Classifies the active phase against 8 dimensions (ambiguity, UI, AI, architecture, security, core-logic, CRUD, milestone-boundary).
Composes the exact ordered list of gsd-* commands across 4 tiers — baseline + pre-plan + pre-verify + (optional) milestone boundary.
Embeds agent-team orchestration into each step — when to run a single agent, when to fan out in parallel, when to invoke GAN-style adversarial loops, when to bring in external AI via Codex.
Bridges GSD's gsd-ui-phase (which only produces UI-SPEC.md contracts) with the user's actual front-end 5-skill design combo (ux-principles + prototyping-ui-directions + taste-skill + anchor-prototype-wave + luxury-editorial-site-builder).
Prints the plan for user confirmation before executing, then dispatches commands serially. Stops on first failure, reports the failed command + artifact + next-action exactly.

This is a meta-orchestrator. It does NOT do work itself — it picks which GSD commands and agent patterns do the work.

Execution mode: SKILL-direct only (2026-05-29 user lock) — GSD pipeline orchestrator does NOT migrate to Workflow tool workflow-spec mode. Reason: 33 gsd-* agents already orchestrate well via SKILL main thread + slash commands; the multi-cycle checkpoint patterns (gsd-debug-session-manager), human-in-the-loop pauses (gsd-ui-checker BLOCK/FLAG verdicts), and parallel-research-to-synthesis fan-ins are inherently NOT single-pass DAGs and would lose semantics if forced into workflow-spec. Continue dispatching via SKILL Skill(...) / Agent(...) / SlashCommand. See ~/.claude/CLAUDE.md §3.5 for migration scope decision.

<when_to_invoke>

Required invocation (auto):

gsd-pipeline-orchestrator