cmd/

Command aclnegotiate shows multiagent NEGOTIATION (pkg/acl, Contract Net): an initiator calls for proposals on a task, agents bid or refuse, and the best bid is awarded — the conversation printed performative by performative.

Command aicam streams an AI video source.

Command aicards renders a glyph "card" message + a truecolor effect via pkg/terminal.Frame.

Command aidraw renders to the terminal via pkg/scene.

Command aiimg renders a raster image into the terminal via the repo's own internal/codec/babe.ImageToFrame.

Command anchordemo is a hand-made bridge to the MDLM world-model paper (Patronus AI, 2026): it shows, on our symbolic grid, why ANY-ORDER decoding with ANCHORS beats strict LEFT-TO-RIGHT.

Command anyorderdemo shows the MDLM "any-order with anchors" idea on the tangram address book: to identify a figure, reveal the most informative cell next (any-order) instead of scanning left-to-right.

Command areciboframe renders the full Lingua-Cosmica plaque (pkg/arecibo): one frame that teaches its own alphabet.

Command arena runs an "origami Warcraft" skirmish: two factions of tangram units, each driven by the reference Commander, charge across an isometric terrain board.

Command arenaai runs the origami-Warcraft skirmish with each side commanded by a tvcp-ai brain (game "rpg"): by default the reference commander on both, or a live model per side via -brain0 / -brain1.

Command arenacall is the project's "three bodies into one core" demo: a live arena match (the agent-world) is rendered as marks pseudo-images and streamed through pkg/semvideo — the very codec TVCP uses for a semantic video call — so a spectator WATCHES THE WAR OF TWO NEURAL NETS AS A FEW-HUNDRED-BYTES/S CALL that survives packet loss.

Command arenacast broadcasts an arena match as a loss-tolerant semantic stream and renders the SPECTATOR's view: keyframe+delta packets over Reed-Solomon are dropped and corrupted on a lossy channel, the spectator freezes on loss and resyncs at the next keyframe.

Command arenalive is the capstone: a LIVE war of two models, broadcast over a lossy channel to a spectator.

Command arenammo is the capstone of the origami-MMORPG: armies are SUMMONED from tangram figures, each side is commanded by a LIVE tvcp-ai brain (-brain0/-brain1), the match is broadcast as RS packets over real UDP to several networked spectators (with loss), and one spectator's reconstructed view is rendered to a GIF.

Command arenaplan pits the lookahead Planner (blue) against the reactive RefCommander (red) on a symmetric board with a central ridge, prints the tick-by-tick score, and renders a top-down GIF (via pkg/arenacast) with a survivor bar.

Command arenaserver is a persistent tick-server with multiple networked spectators: it runs an arena match and broadcasts each tick as an RS-protected packet over UDP (real localhost loopback) to several spectator sockets, dropping ~15% of sends per viewer.

Command arenasummon demonstrates crafting: a tangram figure is classified and, if it names a creature in the bestiary, a unit is summoned into the arena - "the address is the summoning spell".

Command arenatale plays an arena match, distils it into a few-byte chronicle (pkg/chronicle), and retells that byte-log as prose — the project's chain number -> image -> world -> story, end to end.

Command autodemo shows the auto triangle/quadrant chooser (pkg/automode).

Command blazondemo renders example coats of arms (heraldic blazons) to PNGs, proving that a medieval visual-identity language compiles straight into the pseudo-image pipeline.

Command bmeaning demonstrates bidirectional B-frames of meaning (pkg/semvideo): a constant-shape marks "video" is coded as anchors (keyframe + P-frames) with B-frames in between, each predicted from the nearer anchor.

Command braindemo proves the open tvcp-ai/1 format end-to-end over real HTTP: it starts the reference brain on localhost, then (1) plays a full tic-tac-toe where X's moves come from the brain via the protocol, O is random; and (2) asks the brain to draw a scene and renders the returned draw-DSL.

Command brainserver serves the reference tvcp-ai/1 brain over HTTP, so any client (game, draw) can use it.

