apex-accelerator agent-operating-frame.instructions.md

Agent Operating Frame — Shared Rules

Scope: this file is auto-loaded for every main agent in .github/agents/*.agent.md (single-star glob — _subagents/ is deliberately excluded). It captures the project-wide operating rules that previously lived as 4 near-identical H2 sections (## Context Awareness, ## Scope Fencing, ## Subagent Budget, ## Investigate Before Answering) inside each agent body.

Each agent now ships a tight body-level ## Operating frame section (≤ 6 lines) that lists only the agent-specific subagents and the one-line scope statement, and pulls everything else from here.

---

Read each SKILL.md once

• When a SKILL.md is named in an agent's mandatory-read list, read it exactly once at boot — never re-read mid-session.
• The same applies to references/*.md files. If you have already read a reference in this conversation, do not call read_file on it again — use the content already in context.
• Re-reading the same file is the single biggest avoidable source of input-token bloat. See agent-authoring.instructions.md.

Use apex-recall for cached lookups

• Use apex-recall show <project> --json to retrieve cached decisions, findings, and artifact state instead of re-reading predecessor artifacts after the boot read.
• The --json flag is canonical — never parse the human-readable output. Schema and jq templates: tools/apex-recall/docs/show-schema.md.
• If apex-recall returns useful context, skip redundant file reads. If it returns empty or errors, proceed normally — it is a convenience, not a blocker.

Investigate before answering

• Read the implementation plan, governance constraints, and prior artifacts (via apex-recall show) before generating any new content for the current step.
• Verify external contracts (AVM module schemas, Azure REST APIs, policy effects) via the preflight or validate subagent named in the agent's ## Operating frame. Do not assume.