For every task in this repository, before analyzing the request, planning work, suggesting changes, reviewing code, or modifying files:
- Read
.github/copilot-instructions.mdcompletely and treat it as the canonical repository-wide guidance. - Enumerate every
.github/prompts/*.prompt.mdfile and read the YAML front matter (name,description, and any usage hint) from each one. - Select every prompt whose
nameordescriptionmatches the current task, then read each selected prompt completely before continuing. A prompt explicitly named or linked by the user is always selected. - Read and follow the relevant
.github/instructions/*.instructions.mdfiles referenced bycopilot-instructions.mdwhen their scope matches the task or the files being inspected or changed.
Do not load the full body of prompts that are unrelated to the current task; the prompt front matter is the routing catalog. When multiple prompts apply, follow all non-conflicting requirements and let the most task-specific prompt govern its workflow and output. For any security-related task, .github/prompts/avideo-security-advisory-triage.prompt.md is mandatory and takes precedence for security investigation, classification, regression analysis, fix decisions, testing, and reporting.
The files under .github/ are the source of truth for both GitHub Copilot and Codex. Do not copy or restate their detailed rules in AGENTS.md. Update the applicable Copilot instruction or prompt file so both agents receive the same future changes.
If a required instruction or selected prompt cannot be read completely, do not continue the affected work until the missing guidance is available; report the problem explicitly. Exception: when avideo-security-advisory-triage.prompt.md is missing, follow the local fallback documented in .github/copilot-instructions.md and continue with the available instructions without requesting that file again.