openpencil/apps/web/server/utils/gemini-client.ts
Kayshen Xu 2b06d95cc1 V0.5.0 (#67)
* docs: add image search & generation design spec and implementation plan

- Spec: dual-source image search (Openverse + Wikimedia), multi-provider image generation
- Plan: 16 tasks covering types, server endpoints, settings UI, property panel, auto-search pipeline, MCP integration

* feat(types): add image service types and imagePrompt to ImageNode

* feat(server): add image service API key validation endpoint

Adds POST /api/ai/image-service-test that validates credentials for
openverse (client_credentials), openai/custom (Bearer + /v1/models),
gemini (API key + v1beta/models), and replicate (Bearer + /v1/models).

* feat(server): add multi-provider image generation endpoint

* feat(server): add dual-source image search endpoint (Openverse + Wikimedia)

POST /api/ai/image-search searches freely-licensed images via Openverse
with automatic fallback to Wikimedia Commons on 429 rate-limit responses.
Supports optional OAuth credentials for authenticated Openverse requests.

* feat(store): add imageSearchStatuses to canvas store for runtime status tracking

* feat(store): add image generation config and Openverse OAuth to agent settings

* feat(editor): add Images tab to agent settings dialog

Adds Popover primitive, ImagesPage component with Image Search (Openverse OAuth, test) and Image Generation (provider select, API key, model, base URL) sections, and wires them into the settings dialog sidebar.

* feat(panels): add image search popover with Openverse/Wikimedia results grid

* feat(panels): add image generate popover with multi-provider support

* feat(panels): add Search and Generate buttons to image property section

* feat(ai): update prompts to use imagePrompt instead of src for image nodes

* feat(ai): add auto-search pipeline with Openverse/Wikimedia fallback

* feat(ai): trigger auto image search after design generation completes

* feat(mcp): implement G() operation for image search in batch design DSL

Adds the G(parent, mode, prompt) operation to batch_design DSL that creates
an image node and optionally fetches a real image URL via the image-search
API when mode is "search". Converts executeLine to async to support the
network call.

* feat(mcp): auto-fill images after design refinement in layered pipeline

* feat(ai): split imageSearchQuery and imagePrompt for search vs generation

- ImageNode now has both imageSearchQuery (short keywords for search)
  and imagePrompt (long description for AI image generation)
- AI prompts instruct LLM to generate both fields
- Search pipeline and popovers use imageSearchQuery
- Generate popover uses imagePrompt
- Server-side simplifySearchQuery kept as fallback for manual input

* fix(ai): hook auto image search into orchestrator completion path

The primary generation path uses executeOrchestration -> insertStreamingNode,
not applyNodesToCanvas/animateNodesToCanvas. Added scanAndFillImages call
to orchestrator.ts after all sub-agents complete. Added debug logging.
Removed plan/spec docs from git.

* style(editor): remove provider names from image search ready status

* fix(panels): clean up image gen error display and settings UI

- Parse API error response to show concise message instead of raw JSON
- Limit error text to 2 lines with line-clamp
- Fix image gen test button sending wrong service name
- Inline Image Search ready indicator with section header
- Remove debug logging from image search pipeline

* style(panels): allow up to 4 lines for image gen error message

* fix: avoid 1-frame delay when resizing canvas (#60)

rAF callbacks run before ResizeObserver in the same frame.
Scheduling render in ResizeObserver via rAF defers it to the next frame.

Invoke render() synchronously to leverage ResizeObserver's pre-paint timing
and ensure immediate visual update.

* feat(electron): implement desktop application structure and auto-updater

- Introduced a new Electron desktop application with a structured directory for apps and packages.
- Added auto-updater functionality to manage application updates seamlessly.
- Created a comprehensive menu system for the desktop app.
- Implemented logging capabilities for better debugging and error tracking.
- Configured build settings for various platforms (macOS, Windows, Linux) using electron-builder.
- Established TypeScript configurations for both the desktop and web applications.
- Integrated Vite for the web application with support for React and Tailwind CSS.
- Added icons and assets for the desktop application.

* chore: update package versions to 0.5.0 across all package.json files and add pre-commit hook for version synchronization

- Bumped version to 0.5.0 in package.json files for the main project, desktop app, web app, and all packages.
- Introduced a pre-commit hook to automatically sync version numbers from branch names to all package.json files.

