@anionex/dsh-vision-toolkit 0.1.30 → 0.1.32

package/patches/vision-tools-dsh.patch

@@ -16,7 +16,8 @@ index 8f42552..f2ee3f2 100644
 -  image yourself when a description you were given lacks a detail.
 ----
 -
- # vision-tools
+-# vision-tools
++# vision-skills
  
 -Five local CLIs that give a text-only agent eyes. They read one shared
 -vision config (`VISION_API_KEY` / `VISION_BASE_URL` / `VISION_MODEL` /
@@ -29,7 +30,7 @@ index 8f42552..f2ee3f2 100644
 +
 +The visual execution schemas are mounted only for the current Agent after this
 +Skill is loaded. A normal `skill` call activates them for the next model step.
-+If this content arrived through a direct `/vision-tools` invocation and the
++If this content arrived through a direct `/vision-skills` invocation and the
 +visual tools are still absent, call `vision_toolkit_activate` once. Do not call
 +that bootstrap when the visual tools are already present.
  
@@ -802,10 +803,10 @@ index 3261fcf..68025be 100644
     nodes, labels, groups, edges, directions, and edge labels.
  6. Generate the Mermaid, Graphviz, or requested structure from that inventory.
 diff --git a/references/restore-ui.md b/references/restore-ui.md
-index f9ec11a..de97a02 100644
+index f9ec11a..69627895161fc1b57c6241665255fdf86ae1c1e5 100644
 --- a/references/restore-ui.md
 +++ b/references/restore-ui.md
-@@ -28,16 +28,16 @@ typography, shadows, decorative details, and icon geometry.
+@@ -28,17 +28,18 @@
  1. Inspect the existing frontend stack, component library, icon set, source
     assets, and design tokens before writing code, but stop searching as soon as
     a usable local primitive is found.
@@ -824,12 +825,27 @@ index f9ec11a..de97a02 100644
     starting the next call. Normally one or two rounds are enough.
 -4. Do not use `trace`, foreground extraction, repeated color sampling, or
 -   iterative `pixel_diff` work in fast mode. Those are fidelity tools and will
+-   consume the delivery window.
 +4. Do not use `vision_trace`, foreground extraction, repeated color sampling,
-+   or iterative `vision_pixel_diff` work in fast mode. Those are fidelity tools and will
-    consume the delivery window.
++   iterative `vision_pixel_diff` work, or hand-written SVG in fast mode. Those
++   are fidelity tools and will consume the delivery window. Icons and decorative
++   marks stay library-based or screenshot-backed; never hand-write SVG code.
  
  ### Build the approximation
-@@ -57,7 +57,7 @@ typography, shadows, decorative details, and icon geometry.
+ 
+@@ -47,8 +48,9 @@
+    when zoomed in.
+ 2. Reuse the project's existing components and CSS tokens. If its frontend or
+    icon library contains a reasonably similar component or icon, use it
+-   directly instead of recreating the reference. Prefer an approximate library
+-   icon over cropping, tracing, or hand-drawing a new one.
++   directly instead of recreating the reference. Never hand-write SVG in fast
++   mode: use an approximate library icon, or extract the original pixels as a
++   screenshot-backed asset.
+ 3. Use nearby existing palette tokens or visually similar CSS values. Exact
+    sampled hex values, gradients, subtle borders, and shadow opacity are out of
+    scope unless one of them defines the whole composition.
+@@ -57,7 +59,7 @@
  
  ### Render once, fix once, deliver
  
@@ -838,7 +854,7 @@ index f9ec11a..de97a02 100644
     existing browser setup.
  2. Inspect the screenshot once. If there is an obvious structural failure such
     as a missing major region, broken wrapping, or a wildly wrong scale, make
-@@ -66,7 +66,7 @@ typography, shadows, decorative details, and icon geometry.
+@@ -66,7 +68,7 @@
     small color, icon, font, shadow, radius, or spacing differences.
