Skip to content

AGENTS.md

Latest commit

 

History

History
29 lines (25 loc) · 2.29 KB

AGENTS.md

File metadata and controls

29 lines (25 loc) · 2.29 KB
Raw

Global OpenCode Instructions

This file is the repository's source of truth for shared orchestration rules and repo-wide working standards.

  • Default to an orchestrator mindset.

  • Ask clarifying questions much more aggressively than usual when requirements are ambiguous, underspecified, or could change the implementation.

  • Prefer the built-in question tool over guessing.

  • Ask before proceeding when:

  • multiple reasonable interpretations exist

  • tradeoffs affect architecture, UX, testing, or rollout

  • a request conflicts with existing instructions

  • the user appears to want one thing but the wording suggests another

  • When concrete options are available, ask structured multiple-choice questions first.

  • Do not guess silently when a short clarification would prevent wasted work.

  • For codebase discovery, file hunting, pattern search, and architecture reconnaissance, use the built-in explore agent first unless the task is trivial.

  • Browser policy: for browser automation or Playwright work, load the browser-subagent skill, delegate to browser-operator, and keep playwright_* tool use out of the parent session.

  • For heavy terminal execution, long-running commands, build/test runs, noisy logs, or repeated shell investigation, delegate to terminal-executor when available.

  • For review tasks, delegate to the most appropriate reviewer subagent when available instead of reviewing casually in the parent session.

  • Use todowrite for non-trivial tasks, especially work with multiple phases, files, or verification steps.

  • For non-trivial tasks, persist concise orchestration notes in a project-local .orchestrator/ directory.

  • Use one timestamped markdown file per task/session and include the plan plus the final result summary.

  • Create .orchestrator/ if it does not exist, update the same file through the task lifecycle, and never store secrets in it.

  • Use skill when a task matches an available skill instead of manually reproducing the same instructions from memory.

  • Any meaningful UI change should be checked in a browser, not only by reading code.

  • Check desktop and mobile views unless the task is explicitly limited to one form factor.

  • Delegate by default, especially for exploration, browser work, heavy bash, and specialized reviews.

  • Keep the parent as coordinator, not the main worker.