v0.4.8

# v0.4.8 — LLM Tool & Plugin Fix

---

## Features

- **`call_llm` tool** — Invoke a secondary LLM for focused subtasks like summarization, classification, and structured extraction. Supported across all three backends: Claude, Codex, and Copilot. Works with both API key (full features) and OAuth (basic features). Supports parallel calls, file attachments, and structured JSON output (eb0a3476, fa91af4c, b6e5c85f)

## Bug Fixes

- **Plugin name resolution** — Skills failed to resolve when the workspace directory name didn't match the SDK plugin name. Now reads the actual plugin name from `.claude-plugin/plugin.json` (b7904cb7)
- **Skill live reload** — Adding a workspace skill caused global and project skills to disappear until restart. All reload paths now use `loadAllSkills` to return the full three-tier list (4172fd82)
- **Codex event queue race condition** — Tool results and assistant text could be lost when async `item/completed` handlers were still running at `turn/completed`. Fixed by deferring queue completion until all handlers finish (fa91af4c)

## Internal

- Copilot `runMiniCompletion` now functional — enables title generation on the Copilot backend (fa91af4c)
- Copilot event adapter suppresses reasoning/intent events (fa91af4c)
- `call_llm` model badge shown in TurnCard activity rows (8bb4bcd2)

Author: github-actions[bot]

apps/electron/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@craft-agent/electron",
-  "version": "0.4.7",
+  "version": "0.4.8",
   "description": "Electron desktop app for Craft Agents",
   "main": "dist/main.cjs",
   "private": true,

