ONI MCP Control

Reference routing

Keep this file loaded for every control task. Load only the reference needed for the current operation:

• Read references/world-editor.md before any world_editor map edit, off-screen framing, operation-file command, management edit, plan, blueprint, or batch. It contains the exact virtual-file protocol and examples. 中文版: references/world-editor.zh.md
• Read references/tool-reference.md when selecting among aggregate tools, resources, prompts, parameters, tool categories, or standard multi-tool workflows. It contains the detailed tables and quick reference. 中文版: references/tool-reference.zh.md

Do not duplicate those references into the working context unless the task requires them.

Control contract

Run a complete control loop:

1. Observe the smallest relevant live state.
2. Orient around hazards, reachability, dependencies, and the user's scope.
3. Decide on a bounded target state and verification method.
4. Act through a semantic aggregate tool or reviewed virtual-file edit.
5. Verify with an independent targeted read.

Never act from stale assumptions. Never report success from a write response alone.

For autonomous play, keep the game paused while reading, planning, or issuing commands. Resume only for a short bounded window, pause again, and verify. Direct user instructions override viewer suggestions; gameplay safety overrides speed.

Scope and authorization

• Distinguish advice from control. Read-only inspection is allowed for diagnosis; do not mutate the save unless the user authorizes gameplay action.
• Respect marked boundaries, reserved areas, and "do not move the view" or "do not code" instructions literally.
• Do not use sandbox/debug spawning or cheat resources during normal play.
• Printing Pod rewards default to items; do not select a new duplicant unless the user explicitly asks.
• For liquid-adjacent digs, trapped duplicants, broad rectangles, or irreversible edits, pause and inspect exact cells before acting.
• If livestream or OBS health fails during a live session, pause gameplay and restore the stream before continuing.

Public tool policy

Use the compact aggregate surface:

• colony_control: colony snapshots, diagnostics, management, research, schedules.
• dupes_control: duplicant details, priorities, skills, equipment, commands.
• read_control: resources, buildings, infrastructure, world maps, reusable areas.
• building_control: planning, materials, configuration, storage, automation, special buildings.
• orders_control: dig, mop, sweep, disinfect, cancel, harvest, deconstruct, capture, conduit cuts.
• game_control: pause, speed, save, DLC, state, supported sandbox actions.
• navigation_control: camera, overlays, and screenshots only.
• search_control: semantic world/object search and discovery execution.
• server_control: discovery, diagnostics, and batching.
• world_editor: virtual-file reads and edits.

coordinate_control is not public. Ordinary aggregate tools reject raw coordinates. Use semantic areaId/query/plan inputs, or exact coordinates inside supported world_editor map and typed-operation files.

For utilities, cut the exact layer with:

orders_control domain=designation action=cut_conduits type=wire|liquid|gas|solid|logic|travel_tube ... confirm=true

For a normal building, prefer orders_control domain=designation action=deconstruct id=<instanceId> confirm=true. If a cell contains multiple candidates, list or dry-run first.

Mode selection

User intent	First move	Result
Diagnose a problem	Minimal snapshot plus one targeted diagnostic read	Ranked issues and evidence
Ask what to do	Catalog guide plus bounded current-state reads	Short plan with verification
Execute an authorized action	Dry-run or preflight, then confirm	Applied result plus verification
Monitor	Minimal delta snapshot with narrow watches	State change only

If execution fails, return to diagnosis. Do not repeat the same zero-effect call.

Observation discipline

Start compact:

colony_control domain=snapshot action=get profile=minimal delta=true watch=stress,food_kcal,red_alert,alerts

Escalate only the flagged domain:

• food: read_control domain=resources action=food
• power: read_control domain=infrastructure action=power_summary
• rooms: read_control domain=infrastructure action=rooms
• heat: read_control domain=world action=thermal_overheat_risk
• duplicant health/pathing: dupes_control domain=info action=status_check
• exact terrain or utilities: virtual map or read_control domain=world action=area_snapshot|text_map

Do not call the same read with identical arguments twice in a row. State what is missing and switch to a narrower read or discovery call.

Mandatory glyph lookup

Before interpreting a map glyph, generating a glyph from a building/element/entity name, or reading connection/overlay glyphs, announce in commentary that this skill requires an authoritative glyph lookup. Then call:

search_control domain=glyphs queries=["<all unknown glyphs or names>"] direction=auto

• Batch every unknown glyph/name from the current map in one call; do not issue one lookup per cell.
• Do not guess from Chinese characters, an old legend, or memory. Unknown results with count=0 remain unknown.
• Use view=<overlay> to disambiguate context-sensitive overlay meanings.
• Reuse results already queried in the same runtime and same turn.

