aitester-bdd 0.2.0

LLM-driven BDD test authoring for Robot Framework — turn an intention + live app into a .robot suite

aitester-bdd

LLM-driven BDD test authoring for Robot Framework. Give it a story and a live web app; an agent explores the target via the agent-browser CLI, then writes a deterministic .robot suite with selectors grounded in the actual DOM it observed — or files a bug report when the system is broken in a way that prevents authoring.

What it is

A Robot Framework library that turns a plain-English intention into a deterministic, executable .robot test suite. Run-time has no LLM in the loop — the authored suite is plain RF code that runs reproducibly, no tokens consumed on PR gates.

What's novel

	aitester-bdd
Intention → .robot suite	An agent loop drives the live target via shell-out to agent-browser (Playwright under the hood), writes a Robot Framework suite with selectors grounded in the snapshots it actually took.
Bug-report exit channel	When the SUT is broken in a way that prevents authoring (missing UI, broken auth flow, untestable terminal state), the agent writes triage/<story>.md rather than inventing selectors.
Three pluggable runtime backends	agent-browser (default, zero-install) / playwright (in-process speed) / nodriver (bot-detection-resistant). Same .robot runs on any.
AOP failure aspect	Each failed rule ships with an AI-written natural-language diagnosis (SUT-vs-test classification) plus a full MDP trajectory in walk_log.jsonl.
Rule DAG with parent-child composition	Ported from WISE RPA BDD. Position-determined state checks (guard vs observation), retry-with-redo, scope inheritance — all expressed via Given/When/Then.

Status

Alpha. Authoring verified end-to-end on public sites (example.com, en.wikipedia.org, the-internet.herokuapp.com) and on a real internal SPA (login + chat + tool-rendering verification).

How fast is it?

Authoring is headless DeepAgents on Claude Opus 4.7 shelling out to the agent-browser CLI. Typical wall-time for a single suite:

Site / scope	Steps	Wall time
example.com smoke (heading + link)	9	~27s
en.wikipedia.org search + article check (5 assertions)	27	~70s
Real SPA login + chat + multi-rule verification	50-80	2-3 min

The agent batches multiple agent-browser subcommands per shell call (open && snapshot && get count ...) so ~1 LLM round-trip handles 2-4 browser ops. Most remaining wall-time is SUT-bound (waiting for the app's own LLM to stream a response) — not authoring overhead.

Quick start

# 1. Install
pip install aitester-bdd
npm i -g agent-browser

# 2. Point at an LLM endpoint. Defaults assume claude-code-proxy:
export AITESTER_LLM_MODEL=cc/claude-opus-4-7
export OPENAI_BASE_URL=http://localhost:20128/v1
export OPENAI_API_KEY=placeholder

# 3. Author a suite from a story
aitester author \
  --story "Open the homepage, search for 'BDD', verify the article heading and a paragraph containing 'BDD'." \
  --base-url https://en.wikipedia.org \
  --out wiki_smoke.robot

# 4. Run it (no LLM at run time)
aitester run wiki_smoke.robot

For visibility into the agent's exploration, add --debug to aitester author. Every LLM turn and shell call streams to stderr with timestamps.

Output sidecar files at <output_dir>/:

• walk_log.jsonl — every MDP transition (rule_enter / before_action / after_action / state_check / dismiss / emit / rule_exit)
• failures.jsonl — failure context + AI diagnosis for every failed rule
• emit.jsonl — explicit And I emit "..." captures (intention-driven; only when the story is a diagnostic probe)

Three runtime backends, one authored suite

AITESTER_BROWSER= picks the driver at run-time:

Backend	Default?	Setup	Best for
agent-browser	✓	none — CLI ships its own browser	most cases; same driver author + run, zero install friction
playwright		aitester init-browser once	action-heavy tests where subprocess latency matters
nodriver		pip install aitester-bdd[stealth] + Edge/Chrome	bot-detected sites (DataDome / Cloudflare BM / etc.)

Same .robot runs on any of the three.

Architecture (one paragraph)

The LLM is the author, not the runtime. At authoring time, a DeepAgents/LangGraph agent reads SKILL.md as its system prompt, drives the live target by shelling out to the agent-browser CLI (via DeepAgents' LocalShellBackend.execute tool), and emits a .robot file with selectors grounded in real snapshots — or writes a bug report when the system is broken in a way that prevents authoring. The agent batches multiple agent-browser subcommands per shell call (open && snapshot && get count …) so each LLM round-trip drives multiple browser ops. At run time, plain Robot Framework executes the suite via one of three pluggable browser backends; no LLM in the loop. Failures fire an AOP diagnose aspect that hands the LLM the MDP trajectory plus snapshot and asks "why?" — short natural-language diagnoses land on RuleResult.ai_diagnosis and failures.jsonl. The walker, gotcha-fixes, and AspectRegistry are ported from the WISE RPA BDD skill.

License

MIT