vibecomfy-executor

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.