apps/electron/resources/docs/llm-tool.md

@@ -0,0 +1,227 @@
+# LLM Tool (`call_llm`)
+
+Invoke a secondary LLM for focused subtasks. The tool loads file content automatically from paths you provide.
+
+## When to Use
+
+| Use Case | Model | Features |
+|----------|-------|----------|
+| Summarize large file | haiku | `attachments` |
+| Classify content | haiku | `outputFormat: "classification"` |
+| Extract structured data | haiku | `outputSchema` |
+| Deep analysis | sonnet/opus | `thinking: true` (API key only) |
+| Parallel processing | any | Multiple calls in one message |
+
+## Authentication Paths
+
+Features depend on how you authenticate:
+
+| Feature | API Key | OAuth |
+|---------|---------|-------|
+| Text attachments | Yes | Yes |
+| Image attachments | Yes | No |
+| Structured output | Guaranteed (tool_choice) | Prompt-based |
+| Extended thinking | Yes | No |
+| All models | Yes | Yes |
+
+- **API key**: Full features via direct Anthropic SDK
+- **OAuth**: Basic features via agent-native callback
+
+## Parameters
+
+| Parameter | Type | Description |
+|-----------|------|-------------|
+| `prompt` | string | Instructions for the LLM (required) |
+| `attachments` | array | File/image paths to include |
+| `model` | string | Any model from the model registry. Defaults to Haiku (fastest) |
+| `systemPrompt` | string | Optional system prompt |
+| `maxTokens` | number | Max output tokens (1-64000, default 4096) |
+| `temperature` | number | Sampling temperature 0-1 (ignored if thinking=true) |
+| `thinking` | boolean | Enable extended thinking (API key only) |
+| `thinkingBudget` | number | Token budget for thinking (1024-100000, default 10000) |
+| `outputFormat` | enum | Predefined output format |
+| `outputSchema` | object | Custom JSON Schema |
+
+## Attachments
+
+```typescript
+// Simple file
+attachments: ["/src/auth.ts"]
+
+// Large file - use line range (required for files >2000 lines)
+attachments: [{ path: "/logs/app.log", startLine: 1000, endLine: 1500 }]
+
+// Mix of files and images (API key only for images)
+attachments: ["/src/component.tsx", "/designs/mockup.png"]
+```
+
+### Line Ranges
+
+For files larger than 2000 lines or 500KB, you must specify a line range:
+
+```typescript
+{ path: "/path/to/large-file.log", startLine: 100, endLine: 600 }
+```
+
+The tool will tell you the file structure if you try to load a file that's too large.
+
+### Supported Formats
+
+- **Text files**: Any UTF-8 encoded file (`.ts`, `.js`, `.py`, `.md`, `.json`, etc.)
+- **Images**: `.png`, `.jpg`, `.jpeg`, `.gif`, `.webp` (max 5MB each, API key only)
+
+## Output Formats
+
+| Format | Returns |
+|--------|---------|
+| `summary` | `{ summary, key_points[], word_count }` |
+| `classification` | `{ category, confidence, reasoning }` |
+| `extraction` | `{ items[], count }` |
+| `analysis` | `{ findings[], issues[], recommendations[] }` |
+| `comparison` | `{ similarities[], differences[], verdict }` |
+| `validation` | `{ valid, errors[], warnings[] }` |
+
+## Parallel Processing
+
+Call multiple times in a single message for parallel execution:
+
+```
+call_llm(prompt: "Summarize", attachments: ["/file1.ts"])
+call_llm(prompt: "Summarize", attachments: ["/file2.ts"])
+call_llm(prompt: "Summarize", attachments: ["/file3.ts"])
+// All run simultaneously - ~3x faster than sequential!
+```
+
+Use cases for parallel calls:
+- Analyze multiple files at once
+- Get multiple perspectives on the same code
+- Process batch data
+- Generate variations
+
+## Examples
+
+### Summarize a File (Cheap)
+
+```typescript
+call_llm({
+  prompt: "Summarize the main functionality",
+  attachments: ["/src/auth/handler.ts"],
+  model: "claude-haiku-4-5-20251001"
+})
+```
+
+### Extract Structured Data
+
+```typescript
+call_llm({
+  prompt: "Extract all API endpoints from this file",
+  attachments: ["/src/routes.ts"],
+  outputSchema: {
+    type: "object",
+    properties: {
+      endpoints: {
+        type: "array",
+        items: {
+          type: "object",
+          properties: {
+            method: { type: "string" },
+            path: { type: "string" },
+            handler: { type: "string" }
+          }
+        }
+      }
+    },
+    required: ["endpoints"]
+  }
+})
+```
+
+### Classify Content
+
+```typescript
+call_llm({
+  prompt: "Classify this support ticket by urgency and category",
+  attachments: ["/tickets/latest.txt"],
+  outputFormat: "classification"
+})
+// Returns: { category: "billing", confidence: 0.92, reasoning: "..." }
+```
+
+### Deep Analysis with Thinking (API Key Only)
+
+```typescript
+call_llm({
+  prompt: "Analyze this algorithm for edge cases and potential bugs",
+  attachments: ["/src/sorting.ts"],
+  model: "claude-sonnet-4-5-20250929",
+  thinking: true,
+  thinkingBudget: 15000
+})
+```
+
+### Analyze Screenshot (API Key Only)
+
+```typescript
+call_llm({
+  prompt: "Describe the UI issues in this screenshot",
+  attachments: ["/screenshots/bug-report.png"],
+  model: "claude-sonnet-4-5-20250929"
+})
+```
+
+## Constraints
+
+| Constraint | Limit |
+|------------|-------|
+| Max attachments | 20 per call |
+| Max text file size | 2000 lines or 500KB |
+| Max image size | 5MB |
+| Max total content | 2MB across all attachments |
+| thinking + structured output | Mutually exclusive |
+| thinking + haiku | Not supported (use Sonnet/Opus) |
+| thinking + OAuth | Not supported (use API key) |
+| images + OAuth | Not supported (use API key) |
+
+## Error Handling
+
+The tool provides detailed error messages with recovery suggestions. Common errors:
+
+### Attachment Errors
+
+| Error | Cause | Solution |
+|-------|-------|----------|
+| File not found | Path doesn't exist | Check path spelling, use absolute paths |
+| File too large | >2000 lines or >500KB | Use line range: `{ path, startLine, endLine }` |
+| Line range too large | Range exceeds 2000 lines | Reduce range or split into multiple calls |
+| Binary file detected | Non-text file without image extension | Use only text files or supported images |
+| Permission denied | Cannot read file | Check file permissions |
+| Empty file | File has no content | Skip empty files |
+| Broken symlink | Symlink target missing | Fix or remove symlink |
+
+### Parameter Errors
+
+| Error | Cause | Solution |
+|-------|-------|----------|
+| thinking + outputFormat | Incompatible modes | Remove one option |
+| thinking + haiku | Haiku doesn't support thinking | Use Sonnet or Opus |
+| thinking + OAuth | OAuth doesn't support thinking | Use API key or remove thinking |
+| images + OAuth | OAuth doesn't support images | Use API key or remove images |
+| thinkingBudget without thinking | Missing thinking=true | Add `thinking: true` |
+| Invalid line range | startLine > endLine | Fix range values |
+| Unknown model | Model not in registry | Check available models in settings |
+
+### API Errors (API Key Path)
+
+| Status | Meaning | Recovery |
+|--------|---------|----------|
+| 401 | Invalid API key | Check/refresh credentials |
+| 403 | Access denied | Verify model access on your plan |
+| 429 | Rate limited | Reduce parallel calls, wait before retry |
+| 500/502/503 | API unavailable | Retry in a few seconds |
+
+## When NOT to Use
+
+- You can reason through it yourself without needing a cheaper model
+- The task requires your conversation context
+- You need tools (Read, Bash, Glob) - use Task tool with subagents instead
+- Simple one-liner responses that don't need isolation

