Skip to content

CHANGELOG.md

Latest commit

 

History

History
354 lines (290 loc) · 20.2 KB

CHANGELOG.md

File metadata and controls

354 lines (290 loc) · 20.2 KB

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.

Unreleased

[1.6.0] - 2026-03-17

Added

  • Transparent reasoner routing: all deepseek-reasoner requests are now routed through deepseek-chat + thinking mode for full feature support including function calling
  • Model-aware parameter filtering prevents sending unsupported sampling params to thinking mode

Fixed

  • Function calling now works with deepseek-reasoner (via transparent routing to deepseek-chat + thinking)
  • Sampling parameters (temperature, top_p, frequency_penalty, presence_penalty) properly filtered for reasoning requests
  • Thinking parameter now passed as top-level property instead of extra_body (OpenAI SDK v6 compatibility)
  • Corrected model capabilities in deepseek://models resource

Changed

  • deepseek-reasoner capability list updated: added function_calling, removed thinking_mode
  • Both deepseek-chat and deepseek-reasoner run DeepSeek V3.2. The reasoner model is now transparently handled as deepseek-chat with thinking: {type: "enabled"}, which provides identical reasoning behavior plus function calling support

1.5.2 - 2026-03-11

Fixed

  • Worker hang on non-POST requests to /mcp endpoint. External GET requests caused the stateless MCP transport to hang indefinitely, triggering Cloudflare Worker "code hung" errors. Now returns 405 for non-POST methods.

Changed

  • Worker VERSION synced to 1.5.2
  • Added websiteUrl to server.json

1.5.1 - 2026-03-09

Fixed

  • Update @modelcontextprotocol/sdk to ^1.27.1 to patch cross-client data leak vulnerability (CVE-2026-25536)
  • Update wrangler to ^4.71.0 to patch OS command injection vulnerability (CVE-2026-0933)

Changed

  • Circuit breaker is now per-model instead of shared across all models. If one model becomes unhealthy, the other remains available independently.

1.5.0 - 2026-03-07

Added

  • Remote Endpoint: Hosted BYOK (Bring Your Own Key) endpoint at https://deepseek-mcp.tahirl.com/mcp. Users send their own DeepSeek API key as Bearer token — no server-side key stored.
  • Cloudflare Worker: Stateless MCP server on Cloudflare Workers edge network. Zero cold start, global distribution, free tier (100K requests/day).
  • Remote Quick Start: Claude Code, Cursor, Windsurf can connect to the remote endpoint with --transport http — no npm install or Node.js required.

Changed

  • README restructured: Remote (no install) as primary quick start, local stdio as secondary.
  • server.json: Added streamableHttp transport package for remote endpoint.
  • llms.txt/llms-full.txt: Updated with remote endpoint documentation.

1.4.3 - 2026-03-07

Fixed

  • MCP Registry: Shortened server.json description to meet 100-char limit

1.4.2 - 2026-03-07

Added

  • Streamable HTTP Transport: New TRANSPORT=http mode with Express-based Streamable HTTP server. Per-session McpServer instances with shared DeepSeekClient. Supports POST/GET/DELETE on /mcp endpoint and SSE streaming.
  • Health Endpoint: GET /health returns server status, version, uptime, and transport type.
  • Docker Support: Multi-stage Dockerfile, docker-compose.yml, and .dockerignore for containerized deployment. Default transport is HTTP in Docker.
  • Docker CI: GitHub Actions job for Docker build smoke test with health endpoint verification.
  • New Config Variables: TRANSPORT (stdio|http, default: stdio), HTTP_PORT (default: 3000).
  • 253 Tests: Up from 241, with 12 new tests covering HTTP transport endpoints, config transport/port parsing, and session lifecycle.

Changed

  • Entry point (index.ts) now branches on config.transport — stdio (default) or HTTP mode.
  • CI dist check includes transport-http.js.

1.4.1 - 2026-03-07

Added

  • MCP Registry: mcpName for official MCP Registry (registry.modelcontextprotocol.io) publishing
  • Glama Registry: glama.json for glama.ai server listing
  • Smithery Compatibility: CJS bundler fallback in server.ts, createSandboxServer() export

Changed

  • server.json: Updated to official MCP Registry schema format
  • README: Added Glama badge

1.4.0 - 2026-03-07

Added

  • Model-Aware Pricing: MODEL_PRICING map with per-model pricing, getPricing(model?) helper. calculateCost() now accepts optional model parameter for accurate per-model cost calculation.
  • Multimodal Content Types: TextContentPart, ImageContentPart, ContentPart types. Message content now supports string | ContentPart[] (OpenAI-compatible format).
  • Multimodal Schema Validation: ContentPartSchema (discriminated union), ContentSchema, ExtendedMessageSchema with array content support.
  • ENABLE_MULTIMODAL Config: Feature flag (default: false) to enable multimodal input. Rejects array content when disabled with clear error message.
  • getTextContent() Helper: Extracts text from string | ContentPart[] for validation and JSON mode checks.
  • MCP Registry Metadata: server.json file for official MCP Registry submission with tools, resources, prompts, and configuration metadata.
  • 241 Tests: Up from 212, with 29 new tests covering model-aware pricing, multimodal schemas, content part validation, getTextContent, multimodal guard, and array content validation.

