Horizon assist

Unofficial Minecraft integration. Neither Mojang nor Microsoft approves or is associated with this project.

Publisher: LychD contributors · Contact: hexanomicon@proton.me.

An operator speaks a short instruction while playing; a separately controlled Horizon client performs one bounded action and reports what that client observed. This experimental attached profile develops that interaction before the managed inhabitant fixture is available. Playing with a companion turns that boundary into user stories, experience criteria and a staged playtest, including useful feedback and quiet idle behavior.

Execution contract

blockworld.horizon_assist@1 is a separate experimental score, not an implementation or migration of blockworld.bounded_mission@2. Its records cannot satisfy PlacementReceiptV1. The managed, online-mode, dedicated non-operator bot fixture and its server-authoritative receipts remain the proving target for the bounded mission.

Admission pins one explicitly supplied local control endpoint, expected client/world identity, finite permitted actions, coordinate and time bounds, and an external private workspace. One session exclusively owns that endpoint and permits only one active action. A human player's endpoint is not implicitly a companion. A companion needs its own client, identity and endpoint; the pinned Horizon client's authentication limits must be reported before connecting it.

An explicit singleplayer_lab binding may qualify the control mechanism in a disposable local world. It requires the pinned process/socket and exact world/dimension with no remote server; it supplies no second player or multiplayer evidence. A remote binding instead requires the exact remote client/server observations. The two modes must refuse each other's state.

The world field pins the client's reported cosmetic label. A fresh remote Horizon client can report an empty string; an explicitly pinned empty label is accepted as that observation, never replaced with an invented server world name. The exact server address, client process generation, own-client name, dimension and loaded/joined state remain required. These observations do not authenticate a server world or detect every server-side world replacement at the same address.

The first action passage is attributed transcript or explicit text, a closed typed command, validation, one serial GraphStep MapStep Map action, and a client-observed result. Unrecognized or ambiguous instructions refuse. Text never becomes an arbitrary control command, shell command, administrator command or MCP tool name. Speech capture, transcription and synthesis retain their own Echo boundary.

The adapter exposes status as a closed operational projection: client screen/dimension, creative mode, frame/tick/chunk counters and finite performance measurements, selected renderer identity/memory fields, and the pinned server's address/name/joined/loaded fields. Each exposed leaf keeps its expected scalar or fixed-size numeric shape; an allowed key cannot carry an arbitrary nested payload. Missing server data remains missing and malformed server identity cannot become a single-player null binding. The raw status reply's debug/clipboard, keyboard state, name tags and unrecognized fields never become returned observations, GraphStep Map state or events. Binding checks and explicit status queries use the same projection boundary.

The adapter checks current session authority immediately before sending. It records an attempt before the physical send, never automatically replays an action, and distinguishes refusal, client-reported completion, cancellation and unknown outcome. A lost reply or process crash does not prove that an effect failed. Process-local execution supplies no restart continuation, durable PhylacteryMemory VaultMemory Vault custody or server-authoritative world proof.

Treat each RunJob RecordJob Record as one attempt. Once any claim is fenced, this adapter permanently retires that RunJob Record identity, including a claim with a nonzero delivery sequence. It admits no successor worker claim. Resuming or replaying a failed claim would require a different profile with an explicit claim-aware handoff.

The local ownership record also remembers whether the client generation was interrupted or left unknown. Starting another controller or changing the socket path cannot reopen it. To admit a successor, prove that the prior client process generation ended or reconcile it explicitly outside this profile. The record prevents unsafe reuse; it cannot resume a RunJob Record, reconstruct a lost world outcome, or turn observations into placement receipts.

Cancellation fences new sends and uses a separate control connection to stop the owned client's active action and release movement. Closing a waiting socket alone does not stop Horizon. Cancellation completes only after the stop boundary is observed; otherwise the outcome remains unknown. Stopping cannot undo a block already placed or movement already performed.

Following and building

Following requires a separately bound, read-only position feed for the exact person being followed, finite duration and distance bounds, and stop-on-lost-target behavior. An unordered list of player positions cannot establish that identity. Building requires an admitted finite placement plan, plot and inventory limits, and honest per-action outcomes; a spoken request for a house is not a blueprint or proof of completion.

