Agent platform roadmap

Extend comfyui-mcp orchestrator + panel with multi-provider agents, cross-provider image sourcing, and autonomous infra management.

Phase 0 — Per-workflow agent (done in comfyui-mcp)

WorkflowTargetStore + workflow_path injection on graph commands
MCP: panel_get_workflow_target, panel_set_workflow_target
Bridge: set_workflow_target / workflow_target sync
Panel follow-up: honor workflow_path in graph executors (workflow-target.md)

Phase 1 — Grok backend (P0)

GrokBackend implementing AgentBackend (OAuth via Grok Build / xAI CLI)
Add grok to BackendId, readiness probe, panel provider chip
Spike: auth token location, chat/tool wire format, MCP over HTTP parity

Phase 2 — Cross-provider concept images (P1)

fetch_concept_image — Grok Imagine / Google image gen → temp files
apply_reference_to_workflow — wire refs into LoadImage / Qwen edit / Krea img2img nodes
Skill: structure-map → Qwen/Krea pipeline

Phase 3 — AI Toolkit supervisor (P2)

Process supervisor for local AI Toolkit installs (start/stop, job queue, health)
MCP tools: toolkit_status, toolkit_run_job, toolkit_list_models

Phase 4 — RunComfy connector (P3)

API recon for RunComfy pod lifecycle
Tools: runcomfy_list_pods, runcomfy_sync_workflow, runcomfy_queue

Phase 5 — Multi-workflow projects (P4)

YAML project manifest (ordered workflows, shared assets)
run_workflow_pipeline — execute stages with per-workflow pins

Notes

Google AI Pro: Gemini agent backend exists; image gen is a separate provider layer, not a new agent backend.
Panel repo: UI for workflow picker binds to set_workflow_target; graph executors must respect workflow_path.
Do not replace comfyui-mcp-panel — it remains the UI; this repo owns orchestration + MCP tools.