* chore: update package versions to 0.5.0 and refactor Skia components

- Bumped version to 0.5.0 in bun.lock and all relevant package.json files.
- Refactored Skia components to utilize shared functionality from @zseven-w/pen-renderer, including image loading, hit testing, and path utilities.
- Removed redundant code and improved modularity by re-exporting necessary functions and classes from the renderer package.

* fix(panels): handle string fill values in icon nodes (#61)

AI-generated icon/path nodes may have fill stored as a raw string
instead of a PenFill[] array, causing "Cannot use 'in' operator"
crash when selecting the node in the property panel.

* chore: update documentation and project structure for monorepo organization

- Added a new version bump command to synchronize all package.json files.
- Updated the project structure to reflect a monorepo setup with organized workspaces for apps and packages.
- Enhanced README files in multiple languages to include the new structure and commands.
- Adjusted image paths in documentation to point to the correct locations for the desktop application.

* feat(ai): incremental image search and improved image generation prompts

- Refactor image search from batch post-generation to incremental queue:
  enqueueImageForSearch() triggers as each image node is inserted during
  streaming, so images appear progressively instead of all at once after
  generation completes. scanAndFillImages() remains as a final sweep.
- Update imagePrompt guidance to avoid "transparent background" and
  similar phrases that many models cannot reliably produce.
- Pass node width/height from image panel to generation endpoint for
  aspect-ratio-aware output (Gemini aspect ratio mapping, OpenAI size
  selection, Replicate dimensions).

* feat(ai): multi-profile image generation config and cleaner error messages

- Support multiple image generation profiles with active selection;
  first configured profile becomes default. Old single-config migrated
  automatically on hydrate.
- Fix Gemini aspect ratio: move to generationConfig.imageConfig per API spec.
- Extract clean error messages from provider JSON responses (Gemini
  error.message, OpenAI error.message, Replicate detail) instead of
  returning raw JSON text.
- Remove destructive client-side regex that mangled error display.

* feat(design-md): integrate design system panel and functionality

- Added a new DesignMdPanel component for managing design system specifications.
- Implemented functionality to toggle the design system panel in the editor layout and toolbar.
- Introduced new commands for importing, exporting, and auto-generating design.md content.
- Updated AI chat handlers to utilize design.md data for enhanced design generation.
- Enhanced localization support for design system features across multiple languages.

* perf(canvas): skip draw calls for nodes outside the viewport (#64)

Add viewport culling in render() to avoid issuing CanvasKit draw calls
  for off-screen nodes. A 64px screen-space buffer is kept around the
  viewport edges so nearby nodes are pre-rendered, preventing pop-in
  during fast panning.

* feat(utils): enhance Windows process spawning for CLI scripts

- Updated the buildSpawnClaudeCodeProcess function to handle .cmd and .ps1 scripts appropriately.
- Implemented PowerShell invocation for .ps1 files and ensured safe defaults for .cmd and .exe files.
- Improved handling of command execution to avoid limitations of cmd.exe.

* feat(ai): add support for Gemini CLI integration

- Extended the AI provider options to include 'gemini' across various components and APIs.
- Implemented functions for generating, validating, and connecting to the Gemini CLI.
- Added Gemini-specific error handling and model fetching logic.
- Updated UI components to display Gemini as a selectable provider with appropriate icons and labels.
- Enhanced localization support for Gemini-related features in multiple languages.

* feat(editor): warn before closing with unsaved changes

Intercept window/tab close when isDirty is true:
- Electron: native dialog with Save / Don't Save / Cancel
- Web: beforeunload handler + confirm on New/Open actions
- i18n: close-confirm strings for all 15 locales

* feat(ipc): extract IPC handlers to a dedicated module

- Moved IPC dialog handling and updater functions from main.ts to ipc-handlers.ts for better organization and maintainability.
- Implemented file open/save dialogs, theme setting, and preferences management through IPC.
- Enhanced updater functionality with state management and auto-update settings.
- Improved code structure by separating concerns, making it easier to manage IPC-related logic.

* feat(docs): update CLAUDE documentation and add new files for desktop and web apps

- Enhanced CLAUDE.md with detailed module documentation references for `packages/` and `apps/`.
- Updated `pen-core` description to include clone utilities in `pen-core`.
- Added new documentation files for the desktop and web applications, outlining their structure, components, and functionalities.
- Included IPC handler details in the desktop app documentation for better clarity on file dialogs and theme synchronization.

