browser-use/browser-worlds-tinker-rl

Latest recorded source activity: August 8, 2026 at 10:12

SkillsMP catalog refreshed: November 5, 2026 at 19:57

skills collected: 1

GitHub stars: 4
GitHub forks: 3

Install with an AI assistant

Copy this prompt into the AI assistant you're using.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/browser-use/browser-worlds-tinker-rl.

Repository README

Repository README source: README.md

Read the source README

Expand full README

# Browser Worlds × Tinker RL This repository will demonstrate specialist browser-agent improvement with reinforcement learning on deterministic Browser Use Worlds. Before changing the environment, it reproduces Tinker Cookbook's default Harbor RL recipe on Terminal-Bench using Inkling. No browser-agent training code is present yet. ## Pinned baseline Tinker Cookbook is pinned in `pyproject.toml`; see [UPSTREAM.md](UPSTREAM.md). The baseline uses: - the upstream Harbor task loader and bash-tool agent; - Terminal-Bench 2.0; - Modal sandboxes; - each task's `tests/test.sh` reward; - `thinkingmachines/Inkling` instead of the recipe's original model default. ## Setup Place `TINKER_API_KEY` in `.env` and authenticate Modal. Then: ```bash uv sync uvx harbor datasets download terminal-bench@2.0 \ -o ~/.cache/harbor/tasks/terminal-bench-2.0 ``` Load `.env` without passing credentials as command arguments: ```bash set -a source .env set +a ``` ## One-step baseline Run the smallest real training step first: ```bash uv run python scripts/train_terminal_bench_smoke.py ``` This runs one Terminal-Bench task with two Inkling rollouts and disables the upstream iteration-zero evaluation over all 89 tasks. The Harbor agent, sandbox, grader, and training loop are unchanged. Validated on August 8, 2026: one batch and two rollouts completed, the optimizer step succeeded, and final Tinker state and sampler checkpoints were produced. Both rollouts reached their token limit and received `-0.1`, so this validates the baseline pipeline rather than task performance. ## Full upstream example After the one-step run succeeds: ```bash uv run python -m tinker_cookbook.recipes.harbor_rl.scripts.train_terminal_bench \ model_name=thinkingmachines/Inkling \ learning_rate=1e-5 ``` The learning rate is an explicit initial experiment value; Inkling does not publish a universal recommended RL learning rate. ## Zenith sandbox The browser stack uses one reusable Daytona base snapshot containing Chromium and Browser Harness. Each rollout creates one ephemeral sandbox, materializes the exact Zenith World binary, and runs the World, Chromium, and Browser Harness locally together. Browser Use Cloud is not involved. ```bash uv run python scripts/verify_zenith_daytona_sandbox.py \ --world-binary /tmp/zenith-world-server \ --customer-repo ~/work/browser-rl-daytona-harbor \ --output /tmp/zenith-daytona-browser-harness ``` The verifier checks the Zenith task result, records output-file and screenshot hashes, deletes the sandbox, and retains the shared snapshot for later rollouts.

Loading the repository README…

Skills in this repository

Software Developers

1 occupation category · 100% classified

skill

occupation

description

updated

browser-harness

Software Developers

Always use browser-harness for any web interaction: automation, scraping, testing, or site/app work.

2026-08-08

Showing 1 of 1 collected skill.