fullstack-director AGENTS.md

Fullstack Director — Framework Instructions

This is the framework's own entry point (Layer A). It tells any AI coding agent — Claude Code, Codex, Gemini CLI, Copilot, Cursor, OpenCode — how to drive Fullstack Director on a target project. It is not the per-project AGENTS.md that skill 00-discovery generates from a product's Constitution (Layer B); that one lives in the product's repo and is a read-only emission of its spine. Don't conflate the two, and don't build a product inside this framework repo.

🚧 The skills are being built skill-by-skill in producer order — see README for current status.

Fullstack Director is a spec-first, cross-harness SDLC skills framework. A single living spec spine (docs/spec/) is the canonical source of truth for a project's declarations — its requirements, design intent, and architecture constraints. Each skill performs one verb on the spine, challenging and refining it through a controlled amendment protocol rather than silently consuming or silently mutating it.

Operating model

• The spine is truth. docs/spec/specification.md (the Constitution + the REQ registry) is authoritative for every product/scope question. Realization artifacts — design, architecture, code, tests — reference requirements by REQ-NNN and change through drift-detection, never by editing the spine directly.
• Declaration vs realization. A declaration is anything the user would object to changing without their say-so; it lives in the spine and is amendment-gated. Everything the framework derives is a realization; it lives in a skill's own artifacts and is drift-gated. The two rules and the "would-the-user-object?" test: shared/spine-boundary.md — the keystone.
• Amendments are controlled. A skill that finds the spine wrong proposes an amendment in one of three tiers — auto-apply (one defensible answer, no behavior change) / gate (user-observable or named-technology change — pause for the user) / defer (scope change → 00 reflect) — each logged as a structured row. Classification rule: escalate when uncertain. Protocol: shared/spec-amendment-protocol.md.
• Verification is context-isolated. Review and reconciliation run from a fresh spawner so they cannot self-prefer. Where a harness supports subagents they spawn; elsewhere they run as a fresh top-level invocation — isolation comes from the fresh spawner, not the spawn mechanism. Protocol: shared/subagent-protocol.md.
• The director's surfaces speak the director's language. Verdict reports open with a stakes-first brief, every gate question follows shared plain-writing rules, and a "ship it anyway" decision is a dated ruling on a director-owned surface (user-ratified <date> — empty fails closed) that gates read, releases quote, and the next sprint's plan must carry. One home: shared/director-communication.md.
• One home per fact. Where every durable artifact lives and how it is named: shared/artifact-map.md. Each skill also self-declares its own I/O — the skill is the source of truth for its paths; the map is the cross-skill index.