Changed

  • Flexible Fallback Chain: FALLBACK_ORDER (Record<string, string[]>) replaces FALLBACK_MODEL — supports 3+ model fallback chains for future model additions.
  • Models Resource: buildModelsData() now uses getPricing() from cost.ts instead of static pricing constants.
  • Tool Handler: calculateCost() now receives response.model for model-aware pricing. validateMessageLength uses getTextContent() for multimodal content.
  • Config Resource: Exposes enableMultimodal field.

1.3.3 - 2026-03-07

Added

  • Streaming Fallback Tests: 4 new tests covering streaming fallback success, both-models-fail, non-retryable skip, and disabled fallback scenarios
  • Configurable Circuit Breaker: CIRCUIT_BREAKER_THRESHOLD (default: 5) and CIRCUIT_BREAKER_RESET_TIMEOUT (default: 30000ms) environment variables
  • Session Message Limit: MAX_SESSION_MESSAGES (default: 200) environment variable prevents unbounded memory growth with sliding window eviction
  • 212 Tests: Up from 208, covering streaming fallback edge cases

Fixed

  • Session Tool Calls: ChatMessage interface now includes tool_calls field — multi-turn function calling no longer loses tool context between requests
  • Default Model Config: DEFAULT_MODEL env variable now correctly applies to the deepseek_chat tool schema default value
  • Double Config Load: Removed redundant loadConfig() call in index.ts — config is now loaded once and retrieved with getConfig()
  • Circuit Breaker Hardcoded Params: Circuit breaker threshold and reset timeout are now read from config instead of being hardcoded (5, 30000)
  • Config Resource: deepseek://config resource now exposes circuitBreakerThreshold, circuitBreakerResetTimeout, and maxSessionMessages fields

1.3.2 - 2026-03-06

Changed

  • OpenAI SDK v6: Upgraded from v4.104.0 to v6.27.0. No breaking changes for our usage — chat.completions.create() API unchanged.
  • Zod v4: Upgraded from v3.25.76 to v4.3.6. Fixed z.record() call for v4 compatibility.
  • New Config: DEFAULT_MODEL env variable (default: deepseek-chat) for configurable default model.
  • Keywords: Added gemini-cli, mcp-server to npm package keywords.

1.3.1 - 2026-03-06

Added

  • Test Coverage: deepseek_sessions tool tests (8 tests, 0% → 90%+ coverage)
  • Test Coverage: Tool registration aggregator tests (2 tests)
  • 208 Tests: Up from 198, covering session tool actions and tool registration

Fixed

  • Security: Updated transitive dependencies via npm audit fix — resolved 5 vulnerabilities (3 high) in hono, @hono/node-server, rollup, ajv, qs

1.3.0 - 2026-03-04

Added

  • Multi-Turn Sessions: session_id parameter on deepseek_chat for multi-turn conversations. Session history is stored in memory and automatically prepended to requests. Sessions have configurable TTL and max count.
  • Session Management Tool: deepseek_sessions tool with list, delete, and clear actions for managing active sessions.
  • Circuit Breaker: Protects against cascading API failures. After 5 consecutive failures, fast-fails for 30 seconds, then probes for recovery.
  • Model Fallback: Automatic fallback between deepseek-chat and deepseek-reasoner on retryable errors (429, 503, timeout). Configurable via FALLBACK_ENABLED.
  • MCP Resources: 3 read-only resources following MCP Resources spec:
    • deepseek://models: Model list with capabilities, context limits, and pricing
    • deepseek://config: Current server configuration (API key masked)
    • deepseek://usage: Real-time usage statistics (requests, tokens, costs, sessions)
  • Usage Tracker: Global usage statistics tracking across all requests.
  • New Error Classes: FallbackExhaustedError, CircuitBreakerOpenError for resilience error handling.
  • New Config: SESSION_TTL_MINUTES (default: 30), MAX_SESSIONS (default: 100), FALLBACK_ENABLED (default: true).
  • 198 Tests: Up from 150, covering sessions, circuit breaker, fallback, and MCP resources.

Changed

  • deepseek-client.ts now wraps API calls in circuit breaker and supports automatic model fallback.
  • deepseek_chat tool description updated to mention sessions, fallback, and circuit breaker.
  • tools/index.ts now registers both deepseek_chat and deepseek_sessions tools.
  • index.ts now registers MCP resources via registerAllResources().

