
TinyClaw exploits this: a driver script handles everything the model can't (file extraction, state management, verification, git), and the model just generates code edits.
User request --> Big model decomposes into atomic tasks (QUEUE.md)
|
Driver reads next task
|
Extracts target function (Python AST)
|
Builds minimal prompt (~500 tokens)
|
Tiny model returns edited function
|
Verifier checks syntax/imports
|
Pass? --> Splice back, git commit, next task
Fail? --> Retry (max 3), then skip
Usage
# Process a task queue
tinyclaw run --queue QUEUE.md --model qwen/qwen3.5-9b
# Check progress
tinyclaw status
# Auto-decompose a feature (uses big model)
tinyclaw plan "add user authentication with JWT tokens"
Queue Format (QUEUE.md)
EDIT auth.py validate_login "reject empty password strings"
EDIT main.py handle_request "add validate_token() call before processing"
CREATE config.py - "define AUTH_SECRET from environment variable"
TEST auth.py - "verify login rejects empty passwords"
One task per line: ACTION file function "instruction"
Install
pip install -e .
Design Principles
- The model is the hands, not the brain
- One function per task, one task per commit
- Under 500 tokens per prompt (fits in 2k context)
- Verify everything, trust nothing
- Auto-revert on failure
Status
Early development. Milestone 1 (core loop) in progress.
Description
Orchestrator that makes 1-4B models capable of real project work. Big model plans, tiny model executes.
70 KiB
0 Stars
1 Watchers
0 Forks
Languages
Python
74.5%
HTML
25.5%