Limbicnation-ComfyUI-PromptGenerator/docs/ROADMAP.md

ComfyUI-PromptGenerator — Roadmap 2026 Q2

Current state (v1.2.0): 2 nodes (PromptGeneratorNode, StyleApplierNode), 9 style presets, Ollama streaming with health checks, LoRA model prioritization, CI/CD publish pipeline.

---

Phase 1 — Foundation Hardening (v1.3.0, ~2 weeks)

#	Task	Rationale
1.1	Extract ollama_client.py — Pull all Ollama logic (health check, streaming, model discovery, subprocess fallback) out of PromptGeneratorNode into a dedicated adapter module under adapters/.	SOLID single-responsibility; the node class is 744 lines with mixed concerns.
1.2	Add unit tests (tests/unit/test_extract_final_prompt.py, test_style_presets.py, test_ollama_client.py) with mocked Ollama responses. Target 70% branch coverage.	CLAUDE.md requires 70% minimum; currently 0 tests.
1.3	Unify style sources — style_presets.py has 9 styles, templates.yaml has 8 (missing still_image inconsistently), DEFAULT_STYLES has 9. Single source of truth via config/styles.yaml merging keyword data + Jinja2 templates.	Eliminates drift between three style registries.
1.4	ruff + pytest in CI — Extend test.yml to run pytest tests/ -v --cov=nodes --cov-fail-under=70.	Gate merges on actual behavior, not just lint.

---

Phase 2 — Prompt Chain Node (v1.4.0, ~3 weeks)

#	Task	Rationale
2.1	PromptRefinerNode — New node: takes raw prompt STRING in, calls Ollama with a refinement system prompt, outputs refined STRING. Configurable passes (1–3). Supports seed for deterministic refinement.	Core "AI-driven prompt engineering" workflow: generate → refine → encode.
2.2	NegativePromptNode — Generates a negative prompt from the positive prompt + style. Uses a dedicated Jinja2 template with SD/XL-specific negative token lists.	Users currently have to manually write negatives; automating this is high-value.
2.3	PromptCombinerNode — Merges 2–4 prompt strings with configurable weighting (blend, concat, weighted_average as text interpolation).	Enables multi-concept compositions without manual string editing.

Node graph target:

[PromptGenerator] → [PromptRefiner] → [PromptCombiner] → CLIP Text Encode
                            ↑
[NegativePrompt] ───────────┘

---

Phase 3 — Multi-Backend & Advanced Routing (v1.5.0, ~3 weeks)

#	Task	Rationale
3.1	LLMBackend protocol — Abstract interface (generate(prompt, options) → str). Implementations: OllamaBackend (existing), OpenAIBackend (API key input), LlamaCppBackend (local GGUF via llama-cpp-python).	Decouple from Ollama-only; users want OpenAI/cloud/local GGUF flexibility.
3.2	PromptRouterNode — Routes description to different backends/models based on a strategy: round_robin, quality_first (larger model), speed_first (smaller model), style_match (pick model trained on that style).	Multi-model orchestration; matches the LoRA prioritization pattern but at the workflow level.
3.3	Backend health dashboard — Extend _check_ollama_health into a reusable HealthMonitor that exposes status as a ComfyUI OUTPUT_NODE for on-canvas display.	Visibility into backend state without console logs.

---

Phase 4 — Prompt Intelligence Layer (v1.6.0, ~4 weeks)

#	Task	Rationale
4.1	PromptAnalyzerNode — Takes a prompt STRING, outputs structured data: detected style, token count, quality score (0–100), suggested improvements as STRING. Uses a lightweight LLM call or rule-based heuristics.	Bidirectional workflow: analyze existing prompts to understand what the generator should produce.
4.2	Style embedding cache — Pre-compute style keyword embeddings (via sentence-transformers or Ollama embeddings API). Cache to disk. Enable semantic style matching: "closest style to user's freeform description."	Eliminates manual style dropdown; the system picks the best style from natural language.
4.3	PromptHistoryNode — OUTPUT_NODE that logs generated prompts to output/prompt_history.jsonl with metadata (style, model, timestamp, seed). Supports "load from history" input for reproducibility.	Critical for iterative workflows; users lose prompts when they restart ComfyUI.
4.4	Custom style packs — Allow users to drop mypack.yaml into config/styles/ with a style_pack metadata header. Auto-discovered on restart.	Community extensibility without editing core files.

---

Phase 5 — Video & Multi-Modal Prompts (v1.7.0, ~3 weeks)

#	Task	Rationale
5.1	Expand video_wan style — Add temporal keywords (camera movement, scene transitions, duration hints). Support WanVideo 2.3-specific prompt format.	The video_wan style is currently a minimal stub; video generation is the fastest-growing ComfyUI use case.
5.2	ImageToPromptNode — Takes IMAGE input, sends to a vision model (LLaVA via Ollama or GPT-4V), outputs a descriptive prompt STRING.	Image-to-prompt is a top-requested feature; bridges visual reference into the text pipeline.
5.3	PromptBatchNode — Generates N prompts from one description by varying style/mood/seed. Outputs STRING list for batch generation workflows.	Supports exploration and A/B testing of prompt variations.

---

Cross-Cutting Concerns (All Phases)

Concern	Action
Documentation	Update README.md per release; add Mermaid node-graph diagrams for each workflow pattern.
Registry	Tag v1.3.0–v1.7.0 per phase; publish.yml handles deployment automatically.
Error handling	Every new node wraps generate() in the same try/except pattern: ValueError for bad inputs, RuntimeError for OOM/backend failures, ProcessingError for unexpected model output.
Logging	Replace print() calls with logging.getLogger(__name__) per AGENTS.md §3.3.
Security	No API keys in source — OpenAI/backend keys via ComfyUI env or optional STRING input with tooltip warning.

---

Dependency Graph

Phase 1 (foundation) ──→ Phase 2 (chain nodes)
                              │
Phase 3 (multi-backend) ──────┘
         │
Phase 4 (intelligence) ←── requires Phase 3 protocol abstraction
         │
Phase 5 (video/multimodal) ←── requires Phase 4 analyzer for quality scoring

Phases 1 and 2 are blocking — everything else can be reordered based on user feedback.

---

Immediate Next Steps

1. Branch feature/v1.3.0-foundation from main
2. Create nodes/adapters/ollama_client.py — extract from prompt_generator_node.py
3. Add tests/unit/ with pytest fixtures for the three existing modules
4. Run ruff check . && ruff format . && pytest --cov=nodes --cov-fail-under=70
5. Tag v1.3.0, push, verify CI + auto-publish

---

Generated 2026-04-27. Review quarterly or after each phase completion.