CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[2.15.3] - 2026-05-15

Changed

• CI workflow reverted to strict mode: tag pushes run the full quality job before docker. v2.15.2 had skipped quality on tag pushes to avoid duplicate CI runs, but that left a security gap — a tag pointing at an unvalidated commit (e.g. git tag v9.9.9 some-sha directly) could trigger a Docker push without going through type check / lint / tests. Each release now runs quality twice (once on branch push, once on tag push) but guarantees Docker images are only built from validated commits

[2.15.2] - 2026-05-15

Fixed

• CI: lint failures on mainStarted/warmupCount (unused locals in benchmarkEngine.run.test.ts) and DEFAULT_CONFIRM_DELAY_MS (unused fallback in alertNotifier.ts). Renamed the latter to _DEFAULT_CONFIRM_DELAY_MS and dropped the test locals — verification was already covered by expect(execute).toHaveBeenCalledTimes(...)
• CI workflow: quality job now skips on tag pushes (the same commit was already validated on the branch push). Eliminates the duplicate CI runs that fired on every release — one for push branches:main, one for push tags:v*. The docker job still triggers on tag pushes and no longer depends on quality (the underlying commit was already validated)

[2.15.1] - 2026-05-15

Fixed

• CI: supertest and @types/supertest moved from root package.json to backend/package.json. Local tests passed because TypeScript resolves up the directory tree to root node_modules, but CI installs each sub-package independently and could not find the module. Type check now passes in CI

[2.15.0] - 2026-05-15

Changed

• Alert confirmation now uses K-of-N voting instead of "N consecutive failures or one ok abandons cycle". A new alertConfirmFailThreshold setting (default N - 1, e.g. 4-of-5) controls how many of the N attempts must fail to fire an alert. A single transient ok no longer drops the entire confirmation chain — fixes the case where a flaky upstream that returns one healthy response between failures suppressed real outage alerts for 30+ minutes
• Confirmation cycles now exit early on both directions: alert fires the moment failCount reaches the threshold (no longer waits for the full N attempts), and the cycle abandons the moment failThreshold becomes mathematically unreachable
• Health-check probe timeout reduced from 180 s (streaming) / 120 s (non-streaming) to 90 s for monitor probes and the "Test Connection" endpoint. Playground/benchmark calls retain their longer timeouts
• Confirmation probes within the same provider now run in parallel (independent API calls). Previously serialized — a single hung confirmation could delay every other model's confirmation in the same minute

Fixed

• Race in confirmation queue: between pendingConfirmations.delete(key) and the re-add after await confirmProbe, the dedup gate has(key) returned false, allowing a scheduled probe landing in that 60-180 s window to spawn a duplicate parallel confirmation cycle. Added an inFlight token map so the gate covers the await window, and any cycle whose token is overwritten/cleared mid-await drops its result
• Recovery alert no longer leaves a zombie down cycle in flight: when a scheduled probe sees healthy/slow after down, the recovery alert now explicitly cancels any pending or in-flight confirmation for that target, preventing a delayed redundant down alert
• Webhook delivery failures now retry instead of silently consuming the alert (bug #4): when the Feishu webhook returns non-2xx, sendFeishuAlert throws instead of just logging. The fire path catches the throw and re-queues the confirmation cycle rather than recording lastAlertAt — previously a failed delivery still recorded the alert, suppressing all retries for 6 hours
• useMonitor.saveConfig no longer "phantom-saves" (bug #8): UI no longer mirrors the new config into local state on non-2xx responses. Returns boolean so callers can detect failures
• usePlaygroundHistory no longer "phantom-deletes" (bug #2): deleteEntry and clearAll now check res.ok before mutating local state — failed server deletions no longer hide items locally
• useWorkflow mutations surface server errors (bug #1): cancelWorkflow / deleteWorkflow / duplicateWorkflow now check res.ok and propagate the server's error message into state.error instead of silently returning false/null
• useBenchmark rejects malformed responses (bug #3): fetchBenchmarks validates that the body is an array, fetchBenchmark validates it's a plain object. Non-2xx and shape mismatches set an error rather than polluting React state with {error:'…'} placeholders
• PUT /api/monitor/targets now accepts an empty array (bug #7): MonitorTargetsArraySchema dropped .min(1), letting users clear the monitor list entirely
• startWorkflow correctly toggles isRunning (bug #6): set true at the start of the try-block so the catch-branch's setIsRunning(false) is no longer a no-op
• providerStore.create / update reject duplicate model id/name within a provider (bug #5): collisions previously corrupted monitor target tracking. Routes return 400 with the conflict message