  
  A practical time box is roughly 30 seconds for project inspection plus
@@ -847,7 +863,7 @@ index f9ec11a..de97a02 100644
  rendering, one correction, and screenshot delivery. Dependency installation or
  a project that does not already run may extend that target; do not compensate
  by silently switching back to a long precision loop.
-@@ -112,18 +112,19 @@ fonts, icons, images, and design tokens before writing replacements.
+@@ -112,18 +114,19 @@
  ### 2. Establish the coordinate system
  
  Record the reference image dimensions and the target viewport. Screenshots may
@@ -875,7 +891,7 @@ index f9ec11a..de97a02 100644
  
  Treat model boxes as approximate handles. Use them to organize the page, but
  do not mistake their final few pixels for measured boundaries.
-@@ -141,15 +142,11 @@ Use a padded crop only to inspect and center the target. Before foreground
+@@ -141,15 +144,11 @@
  extraction, tighten the final region to the visual's own ink; adjacent text or
  rules inside the region can become foreground components too.
  
@@ -896,7 +912,7 @@ index f9ec11a..de97a02 100644
  ```
  
  Inspect the transparent result before use. Confirm that the whole visual is
-@@ -162,8 +159,8 @@ sheet instead of checking them from filenames alone.
+@@ -162,8 +161,8 @@
  - Follow the project's existing framework and component patterns.
  - Build content, interaction, responsive layout, borders, shadows, and simple
    shapes natively.
@@ -907,7 +923,7 @@ index f9ec11a..de97a02 100644
  - Place extracted assets with explicit logical dimensions. Preserve their
    aspect ratio and avoid baking surrounding whitespace into the asset.
  - Match structure and proportions before tuning small spacing.
-@@ -172,10 +169,10 @@ sheet instead of checking them from filenames alone.
+@@ -172,10 +171,10 @@
  
  For an existing implementation, start here.
  
@@ -920,7 +936,7 @@ index f9ec11a..de97a02 100644
     spot or explain.
  3. Fix material discrepancies: missing or wrong content, incorrect hierarchy,
     broken wrapping, visibly wrong alignment or scale, wrong component state,
-@@ -185,7 +182,7 @@ For an existing implementation, start here.
+@@ -185,7 +184,7 @@
     rasterization, antialiasing, subpixel placement, or another imperceptible
     rendering detail.
  

package/src/exposure.ts

@@ -11,12 +11,18 @@ import type { ContentBlock } from '@deepseek-ai/dsh-llm'
 import type { Session } from '@deepseek-ai/dsh-session'
 import { defineTool, type ToolDefinition } from '@deepseek-ai/dsh-tools'
 import type { Context } from '@deepseek-ai/cordis'
-import { VISION_TOOLS_SKILL_CONTENT, VISION_TOOLS_SKILL_NAME } from './skill.ts'
+import { VISION_SKILLS_CONTENT, VISION_SKILLS_NAME } from './skill.ts'
 import { VISION_TOOL_NAMES } from './tools.ts'
 
 /** Small bootstrap tool retained only until the current Agent gains visual tools. */
 export const VISION_TOOLKIT_ACTIVATE = 'vision_toolkit_activate'
 
+/** Skill name used by releases before the rename to vision-skills. */
+export const LEGACY_VISION_TOOLS_SKILL_NAME = 'vision-tools'
+
+/** Unique pre-rename line in bundled instructions, kept for Session restore. */
+export const LEGACY_VISION_TOOLS_SKILL_MARKER = 'If this content arrived through a direct `/vision-tools` invocation and the'
+
 interface AgentExposure {
   active: boolean
   liftRestriction?: () => void
@@ -38,8 +44,16 @@ function isRecord(value: unknown): value is Record<string, unknown> {
   return value !== null && typeof value === 'object' && !Array.isArray(value)
 }
 
+function isBundledSkillName(name: unknown): boolean {
+  return name === VISION_SKILLS_NAME || name === LEGACY_VISION_TOOLS_SKILL_NAME
