Fara-1.5: single Fara15Agent, self-contained env/core, updated CLI + README

[husseinmozannar]

Summary

Ports the repo to Fara-1.5 and makes the agent stack self-contained. The user-facing fara-cli interface is unchanged; the changes are the new agent, a vendored browser/eval stack, and a refreshed README.

What's in here

Fara-1.5 agent

• New single Fara15Agent (src/fara/agents/fara/fara15_agent.py): the browser observe–think–act loop with the expanded FaraNext action space (double_click, right_click, triple_click, left_click_drag, hscroll, read_page_answer_question, ask_user_question, …).

Self-contained stack (vendored + adapted)

• environments/ — the PlaywrightEnvironment browser stack.
• core/ — DataPoint / RunContext trajectory model + incremental persistence (trimmed the verifier-only result schemas and screen_resolutions, which the agent never uses).
• clients/ — a minimal ChatCompletionClient over AsyncOpenAI.
• _prompts.py — the FaraNext browser tool schema + identities + critical points, reproducing the exact training prompt.

Previous generation preserved

• Fara-7B moved under fara_7b/, still runnable via fara-cli --fara-7b. from fara import FaraAgent, FARA_ACTION_DEFINITIONS still resolves.

CLI

• Defaults to Fara-1.5; --fara-7b runs the old agent; saves a full trajectory (data_point.json + per-step screenshots) to --output_folder; supports the ask_user_question critical-point resume flow.

Other

• README retitled and rewritten for the Fara-1.5 family; documented --fara-7b and trajectory output.
• vLLM pin updated to the Qwen3.5 serving stack (vllm==0.19.1, transformers>=5.2.0, nvidia-cutlass-dsl==4.5.2).
• tests/test_fara15.py: prompt byte-identity guardrail, action-space consistency, coord scaling, DataPoint round-trip.

Testing

• Unit tests pass (8/8), ruff clean.
• Smoke-tested fara-cli end-to-end against a hosted fara-1.5-9b vLLM endpoint (multi-step web tasks + the ask_user_question back-and-forth).