Added

• New monitor config field alertConfirmFailThreshold (range 1-20, clamped to [1, alertConfirmCount] server-side)
• Settings UI exposes the K threshold as a K / N selector that auto-adjusts options when N changes
• Comprehensive test coverage expansion: 906 total tests (712 backend + 194 frontend) covering alert state coordination, K-of-N decision math, multi-provider streaming token fields, route HTTP semantics via supertest, store CRUD with sqlite migrations, and full executeWorkflow integration with real benchmarkEngine
• "Writing tests" discipline section in CLAUDE.md capturing the lesson from the May 2026 reverse-review: 8 bugs were silently rationalized by tests that matched current behavior instead of expected behavior

[2.14.0] - 2026-05-13

Added

• Configurable alert confirmation: number of consecutive failures (default 5, range 1-20) and delay between checks (default 1 min, range 1-60) before sending alerts, replacing the previous fixed single 1-minute re-check
• Monitor settings UI exposes confirm count and confirm delay alongside language and reminder interval

Fixed

• Alert reminder interval ignored: every save of monitor settings was wiping last_alert_at because setTargets/addTarget rebuilt the row without preserving the column, so reminders fired roughly every probe interval instead of every 6 hours
• Down/very_slow status oscillation triggered spurious "new failure" alerts instead of reminders; wasDown now treats both as the same down state
• Backend dev watcher missed source edits made by atomic-replace writes (inode changes); switched from tsx watch to nodemon --legacy-watch polling
• Frontend dev watcher hardened with usePolling for parity
• PUT /api/monitor/config silently dropped alertConfirmCount and alertConfirmDelayMinutes from the request body, so UI changes were not persisted
• Alert confirmation probe now records a ping on error (previously failed probes left no DB trace) and re-queues on transient failures instead of silently dropping the confirmation

[2.13.1] - 2026-05-11

Changed

• Alert confirmation: down/reminder alerts now require a second probe after 1 minute to reduce false positives
• Recovery alerts are still sent immediately without confirmation
• Switch docker-compose.yml to use Docker Hub image (idemerge/llm-api-bench)
• Remove unused variables flagged by code quality analysis

[2.13.0] - 2026-05-11

Added

• Full i18n support with Chinese/English language switcher (react-i18next)
• Feishu webhook alert notifications for monitor
  • Per-target alert enable/disable toggle
  • Status change detection: new failure, repeated failure (configurable interval), recovery
  • DB-persisted alert state (survives restarts)
  • Optional webhook signature verification
  • Configurable notification language (en/zh, default en)
• Alert bell indicator on monitor model cards (color-coded by health status)

Changed

• All hardcoded UI strings replaced with i18n translation keys
• Monitor settings modal now includes alert configuration section (webhook URL, secret, language, reminder interval)

[2.12.1] - 2026-04-28

Fixed

• Touch targets undersized: removed size="small" from Settings buttons, increased model tag padding
• Heading scale too flat: increased H1 from 20px to 24px
• Capability tags (T/S/V) nearly illegible: increased font from 8px to 10px with larger padding
• Mobile parameter labels overflow: responsive grid for Core Parameters section
• Playground history panel overlaps form on mobile: full-screen overlay on mobile
• Grammar: "1 models" now correctly pluralized across Monitor and History pages
• antd deprecation: replaced Alert message prop with title (5 instances)
• History page duplicate heading: removed redundant H2 title (topbar already shows page name)

[2.12.0] - 2026-04-28

Added

• Naming validation rules for Provider name, Model ID, and DisplayName (backend + frontend)
  • Provider name: alphanumeric/dash/underscore, no spaces, 1-64 chars
  • Model ID: alphanumeric/dash/underscore/dot/slash, 1-64 chars (LiteLLM compatible)
  • DisplayName: alphanumeric/space/dash/underscore/dot, 1-64 chars
• Frontend real-time validation with error hints on Settings provider form
• Frontend validation unit tests (16 cases)
• Backend validation boundary tests (4 cases)

Changed

• Renamed project from LLM API Radar to LLM API Bench (repo, UI, docs, Docker image)
• Playground history sidebar now shows ProviderName/DisplayName instead of raw model ID
• Backend stores model displayName in playground history for friendly display
• Adaptive QuickButtons sizing: auto-shrink when >7 options to prevent line wrapping