+}
+
+function isBundledSkillContent(text: string): boolean {
+  return text.includes(VISION_SKILLS_CONTENT) || text.includes(LEGACY_VISION_TOOLS_SKILL_MARKER)
+}
+
 function isVisionSkillArguments(value: unknown): boolean {
-  return isRecord(value) && value.name === VISION_TOOLS_SKILL_NAME
+  return isRecord(value) && isBundledSkillName(value.name)
 }
 
 function nativeSkillCall(raw: string): boolean {
@@ -54,13 +68,14 @@ function containsBundledSkillContent(blocks: readonly unknown[]): boolean {
   return blocks.some(block => isRecord(block)
     && block.type === 'text'
     && typeof block.text === 'string'
-    && block.text.includes(VISION_TOOLS_SKILL_CONTENT))
+    && isBundledSkillContent(block.text))
 }
 
 function isBundledSkillResult(value: unknown): boolean {
   return isRecord(value)
-    && value.name === VISION_TOOLS_SKILL_NAME
-    && value.content === VISION_TOOLS_SKILL_CONTENT
+    && isBundledSkillName(value.name)
+    && typeof value.content === 'string'
+    && isBundledSkillContent(value.content)
 }
 
 /** Whether durable history proves that this Session loaded the bundled Skill. */
@@ -70,7 +85,7 @@ function hasLoadedVisionSkill(session: Session): boolean {
     if (event.type === 'user/message') {
       const source = event.data.source
       if (source.kind === 'skill-invocation'
-        && source.name === VISION_TOOLS_SKILL_NAME
+        && isBundledSkillName(source.name)
         && containsBundledSkillContent(event.data.content)) return true
       continue
     }
@@ -119,7 +134,7 @@ export class VisionToolExposure {
     this.activationTool = defineTool({
       name: VISION_TOOLKIT_ACTIVATE,
       description: `Activate the independent Vision Toolkit execution tools for this Agent: ${Object.values(VISION_TOOL_NAMES).join(', ')}. `
-        + `Loading the ${VISION_TOOLS_SKILL_NAME} Skill normally activates them automatically; call this once when the visual tools are still absent, then use them for image understanding, OCR, UI detection, and related tasks. `
+        + `Loading the ${VISION_SKILLS_NAME} Skill normally activates them automatically; call this once when the visual tools are still absent, then use them for image understanding, OCR, UI detection, and related tasks. `
         + 'It is safe to call before the Skill is loaded, and this activation tool disappears after success.',
       parameters: {},
       output: {

package/src/image-input-variants.ts

@@ -41,6 +41,12 @@ export const VARIANT_SUFFIX = ' (Vision Toolkit)'
 /** Promise-cache bound for image descriptions, so a long-lived Web profile cannot hoard evidence text. */
 const EVIDENCE_CACHE_LIMIT = 64
 
+/** Periodic self-heal sweep for image-input variant routes. */
+const VARIANT_SWEEP_INTERVAL_MS = 10_000
+
+/** How long an upstream must stay absent before its wrapper is released. */
+const VARIANT_RELEASE_GRACE_MS = 10_000
+
 /**
  * Media types the Vision Toolkit glance pipeline accepts, by declared media
  * type. Narrower than the paste-to-workspace route (which stores any image
@@ -835,6 +841,12 @@ export function installImageInputVariants(
   getRuntime: () => VisionToolkitRuntime | undefined,
 ): { dispose: () => void; reconcile: () => void } {
   const registrations = new Map<string, () => void>()
+  // Upstream ids observed missing and the timestamp of the first missing
+  // observation. A wrapper is only released after its upstream stays absent
+  // for the full grace period, so a transient registry gap (adapter
+  // re-registration, HMR reload) cannot drop a wrapper that a live session is
+  // still pointed at, no matter how many sweeps fire in between.
