idea: goal-oriented orchestrator plugin — Research → Design → Plan → Implement → Review loop

[Luis85]

Concept

A Claude Plugin that ships an orchestrator agent and a set of specialised subagents to drive a structured, goal-oriented resolution loop for a user-submitted issue or problem statement.

The orchestrator acts as the single point of contact and coordinates the full lifecycle — from scoping the problem through to presenting a verified solution — without requiring the user to drive each stage manually.

---

Motivation

Complex engineering or product problems benefit from sequential, evidence-backed stages. Today, users must manually chain together research, design, planning, and review. This plugin encodes that discipline as a first-class orchestrator loop so that any issue becomes the entry point into a rigorous, parallel-capable workflow.

---

Proposed Flow

User submits issue
        │
        ▼
┌──────────────────┐
│  Orchestrator    │  ← Grills user (structured intake via AskUserQuestion)
│  (Scope phase)   │    to fully define goals, constraints, acceptance criteria
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│  Research phase  │  ← Dispatches N parallel Researcher subagents
│                  │    Each collects evidence, prior art, risks, context
│                  │    Results merged + summarised by orchestrator
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│  Design phase    │  ← Orchestrator synthesises a solution design
│                  │    User reviews and approves (or rejects with feedback)
└────────┬─────────┘
         │ (approved)
         ▼
┌──────────────────┐
│  Plan phase      │  ← Planner subagent decomposes design into
│                  │    parallel-safe task DAG (tasks with deps, owners)
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│  Implement phase │  ← Orchestrator dispatches dedicated subagents
│                  │    per task/group; runs parallel where DAG allows
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│  Review phase    │  ← Review subagent validates output vs. acceptance criteria
│                  │    Orchestrator summarises session + presents solution
└──────────────────┘

---

Key Design Decisions (to be resolved in refinement)

#	Decision	Options	Default
D1	Scope intake format	Free text / structured form / EARS clauses	TBD
D2	Researcher subagent count	Fixed N / dynamic based on scope	TBD
D3	Design presentation	Inline chat / generated artifact / both	TBD
D4	Plan format	Task DAG markdown / JSON / tasks.md	TBD
D5	Parallel execution model	Worktrees / background agents / sequential fallback	TBD
D6	Review criteria source	Acceptance criteria from intake / auto-derived	TBD
D7	Plugin packaging	Standalone manifest / Specorator plugin group	TBD

---

Relationship to Existing Workflow

• Overlaps conceptually with Specorator Stages 1–9 but is packaged as a single-command plugin entry point.
• Distinct from /spec:start in that it is designed for issue-resolution (bounded, outcome-defined problems) rather than open-ended feature development.
• Reuses agent patterns from .claude/agents/ where applicable; does not duplicate logic unnecessarily.

---

Open Questions

• Should the orchestrator surface as a /plugin:goal-loop slash command or as a natural-language trigger?
• How does the plugin handle multi-file codebases vs. single-file scripts? Does it need worktree support?
• What is the minimum viable scope for a first plugin release (e.g., Research + Plan only)?
• Should the design review step be synchronous (blocks until user approves) or async (creates a PR for review)?
• How does this compose with the existing /issue:tackle track?

---

Acceptance Criteria (draft — to be refined)

• User can submit a natural-language problem statement and receive a scoped, parallelised resolution session.
• Orchestrator correctly gates on user approval before proceeding from Design → Plan and Plan → Implement.
• Researcher subagents run in parallel and their outputs are merged without duplication.
• Session ends with a human-readable summary of decisions made, evidence used, and artifacts produced.

---

Next Steps

1. Run a Discovery sprint (/discovery:start) or Requirements pass (/spec:requirements) to formalise scope.
2. Resolve Decision table above.
3. Prototype orchestrator loop as a Specorator skill before packaging as a plugin.

---

Status: Idea — needs refinement
Track: Plugin concept → /specorator:update or /spec:idea to proceed