Fixed

• Getting Started hint no longer flashes on page refresh (waits for data load)
• Playground provider/model selectors no longer flash raw IDs before names load
• Playground history correctly resolves model displayName from provider data

[2.11.3] - 2026-04-28

Changed

• Raised max concurrency from 1000 to 5000 (frontend InputNumber, backend validation schemas, route caps)
• Raised max iterations from 1M to 10M (frontend InputNumber, backend validation schemas, route caps)
• Added quick-select buttons for 2K/5K concurrency and 5M/10M iterations
• Updated README (EN/CN) with corrected cd path and new concurrency/iterations limits
• Fixed Quick Start instructions: cd llm-benchmark → cd llm-api-bench

[2.11.2] - 2026-04-27

Changed

• Demo mode now masks vendor-prefixed model names (e.g. z-ai/glm-4.7 → ProviderX/glm-4.7) and workflow providerSummaries, sharing a single id-stable letter namespace across providers and vendors
• Masking is fully applied at the React hook fetch boundary (useWorkflow, useMonitor, usePlaygroundHistory, useProviders); the legacy DOM regex redactor is now a deprecated no-op safety net
• Regenerated all 6 README screenshots and docs/demo.gif under VITE_DEMO_MODE=true

Fixed

• Workflow result table no longer leaks raw provider names through summary.providerSummaries[*].provider (previously masked only by the DOM regex layer)

[2.11.1] - 2026-04-27

Added

• Sensitive info redaction module (scripts/redact-sensitive.mjs) for screenshots and GIF recording — provider names, API URLs, and keys are automatically replaced with generic labels
• Screenshot script (take-screenshots.mjs) now calls redactPage() before each capture
• Demo recorder (record-demo.mjs) installs a persistent MutationObserver to redact text as React re-renders during screencast

Fixed

• Playground: disable image upload button for non-vision models and clear uploaded images when switching to a non-vision model
• Workflow SSE: fix race condition where activeRunIdRef was cleared after fetchWorkflow, causing stale state — now fetches final workflow state directly before clearing ref

Changed

• Regenerated all 6 screenshots and demo GIF with redacted sensitive information
• Removed prettier from frontend and backend devDependencies (unused)

[2.11.0] - 2026-04-27

Added

• Workflow page now shows the same Mission Control header (status, duration, edit) and Live Metrics strip (avg RT, TPS, last RT) + cooldown countdown that History Detail had — exposed via liveMetrics and cooldown from useWorkflow
• New shared WorkflowHeader component used by both the active Workflow page and History Detail

Changed

• Refactored HistoryDetailPage to compose WorkflowHeader instead of duplicating header markup (~280 line reduction)
• Tightened pre-commit lint gate: lint-staged now runs eslint --max-warnings 0 on staged frontend files

Fixed

• CI lint failure on v2.10.1: removed empty catch {} block in HistoryDetailPage and silenced react-hooks warnings via targeted disables (no behavior change)
• Various react-hooks lint warnings across ConfigPanel, MonitorPage, PlaygroundPage, WorkflowConfigPanel, WorkflowProgress

[2.10.1] - 2026-04-23

Changed

• Raised concurrency limit from 200 to 1000 and iterations limit from 2000 to 1M (frontend InputNumber + backend Math.min caps)
• Updated quick-select buttons: concurrency adds 500 and 1K options, iterations adds 10K, 100K, and 1M options

[2.10.0] - 2026-04-23

Added

• Workflow name inline editing with PATCH endpoint and edit UI in History Detail header
• Running workflow "Mission Control" experience: live metrics strip (avg RT, TPS, last RT), cooldown countdown timer between tasks, real-time elapsed timer
• Completed workflow stat-card dashboard: Duration, Tokens, Best Avg RT, Success Rate, Total T/s in a 6-column grid
• History list redesign: colored status icons, config chips (concurrency × iterations × tokens + cache rate + stream), dedicated Models column with provider-colored tags, Duration and Tokens columns
• Monitor Settings as Modal dialog (replaces inline collapsible panel) with scrollable Targets area
• CSS design system additions: stat-card / stat-value / stat-label, section-header with color variants, running-card-glow animation, running-row-active styling, Ant Design overrides for tables, tooltips, and popconfirm