+  const stale = new Map<string, number>()
   let disposed = false
   // Serialize sweeps: a registration itself announces llm/adapters-updated,
   // and two interleaved sweeps must never probe the same route concurrently.
@@ -843,13 +855,29 @@ export function installImageInputVariants(
   // not one per notification (a single registration emits a notification).
   let sweepQueued = false
 
+  const release = (upstream: string): void => {
+    const dispose = registrations.get(upstream)
+    if (dispose === undefined) return
+    registrations.delete(upstream)
+    stale.delete(upstream)
+    try {
+      dispose()
+    } catch (error) {
+      ctx.logger.warn(
+        'dsh-vision-toolkit: image-input variant release failed for "%s": %s',
+        upstream,
+        messageOf(error),
+      )
+    }
+  }
+
   const releaseAll = (): void => {
-    for (const dispose of [...registrations.values()]) dispose()
-    registrations.clear()
+    stale.clear()
+    for (const upstream of [...registrations.keys()]) release(upstream)
   }
 
   const sweep = (): void => {
-    if (sweepQueued) return
+    if (disposed || sweepQueued) return
     sweepQueued = true
     queueMicrotask(() => {
       sweepQueued = false
@@ -874,12 +902,33 @@ export function installImageInputVariants(
       } catch {
         return
       }
+      const live = new Set(providers.map(provider => provider.id))
+      // Release restrict-narrowed routes immediately (a configuration change is
+      // authoritative). For routes merely missing from the live registry, keep
+      // the wrapper until the absence outlives the grace period.
+      for (const upstream of [...registrations.keys()]) {
+        if (restrict.size > 0 && !restrict.has(upstream)) {
+          release(upstream)
+          continue
+        }
+        if (live.has(upstream)) {
+          stale.delete(upstream)
+          continue
+        }
+        const missedAt = stale.get(upstream)
+        if (missedAt !== undefined && Date.now() - missedAt >= VARIANT_RELEASE_GRACE_MS) {
+          release(upstream)
+        } else if (missedAt === undefined) {
+          stale.set(upstream, Date.now())
+        }
+      }
       for (const provider of providers) {
         const upstream = provider.id
         if (restrict.size > 0 && !restrict.has(upstream)) continue
         // Our own variant routes declare image input and are never wrapped;
         // probing them would just re-probe their upstream route.
         if (upstream.startsWith(VARIANT_PROVIDER_PREFIX)) continue
+        const variantId = variantProviderId(upstream)
         let models: readonly LlmModelInfo[]
         try {
           models = await llm.listModels(upstream)
@@ -888,18 +937,26 @@ export function installImageInputVariants(
         }
         const eligible = models.some(shouldWrapModel)
         const registered = registrations.has(upstream)
-        if (!eligible && registered) {
-          // The route lost its eligible models: release the stale variant.
-          const dispose = registrations.get(upstream)
-          dispose?.()
-          registrations.delete(upstream)
-          continue
+        if (registered && live.has(variantId)) {
+          // Re-read the live registry after the await: a rebuild inside the
+          // probe window can drop our wrapper without touching our map.
+          const liveNow = new Set(llm.listProviders().map(provider => provider.id))
+          if (liveNow.has(variantId)) {
+            // Healthy handle: only react when the route lost every eligible model.
+            if (!eligible) release(upstream)
+            continue
+          }
+        }
+        if (registered) {
+          // Dead handle from a registry rebuild/reset: drop it and register a
+          // fresh wrapper below.
+          release(upstream)
         }
-        if (!eligible || registered) continue
+        if (!eligible) continue
         if (disposed) return
         try {
           const dispose = llm.registerAdapter(
-            [variantProviderId(upstream)],
+            [variantId],
             new ImageInputVariantAdapter(
               ctx,
               llm,
@@ -910,6 +967,7 @@ export function installImageInputVariants(
             ),
           )
           registrations.set(upstream, dispose)
+          stale.delete(upstream)
         } catch (error) {
           ctx.logger.warn(
             'dsh-vision-toolkit: image-input variant registration skipped for "%s": %s',
@@ -918,16 +976,6 @@ export function installImageInputVariants(
           )
         }
       }
-      const live = new Set(providers.map(provider => provider.id))
-      for (const [upstream, dispose] of [...registrations]) {
-        // A wrapper is released when its upstream route vanished OR the
-        // current configuration no longer allows that route (restrict
-        // narrowing must not leave stale variants behind).
