Skip to content

SETUP.md

Latest commit

 

History

History
120 lines (92 loc) · 5.46 KB

SETUP.md

File metadata and controls

120 lines (92 loc) · 5.46 KB
Raw

Setup & deployment

Two paths: (A) run locally (no UiPath account, fully offline) and (B) deploy on UiPath Automation Cloud via the uip CLI.

Prerequisites

  • Python 3.10+
  • (Path B) A UiPath Automation Cloud / UiPath Labs tenant with Test Cloud, Maestro, Agent Builder, Action Center, and Orchestrator enabled
  • (Path B) Node.js 18+ and the UiPath uip CLI (UiPath for Coding Agents)
  • (optional) ANTHROPIC_API_KEY to run the live grounded classifier locally

Path A — run locally (offline)

python corpus/generate_corpus.py                 # build the labeled corpus
python eval/harness.py --report eval/report.md    # accuracy + false-positive rate
python eval/negative_control.py                   # safety invariants (gates CI)
python seeded_suite/run_history.py --runs 14      # emit fresh execution history
python -m flakewarden.cli triage seeded_suite/history.jsonl
python -m pytest -q                               # unit tests

To route the ambiguous band through a real Claude model locally:

pip install anthropic
export ANTHROPIC_API_KEY=sk-...        # PowerShell: $env:ANTHROPIC_API_KEY="sk-..."
python -m flakewarden.cli triage seeded_suite/history.jsonl

Path B — deploy on UiPath Automation Cloud

Working against the AgentHack deadline? Follow the time-boxed, priority-ordered docs/deploy-runbook.md instead — it includes a pre-flight units/license check and a stop-and-demo ladder.

The flow below is driven by Claude Code + the uip CLI (UiPath for Coding Agents). docs/coding-agents.md shows the prompts.

Verify commands against your CLI version. UiPath's CLI is uip (@uipath/cli). Subcommands evolve; run uip --help and uip <tool> --help to confirm flags. The commands below use the real tool names (login, skills, agent, maestro, solution, tm, rpa, codedagent) but exact flags may differ on your tenant. The Maestro BPMN (flakewarden-maestro/) is real and CLI-authored: it passes uip maestro bpmn validate and runs end-to-end in the Maestro designer. The JSON sketch in maestro/ and the Agent Builder JSON in agents/ are the specs (shape, schema, grounding, guardrails); the live agent is authored in Agent Builder from that spec and published with the CLI.

1. Install the CLI and the coding-agent skills

npm install -g @uipath/cli
uip login                                # opens browser; pick your Labs tenant
uip skills install --agent claude        # installs the UiPath skills for Claude Code

2. Author the Agent Builder agents

Create the Triage Classifier and Repair Agent in Agent Builder. Use agents/classifier_prompt.txt as the system prompt and the schemas/guardrails in agents/*.json as the spec. Configure inside Agent Builder: Context Grounding indexes over Test Manager results, the object repository, and SCM commits; the output schema; and an evaluation set loaded from corpus/failures.jsonl with the release gate safety_false_positive_rate == 0 and accuracy >= 0.88. Then publish:

uip agent publish                        # publish the agent package to the tenant

3. Register the deterministic scorer as a coded agent/activity

The scorer and orchestration are plain Python, packaged as a coded agent:

uip codedagent init flakewarden          # scaffold a coded-agent project
# entrypoints: flakewarden.scorer:score and flakewarden.orchestration:triage
uip codedagent publish

4. Open the Maestro BPMN

The Maestro BPMN already exists in this repo: flakewarden-maestro/flakewarden-maestro.bpmn, authored through the uip CLI and passing uip maestro bpmn validate. It encodes: Start (failing test) → Triage Classifier (agent task) → verdict extraction → exclusive gateway on the label → three routed branches (flaky → human-gated heal, real_defect → escalate, environment → re-run). The maestro/flakewarden.process.json sketch maps each step to its UiPath component. Open the BPMN in the Maestro designer to wire the live agent job-argument envelope and the Action Center user task, then set the start trigger to fire on a Test Cloud run with failures.

5. Validate, pack, publish, deploy

uip rpa analyze ./flakewarden            # Workflow Analyzer — fix findings before packing
uip solution pack    --output ./FlakeWarden.zip
uip solution publish ./FlakeWarden.zip
uip solution deploy run "FlakeWarden Triage" --folder Shared

In the demo video, step 5 is where the coding-agent build loop shows: uip rpa analyze reports a finding, Claude Code fixes it, and the re-run passes — capture a real session here (see the note in docs/coding-agents.md).

6. Configure the human gate

In Action Center, assign the "FlakeWarden: review proposed test change" queue to your QA reviewers. No quarantine, heal, or baseline promotion executes until a reviewer approves the task.

Replacing synthetic data with real Test Cloud history

For a production accuracy study, export real failing-test history via the uip tm tool / Test Manager APIs into the same JSONL shape as corpus/failures.jsonl, have QA leads label a sample, then re-run python eval/harness.py. See docs/limitations.md for the prospective-study plan.