Changed

• History Detail running state: animated amber border glow, live metrics from SSE latestResults, per-task completed summaries showing fastest RT and highest TPS providers
• History Detail completed state: stat-card grid replaces flat text metrics for visual impact
• History Panel: complete rewrite with richer row content and consistent visual hierarchy
• Monitor: summary bar uses stat-card CSS class, threshold inputs use Ant Design InputNumber, chart tooltip uses CSS variables, removed redundant tok/s display, unified TTFT/TPS status coloring
• Playground: MetricsRow uses stat-card with provider-colored accents, provider label uses getProviderColor
• WorkflowResults: removed bar charts (MetricBarChart, TaskCharts) — cleaner table-only layout
• WorkflowProgress: added live metrics strip, cooldown timer, elapsed timer, completed task summary pills
• WorkflowConfigPanel: cache hit rate input width narrowed for compact layout

[2.6.0] - 2026-04-23

Added

• Output Scope selector for long-context presets (16K/64K/150K/256K): controls how many documents the model reads, limiting output length (~500 tokens for 3 docs, unlimited for All docs)
• Output Scope available in Benchmark, Workflow, and Playground pages with persistent selection via localStorage
• Input/Output/Total throughput metrics in Workflow Detail: calculated as concurrency × avg tokens per request / avg response time
• Throughput columns (In T/s, Out T/s, Total T/s) in provider comparison tables
• Throughput summary in workflow header and results summary bar
• Tooltips on all metric labels, table column headers, and parameter controls across all pages (WorkflowResults, ResultsPanel, ConfigPanel, PlaygroundPage, HistoryDetailPage)

Changed

• Long-context 64K preset prompt suffix updated to support configurable output scope

[2.5.1] - 2026-04-19

Added

• Workflow task editor: duplicate button to clone an existing task with all its configuration

Fixed

• Duplicating, deleting, or reordering tasks now correctly preserves heavy prompts (>10K chars) instead of silently truncating them

[2.5.0] - 2026-04-18

Added

• History list: show concurrency and iteration count columns
• History detail: show input/output token counts and ratio (In:Out)
• History detail: real-time iteration progress bar for running workflows via SSE
• Backfill input/output token stats for older workflows on first access

Changed

• Long context preset prompts: balanced for ~40:1 input-to-output token ratio with "Don't overthink this" guidance

[2.4.6] - 2026-04-18

Fixed

• Cache hit rate: reduced sliding window from concurrency-sized (e.g. 50) to fixed 5, keeping KV cache memory pressure realistic for large prompts

[2.4.5] - 2026-04-18

Fixed

• Cache hit rate: reuse now picks from a sliding window of recent prefixes (sized to concurrency) instead of the entire pool, avoiding stale entries that inference engines (SGLang, vLLM) may have evicted under memory pressure

[2.4.4] - 2026-04-18

Changed

• Cache hit rate: replaced fixed-K-prefixes + shuffle with per-request Bernoulli scheduling — each request independently rolls miss/hit with the target probability, producing a uniform distribution throughout the run instead of clustering all misses at the start

[2.4.3] - 2026-04-18

Fixed

• Cache hit rate: targetCacheHitRate was silently dropped by both the benchmark and workflow route handlers — the field was validated but never passed to the engine, so the feature had no effect

[2.4.2] - 2026-04-18

Fixed

• Cache hit rate: prefix size now adapts to prompt length (~5%, clamped 128–4096 chars) to avoid inflating short prompts — previously a fixed ~4 KB prefix would double a 1K-token input

[2.4.1] - 2026-04-18

Fixed

• Cache hit rate: replaced short UUID prefix (~5 tokens) with ~1024-token random prefix to reliably bust block-level KV cache on inference engines (vLLM, SGLang, etc.)
• Cache hit rate: replaced round-robin variant assignment with Fisher–Yates shuffled schedule so cache misses are spread evenly across the run instead of clustered at the start

[2.4.0] - 2026-04-18

Added

• Cache hit rate control (targetCacheHitRate): prepends unique UUID prefixes to each request to simulate realistic multi-user traffic with configurable prefix-cache hit rate (0–99%). Available as a toggle + percentage input in the WorkflowConfigPanel Advanced section. Formula: K = iterations × (1 − rate) unique variants, cycled round-robin.

[2.3.0] - 2026-04-18

Changed

• Raised concurrency limit from 50 to 200 and iterations limit from 1000 to 2000
• Replaced batch-based concurrency with sliding-window worker pool to maintain steady in-flight request count — previously, requests that completed early left slots idle causing actual concurrency to drop over time; now a new request starts immediately whenever one finishes

[2.2.0] - 2026-04-17

Added

• Long Context 150K preset: a new built-in prompt preset (~150,000 tokens) bridging the gap between the existing 64K and 256K presets. Available in ConfigPanel, WorkflowConfigPanel, and PlaygroundPage. Loaded on demand via dynamic import to avoid bundle size impact.