-        if (!live.has(upstream) || (restrict.size > 0 && !restrict.has(upstream))) {
-          dispose()
-          registrations.delete(upstream)
-        }
-      }
     } catch (error) {
       ctx.logger.warn('dsh-vision-toolkit: image-input variant sweep failed: %s', messageOf(error))
     }
@@ -937,9 +985,14 @@ export function installImageInputVariants(
     ctx.on('llm/adapters-updated', () => { sweep() })
   }
   sweep()
+  // Self-heal safety net: re-verify periodically so a wrapper dropped by an
+  // unobserved cause (registry rebuild without an event we can see) is restored
+  // within one interval.
+  const timer = setInterval(() => { sweep() }, VARIANT_SWEEP_INTERVAL_MS)
   return {
     dispose: () => {
       disposed = true
+      clearInterval(timer)
       releaseAll()
     },
     reconcile: () => { sweep() },

package/src/index.ts

@@ -2,7 +2,7 @@
  * @anionex/dsh-vision-toolkit — DSH Vision Toolkit profile bundle.
  *
  * Plugin lifecycle follows the documented readiness chain: verify the pinned
- * upstream checkout, publish the vision-tools Skill and its one-shot bootstrap,
+ * upstream checkout, publish the vision-skills Skill and its one-shot bootstrap,
  * then mount the execution tools only in Agents that load that Skill or invoke
  * the bootstrap. Any
  * failure leaves no model capability behind, and disposal unregisters every
@@ -24,7 +24,7 @@ import {
 import { VisionToolExposure } from './exposure.ts'
 import { createPasteTakeoverResolver, installImageInputVariants } from './image-input-variants.ts'
 import { VisionToolkitRuntimeManager } from './runtime-manager.ts'
-import { VISION_TOOLS_SKILL } from './skill.ts'
+import { VISION_SKILLS_SKILL } from './skill.ts'
 import { createVisionTools } from './tools.ts'
 import { PLUGIN_VERSION } from './version.ts'
 import { installVisionToolkitWeb, VisionToolkitWebBackend } from './web.ts'
@@ -65,7 +65,7 @@ export async function apply(ctx: Context, config: VisionToolkitConfig = {}): Pro
     let skill: (() => void) | undefined
     try {
       activationTool = ctx.tools.register(exposure.activationTool)
-      skill = ctx.skills.register(VISION_TOOLS_SKILL)
+      skill = ctx.skills.register(VISION_SKILLS_SKILL)
       exposureDisposer = exposure.install()
       operationalDisposers = { activationTool, exposure: exposureDisposer, skill }
       const info = manager.current().upstreamVersion
@@ -90,7 +90,7 @@ export async function apply(ctx: Context, config: VisionToolkitConfig = {}): Pro
   } catch (error) {
     const message = error instanceof Error ? error.message : String(error)
     ctx.logger.error(
-      'dsh-vision-toolkit %s: runtime not ready; the vision-tools skill, activation bootstrap, and Agent-scoped visual tools are NOT registered. Settings remain available for repair. %s',
+      'dsh-vision-toolkit %s: runtime not ready; the vision-skills skill, activation bootstrap, and Agent-scoped visual tools are NOT registered. Settings remain available for repair. %s',
       PLUGIN_VERSION,
       message,
     )

package/src/plugin-update.ts

@@ -822,6 +822,13 @@ export class VisionToolkitPluginUpdateService {
     profile: ProfileInstall,
     pnpmPath: string,
   ): Promise<CommandResult> {
+    // Windows batch shims cannot be spawned directly; route them through cmd.exe.
