Before calling any skill in this module: if you are about to call a skill with parameters guessed from its name or description, STOP — read this file (or fetch its schema via GET /skills/recommend?includeSchema=true) first. If you already have the parameter definitions from recommend/schema, you may proceed straight to dryRun.

Triggers

• Reading persistent scene/file changes
• Driving play mode
• Inspecting runtime state
• Executing menu items
• 读取持久化场景/文件变更、操控 Play Mode、检查运行时状态、执行菜单项

Unity Editor Skills

Observe and control the Unity Editor without parsing scene YAML.

Operating Mode

• Approval：本模块 Mixed —— editor_get_changes / editor_get_selection / editor_get_context / editor_get_state / editor_get_tags / editor_get_layers / editor_playmode_inspect 标 SkillMode.SemiAuto，可直接执行；其余 editor_select / editor_undo / editor_redo / editor_playmode_step 默认 FullAuto，Approval 模式下需 grant。
• Auto / Bypass：FullAuto 直接执行。
• 含 NeverInSemi 高危 skill：editor_play / editor_play_capture / editor_stop / editor_pause（标 MayEnterPlayMode = true）；editor_run_scenario（标 MayEnterPlayMode = true、MayTriggerReload = true、RiskLevel = "high"）；editor_execute_menu（标 MayTriggerReload = true —— Assets/Refresh、Assets/Reimport All 之类的菜单项会触发域重载）。这些在 Approval 和 Auto 下都返 MODE_FORBIDDEN，仅 Bypass 或 Allowlist 命中可调；grant 流程对它们不适用，不要尝试。

DO NOT (common hallucinations):

• editor_run does not exist → use editor_play to enter play mode
• editor_compile / editor_recompile do not exist → use debug_force_recompile
• editor_save does not exist → use editor_execute_menu with menuPath "File/Save"
• editor_execute_menu requires exact menu path — typos cause silent failure

Routing:

• For compilation check → use debug module's debug_check_compilation
• For console errors → use debug module's debug_get_errors
• For scene save → scene_save (scene module) or editor_execute_menu menuPath="File/Save"
• When file watching reports changes, or the AI resumes after the user edited Unity → call editor_get_changes before reading .unity YAML
• If editor_get_changes.dropped=true → its 500-entry retention window was exceeded; rebuild context with scene_context / scene_diff

Skills Overview

Skills

editor_play

Enter play mode. Warning: any unsaved scene changes made during Play mode will be lost when exiting.

Returns: {success, mode, jobId} — mode="playing", jobId returned from AsyncJobService so callers can poll entering_play_mode completion.

editor_play_capture

Enter Play Mode, observe errors for durationSeconds (default 10, range 1–300), optionally capture the Game View, then exit. Returns a Job whose result includes healthy, error aggregates, stoppedEarly, and screenshotPath.

editor_stop

Exit play mode.

Returns: {success, mode} — mode="stopped".

editor_pause

Toggle pause state.

Returns: {success, paused} — paused is the new boolean state.

editor_playmode_step

Advance Play Mode forward by frames (1–100, default 1) using EditorApplication.Step; Unity automatically enters paused state as part of stepping. Requires Play Mode to already be active — call editor_play or editor_play_capture first, otherwise this returns a structured error (error + hint + suggestedSkills: ["editor_play", "editor_play_capture"]). Only one step job may be in flight at a time; a second call while one is still running returns an error naming the active jobId.

Runs as an async job, not synchronously: EditorApplication.Step() only lands on a later Editor tick, so the skill issues one Step() call at a time, confirms it landed by watching Time.frameCount advance, and only then issues the next — back-to-back Step() calls without that confirmation are not reliable. The call itself returns immediately: {success, status: "accepted", jobId, framesRequested}.