Command capsdemo shows tvcp-ai/1 capability negotiation (pkg/brain): a brain serves /v1/capabilities and /v1/decide on one endpoint; a client fetches the capabilities and checks support before sending.

Command conform checks whether a tvcp-ai brain satisfies the protocol contract.

Command connect4demo shows a SECOND game on the same session runner (pkg/session): Connect Four, a win/block tactician bot against a naive leftmost bot.

Command debate stages a judged AI debate (pkg/debate): two debaters argue a topic over several rounds and a judge scores each one.

Command dtncast streams an arena match over a DELAY-TOLERANT (deep-space) link (pkg/dtn over pkg/semvideo): the world's frames become bundles forwarded only during a few short contact passes, with a time-to-live.

Command foldcall sends an origami unfold of a tangram figure as a stream of per-piece fold angles (keyframe + deltas) over the Reed-Solomon channel, then corrupts a burst of wire bytes, recovers, and renders the reconstructed animation to a GIF - a pseudo-3D fold that travels in a few dozen bytes and survives loss, exactly like a semantic video call.

Command folddemo folds flat tangram7 figures into a pseudo-3D isometric relief: it writes a hero still and an animated GIF of the figure rising out of the plane, and prints the tiny FoldSpec that describes the whole animation.

Command glyphdemo renders the digitized alphabet chart, and compares the quadrant codec with the new diagonal/triangle codec on the same image.

Command glyphqrdemo shows the glyph-QR with Reed-Solomon error correction: encode a message into a glyph grid, damage several cells, and watch it decode back to the exact text — the self-correcting "crossword x QR" idea, applied to our glyph alphabet.

Command glyphqrscan demonstrates the optical glyph-QR: a framed code with finder marks in three corners decodes from ANY rotation.

Command imagine is the "protocol of shared imagination": one agent-world is broadcast ONCE as a semantic-video stream (pkg/arenacast + pkg/semvideo), and TWO independent receivers — with their own packet loss — each reconstruct it LOCALLY.

Command linguaframe builds a self-describing "Lingua Cosmica" frame for a tangram figure and then decodes it back using nothing but the bytes — the receiver learns to count, then learns the radix, the grid, the cells and the checksum, all from the stream itself (pkg/lincos).

Command orbit3d turns the pseudo-3D world into a turntable: it renders a relief voxel cat across a full 360-degree camera-yaw sweep (fold.RenderCam) and ships the camera path as a one-byte-per-frame deltastream - rotation costs nothing in the codec, it is just one more number in the record.

Command paint shows the on-device painter (pkg/painter): with no model and no network it reads a prompt into a pseudo-image grid, which the deterministic pseudo-diffusion renderer turns into a painterly picture.

Command planbrain serves the arena lookahead Planner as a tvcp-ai/1 endpoint, so any client can point -brain at it by URL.

Command pseudo3d is the synthesis of the three pseudo-3D paths in one scene and one stream.

Command pseudodemo renders the same scene ("a calm harbor at dawn") in every pseudo-image format, proving that an ordinary text model — one that can only emit JSON, no diffusion head — can paint pictures for the terminal.

Command ray3d renders a small CPU ray-traced scene — spheres on a checkerboard floor with Lambert/Blinn-Phong shading, hard shadows, a mirror and glass — into the terminal using truecolor glyph modes (no ASCII ramp).

Command rayagent puts the robots under a tvcp-ai/1 brain: whenever a robot is ready for a new destination the brain picks one of its stations, given the robot's position and status.

Command rayarena stages a 3-D "arena": a tvcp-ai brain authors a moving world (orbiting, breathing spheres) AND flies the camera, the whole evolving scene is broadcast as independent Reed-Solomon packets (one per tick), and a spectator reconstructs it over a lossy channel — rendering whatever arrives and freezing on loss.

Command rayask asks a brain to author a world from a prompt and path-traces it.

Command raybench evaluates a rayscene director (the reference brain, or a live model via -url): it asks for many scenes and reports how renderable, rich and varied they are — a quick objective read before a live session.