* feat(docker): add Gemini CLI support and update documentation

- Introduced a new Docker build stage for the Gemini CLI, allowing users to install and run it.
- Updated the Dockerfile to include the installation of the Gemini CLI alongside existing CLI tools.
- Enhanced README files in multiple languages to document the new `openpencil-gemini` image variant.
- Added Gemini CLI connection instructions to the main README for better user guidance.

* feat(docs): add Gemini CLI connection instructions to multiple language READMEs

- Updated README files in German, Spanish, French, Hindi, Indonesian, Japanese, Korean, Portuguese, Russian, Thai, Turkish, Vietnamese, and both Traditional and Simplified Chinese to include connection instructions for the Gemini CLI.
- Enhanced documentation to improve user guidance for connecting the Gemini CLI in agent settings.

* perf(renderer): replace count-based text cache limits with memory-based eviction (#66)

previous limits (PARA_CACHE_MAX=200, TEXT_CACHE_MAX=300) were too small
  for scenes with many nodes, causing constant cache churn and paragraph
  rebuilds every frame, which dropped FPS significantly during canvas pan.

  - switch to byte-budget limits (64 MB paragraphs, 256 MB bitmaps)
  - bitmap size measured exactly as cw*ch*4; paragraph WASM heap estimated
    as content.length*64+4096
  - eviction uses Map insertion order (FIFO) instead of a separate string[]
    array, replacing O(n) array.shift() with O(1) Map.entries().next()
  - evict before insert so the budget check includes the incoming entry

---------

Co-authored-by: Fini <fini.yang@gmail.com>
Co-authored-by: leinaldo <60176594+leinaldo@users.noreply.github.com>
2026-03-22 09:44:04 +08:00