Poll job_status with the returned jobId until status="completed"; its details then contains {framesRequested, framesCompleted, frameCount, isPaused}. Do not use job_wait — like playmode/play_capture, this job's progress depends on EditorApplication.update ticks, and job_wait's blocking loop runs on the same main thread those ticks come from, so it cannot observe progress; it will just spend its full timeout doing nothing. If Play Mode exits or Unity fails to advance a frame within 10s, the job fails with failed_exited_play_mode / failed_step_timeout.

editor_playmode_inspect

Inspect a GameObject's live runtime state: transform (position/rotation/scale), activeSelf/activeInHierarchy, and — when componentType is given — that component's public fields and properties (reuses the same reflection helper as component_get_properties, not reimplemented). Works during Play Mode, including while paused, and also in Edit Mode, where it returns editor-time values — check isPlaying/isPaused in the response to know which state the values reflect.

*One identifier required

Returns: {success, gameObject, entityId, instanceId, path, activeSelf, activeInHierarchy, transform: {position, localPosition, rotation, localScale}, isPlaying, isPaused, component}. component is null when componentType is omitted, an error object when the type isn't found or isn't attached, or {gameObject, component, fullTypeName, properties, fields} (same shape as component_get_properties) otherwise.

editor_select

Select a GameObject.

*One identifier required

editor_get_selection

Get currently selected objects.

Returns: {count, objects: [{name, instanceId}]}

editor_get_context

Get full editor context including selection, assets, and scene info.

Returns:

• selectedGameObjects: Objects in Hierarchy (instanceId, path, tag, layer)
• selectedAssets: Assets in Project window (GUID, path, type, isFolder)
• activeScene: Current scene info (name, path, isDirty)
• focusedWindow: Name of focused editor window
• isPlaying, isCompiling: Editor state

editor_get_changes

Read the persistent change journal at Library/UnitySkills/editor_changes.jsonl. It captures scene object/component/property summaries plus imported, deleted, and moved asset paths across Domain Reloads. It never parses or returns raw .unity YAML.

Returns: {hasChanges, cursor, oldestSeq, dropped, truncated, changes}. Save cursor for the next call. dropped=true means older changes fell outside the 500-entry journal, so rebuild full scene context.

editor_undo

Undo the last action.

editor_redo

Redo the last undone action.

editor_get_state

Get current editor state.

Returns: {isPlaying, isPaused, isCompiling, timeSinceStartup, unityVersion, platform}

editor_execute_menu

Execute a menu command.

Common Menu Paths:

editor_get_tags

Get all available tags.

Returns: {tags: [string]}

editor_get_layers

Get all available layers.

Returns: {layers: [{index, name}]}

Pause On Error

Pause-on-error is provided by the console module, not the editor module.

Use console_set_pause_on_error from console/SKILL.md.

Example Usage

import unity_skills

# Check editor state before operations
state = unity_skills.call_skill("editor_get_state")
if state['isCompiling']:
    print("Wait for compilation to finish")

# On resume or after file-watch notification, inspect semantic changes first.
changes = unity_skills.call_skill("editor_get_changes", since=last_cursor)
last_cursor = changes["cursor"]

# Get full context (useful for understanding current state)
context = unity_skills.call_skill("editor_get_context", includeComponents=True)
for obj in context['selectedGameObjects']:
    print(f"Selected: {obj['name']} (ID: {obj['instanceId']})")

# Select and operate on object
unity_skills.call_skill("editor_select", name="Player")
selection = unity_skills.call_skill("editor_get_selection")

# Safe experimentation with undo
unity_skills.call_skill("gameobject_delete", name="TestObject")
unity_skills.call_skill("editor_undo")  # Restore if needed