[2.1.0] - 2026-04-16

Added

• maxQps parameter for workflow tasks: global token bucket rate limiting across all concurrent slots. Set to a positive integer to cap requests per second; 0 means unlimited. Available in the WorkflowConfigPanel Advanced section with quick-select buttons (Off / 1 / 5 / 10).
• Token bucket implementation in the benchmark engine with cancellation support — rate limiting integrates cleanly with existing cancel flow and does not affect requestInterval or concurrency behavior.

[2.0.0] - 2026-04-13

Security

• Eliminated all hardcoded secrets: JWT secret, encryption key, and salt are now auto-generated and persisted to data/ directory
• Force password change on first login with default credentials (changeme)
• Restricted CORS to configured origin (default: same-origin only)
• Added login rate limiting (5 attempts per 5 minutes per IP)
• Added Helmet security headers with Content Security Policy
• Moved auth verify and change-password endpoints behind authentication middleware
• Replaced JWT-in-query-string with short-lived one-time tokens for SSE and download URLs
• New password must differ from current password when changing
• JWT token storage moved from localStorage to sessionStorage

Added

• Zod schema validation for all API request bodies with descriptive error messages
• ProviderConfigUpdateSchema for partial provider updates
• POST /api/auth/change-password endpoint
• POST /api/auth/sse-token endpoint for one-time token exchange
• Shared SQLite connection singleton with WAL mode and busy timeout
• CSV escaping utility to prevent injection in exports
• Express Request type augmentation (req.user)
• 5 new test suites: encryption, auth middleware, validation schemas, benchmark engine, store sync (65 backend tests total)

Changed

• SQLite-first write pattern across all stores: DB writes before in-memory Map updates to prevent inconsistency on failures
• Encryption migration runs synchronously before server startup to prevent race conditions
• Monitor cleanup runs daily at 3am with 7-day retention
• cancelledRuns cleanup in benchmark engine on both success and error paths
• PRAGMA table_info migration pattern replaces try/catch ALTER TABLE
• apiKeys field in benchmark and workflow schemas is now optional with empty default
• supportsVision/supportsTools use nullish coalescing (??) instead of logical OR

Fixed

• Encryption migration race condition — server could accept requests before migration completed
• store.delete() violated SQLite-first pattern (deleted Map before DB)
• PUT /api/providers/:id had no input validation
• Monitor error responses returned HTTP 200 instead of 500
• Provider test endpoints could crash the server on connection failure (now returns 502)
• Workflow error recovery could overwrite cancellation status
• Frontend infinite re-render loop on History page with running workflows
• EventSource not cleaned up on component unmount in useWorkflow hook
• Missing pageConfig fallback for unknown routes
• Redundant method/action attributes on login form

Removed

• Dead code: backend/src/services/store-old.ts

[1.3.3] - 2026-04-13

Changed

• Merged Docker publish into CI pipeline — Docker image build now requires all quality checks to pass first
• CI workflow also triggers on version tags so quality gate runs before Docker push

Fixed

• CI lint warnings: cleaned up unused imports/variables across frontend and backend
• Upgraded GitHub Actions to v5 (Node.js 24 compatible)
• Root docs (CHANGELOG, README, docker-compose) excluded from Prettier formatting
• lint-staged glob expanded to cover config .js files

[1.3.2] - 2026-04-13

Added

• Vitest test framework with initial test suites (frontend: tokenCount, MonitorPage helpers; backend: MonitorStore CRUD)
• Prettier for consistent code formatting across frontend and backend
• ESLint for backend (flat config, typescript-eslint)
• Husky pre-commit hook with lint-staged (auto-format on commit)
• GitHub Actions CI pipeline: typecheck, lint, format check, tests, build
• scripts/release.sh for automated version bump, changelog update, tag, and push

Fixed

• 11 TypeScript type errors across the codebase (ResultsPanel JSX, App.tsx, WorkflowResults, etc.)
• CI error messages now show specific failures and actionable fix instructions

[1.3.1] - 2026-04-10

Added

• Monitor trend charts: expandable TTFT, TPS, and Latency time-series graphs per model with 1h/6h/24h range selector and threshold reference lines

Fixed

• Playground history: clicking a failed history entry no longer crashes (undefined metrics guard)
• Playground history: selecting a record whose provider was deleted no longer causes a blank screen (graceful fallback to empty provider selection)