394 lines
11 KiB
TypeScript

import { spawn } from 'node:child_process'
import { resolveGeminiCli } from './resolve-gemini-cli'
type ThinkingMode = 'adaptive' | 'disabled' | 'enabled'
type ThinkingEffort = 'low' | 'medium' | 'high' | 'max'
export interface GeminiExecOptions {
model?: string
systemPrompt?: string
thinkingMode?: ThinkingMode
thinkingBudgetTokens?: number
effort?: ThinkingEffort
timeoutMs?: number
}
interface GeminiCliResult {
text?: string
error?: string
}
const DEFAULT_GEMINI_TIMEOUT_MS = 15 * 60 * 1000
/**
* Allowlist-based env filter for Gemini CLI subprocess.
* Passes through safe system vars and Google/Gemini-specific prefixes.
*/
const GEMINI_ENV_ALLOWLIST = new Set([
'PATH', 'HOME', 'TERM', 'LANG', 'SHELL', 'TMPDIR',
// Windows-essential
'SYSTEMROOT', 'COMSPEC', 'USERPROFILE', 'APPDATA', 'LOCALAPPDATA',
'PATHEXT', 'SYSTEMDRIVE', 'TEMP', 'TMP', 'HOMEDRIVE', 'HOMEPATH',
])
function filterGeminiEnv(
env: Record<string, string | undefined>,
): Record<string, string | undefined> {
const result: Record<string, string | undefined> = {}
for (const [k, v] of Object.entries(env)) {
if (
GEMINI_ENV_ALLOWLIST.has(k)
|| k.startsWith('GOOGLE_')
|| k.startsWith('GEMINI_')
|| k.startsWith('GCLOUD_')
) {
result[k] = v
}
}
return result
}
/**
* Run Gemini CLI in non-interactive mode with JSON output.
* Passes prompt via stdin to avoid command-line length limits.
* The CLI handles its own authentication (OAuth or API key).
*/
export async function runGeminiExec(
userPrompt: string,
options: GeminiExecOptions = {},
): Promise<GeminiCliResult> {
const binPath = resolveGeminiCli()
if (!binPath) {
return { error: 'Gemini CLI not found. Install it first.' }
}
const prompt = buildPrompt(options.systemPrompt, userPrompt)
const args = [
'-o', 'json',
'--approval-mode', 'plan',
]
if (options.model) {
args.push('-m', options.model)
}
// Use -p with a minimal marker; full prompt piped via stdin.
// Gemini CLI appends -p value after stdin content.
args.push('-p', ' ')
try {
const result = await executeGeminiCommand(
binPath,
args,
options.timeoutMs ?? DEFAULT_GEMINI_TIMEOUT_MS,
prompt,
)
return result
} catch (error) {
return { error: error instanceof Error ? error.message : 'Gemini execution failed' }
}
}
/**
* Stream Gemini CLI output in real-time using `stream-json` format.
* Passes prompt via stdin. Yields text deltas as they arrive.
*/
export function streamGeminiExec(
userPrompt: string,
options: GeminiExecOptions = {},
): {
stream: AsyncGenerator<{ type: 'text' | 'error' | 'done'; content: string }>
kill: () => void
} {
const binPath = resolveGeminiCli()
if (!binPath) {
return {
stream: (async function* () {
yield { type: 'error' as const, content: 'Gemini CLI not found.' }
})(),
kill: () => {},
}
}
const prompt = buildPrompt(options.systemPrompt, userPrompt)
const args = [
'-o', 'stream-json',
'--approval-mode', 'plan',
]
if (options.model) {
args.push('-m', options.model)
}
// Use -p with minimal marker; full prompt piped via stdin.
args.push('-p', ' ')
const child = spawn(binPath, args, {
env: filterGeminiEnv(process.env as Record<string, string | undefined>),
stdio: ['pipe', 'pipe', 'pipe'],
...(process.platform === 'win32' && { shell: true }),
})
// Pipe prompt via stdin
if (child.stdin) {
child.stdin.write(prompt)
child.stdin.end()
}
const timeoutMs = options.timeoutMs ?? DEFAULT_GEMINI_TIMEOUT_MS
const timer = setTimeout(() => {
child.kill('SIGTERM')
}, timeoutMs)
async function* generateStream(): AsyncGenerator<{ type: 'text' | 'error' | 'done'; content: string }> {
let buffer = ''
child.stderr?.on('data', () => { /* discard stderr */ })
try {
for await (const chunk of child.stdout!) {
buffer += chunk.toString('utf-8')
let idx = buffer.indexOf('\n')
while (idx >= 0) {
const line = buffer.slice(0, idx).trim()
buffer = buffer.slice(idx + 1)
if (line) {
const event = parseStreamJsonLine(line)
if (event) yield event
}
idx = buffer.indexOf('\n')
}
}
// Flush remaining buffer
const tail = buffer.trim()
if (tail) {
const event = parseStreamJsonLine(tail)
if (event) yield event
}
yield { type: 'done', content: '' }
} catch (error) {
const msg = error instanceof Error ? error.message : 'Stream error'
yield { type: 'error', content: msg }
} finally {
clearTimeout(timer)
}
}
return {
stream: generateStream(),
kill: () => {
clearTimeout(timer)
child.kill('SIGTERM')
},
}
}
function buildPrompt(systemPrompt: string | undefined, userPrompt: string): string {
const userText = userPrompt.trim()
if (!systemPrompt?.trim()) return userText
return [
'You are a design generation assistant. Follow the guidelines below to produce the requested output.',
'',
'--- GUIDELINES ---',
systemPrompt.trim(),
'',
'--- TASK ---',
userText,
].join('\n')
}
async function executeGeminiCommand(
binPath: string,
args: string[],
timeoutMs: number,
stdinText?: string,
): Promise<GeminiCliResult> {
return await new Promise((resolve, reject) => {
const child = spawn(binPath, args, {
env: filterGeminiEnv(process.env as Record<string, string | undefined>),
stdio: [stdinText ? 'pipe' : 'ignore', 'pipe', 'pipe'],