+    const program = this.platform === 'win32' && /\.(?:cmd|bat)$/i.test(pnpmPath)
+      ? process.env.COMSPEC ?? 'cmd.exe'
+      : pnpmPath
+    const argv = program === pnpmPath
+      ? [pnpmPath, ...args]
+      : [program, '/d', '/s', '/c', pnpmPath, ...args]
     const controller = new AbortController()
     let timedOut = false
     const timeout = setTimeout(() => {
@@ -830,7 +837,7 @@ export class VisionToolkitPluginUpdateService {
     }, timeoutMs)
     try {
       const handle = this.ctx.subprocess.spawn({
-        argv: [pnpmPath, ...args],
+        argv,
         cwd: profile.profileDir,
         stdio: {
           stdin: 'ignore',

package/src/runtime.ts

@@ -56,7 +56,9 @@ const VISION_MODEL_TEST_IMAGE = fileURLToPath(new URL('../assets/vision-model-te
 const VISION_MODEL_TEST_PROMPT = 'This is an explicit service readiness test. Reply with one short sentence confirming that you received the image.'
 
 /** Bump when the Pillow compression ladder changes so stale cache entries are ignored. */
-const COMPRESSED_IMAGE_CACHE_VERSION = 'v1'
+const COMPRESSED_IMAGE_CACHE_VERSION = 'v2'
+/** Cache keys carry 64-bit digests so Windows paths stay below MAX_PATH; the full file sha256 is computed on read and compared against this prefix. */
+const COMPRESSED_IMAGE_CACHE_KEY_DIGEST_LENGTH = 16
 const COMPRESSED_IMAGE_CACHE_MAX_ENTRIES = 200
 const COMPRESSED_IMAGE_CACHE_MAX_BYTES = 512 * 1024 * 1024
 const COMPRESSED_IMAGE_CACHE_STALE_PARTIAL_MS = 60 * 60 * 1000
@@ -883,7 +885,7 @@ export class VisionToolkitRuntime {
   private async readCacheCandidate(
     root: string,
     name: string,
-    expectedOutDigest: string,
+    expectedOutDigestPrefix: string,
     maxBytes: number,
     maxPixels: number,
     operation: OperationContext,
@@ -909,7 +911,8 @@ export class VisionToolkitRuntime {
     } catch {
       return undefined
     }
-    if (bytes.length !== info.size || createHash('sha256').update(bytes).digest('hex') !== expectedOutDigest) return undefined
+    const digest = createHash('sha256').update(bytes).digest('hex')
+    if (bytes.length !== info.size || !digest.startsWith(expectedOutDigestPrefix)) return undefined
     let probed: { width: number; height: number; format: string } | undefined
     try {
       probed = await this.adapter.probeImageSize(real, { signal: operation.signal })
@@ -929,7 +932,7 @@ export class VisionToolkitRuntime {
 
   private cacheEntryOutDigest(entry: string, prefix: string): string | undefined {
     const tail = entry.slice(prefix.length + 1)
-    return /^[0-9a-f]{64}-/u.test(tail) ? tail.slice(0, 64) : undefined
+    return /^[0-9a-f]{16}-/u.test(tail) ? tail.slice(0, 16) : undefined
   }
 
   private async pruneCompressedCache(root: string): Promise<void> {
@@ -997,21 +1000,21 @@ export class VisionToolkitRuntime {
     if (bytes.length !== image.bytes) {
       throw new VisionToolkitError('input', `image changed while preparing the vision request: ${image.path}`)
     }
-    const digest = createHash('sha256').update(bytes).digest('hex')
+    const digest = createHash('sha256').update(bytes).digest('hex').slice(0, COMPRESSED_IMAGE_CACHE_KEY_DIGEST_LENGTH)