Command raycall is the semantic video call, end to end, over a real (loopback) UDP socket: it joins the pieces this project built separately —

Command raycamp runs a robot logistics yard — Part 1's "logistics in a camp".

Command rayclimate visualises the hexagram-cellular-automaton climate (block: hexca -> weather).

Command raycurator is a curator agent that searches the Q6 hypercube for the world a viewer loves MOST (Block K: curator = engagement × maze × agent).

Command raydebate stages adversarial co-direction (the "debate × director" combination): two brains each author a region for a prompt, the rayscene validator referees, and the richer valid region wins — plus a MERGED region that unions both directors' forms.

Command raydetect closes the semantic-video chain: a vision model's detections become a rayscene, which is then path-traced.

Command raydirect shows meaning-driven direction in both directions (Block C).

Command raydungeon plays the Q6 maze as a LIVING dungeon (Block E: alife × rogue).

Command rayexplore is the experience: you walk, in the terminal, through a 3-D world the AI director dreams up and extends on the fly.

Command rayface renders a meeting INSIDE the world: participant avatars whose faces are drawn from keypoints (raydir.AvatarFace — the landmarks internal/avatar sends ~35 kbps over the wire) stand together in a shared world and look at you.

Command rayfilm flies a cinematic camera through a brain-authored world and prints a contact sheet of the shots.

Command rayfleet watches a fleet of machines the way info150's triage4-drive watches a driver: it ingests per-unit signal readings, fuses them into a severity, tracks each unit against its own baseline, and reacts — printing short operator cues (text) and rendering the fleet in 3-D with an alarm glow (graphics), plus the relationship graph of shared faults.

Command rayfx activates three renderers that shipped in pkg/raytrace fully tested but wired to no command (the "dormant code" of the audit): photon-mapped caustics, adaptive sampling, and temporal reprojection.

Command raygallery browses a directory of saved worlds (.rwld) and recordings (.rrec), printing a summary of each and how to open it — a little gallery of shareable worlds and walks, each a few KB of meaning (not pixels).

Command raygates runs the sim-to-real conformance gates — Part 1's variant E. The same demo logistics scenario is assessed twice: once as clean "sim" readings and once perturbed by a sim-to-real gap (bias + noise).

Command raygather is a meeting INSIDE a shared world (Block A: metaverse): N participants, each an avatar with a face (raydir.AvatarFace from keypoints), sit in a circle in a hexagram-authored world and look at one another.

Command rayhard shows the integrator "renderer brain" (block L): on a caustic dream scene where the unidirectional path tracer is noisy, RenderBest probes the difficulty and switches to the bidirectional path tracer — BDPT, shipped and tested but previously reachable only from the ray3d demo.

Command rayimagine reads a picture and renders the 3-D world it implies — the project's idea in reverse: meaning extracted FROM pixels.

Command raylife runs a small artificial-life world that lives WITHOUT a player (Block B: alife).

Command raymandala activates raydir.Mandala — dihedral (D_n) replication of a motif — which shipped with tests but no command.

Command raymeet is the shared-world experience for a group: several people (and AIs) "call in" and walk the SAME 3-D world together, in the terminal, each seeing everyone else's avatar and able to talk.

Command raymetaverse is a real networked meeting INSIDE a shared world (Block F: metaverse over the wire).

Command raypipe shows the closed triangle end to end, offline and without a server: an AI brain authors a world from a prompt (2 — tvcp-ai/1), the world crosses the "wire" as MEANING — a compact SceneSpec, exactly the bytes raymeet ships over UDP (1 — the network), and is rebuilt and ray-traced locally (3 — the renderer).

Command rayplay replays a recorded shared-world session (see raymeet -record): it reconstructs the world, the walkers and the chat over time from a tiny file of timestamped events — meaning, not pixels — and plays it back in the terminal from a participant's point of view.

Command rayquest plays the Q6 hypercube as a roguelike (Block D).