apps/electron/resources/release-notes/0.4.8.md

@@ -0,0 +1,19 @@
+# v0.4.8 — LLM Tool & Plugin Fix
+
+---
+
+## Features
+
+- **`call_llm` tool** — Invoke a secondary LLM for focused subtasks like summarization, classification, and structured extraction. Supported across all three backends: Claude, Codex, and Copilot. Works with both API key (full features) and OAuth (basic features). Supports parallel calls, file attachments, and structured JSON output (eb0a3476, fa91af4c, b6e5c85f)
+
+## Bug Fixes
+
+- **Plugin name resolution** — Skills failed to resolve when the workspace directory name didn't match the SDK plugin name. Now reads the actual plugin name from `.claude-plugin/plugin.json` (b7904cb7)
+- **Skill live reload** — Adding a workspace skill caused global and project skills to disappear until restart. All reload paths now use `loadAllSkills` to return the full three-tier list (4172fd82)
+- **Codex event queue race condition** — Tool results and assistant text could be lost when async `item/completed` handlers were still running at `turn/completed`. Fixed by deferring queue completion until all handlers finish (fa91af4c)
+
+## Internal
+
+- Copilot `runMiniCompletion` now functional — enables title generation on the Copilot backend (fa91af4c)
+- Copilot event adapter suppresses reasoning/intent events (fa91af4c)
+- `call_llm` model badge shown in TurnCard activity rows (8bb4bcd2)

apps/electron/src/main/lib/config-watcher.ts

@@ -39,7 +39,7 @@ import {
 import { permissionsConfigCache, getAppPermissionsDir } from '@craft-agent/shared/agent';
 import { getWorkspacePath, getWorkspaceSourcesPath, getWorkspaceSkillsPath } from '@craft-agent/shared/workspaces';
 import type { LoadedSkill } from '@craft-agent/shared/skills';
-import { loadSkill, loadWorkspaceSkills, skillNeedsIconDownload, downloadSkillIcon } from '@craft-agent/shared/skills';
+import { loadSkill, loadAllSkills, skillNeedsIconDownload, downloadSkillIcon } from '@craft-agent/shared/skills';
 import {
   loadStatusConfig,
   statusNeedsIconDownload,
@@ -323,7 +323,6 @@ export class ConfigWatcher {
     debug('[ConfigWatcher] Setting up workspace watcher for:', this.workspaceDir);
     try {
       const watcher = watch(this.workspaceDir, { recursive: true }, (eventType, filename) => {
-        debug('[ConfigWatcher] RAW FILE EVENT:', eventType, filename);
         if (!filename) return;
 