...(process.platform === 'win32' && { shell: true }),
})
// Pipe prompt via stdin
if (stdinText && child.stdin) {
child.stdin.write(stdinText)
child.stdin.end()
}
let stdoutBuffer = ''
let stderrBuffer = ''
const timer = setTimeout(() => {
child.kill('SIGTERM')
reject(new Error(`Gemini request timed out after ${Math.round(timeoutMs / 1000)}s.`))
}, timeoutMs)
child.stdout!.on('data', (chunk: Buffer) => {
stdoutBuffer += chunk.toString('utf-8')
})
child.stderr!.on('data', (chunk: Buffer) => {
stderrBuffer += chunk.toString('utf-8')
})
child.on('error', (err) => {
clearTimeout(timer)
reject(err)
})
child.on('close', (code) => {
clearTimeout(timer)
// Parse JSON output — Gemini CLI always outputs a JSON object at the end of stdout.
// Error text / stack traces may appear before it.
const parsed = parseGeminiJsonOutput(stdoutBuffer)
if (parsed) {
if (parsed.response) {
resolve({ text: parsed.response })
return
}
if (parsed.errorMessage) {
resolve({ error: friendlyGeminiApiError(parsed.errorMessage) })
return
}
}
if (code !== 0) {
// Extract meaningful error from stderr or stdout
const errorMsg = extractGeminiError(stdoutBuffer, stderrBuffer)
resolve({ error: errorMsg || `Gemini exited with code ${code ?? 'unknown'}.` })
return
}
const raw = stdoutBuffer.trim()
resolve(raw ? { text: raw } : { error: 'Gemini returned no output.' })
})
})
}
/**
* Parse Gemini CLI JSON output.
* The CLI may print error text before the final JSON object.
* We search from the END of stdout for the last valid JSON block.
*/
function parseGeminiJsonOutput(raw: string): { response?: string; errorMessage?: string } | null {
const trimmed = raw.trim()
if (!trimmed) return null
// Search backwards for the last top-level JSON object (starts with `{` at line beginning)
const lines = trimmed.split('\n')
for (let i = lines.length - 1; i >= 0; i--) {
const line = lines[i].trim()
if (!line.startsWith('{')) continue
// Try to parse from this line to the end
const candidate = lines.slice(i).join('\n').trim()
try {
const data = JSON.parse(candidate) as Record<string, unknown>
// Must have session_id to be a valid Gemini CLI response
if (!data.session_id && !data.response && !data.error) continue
const response = typeof data.response === 'string' ? data.response : undefined
// error can be a string or an object { type, message, code }
let errorMessage: string | undefined
if (data.error) {
if (typeof data.error === 'string') {
errorMessage = data.error
} else if (typeof data.error === 'object' && data.error !== null) {
const errObj = data.error as Record<string, unknown>
errorMessage = typeof errObj.message === 'string' ? errObj.message : JSON.stringify(data.error)
}
}
return { response, errorMessage }
} catch { /* not valid JSON from this point */ }
}
return null
}
/** Extract a human-readable error from Gemini CLI stdout/stderr */
function extractGeminiError(stdout: string, stderr: string): string | null {
// Look for quota errors
const quotaMatch = stdout.match(/quota will reset after (\S+)/i)
|| stderr.match(/quota will reset after (\S+)/i)
if (quotaMatch) {
return `Gemini quota exhausted. Resets after ${quotaMatch[1]}.`
}
// Look for TerminalQuotaError or other named errors
const namedError = stdout.match(/(Terminal\w+Error|ApiError|AuthError):\s*(.+)/m)
if (namedError) {
return namedError[2].trim()
}
// Stderr fallback
const stderrTrimmed = stderr.trim()
if (stderrTrimmed) return stderrTrimmed
return null
}
/** Map raw Gemini API errors to user-friendly messages */
function friendlyGeminiApiError(raw: string): string {
if (/quota|exhausted|429|capacity/i.test(raw)) {
const resetMatch = raw.match(/reset after (\S+)/i)
return resetMatch
? `Gemini quota exhausted. Resets after ${resetMatch[1]}.`
: 'Gemini quota exhausted. Please wait and try again.'
}
if (/401|unauthenticated|auth/i.test(raw)) {
return 'Gemini auth expired. Run "gemini" in your terminal to re-authenticate.'
}
if (/\[object Object\]/.test(raw)) {
return 'Gemini API error. Check your quota or try a different model.'
}
return raw
}
function parseStreamJsonLine(
line: string,
): { type: 'text' | 'error' | 'done'; content: string } | null {
// Skip non-JSON lines (e.g. "Loaded cached credentials.")
if (!line.startsWith('{')) return null
let parsed: Record<string, unknown>
try {
parsed = JSON.parse(line) as Record<string, unknown>
} catch {
return null
}
const type = typeof parsed.type === 'string' ? parsed.type : ''
if (type === 'message' && parsed.role === 'assistant') {
const content = typeof parsed.content === 'string' ? parsed.content : ''
if (content) return { type: 'text', content }
}
if (type === 'result') {
// Check for error in result event
if (parsed.status === 'error' && parsed.error) {
const errObj = parsed.error as Record<string, unknown>
const msg = typeof errObj.message === 'string' ? errObj.message : 'Unknown error'
return { type: 'error', content: friendlyGeminiApiError(msg) }
}
return null
}
if (type === 'error') {
const content = typeof parsed.message === 'string' ? parsed.message : 'Unknown error'
return { type: 'error', content: friendlyGeminiApiError(content) }
}
return null
}