Command rayquestai is an agent that PLAYS the Q6 roguelike (Block H: agent × maze).

Command rayread points the inverse reader at the real world (Block G: reader × vision).

Command rayreading casts an I-Ching reading and renders it as a world transition: the primary hexagram's world dissolves into voxel blocks and the relating hexagram's world forms from them (raydir.Reading + raytrace.Materialize) — the oracle's narrative arc A -> B made into a moving image, and a deeper bridge to pro2 than the static Gray walk.

Command rayscore writes the world's adaptive soundtrack to a WAV — activating raydir.ScoreWAV, which shipped with tests but no command.

Command rayseasons runs a living world through the turning year (Block L: seasons).

Command raysound gives the Q6 world a voice (Block J: sonification).

Command rayspectate renders the raymeet shared world populated like an MMORPG: several player avatars and a swarm of patrolling robots in one world, shown both from a spectator's overhead view and through one player's eyes — the same scene raymeet builds (AvatarSpheres over World.SceneWith), so it is the multiplayer world of Part 1, with the robots of block H sharing it.

Command raystream is a call where the wire carries scene DELTAS, not pixels: a sphere orbits a small world over many frames, each frame sent as a rayscene P-frame (raydir.SceneStream), and the receiver reconstructs and ray-traces it locally.

Command rayterrain activates the Delaunay terrain (raytrace.ScatterTerrain / DelaunayMesh / Delaunay) — shipped with tests but reachable from no command.

Command raytour renders the Gray-code grand tour of all 64 hexagrams as cinema (Block I: tour of every world).

Command rayview is an interactive terminal navigator for the ray tracer.

Command rayvoice shows voice prosody steering the world's mood (the audio/reactive -> mood combination): three synthetic voice profiles — quiet, loud-and-steady, loud-and-swinging — are fed to a world's mood sense, and the world authored under the induced mood is rendered.

Command rayvoxel renders a voxel-space landscape — the blocky height-field look the dream hackers describe as how dreams form ("Воксельная графика в сновидениях").

Command raywatch is the camera observer of Part 1 (variant A): a camera frame of a machine goes through the same engine as a telemetry stream — visual signatures (heat, occlusion, glare, blur) are extracted, assessed, and the verdict is composited back onto the frame as a status border and banner, with a bloom/dream pass so the hot motor glows.

Command rayworld flies a ray-traced camera with a tvcp-ai brain and renders the result — neural directives -> camera -> real 3-D. By default the built-in reference brain drives it via game:"world"; -brain URL points at a live model (Ollama / OpenAI / Anthropic adapter); -ref uses the scripted director.

Command rayyard renders the shared world as a robot yard: several robots patrol loops with status beacons green->red, and the world is rendered over time into a contact sheet — proof that robots are first-class world entities.

Command reliefdemo renders tangram grid figures as 2.5D isometric reliefs (path 2): each glyph cell is extruded into a block, the diagonal glyphs becoming lit facets.

Command scenecall animates a 3-D scene of tangram pieces (orbit + spin + bob), ships it as a SceneAnim (keyframe + deltas) over the Reed-Solomon channel, corrupts a contiguous burst, recovers, and renders the reconstructed scene to a GIF - "a real game" that travels in a few dozen bytes per frame.

Command semcall is a REAL semantic video call: it sends a marks-spec stream (keyframes + semframe P-frames, each Reed-Solomon protected) over an actual localhost UDP socket under lossy conditions — packets are dropped and bytes are corrupted in flight.

Command semcodec measures the rate-distortion of MEANING.

Command semdemo shows semantic P-frames: a tiny "video" (a sun drifting across a sky grid) sent as one keyframe + per-frame deltas, and the byte savings vs sending every full frame.

Command semdistdemo prints the rate-distortion-of-MEANING benchmark (pkg/semdist): how many wire bytes buy how much surviving meaning (MarksIoU), as the keyframe interval is varied under packet loss and as Reed-Solomon parity is varied under byte corruption.

Command semvideo is the "closed loop": a semantic video call.