1.2.0 - 2026-02-26

Added

  • Thinking Mode: Enable thinking on deepseek-chat with thinking: {type: "enabled"} parameter. Automatically filters incompatible params (temperature, top_p, etc.) with logged warnings.
  • JSON Output Mode: Structured JSON responses with json_mode: true. Supported by both models. Warns if "json" word missing from prompt.
  • Cache-Aware Cost Tracking: V3.2 API returns prompt_cache_hit_tokens and prompt_cache_miss_tokens. Cost display now shows cache hit ratio and savings.
  • CostBreakdown Interface: calculateCost() returns structured {inputCost, outputCost, totalCost, cacheHitRatio?, cacheSavings?} instead of flat number.

Changed

  • DeepSeek V3.2 Pricing: Unified pricing for both models — cache hit $0.028/1M, cache miss $0.28/1M, output $0.42/1M (replaces old per-model pricing).
  • max_tokens Limit: Updated from 32768 to 65536 (reasoner max). Model-specific warnings: deepseek-chat warns above 8192, deepseek-reasoner warns above 65536.
  • Tool Description: Updated to mention V3.2, thinking mode, JSON mode, and cache-aware cost tracking.
  • 150 Tests: Up from 126, covering thinking mode, JSON mode, cache tokens, and cost breakdown.

Fixed

  • Incorrect Cost Reports: Old flat pricing ($0.14/$0.28 chat, $0.55/$2.19 reasoner) replaced with accurate V3.2 unified pricing.

1.1.1 - 2026-02-11

Added

  • Custom Error Classes (src/errors.ts): BaseError, ConfigError, ApiError, RateLimitError, AuthenticationError, ValidationError, ConnectionError with error cause chaining
  • DeepSeek Type Extensions (src/types.ts): DeepSeekRawResponse, DeepSeekStreamChunk, DeepSeekStreamDelta types and hasReasoningContent(), getErrorMessage() type guards
  • Message Content Length Limit: MAX_MESSAGE_LENGTH config (default: 100K chars) prevents excessive API costs
  • Optional Connection Test: SKIP_CONNECTION_TEST=true env skips startup API call for faster boot
  • AI Discoverability: llms.txt and llms-full.txt for LLM/AI agent consumption
  • New Tests: 126 tests (up from 85) covering errors, server factory, tool handlers, prompt registration

Changed

  • Modular Architecture: Monolithic index.ts (783 lines) split into focused modules:
    • src/server.ts: McpServer factory with auto-version from package.json
    • src/tools/deepseek-chat.ts: Tool handler (extracted from index.ts)
    • src/tools/index.ts: Tool registration aggregator
    • src/prompts/core.ts: 5 core reasoning prompts
    • src/prompts/advanced.ts: 5 advanced prompts
    • src/prompts/function-calling.ts: 2 function calling prompts
    • src/prompts/index.ts: Prompt registration aggregator
    • src/index.ts: Slim bootstrap (~80 lines)
  • DRY Refactoring (deepseek-client.ts): Extracted buildRequestParams() and wrapError() methods (eliminated code duplication)
  • Type Safety: Replaced 16 any casts with proper DeepSeek type extensions and type guards (error: unknown pattern)
  • Config: process.exit(1) replaced with throw ConfigError for testability
  • Version: Single source of truth from package.json via createRequire (no more manual sync)

Fixed

  • Security: Updated @modelcontextprotocol/sdk to fix cross-client data leak (GHSA-345p-7cg4-v4c7)
  • Security: Fixed hono transitive dependency vulnerabilities (XSS, cache deception, IP spoofing)
  • CI dist check updated for new file structure

1.1.0 - 2026-02-10

Added

  • Function Calling Support: Full OpenAI-compatible function calling via tools and tool_choice parameters
    • Define up to 128 tool definitions with JSON Schema parameters
    • Control tool behavior with tool_choice: auto, none, required, or specific function
    • Tool call results formatted in response with call IDs and arguments
    • Streaming + function calling works together (delta accumulation)
    • tool message role for sending tool results back
  • Centralized Config System (src/config.ts)
    • Zod-validated configuration from environment variables
    • DEEPSEEK_BASE_URL: Custom API endpoint (default: https://api.deepseek.com)
    • SHOW_COST_INFO: Toggle cost display in responses (default: true)
    • REQUEST_TIMEOUT: API request timeout in ms (default: 60000)
    • MAX_RETRIES: Maximum API retry count (default: 2)
  • Test Suite: 85 tests with Vitest
    • Config, Cost, Schemas, Client, and Function Calling tests
    • 80%+ code coverage with v8 provider
    • npm test, npm run test:watch, npm run test:coverage scripts
  • 2 New Prompt Templates (total: 12)
    • function_call_debug: Debug function calling issues
    • create_function_schema: Generate JSON Schema from natural language
  • CI coverage job in GitHub Actions

