Navigation: Project README · Documentation home · Previous: Quick start · Next: Pi runtime and model-provider setup
All nodes appear under the Pi Agent category.
Finds Pi from an explicit path, PI_AGENT_EXECUTABLE, or PATH.
Outputs:
- status JSON;
- plain-language message.
Starts Pi in RPC mode, sends a prompt, collects streaming events, and waits for agent_settled.
Inputs include provider, model, project directory, executable, and timeout. An optional
workflow_json_or_path input lets the normal Pi Agent Prompt node recognize and reason
about the supplied workflow without requiring a pack-specific node. Provider credentials
remain in Pi's configuration.
When a workflow is supplied, ComfyUI-Pi injects only a compact structural digest. The full workflow is written to a temporary ComfyUI user-data context file and Pi is told to read it only when graph-level detail is actually needed. Matching node-pack guidance is loaded by the same lazy integration router used by sidebar chat.
Accepts a UI workflow or API prompt graph. Reports nodes, links, input files, model-like filenames, execution order, missing registered classes, issues, and recognized first-class integrations such as MiniMax H3 Director, WhatDreamsCost-ComfyUI, Scene Camera Action, and MiniMax H3 Turbo.
Returns a Boolean and a full report. Strict mode treats warnings as validation failure.
Performs conservative structural repairs only. It does not replace creative nodes or guess missing models.
Reads ComfyUI's registered model categories through folder_paths.
Ranks installed filenames for a requested family and preferred format. A result is a candidate, not proof of architecture compatibility.
Returns the bundled capability registry for image, edit, speech, music, and audio families.
Creates a typed package with prompt, modality, model family, references, preservation rules, and negative constraints.
Creates a versioned reference manifest for character sheets, voices, music, audio, mood boards, storyboards, environments, props, or visual bibles.
Creates a model-family-aware plan for Qwen Image, Qwen Image Edit, Krea 2, Krea 2 Edit, FLUX.2 Klein, or Z-Image. It records references, component requirements, format preference, speed, quality, and validation steps.
Specialized edit plan that emphasizes image-aware conditioning, reference roles, preservation rules, and GGUF companion validation.
Specialized Krea 2 edit plan that records Raw, Turbo, or Identity Edit behavior and the requirement for compatible edit LoRA and adapter nodes.
Detects the separately installed ComfyUI-MiniMaxH3-Director pack through live public node registration and normal custom-node discovery. Reports the four supported public Director node IDs, installed example workflows, source license, and integration profile.
Creates a Director-aware generation plan. auto uses FL2VA when no actual reference media is requested and ref2VA when image, video, or audio references are requested. It checks the H3 frame grid, trained-duration guidance, reference limits, prompt format, preview, Enhance Prompt, and retake options.
Creates a Director workflow from the installed upstream pack's own example workflow. It does not ship or fabricate a copy of the GPL workflow. When the upstream pack is missing, it returns a clear unavailable result instead of guessing schemas or frontend timeline state.
Recognizes the Director, Preview Override, Retake Stitch, and Enhance Prompt nodes and checks Director-specific requirements such as CLIPLoader type minimax, FL2VA/ref2VA checkpoint roles, video/audio VAE roles, serialized reference counts, and safe editing boundaries.
Read MiniMax H3 Director integration.
Detects the separately installed ComfyUI-scene-camera-action pack, its three public nodes, available staging presets, and the upstream scene-staging-builder skill. Detection is lazy and does not import the pack into ComfyUI-Pi at startup.
Creates a previz plan for staging, human/car acting, duration, and optional camera directing. It preserves the upstream 4–15 second Acting duration range and identifies whether the scene should be generated, loaded from a preset, or edited from existing SceneState JSON.
Creates a conservative SceneNode → ActingNode → DirectingNode base graph using only the public socket contract. It deliberately leaves frontend-managed scene editing, recorded actor motion, and camera-cut timeline state to the installed upstream widgets instead of fabricating unknown serialization.
Recognizes SceneNode, ActingNode, and DirectingNode; checks chain connections and actor settings; and validates SceneState block/group JSON when available. SceneState checks include IDs, numeric transforms, spawn-point fields, supported block/group types, and obvious ground-placement problems.
Read Scene Camera Action integration.
Detects ComfyUI-MiniMax-H3-Turbo, its MiniMaxH3TurboLoRA and MiniMaxH3TurboSampler nodes, and its installed example workflows.