The first building qualification is blockworld.horizon_build_step@1: an explicitly admitted, immutable one-cell or three-cell plan of ordinary minecraft:oak_planks, with exact integer coordinates, plot, session, expected client labels/dimension, mode, expiry, placement and read budgets, and a canonical content digest. Each cell is one fresh native RunJob RecordJob Record. It shares the assist service's session authority and attempt budget; a plan grants no selection, inventory rearrangement, item creation, movement, administrator command or arbitrary material.

The initial plan read budget counts logical composition calls: eight per cell and at most 24 for three cells. A cell/inventory inspection performs one fresh state/status pair and its fixed query under a combined timeout. Together with the effect's binding checks, a successful cell can issue 26 read query lines; three cells can issue 78, plus admission reads. The adapter exposes physical query counts separately. Containment has its own finite priority path and is never refused because the plan used its read budget.

Preflight first establishes that the placement is eligible: a loaded exact-air target, plain stone immediately below, an idle healthy client in the admitted mode, and a target within the shared session's maximum distance and conservative eye reach. Ordinary planks must already be selected with enough held inventory; offhand and carried stacks must be empty.

The adapter then emits only place X Y Z. It omits --item because Horizon changes selection before some placement checks. Afterwards, a fresh cell/inventory/state read must match the intended material and expected delta: one consumed item in survival, or zero consumption in explicitly admitted creative mode. Only then may the cursor advance. Existing blocks refuse and are never credited as this bot's work. Any failed, changed, cancelled or unknown step stops the plan without automatic replay or repair.

Horizon's stock placement primitive has no atomic expected-before/material/support comparison. Its click is consumed on a later game tick, and its ok means any target state change. These preflight reads therefore refuse observed changed conditions; they cannot exclude every concurrent world change or interaction. The first live proof requires a dedicated controlled disposable fixture without concurrent human input or entities around the cell. Records retain client observations and partial effects honestly; they are not server receipts or an atomic placement guarantee. A shared-world house plan needs further qualification or a guarded upstream primitive. The explicit local assembly now registers this score only when supplied its shared build service. Portable native-Run tests and real synthetic-game socket tests exercise one and three cells, expected query counts, the last shared action slot, uncertain partial effects and cancellation. Live placement remains a separate qualification.

The supervisor admits a successor cell only after the preceding native delivery reaches DONE and its exact RunJob Record has a known build result. A report may have observed a cell before native settlement; cancellation in that gap fences the whole plan. Operator instructions and capture invalidate building authority synchronously, then settle the previous RunJob Record and effect before continuing. A missing or unknown result disarms the controller. No cell resumes after restart.

Perception and model proposals

The current action and behavior graphs use deterministic policy, without a planning model. Speech recognition and synthesis do not observe the game. A future Mind passage must receive explicitly attributed game observations; a model being available does not give it sight or action authority. This passage can use the existing AnimatorModel RuntimeModel Runtime, DispatcherTraffic ControllerTraffic Controller and Agent contracts without adding a second model runtime or changing the Core execution contract.

The first perception target is a bounded structured sample: pinned client generation, session, sample identity, observation interval, pose, health, held item and exact queried cells. Select only required fields; a raw diagnostic status document may include unrelated host data. Missing cells, unloaded terrain and unobserved players remain unknown. Exact cell queries must also respect the dimension's vertical bounds: a horizontally resident chunk does not establish valid height. Separate state and status replies are observations over an interval, not an atomic world snapshot or authenticated server proof.

A finite reasoning step leases one exact capability and returns a typed proposal tied to the provided sample. It receives no game effect tools, arbitrary command grammar or runtime lifecycle authority. Hostile world labels, chat and other game text remain data. Missing, old or changed observations require fresh observation rather than invented facts. Schema validity establishes shape only; the proposal still needs independent session, freshness, bounds, material and action admission. A model response never calls the effect port directly.

Qualification separates structured perception from pixel vision. A screenshot needs an owned private destination, bounded capture and image validation, including the renderer's actual pixel format. Seeing a captured image during an operator test does not establish native agent image materialization. Before any broader autonomy, test observe, propose, validate, act and observe again, including stale samples, unknown terrain, misleading game text, cancellation and a game reply that reports success without the expected change. Measure reasoning separately; planning does not belong in every idle or follow tick.

Behavior graphs and idle

The intended supervisor selects finite behavior scores from explicit operator policy: handle a spoken instruction, observe while idle, follow the bound target, or advance an admitted building plan. Each invocation pins its trigger, session, deadline, action budget and stop conditions. World events are observations, not instructions or authority. A missing or stale observation stops the dependent behavior; it cannot justify invented player identity or world facts.