     const root = await this.compressedImageRoot(policy)
     await this.pruneCompressedCache(root)
     const prefix = `${COMPRESSED_IMAGE_CACHE_VERSION}-${digest}-b${this.config.maxImageBytes}-p${this.config.maxImagePixels}`
     for (const entry of await readdir(root)) {
       if (!entry.startsWith(`${prefix}-`) || entry.startsWith('.')) continue
-      const outDigest = this.cacheEntryOutDigest(entry, prefix)
-      if (outDigest === undefined) {
+      const outDigestPrefix = this.cacheEntryOutDigest(entry, prefix)
+      if (outDigestPrefix === undefined) {
         await rm(join(root, entry), { force: true }).catch(() => {})
         continue
       }
       const cached = await this.readCacheCandidate(
         root,
         entry,
-        outDigest,
+        outDigestPrefix,
         this.config.maxImageBytes,
         this.config.maxImagePixels,
         operation,
@@ -1037,7 +1040,7 @@ export class VisionToolkitRuntime {
     }
     const extension = compressed.format === 'jpeg' ? 'jpg' : compressed.format
     const stagedBytes = await readFile(staged, { signal: operation.signal })
-    const outDigest = createHash('sha256').update(stagedBytes).digest('hex')
+    const outDigest = createHash('sha256').update(stagedBytes).digest('hex').slice(0, COMPRESSED_IMAGE_CACHE_KEY_DIGEST_LENGTH)
     const finalName = `${prefix}-${outDigest}-${compressed.width}x${compressed.height}.${extension}`
     const finalPath = join(root, finalName)
     const existing = await this.readCacheCandidate(

package/src/skill.ts

@@ -8,28 +8,28 @@ import { fileURLToPath } from 'node:url'
 import type { SkillRegistration } from '@deepseek-ai/dsh-skill'
 
 /** Stable catalog/invocation name shared with progressive tool exposure. */
-export const VISION_TOOLS_SKILL_NAME = 'vision-tools'
+export const VISION_SKILLS_NAME = 'vision-skills'
 
 /** Packaged resource root for the adapted upstream playbooks. */
-export const VISION_TOOLS_SKILL_RESOURCE_BASE = fileURLToPath(
+export const VISION_SKILLS_RESOURCE_BASE = fileURLToPath(
   new URL('../assets/skill/', import.meta.url),
 )
 
 /** Exact bundled instructions used as the progressive-exposure evidence marker. */
-export const VISION_TOOLS_SKILL_CONTENT = readFileSync(
+export const VISION_SKILLS_CONTENT = readFileSync(
   new URL('../assets/skill/SKILL.md', import.meta.url),
   'utf8',
 )
 
 /** Runtime skill registration mounted only after every native tool is ready. */
-export const VISION_TOOLS_SKILL: SkillRegistration = {
-  name: VISION_TOOLS_SKILL_NAME,
+export const VISION_SKILLS_SKILL: SkillRegistration = {
+  name: VISION_SKILLS_NAME,
   description: '把截图或设计图还原为 UI（网页、应用界面或组件），生成 HTML/CSS 或项目现有前端代码；也支持图片问答、长截图/聊天记录 OCR、元素定位与盘点、裁剪、前景提取、像素对比、取色、SVG 描摹和 HTML 转截图。当任务涉及图片理解、UI 还原/重建（rebuild or restore a UI from a screenshot）、视觉回归、像素坐标、基于截图的 GUI 操作、可复用图片/SVG 素材、图表还原或长截图 OCR 时使用。',
   whenToUse: '任务依赖图片文字/内容、像素坐标、截图转 UI 重建、视觉回归、可复用图片/SVG 素材、图表还原、基于截图的 GUI 操作或长截图 OCR 时使用。',
   source: 'runtime',
   resourceBase: {
     kind: 'directory',
-    path: VISION_TOOLS_SKILL_RESOURCE_BASE,
+    path: VISION_SKILLS_RESOURCE_BASE,
   },
-  content: VISION_TOOLS_SKILL_CONTENT,
+  content: VISION_SKILLS_CONTENT,
 }