Creates a Turbo-aware T2V, I2V, or first/last-frame plan with a minimum of four steps, simple scheduler guidance, LoRA strength, and the upstream low_vram sharpness/VRAM trade-off.
Creates from the installed upstream Turbo example workflow, preserving the official H3 joint video/audio pipeline. For I2V or FLF it attaches LoadImage nodes only by confirmed named MiniMaxH3ImageToVideo sockets and marks the placeholder images for replacement before execution. It does not fabricate the complete H3 graph when the validated upstream baseline is unavailable.
Checks that both Turbo nodes are present, the custom Turbo sampler feeds SamplerCustomAdvanced, the LoRA sits in the model path, and the H3 scheduler/joint-AV structure is appropriate. It warns about stock-sampler four-step audio problems, scheduler mismatches, and other unsafe substitutions instead of silently rewriting them.
Read MiniMax H3 Turbo integration.
Reports both lazily matched third-party integration context and the small set of ComfyUI-Pi task procedures selected for the current request. Context preview remains optional.
Shows which optional node-pack integrations match a message and/or workflow. With context preview off, only the tiny registry is consulted. With preview on, matching integration modules and their detailed skills are loaded lazily. Unrelated requests return no integration context.
Detects the separately installed WhatDreamsCost-ComfyUI pack using its public node IDs and normal custom-node discovery. Reports registered nodes and installed upstream example workflows.
Creates an LTX/WhatDreamsCost-aware plan for Director, custom-audio, two-stage FFLF, or three-stage FFLF workflows. It records safetensors/GGUF preference plus Prompt Relay, custom-audio, IC-LoRA, and Retake requirements.
Creates a workflow by copying the closest installed upstream example workflow. GGUF requests prefer the upstream GGUF example. The node does not fabricate the large LTX Director timeline serialization when the pack is unavailable.
Recognizes all public WhatDreamsCost node IDs, checks LTX Director-specific loader/timeline concerns, records the present features it can safely infer, and warns rather than guessing frontend-managed timeline state.
Read WhatDreamsCost-ComfyUI integration and Dynamic integration context.
Creates a plan for ACE-Step 1.5, Qwen3-TTS, or Stable Audio 3. It records mode, duration, language, reference audio, components, and workflow steps.
Creates a structured turnaround, expression, wardrobe, consistency, and output plan.
Creates a role-aware mood-board manifest with categories and source references.
Converts shot-list JSON into planned storyboard panels with stable scene and shot links.
Creates a draft story, series, character, world, location, continuity, visual, voice, music, sound, workflow, or production bible.
Creates a selective build plan for fresh, resume, incremental, repair, rebuild, validation, or dry-run compilation.
Creates a valid Fountain starter document from title, author, source text, and a default location.
Parses title fields, scenes, action, character cues, parentheticals, dialogue, transitions, sections, notes, and page breaks.
Creates scene records with locations, time of day, characters, summaries, and production placeholders.
Creates minimal, standard, or detailed starter coverage for every screenplay scene.
Creates a proportional stage plan and clearly states which stages are optional.
Creates the complete project folder structure, starter documents, Fountain draft when appropriate, bibles, manifests, and NLE handoff documents.
The generated project has separate top-level folders for the story, screenplay, production bibles, mood boards, reference sheets, storyboards, scene and shot plans, prompts, ComfyUI workflows, generated media, editorial media, NLE files, tutorials, quality control, delivery, and archives. Every directory includes a plain-language README.md.
Creates an original Pi skill folder containing SKILL.md and skill.json.
Compiles one or more workflows into a ComfyUI-native tutorial package.
Loads tutorial.json from a compiled tutorial directory.
Compares tutorial requirements with live registered node classes and model filenames.
Selects a stage by one-based number.
Checks the analysis issues stored for a stage.
A normal ComfyUI node used as a canvas instruction card in annotated workflows.
Writes Markdown atomically.
Creates a readable DOCX with headings, paragraphs, lists, and code-style paragraphs using only Python's standard library.
Parses and writes formatted JSON.
Creates a conservative Kdenlive/MLT project, portable timeline data, media map, profile, and assembly guides.
Simple output node for viewing or forwarding text.
Navigate: Project README · Documentation home · Previous: Quick start · Next: Pi runtime and model-provider setup