Idle means bounded observation or waiting for an admitted event. It grants no exploration, building, inventory use or open-ended model loop. Effects during idle require a separately selected finite task. Follow and build yield to explicit stop and new operator instructions; replacing a task first settles or reconciles its active effect.

Scheduling coalesces repeated observations, uses bounded cadence and cooldowns, and avoids model calls when deterministic policy can wait or select a known command. At most one effect owns the client. Each tick or action receives a deadline and freshness bound, with finite retry policy for reads and no replay of uncertain effects. Logs distinguish time waiting, speech inference, reasoning, GraphStep MapStep Map execution and game response. Optimizations must preserve cancellation, authority and outcome semantics, and be supported by measurements of the exact runtime profile.

The first behavior cut uses separate blockworld.horizon_idle_tick@1 and blockworld.horizon_follow_tick@1 scores. Each tick is a fresh native RunJob RecordJob Record with a pinned typed trigger and at most one physical command, using the same assist policy and action budget as operator commands. Nodes never submit a nested RunJob Record. The process-local supervisor owns cadence; an in-memory timer is not a durable GraphStep Map wait or restart schedule.

An initial directive lasts at most 60 seconds and 60 ticks, at a cadence of at least one second. One session admits at most 20 directives by default, with an explicit ceiling of 100. Idle reads the bound bot and reports changes without movement, inventory use or model calls. Follow additionally reads the exact target client's own pose through a query-only port; that port has no action, stop or arbitrary-command operation. Both clients must have distinct pinned processes and endpoints, and explicit remote bindings to the same server. Integrated-host LAN and single-player associations are outside this first follow profile.

The engineering voice assembly explicitly opts its concrete target observer into owned read settlement. If a tick expires or is interrupted during an already started target read, the observer joins that read under its independent observation timeout, discards the late sample and propagates the original cancellation. Replacement waits for settlement; neither the sample nor cleanup extends effect authority. Whole-read timeout is twice the binding timeout, with bounded connection cleanup, so teardown can finish after the behavior deadline.

Default direct observation still retires on cancellation. In both modes, actual timeout, invalid identity, disconnect, malformed response or explicit closure permanently retires the observer. It never reopens itself, discovers another player or reuses a cancelled sample.

Follow uses a stop-distance deadband and segments of at most four blocks, with a game action timeout of at most one second. A target sample carries a local monotonic time, sequence and task generation. Stale samples, health loss, loading, changed world/dimension/server, excessive separation, elapsed task authority or exhausted budgets fence the task. Freshness and generation are checked again immediately before action bytes are sent. A deadline watcher cancels an action that outlives its sample or task; a command already handed off still requires containment.

One arbiter admits operator commands and ticks. It coalesces repeated tick requests into one pending request, leaving no catch-up backlog. When replacing a behavior, it invalidates that generation before yielding, settles the previous RunJob Record and owned effect, and only then admits a successor. Operator commands take priority and clear the prior behavior.

If an outcome is unknown, disarm the controller rather than advance the task. Explicit stop ends the session. The logs retain tick decisions, native RunJob Record events, action/observation counts and monotonic timing. Deterministic idle and follow use zero model calls.

The stock goto dialect stops within 1.5 blocks of its waypoint. Follow accounts for that radius when selecting a waypoint and requires a segment ceiling greater than 1.5 blocks; an already reached waypoint cannot serve as evidence of movement toward the target. A successful response remains a client report, including when pathfinding made no displacement.

A client-reported follow completion still needs a progress check. Read the bot again and compare its pose with the pose validated immediately before sending the command, using the tick's original target sample. Report moved only after at least 0.05 blocks of observed progress toward that sample, with a current in-bounds, healthy, idle client in the same world/dimension. This proves neither that the target stayed there nor that the command caused the displacement.

If there was no progress, stop the finite task instead of spending its remaining budget on repeated segments. Missing, stale or changed post-observation also stops the task, preserving the known client response and leaving movement unverified. Count this extra logical read. Any later operator instruction needs ordinary fresh admission; no automatic retry follows.

The explicit local assembly registers these finite behavior scores only when supplied the shared behavior service. Portable tests cover their native checkpoints and the supervisor's arbitration; live two-client following remains a separate qualification. The initial one-action score alone establishes neither following nor autonomous building. The separately supplied finite build service starts with one-cell and three-cell qualification before a house blueprint.