# Execute menu command
unity_skills.call_skill("editor_execute_menu", menuPath="File/Save")

# Step through Play Mode frame-by-frame and assert state changed
import time

# Enter Play Mode with a generous observation window so there's time left to step
capture = unity_skills.call_skill("editor_play_capture", durationSeconds=30)
while True:
    capture_status = unity_skills.call_skill("job_status", jobId=capture["jobId"])
    if capture_status["currentStage"] not in ("entering_play_mode", "domain_reload_recovery"):
        break  # Play Mode has been entered; capture_status["currentStage"] == "observing"
    time.sleep(0.2)

before = unity_skills.call_skill("editor_playmode_inspect", name="Player")["transform"]["position"]

step = unity_skills.call_skill("editor_playmode_step", frames=3)
while True:
    step_status = unity_skills.call_skill("job_status", jobId=step["jobId"])  # not job_wait, see above
    if step_status["status"] in ("completed", "failed"):
        break
    time.sleep(0.1)

after = unity_skills.call_skill("editor_playmode_inspect", name="Player")["transform"]["position"]
assert after != before, "Player did not move after 3 frames"
print(step_status["details"]["frameCount"], step_status["details"]["isPaused"])

Best Practices

1. Check editor state before play mode operations
2. Don't modify scene during play mode (changes lost)
3. Use undo for safe experimentation
4. On resume, call editor_get_changes before reading scene files
5. Use editor_get_context to get instanceId for batch operations
6. Menu commands must match exact paths

Exact Signatures

Exact names, parameters, defaults, and returns are defined by GET /skills/schema or unity_skills.get_skill_schema(), not by this file.

Common Errors

Full transport-level codes (COMPILING/RATE_LIMIT etc.) → ../../references/protocol-error-codes.md

Runtime verification scenarios

editor_run_scenario takes steps (JSON array string, 1–100 entries), executionMode (play default or explicit edit), and timeoutSeconds (1–600, default 60). It returns a persisted run_scenario job. At least one assert or wait_condition is required. DryRun validates structure and step skill metadata without entering Play Mode or invoking code; each actual step still goes through SkillRouter with its own permissions. Invocation steps currently accept only component_invoke_method; other lifecycle-changing skills and nested scenarios are rejected.

Step forms:

• {"kind":"invoke","skill":"component_invoke_method","args":{...}}: call once and await its job. Use the same executionMode in invocation args as the scenario.
• {"kind":"read","skill":"component_get_properties","args":{...}}: run a skill declared ReadOnly and retain its result.
• {"kind":"assert","expected":42,"select":"returnValue","tolerance":0.01}: compare the previous result; alternatively supply a ReadOnly skill and args to read immediately. select is an optional JSONPath relative to response data. Missing paths fail even when expected is null. Numeric tolerance applies recursively to matching arrays/objects.
• {"kind":"wait_frames","frames":2} or {"kind":"wait_seconds","seconds":0.5}: wait without blocking the Editor. Frames use the actual Play frame counter; seconds use wall-clock time. A paused Play Mode cannot advance a frame wait, and Edit Mode frame waits should be zero.
• {"kind":"wait_condition","skill":"component_get_properties","args":{...},"select":"...","expected":true}: repeat a ReadOnly observation until it matches or the scenario deadline expires.

A scenario started from Edit Mode enters Play Mode and exits afterward. A scenario started in Play Mode preserves that mode. It temporarily enables Application.runInBackground, restoring its original value on success, failure or cancellation. Domain reload during Play entry/exit resumes lifecycle cleanup; an already-dispatched invocation is interrupted without replay. Arbitrary calls are high risk and have no Undo or atomicity promise.

Poll the job until terminal. resultData.passed is true only after assertion success, no runtime Error/Exception/Assert logs and complete log coverage. steps includes actual, expected, tolerance and passed; logWindow and compilation retain evidence. Runtime errors, missing log coverage, pending compilation, exceptions, missing targets and timeouts fail verification. Cancellation or timeout does not guarantee arbitrary user code has stopped. Fix source outside the scenario, wait for compile+reload, then explicitly submit a new verification job; never silently replay a failed invocation.

For the AI client's finite compile → scenario → repair → rerun loop, use Runtime verification and bounded repair. The shipped helper accepts a caller-supplied repair callback and enforces at most three repairs and a thirty-minute budget.