         // Normalize path separators
@@ -714,7 +713,7 @@ export class ConfigWatcher {
       }
 
       // Notify list change
-      const allSkills = loadWorkspaceSkills(this.workspaceDir);
+      const allSkills = loadAllSkills(this.workspaceDir);
       this.callbacks.onSkillsListChange?.(allSkills);
     } catch (error) {
       debug('[ConfigWatcher] Error handling skills dir change:', error);

apps/electron/src/main/sessions.ts

@@ -35,6 +35,7 @@ import {
   migrateLegacyCredentials,
   migrateLegacyLlmConnectionsConfig,
   migrateOrphanedDefaultConnections,
+  MODEL_REGISTRY,
   type Workspace,
 } from '@craft-agent/shared/config'
 import { loadWorkspaceConfig } from '@craft-agent/shared/workspaces'
@@ -76,7 +77,7 @@ import { getCredentialManager } from '@craft-agent/shared/credentials'
 import { CraftMcpClient } from '@craft-agent/shared/mcp'
 import { type Session, type Message, type SessionEvent, type FileAttachment, type StoredAttachment, type SendMessageOptions, IPC_CHANNELS, generateMessageId } from '../shared/types'
 import { formatPathsToRelative, formatToolInputPaths, perf, encodeIconToDataUrl, getEmojiIcon, resetSummarizationClient, resolveToolIcon } from '@craft-agent/shared/utils'
-import { loadWorkspaceSkills, loadAllSkills, loadSkillBySlug, type LoadedSkill } from '@craft-agent/shared/skills'
+import { loadAllSkills, loadSkillBySlug, type LoadedSkill } from '@craft-agent/shared/skills'
 import type { ToolDisplayMeta } from '@craft-agent/core/types'
 import { getToolIconsDir, isCodexModel, getMiniModel, isAnthropicProvider, DEFAULT_MODEL, DEFAULT_CODEX_MODEL } from '@craft-agent/shared/config'
 import type { SummarizeCallback } from '@craft-agent/shared/sources'
@@ -518,6 +519,7 @@ function resolveToolDisplayMeta(
       const internalMcpServers: Record<string, Record<string, string>> = {
         'session': {
           'SubmitPlan': 'Submit Plan',
+          'call_llm': 'LLM Query',
           'config_validate': 'Validate Config',
           'skill_validate': 'Validate Skill',
           'mermaid_validate': 'Validate Mermaid',
@@ -585,7 +587,7 @@ function resolveToolDisplayMeta(
       if (skillSlug) {
         // Load skills and find the one being invoked
         try {
-          const skills = loadWorkspaceSkills(workspaceRootPath)
+          const skills = loadAllSkills(workspaceRootPath)
           const skill = skills.find(s => s.slug === skillSlug)
           if (skill) {
             // Try file-based icon first, fall back to emoji icon from metadata
@@ -1077,8 +1079,8 @@ export class SessionManager {
       onSkillChange: async (slug, skill) => {
         sessionLog.info(`Skill '${slug}' changed:`, skill ? 'updated' : 'deleted')
         // Broadcast updated list to UI
-        const { loadWorkspaceSkills } = await import('@craft-agent/shared/skills')
-        const skills = loadWorkspaceSkills(workspaceRootPath)
+        const { loadAllSkills } = await import('@craft-agent/shared/skills')
+        const skills = loadAllSkills(workspaceRootPath)
         this.broadcastSkillsChanged(skills)
       },
 