Qualification

The explicit checkout tools run as python -m scripts.horizon_assist for one typed/text action, or python -m scripts.blockworld_voice for a supplied WAV or an Enter-to-record session. Their help owns the exact option grammar. Normal application boot selects neither profile.

The private game profile binds binding, policy, workspace and nonempty checkout_roots. The workspace is disjoint from every declared checkout in both directions; its profile and evidence stay inside it. The speech entry additionally receives an exact local speech profile. Both profiles must refer to the same private evidence workspace. The operator pins the actual client PID/start ticks, socket, world/dimension, mode, finite bounds, action budget and expiry.

The initial command grammar supports state/stav, status, goto/choď X Y Z, look-at/pozri X Y Z, and say/povedz TEXT. Coordinates must be explicit finite numbers; unknown wording refuses. This parser uses no LLM and promises no free-form language understanding. stop, quit or SIGINT ends the voice session and fences future sends. A new explicit session is required afterward. Spoken stop supplements the independent keyboard/control stop; transcription quality is not a guarantee that an emergency word will be recognized.

The behavior-enabled voice entry additionally recognizes exact idle/čakaj/čekej and follow me/nasleduj ma/následuj mě phrases. A finite --behavior-seconds duration defaults to 30 seconds. Idle can observe the explicit single-player lab. Follow requires an additional --target-profile inside the same private workspace: a closed JSON object with only binding, using the same binding fields as the bot profile. It grants query access only. Both bindings must be remote with matching expected world labels, dimension and server address, and different clients and endpoints. Starting capture interrupts and settles a previous behavior before the microphone opens.

An optional --build-plan supplies a private JSON BuildPlan inside that same workspace. Duplicate keys, nonfinite numbers and oversized documents refuse. The exact phrase build approved plan or postav schválený plán admits those frozen bytes once; it cannot generate coordinates or materials from speech. The plan pins the existing session and expires within its authority; its finite duration is at most 60 seconds. The game policy must explicitly permit place, and the binding timeout must be at least six seconds for Horizon's fixed five-second placement task. place remains absent from the general typed/text action grammar. The voice source-file mode joins the finite build before closing; interactive capture interrupts it.

The voice entry privately retains the source WAV, attributed transcript, native events and a monotonic stage timing receipt. It presents the retained outcome in fixed Slovak/Czech/English wording, distinguishing task start, client report, refusal and unknown effect. Unchanged idle ticks keep their private evidence without repeated console announcements.

Read timing according to the work being measured. Pre-capture settlement has its own field. For a one-action command, the GraphStep MapStep Map stage includes native admission, game work and event custody. For a background behavior or build, it measures acceptance and scheduling only; each tick or cell records its own native events, elapsed time and process CPU. Reply generation may overlap that background work. Source-file mode joins the finite task before closing, but the utterance's GraphStep Map timing still cannot stand for task completion time.

Optional replies produce a WAV and generation facts with playback not_attempted. Explicit --play-replies, together with --reply-language, additionally invokes a fixed local player on those validated bounded bytes after capture has closed. Player exit is recorded separately as player_exited, not human hearing. Playback has its own timing. Stop joins the owned process group; a late process launch is adopted and retired without audio even after its work deadline. Required OS ownership settlement may exceed that deadline. Unproven player retirement ends the session and keeps capture closed. This helper is not native EchoVoiceVoice playback admission or a physical speaker qualification. The shared local player uses the host's ALSA default output unless pulse is explicitly selected at its helper boundary. There is no device fallback, routing change or volume adjustment.

An optional synthesis or settled player failure preserves the game result and records the failed reply stage without replay. Synthesis observations remain generation facts; later playback is a separate observation. Capture starts only after the person explicitly requests that finite recording, with active/off indicators. Continuous microphone listening, conversational barge-in and automatic music ducking are outside this profile. The shared recorder distinguishes a failed recording with proved group retirement from unknown ownership. The former permits a fresh explicit gesture; the latter reports unknown and fences this session before another capture or reply. Cancellation during final retirement still propagates after settlement. An uncertain recorder never produces a normal off indication.

Portable tests must cover command refusal, endpoint exclusivity, identity and bounds changes, cancel-before-send, stop during a waiting action, lost replies and absence of automatic replay. Live qualification uses a disposable world and a dedicated client. The operator verifies microphone capture, transcript meaning, audible playback, movement, interruption and later collaborative play. State of Work owns delivery evidence; this contract alone does not establish a working adapter.