Skip to content

Latest commit

 

History

History
68 lines (60 loc) · 3.99 KB

AGENTS.md

File metadata and controls

68 lines (60 loc) · 3.99 KB

Agent instructions

This is a share-once demonstration of prompt-discovered UI tests. Help the user adapt it to their application with small, understandable changes. Read the target repository's own instructions before changing it.

Onboard a project

  1. Inspect startup commands, dependencies, authentication, optional extension builds, and existing validation commands. Do not guess credentials or replace a real application with a fixture.
  2. Run agentic-ui init --workspace <project>. Preserve existing configuration, secrets, and unrelated changes. Set the real target URL, credential names, extension path, repair paths, and appropriate prepare/validation argv arrays.
  3. Help the user supply an OpenRouter key in .env.agentic-ui or their shell. The default model is openrouter/z-ai/glm-5.3-flash. Never print secret values, overwrite the target's ordinary .env, or commit secret files.
  4. Start the real app and dependencies. Use host.docker.internal for host services from Docker and a reachable deployed URL remotely. Build the image and run doctor. Public recipe replay needs no model key.
  5. Discover one requested workflow, replay its saved recipe, and inspect the result and video. Report observed failures truthfully. Have the user review recipes before committing; commit nothing without authorization.
  6. Explain repair prerequisites: clean Git checkout, reviewed repair paths, and a project validation command. Publishing a draft PR requires an explicit request and GitHub access; creating a report alone does not require either.

Runtime and evidence

  • OpenHands owns Chromium and every browser interaction inside the container. OpenCode orchestrates discovery and repair analysis. Never add a host-browser execution path or another browser driver.
  • Default local Docker builds and runs to linux/arm64. Honor an explicit AGENTIC_UI_PLATFORM=linux/amd64 override when a user selects an AMD/x86 container.
  • Saved recipes replay deterministically without a model. Local and remote use the same runtime definition; multi-platform immutable images make each architecture-specific runtime reproducible.
  • Exercise real visible UI and real provider responses. Never invent expected data, titles, pages, accounts, or successful outcomes. Report provider failures.
  • Prefer stable semantic selectors and explicit visible assertions. Keep credentials as ${VARIABLE_NAME} references resolved by the browser process.
  • Preserve result JSON, failed-step details, screenshots, and continuous video. Quick recording is 720p/10 fps; demo is 1080p/60 fps with an 800 ms dwell between steps. Inspect the actual video rather than assuming recorder startup proves the artifact is usable.

Repair and publication

  • Read the failure artifact, reproduce it, and distinguish product defects from environment/provider/test failures. A passing reproduction does not justify inventing a defect or claiming it was fixed.
  • Make the smallest root-cause change inside configured repair paths. Do not weaken a recipe to conceal a product bug. Respect unrelated project changes.
  • The controller runs project validation independently. This demo does not automatically start a repaired candidate, so replaying the original URL cannot validate repaired source. State that limitation in the report.
  • Repair agents must not commit, push, create branches, or contact GitHub. The explicit host publish-pr command checks recorded state and publishes a draft PR when requested. Never automatically merge.
  • PR descriptions explain root cause, resulting behavior, actual validation, and uncertainty. Do not claim unexecuted checks passed.
  • OpenCode permissions are advisory. Mounted files are writable and shell commands are powerful; this is not isolation for hostile repositories. Profiles and artifacts can contain secrets and private data.

Keep onboarding concrete. Add project details where needed without introducing a universal framework or claiming support for unexercised integrations.