[1.3.0] - 2026-04-09

Added

• Monitor health classification now based on TPS (tokens per second) instead of raw latency
• Monitor probe prompt upgraded to generate longer responses for accurate TPS measurement
• Provider deletion cascades to monitor targets cleanup
• Provider model rename auto-syncs monitor targets (preserves monitoring config)
• Anthropic streaming fallback: read input_tokens from message_delta for LiteLLM compatibility

Changed

• Monitor health thresholds: latencySlowMs/latencyVerySlowMs replaced with tpsSlowThreshold (default 20) / tpsVerySlowThreshold (default 5)
• Monitor UI tooltips show TPS instead of latency as primary metric
• Workflow templates default warmupRuns changed from 2 to 0
• Quick Benchmark and Workflow config default warmupRuns changed from 2 to 0

Fixed

• formatNumber crash when token values are undefined (WorkflowResults page)
• Workflow detail table showing Tokens as 0 (field mismatch: promptTokens vs totalTokens)
• Orphaned monitor targets remaining after provider model rename or deletion

[1.2.1] - 2026-04-07

Changed

• Playground image input redesigned: inline button at prompt bottom, drag-and-drop, clipboard paste support (matching ChatGPT/Claude UX)
• Playground Run button moved to prompt textarea bottom-right for faster access
• Config row (Max Tokens, Streaming, Thinking) moved above prompt area
• Presets integrated into prompt bottom bar
• Removed URL image input (file upload only)

[1.2.0] - 2026-04-07

Added

• Playground history with SQLite persistence, auto-save on every run
• History sidebar with replay: click any past run to restore prompt, config, and response
• Thinking/reasoning toggle for Anthropic extended thinking and OpenAI reasoning effort
• Copy response button in Playground
• Long Context presets (8K/16K/32K/64K/128K) in Playground
• Backend image validation (size and count limits)

Fixed

• Anthropic extended thinking not working (wrong API version, missing thinking params)
• Anthropic required CLI headers accidentally removed
• Playground upstream API calls not aborted when client disconnects (resource leak)
• Gemini using fake conversation turns instead of native systemInstruction
• Playground /run returning HTTP 200 on errors instead of 502
• Image URLs silently dropped for Anthropic/Gemini (now fetched and converted to base64)
• Backend maxTokens default misaligned (512 vs 4096)
• Flash of empty state on Monitor, Config, and Workflow pages before data loads

Changed

• History sidebar defaults to open for better discoverability
• Playground design polish: label sizes, config row layout, mobile responsiveness

[1.1.0] - 2026-04-04

Added

• Gemini streaming support (streamGenerateContent with alt=sse) for accurate TTFT measurement
• Gemini streaming in Playground with real-time token output
• History page refresh button for manual data reload
• Auto-refresh History page every 30s when running workflows exist
• Reload workflow data when navigating to History page

Fixed

• Gemini TTFT always showing 0 due to missing streaming implementation
• Playground non-streaming TTFT showing fabricated value (responseTime * 0.3) instead of N/A
• Playground Gemini image input not being passed to API (images were silently dropped)
• Playground Gemini format falling back to non-streaming instead of using native streaming
• Frontend TTFT displaying 0ms instead of N/A for non-streaming requests (WorkflowResults, ResultsPanel, PlaygroundPage)
• Non-streaming /run endpoint not passing images parameter to provider

Changed

• backend/public/ added to .gitignore (build artifact)

[1.0.3] - 2026-04-04

Changed

• Renamed project from LLM Benchmark to LLM API Radar
• Updated all UI references, branding, screenshots, and demo GIF

[1.0.0] - 2026-04-04

Initial release.

Features

• Multi-provider benchmark engine (OpenAI, Anthropic, Gemini, OpenAI-Compatible)
• Workflow engine with multi-task sequential execution
• Per-task prompt, concurrency, and iteration configuration
• Warmup runs to eliminate cold-start bias
• Live streaming metrics with per-provider area charts
• Radar comparison across all dimensions
• Persistent run history with full result details
• JSON and CSV export
• Playground page with streaming, vision support, and image upload
• Monitor page with periodic health checks, configurable thresholds, and 24h history
• JWT-based authentication with configurable credentials
• Docker deployment with multi-stage alpine build
• One-click build script (start.sh) and Docker Compose
• GitHub Actions workflow for Docker Hub auto-publish on tag push
• Dark theme UI with Ant Design 5
• SQLite storage with single-file database