vibecomfy-executor

name	vibecomfy-executor
description	Arnold pipeline that classifies a user query, optionally researches via Hivemind/Banodoco, optionally edits a ComfyUI workflow, and always emits a final reply.

vibecomfy-executor

A small decision-graph pipeline for routing user queries through the right mix of research and action.

Topology

classify → research → implement → reply → halt

• classify — decides which downstream steps are needed.
• research — searches the Hivemind/Banodoco corpus when the plan asks for it.
• implement — produces an implementation artifact when the plan asks for it. If a workflow graph is supplied, this stage runs VibeComfy's agent-edit machinery and mutates the workflow.
• reply — always runs; synthesizes the final response.

Execution plan

The classifier outputs a JSON plan with three booleans:

{
  "research": true,
  "implement": true,
  "reply": true
}

• research — query asks about facts, settings, best practices, comparisons, etc.
• implement — query asks to create, edit, fix, build, or change something.
• reply — almost always true; the final textual answer.

CLI usage

python -m arnold run vibecomfy-executor --inputs query="What are the best Wan settings?"

To edit a workflow, pass a graph input that points to a ComfyUI UI JSON file (or an inline JSON object):

python -m arnold run vibecomfy-executor \
  --inputs 'query=Set the KSampler seed to 12345 and the steps to 30,graph=tests/fixtures/agent_edit/flat.json' \
  --plan-dir /tmp/qo-edit-run

The positional input_file argument maps to ctx.inputs["draft"] if provided; --inputs query=... is the recommended way to pass the user query.

Profiles

The pipeline ships a default profile in profiles/default.toml that maps each stage to a Hermes-backed DeepSeek subagent:

[profiles.default]
classify   = "hermes:deepseek:deepseek-v4-flash"
research   = "hermes:deepseek:deepseek-v4-pro"
implement  = "hermes:deepseek:deepseek-v4-pro"
reply      = "hermes:deepseek:deepseek-v4-pro"

At runtime the resolved profile is available in ctx.profile, and each step constructs the matching AIAgent via the canonical key pool. If no worker can be resolved, steps fall back to fast heuristics or placeholders so the pipeline can still be tested without a model backend.

To use a different profile:

python -m arnold run vibecomfy-executor \
  --profile @vibecomfy-executor:default \
  --inputs query="..."

Hivemind integration

The research step calls the public Banodoco PostgREST endpoint (https://ujlwuvkrxlvoswwkerdf.supabase.co/rest/v1) using the anonymous publishable key. It searches unified_feed distillations first, then falls back to messages/resources.

Workflow editing

When graph is provided and the plan selects implement, the pipeline calls vibecomfy.comfy_nodes.agent.edit.handle_agent_edit using the implement profile model as the batch-REPL client. This is the same edit path VibeComfy uses today, including the same render/apply/gate/emit stages.

The edited candidate graph is written to edited_graph.json, the change report to edit_report.json, and a human summary to implementation.md.

Test fixtures

Three example inputs live in tests/fixtures/:

Fixture	Workflow	Query
edit_flat_ksampler.json	tests/fixtures/agent_edit/flat.json	Set KSampler seed to 12345 and steps to 30.
edit_wan_t2v_size.json	ready_templates/sources/official/video/wan_t2v.json	Make the video 512x512 with 49 frames.
edit_qwen_image_scale.json	ready_templates/sources/official/edit/qwen_image_edit.json	Set ImageScaleToTotalPixels megapixels to 2.0 and SaveImage prefix to qwen-edited.

Run a fixture from the VibeComfy repo root with Arnold on PYTHONPATH:

PYTHONPATH=/path/to/arnold python -m arnold run vibecomfy-executor \
  --inputs query="Set the KSampler seed to 12345 and the steps to 30",graph=tests/fixtures/agent_edit/flat.json \
  --plan-dir /tmp/qo-flat-test

Outputs

Each stage writes artifacts under the plan directory:

Stage	Artifact(s)
classify	plan.json
research	research_summary.md, research_tool_calls.json
implement	implementation.md, edited_graph.json¹, edit_report.json¹
reply	reply.md

¹ Only when a graph input was provided and the plan selected implement.

The final reply text is also returned in the run result at state.reply.