@@ -4889,6 +4891,18 @@ To view this task's output:
         // Format tool input paths to relative for better readability
         const formattedToolInput = formatToolInputPaths(event.input)
 
+        // Resolve call_llm model short name (e.g., "haiku") to full ID (e.g., "claude-haiku-4-5-20251001")
+        // for TurnCard badge display. The LLM sends short names but we want the resolved model shown.
+        if (event.toolName === 'mcp__session__call_llm' && formattedToolInput?.model) {
+          const shortName = String(formattedToolInput.model)
+          const modelDef = MODEL_REGISTRY.find(m => m.id === shortName)
+            || MODEL_REGISTRY.find(m => m.shortName.toLowerCase() === shortName.toLowerCase())
+            || MODEL_REGISTRY.find(m => m.name.toLowerCase() === shortName.toLowerCase())
+          if (modelDef) {
+            formattedToolInput.model = modelDef.id
+          }
+        }
+
         // Resolve tool display metadata (icon, displayName) for skills/sources
         // Only resolve when we have input (second event for SDK dual-event pattern)
         const workspaceRootPath = managed.workspace.rootPath
@@ -5516,7 +5530,7 @@ To view this task's output:
    */
   private resolveHookMentions(workspaceRootPath: string, mentions: string[]): { sourceSlugs: string[]; skillSlugs: string[] } | undefined {
     const sources = loadWorkspaceSources(workspaceRootPath)
-    const skills = loadWorkspaceSkills(workspaceRootPath)
+    const skills = loadAllSkills(workspaceRootPath)
     const sourceSlugs: string[] = []
     const skillSlugs: string[] = []
 

apps/electron/src/renderer/App.tsx

@@ -24,7 +24,7 @@ import { useUpdateChecker } from '@/hooks/useUpdateChecker'
 import { NavigationProvider } from '@/contexts/NavigationContext'
 import { navigate, routes } from './lib/navigate'
 import { stripMarkdown } from './utils/text'
-import { extractWorkspaceSlug } from '@craft-agent/shared/utils/workspace'
+import { extractWorkspaceSlugFromPath } from '@craft-agent/shared/utils/workspace-slug'
 import { initRendererPerf } from './lib/perf'
 import {
   initializeSessionsAtom,
@@ -186,7 +186,7 @@ export default function App() {
     if (!windowWorkspaceId) return null
     const workspace = workspaces.find(w => w.id === windowWorkspaceId)
     if (!workspace?.rootPath) return windowWorkspaceId // Fallback to ID
-    return extractWorkspaceSlug(workspace.rootPath, windowWorkspaceId)
+    return extractWorkspaceSlugFromPath(workspace.rootPath, windowWorkspaceId)
   }, [windowWorkspaceId, workspaces])
 
   // LLM connections with authentication status (for provider selection)

apps/electron/src/renderer/components/app-shell/input/FreeFormInput.tsx

@@ -15,7 +15,7 @@ import {
 import { Icon_Home, Icon_Folder } from '@craft-agent/ui'
 
 import * as storage from '@/lib/local-storage'
-import { extractWorkspaceSlug } from '@craft-agent/shared/utils/workspace'
+import { extractWorkspaceSlugFromPath } from '@craft-agent/shared/utils/workspace-slug'
 
 import { Button } from '@/components/ui/button'
 import {
@@ -363,7 +363,7 @@ export function FreeFormInput({
   // SDK expects "workspaceSlug:skillSlug" format, NOT UUID
   const workspaceSlug = React.useMemo(() => {
     if (!workspaceRootPath) return workspaceId // Fallback to ID if no path
-    return extractWorkspaceSlug(workspaceRootPath, workspaceId ?? '')
+    return extractWorkspaceSlugFromPath(workspaceRootPath, workspaceId ?? '')
   }, [workspaceRootPath, workspaceId])
 
   // Shuffle placeholder order once per mount so each session feels fresh