Command sessiondemo shows the unified game session (pkg/session): the same runner plays one Game (tic-tac-toe) between any Participants — here a tvcp-ai brain against another brain, then a brain against a scripted bot.

Command solantype shows two ways to set the myFront4Solan4 font: glyphs printed ADJACENT fuse their box borders into one continuous woven ornament/line, while a SPACE between them separates the letters into legible text.

Command streamdemo shows the streaming tvcp-ai transport (pkg/stream): a brain PUSHES a sequence of responses over Server-Sent Events on one connection.

Command tangram7demo renders the AUTHENTIC seven-piece geometric tangram: the canonical square (verified to tile) and an exploded tray of all seven pieces.

Command tangram7fig renders the authored tangram7 figures (square, cat, swan) to PNGs and a combined menagerie sheet, printing each figure's verification.

Command tangrambench is a spatial-reasoning benchmark for the tvcp-ai protocol.

Command tangramcraft is a terminal "gate" prototype: you build a figure from the glyph ring while the AI (a pose-invariant classifier) names what you are becoming and reads your address; on completion the gate "locks" onto a known creature.

Command tangramdemo proves the tangram idea end to end: ONE small set of sub-cell pieces (◤◥◣◢ ▀▄▌▐ ▞▚ █) composes into many figures, every figure is checked against the tangram rule (pieces meet edge to edge, never stacking), procedural creatures are grown from a seed, and one figure MORPHS into another as a stream of semantic P-frames (cat -> swan), saved as a GIF.

Command tangramgallery renders our figures as a framed "gallery wall": the catalog plus many procedural Creature(seed) figures, each in a wooden frame on a warm wall — the kind of geometric gallery the project naturally generates, where every framed piece is a figure with its own number/address.

Command tangramhard benchmarks silhouette-only ("hard") tangram solving: given just the figure's outline (no per-cell glyphs), recover the tiling.

Command tangramnum gives every tangram figure a unique NUMBER, the way the I Ching numbers its 64 hexagrams: read the figure's glyph grid row-major as a base-16 integer.

Command tangramsong is the third face of the project's "see, hear, summon" triad: it reads a tangram figure cell by cell (the same row-major reading that gives the figure its NUMBER), turns each cell into a note via pkg/sona, and writes BOTH a .wav you can hear and a .gif you can watch — the playhead sweeps the very cells you hear, while a piano-roll of the melody grows beneath.

Command tangramzoo renders the sixteen-figure tangram address book to a sheet and prints each figure's address and canonical id.

Command tournament runs a spectated tic-tac-toe LEAGUE (pkg/tournament over pkg/session): several participants — a minimax brain and a few scripted bots — play round-robin, home and away, and the standings print out.

Command unodemo runs two reference brains through a full game of UNO on the session runner (pkg/session) — the first game on the frame with HIDDEN INFORMATION (each player sees only its own hand) and CHANCE (a seed-shuffled deck).

Command vocabdemo shows the plural, localizable vocabulary (pkg/vocab): one canonical scene spoken in several languages.

Command watch is the project's headline scenario, fused into ONE experience: two neural-or-reference commanders fight an arena match; the match is broadcast as a lossy SEMANTIC VIDEO CALL (pkg/arenacast + pkg/semvideo); a spectator reconstructs it; the events become a byte CHRONICLE (pkg/chronicle) shown as a live caption; and the match's energy is SONIFIED (pkg/sona) to a melody.

Command wordledemo shows the reference brain playing WORDLE on the session runner (pkg/session) — the first SOLITAIRE game on the same frame that runs tic-tac-toe and Connect Four.

Command worldai drives the 3-D world through the tvcp-ai protocol: each tick a brain receives a Brief of the world (game "world") and returns directives {fold,rise,spin,camera}; the world physics applies them and renders the frame.

Command worldloop runs the closed loop: a Brain emits the next 51-byte world state each tick, the states ship as one deltastream over Reed-Solomon, a burst is injected and recovered, and every state is rendered into a 3-D frame.