Changed

  • Project Structure: Modularized codebase
    • src/config.ts: Centralized configuration
    • src/cost.ts: Cost calculation (extracted from index.ts)
    • src/schemas.ts: Zod validation schemas (extracted from index.ts)
  • DeepSeekClient constructor now uses centralized config (no manual apiKey passing)
  • Server version bumped to 1.1.0
  • Updated deepseek_chat tool description to mention function calling

1.0.3 - 2025-02-07

Added

  • Cost tracking for API requests
    • Automatic cost calculation based on token usage
    • USD cost included in response output
    • Cost data available in structuredContent.cost_usd
  • 10 MCP prompt templates for common reasoning tasks
    • debug_with_reasoning: Debug code with step-by-step analysis
    • code_review_deep: Comprehensive code review
    • research_synthesis: Research and synthesize information
    • strategic_planning: Create strategic plans with reasoning
    • explain_like_im_five: Explain complex topics simply
    • mathematical_proof: Prove mathematical statements
    • argument_validation: Analyze arguments for logical fallacies
    • creative_ideation: Generate creative ideas with feasibility analysis
    • cost_comparison: Compare LLM costs
    • pair_programming: Interactive coding assistant
  • Enhanced response format with request information section
    • Token breakdown (prompt + completion)
    • Model name
    • Cost in USD

Changed

  • Updated DeepSeek Reasoner pricing to current rates ($0.55/$2.19 per 1M tokens)

1.0.0 - 2025-01-13

Added

  • Initial release of DeepSeek MCP Server
  • Support for deepseek-chat model
  • Support for deepseek-reasoner (R1) model with reasoning traces
  • Streaming mode support
  • Full TypeScript implementation with type safety
  • OpenAI-compatible API client
  • Comprehensive error handling
  • MCP protocol compliance via stdio transport
  • Tool: deepseek_chat for chat completions
  • Environment variable configuration for API key
  • Detailed documentation and examples
  • MIT License

Features

  • Models:
    • deepseek-chat: Fast general-purpose model
    • deepseek-reasoner: Advanced reasoning with chain-of-thought
  • Parameters:
    • temperature: Control randomness (0-2)
    • max_tokens: Limit response length
    • stream: Enable streaming mode
  • Output:
    • Text content with formatting
    • Reasoning traces for R1 model
    • Token usage statistics
    • Structured response data

Technical

  • Built with @modelcontextprotocol/sdk v1.0.4
  • Uses OpenAI SDK v4.77.3 for API compatibility
  • Zod v3.24.1 for schema validation
  • TypeScript v5.7.3
  • Node.js 18+ required
  • Stdio-based transport for process communication

0.1.0 - Development

Added

  • Initial project setup
  • Basic MCP server structure
  • DeepSeek API integration prototype

Version History

  • 1.5.2 (2026-03-11): Fix worker hang on non-POST /mcp requests, add registry badges
  • 1.5.1 (2026-03-09): CVE patches, per-model circuit breaker
  • 1.5.0 (2026-03-07): Cloudflare Workers remote endpoint (BYOK), deepseek-mcp.tahirl.com
  • 1.4.3 (2026-03-07): MCP Registry description fix
  • 1.4.2 (2026-03-07): Streamable HTTP transport, Docker support, health endpoint, 253 tests
  • 1.4.1 (2026-03-07): MCP Registry, Glama registry, Smithery compatibility
  • 1.4.0 (2026-03-07): Model-aware pricing, multimodal content types, flexible fallback chain, MCP Registry, 241 tests
  • 1.3.3 (2026-03-07): Streaming fallback tests, session tool_calls fix, configurable circuit breaker, session message limit, 212 tests
  • 1.3.2 (2026-03-06): OpenAI SDK v6, Zod v4, DEFAULT_MODEL config
  • 1.3.1 (2026-03-06): Security fixes, session tool tests, 208 tests
  • 1.3.0 (2026-03-04): Sessions, circuit breaker, model fallback, MCP resources, 198 tests
  • 1.2.0 (2026-02-26): DeepSeek V3.2 support — thinking mode, JSON mode, cache-aware pricing, 150 tests
  • 1.1.1 (2026-02-11): Modular architecture, type safety, security fixes, 126 tests
  • 1.1.0 (2026-02-10): Function calling, config system, test suite
  • 1.0.3 (2025-02-07): Cost tracking and prompt templates
  • 1.0.0 (2025-01-13): Initial public release
  • 0.1.0 (Development): Internal development version

Links