Examples:

search_control domain=glyphs queries=["砖","?","┼"] direction=code_to_meaning matchMode=auto view=logic
search_control domain=glyphs queries=["砖块","氧气","研究站"] direction=meaning_to_code matchMode=auto
search_control domain=glyphs queries=["零","■","晒","美","微","低","枯"] direction=auto perQueryLimit=20

Spatial and camera discipline

• Prefer semantic search, virtual-file maps, area_snapshot, or text_map for exact decisions.
• Use navigation_control only when the user asks to look somewhere or a visual overlay/screenshot materially helps.
• Do not move or focus the camera merely to edit an off-screen rectangle. Use the non-camera framing workflow in references/world-editor.md.
• After an intentional camera move, call navigation_control action=get_view when exact confirmation matters.
• Use navigation_control action=switch_view screenshot=true for a visual overlay capture; use view=none for a normal screenshot.
• Screenshots are asynchronous. Honor readyAfterFrames before fetching the HTTP URL. Prefer coordinate_screenshot when later reasoning references cells.
• Structured maps, not screenshots, are authoritative for digging, building, piping, wiring, and deconstruction coordinates.

Build and order discipline

For construction:

1. Resolve the prefab with building_control domain=planning action=search_defs.
2. Resolve valid materials with building_control domain=planning action=materials.
3. Check support, obstruction, and placement constraints with placement_candidates or a map-edit dry-run.
4. Use semantic areaId/plan placement, or exact map tokens through world_editor.
5. Re-read the target area.

For dig, sweep, mop, disinfect, cancel, harvest, capture, deconstruct, and utility cuts, use orders_control or /active/ops/orders.md; never simulate UI clicks.

Use sweep only for debris/pickupables. Use mop for liquid cells and spills.

Sandbox and instant-build behavior

Every world_editor call runs with instantBuild=false by default, even when global debug instant-build is enabled. This default creates ordinary blueprints and uses normal materials. Never request instant build implicitly.

• Read completed building parameters through /active/buildings/index.md, then open the stable /active/buildings/instances/<prefab>-<InstanceID>.md file. Edit only its canonical lines.
• instantBuild=true requires allowSandbox=true confirm=true and applies only to that call.
• Sandbox writes use world_editor command=sandbox allowSandbox=true confirm=true plus the narrow capability: allowTerrainMutation, allowEntitySpawn, allowDestroy, or allowForce.
• Keep sandboxMaxCells small; it defaults to 100 and cannot exceed 1000.
• Never widen these permissions inside a batch child.

Execution gates

• Preview risky or exact edits first: dryRun=true and no confirmation.
• Execute only with dryRun=false confirm=true after reviewing the translated plan.
• server_control domain=batch action=call_many supports up to 20 items. Use requireAllValid=true and stopOnError=true for coordinated writes.
• Do not batch dependent reads whose later arguments require an earlier result.
• Keep response payloads compact: responseMode=summary normally, errors for retry loops, full only for child details.
• A successful request can still be partial. Inspect changedCells, applied, failed, partial, remainingCells, and per-item results.

Verification patterns

Choose verification before acting:

Action	Verify with
Map build/dig/order	Re-read the exact map rectangle or matching layer
Power change	power_summary plus exact power_ports if disconnected
Schedule or policy	Re-list the same management object
Duplicant update	Re-read the duplicant or priority/skill list
Building configuration	Re-list configuration for that building ID
Camera/overlay	get_view or screenshot after readyAfterFrames

Verify the effect, not merely the existence of a chore or blueprint, unless the user explicitly says the designation itself is the test target.

Error recovery

Schema or tool mismatch

1. Read the exact error.
2. Search the current catalog: server_control domain=catalog action=search query=<intent> detail=brief.
3. Follow the returned schema; do not invent a nearby tool name.

Zero-effect or stale write

1. Do not repeat it.
2. Re-read the smallest target state.
3. Resolve changed IDs, coordinates, reachability, materials, or support.
4. Submit a corrected dry-run.

Map preflight mismatch

Re-read and regenerate the patch. For explicit rectangles, use the three-header protocol in references/world-editor.md; do not fall back to guessed viewport offsets.

Timeout

Check task state if a task ID was returned. Otherwise narrow the area, limit, or response detail before retrying.

Confirmation rejection

Add confirm=true only after the user has authorized the mutation and the dry-run is acceptable.

Risk checklist

Before a medium or dangerous action, confirm:

• Current target state was read.
• User authorization covers the mutation.
• Exact IDs, world, rectangle, support, and utility layer are correct.
• Broad or hazardous work was dry-run.
• confirm=true is present only on the execution call.
• A targeted verification read is ready.
• Complex spatial work is paused.

Efficiency rules

• Cache catalog results for the loaded runtime, but invalidate assumptions after a reload.
• Invalidate related state after every write/execute call.
• Reuse bounded areaId handles for repeated reads and actions; forget temporary custom areas afterward.
• Use one-shot minimal snapshots instead of assembling many broad reads.
• Batch independent reads; serialize dependent calls.
• Preserve low token use with compact details, narrow rectangles, and RLE only for read-only scans. Use expanded edit format for patches.