CHANGELOG.md

Changelog

All notable changes to the Converse MCP Server project will be documented in this file.

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

3.2.4 (2026-07-30)

Bug Fixes

• triple provider and transport timeouts for long-running agentic models (d59e472)

Dependencies

• deps: bump @lydell/node-pty from 1.2.0-beta.12 to 1.2.0-beta.14 (#58) (75571fd)

3.2.3 (2026-07-26)

Dependencies

• drop the unused ai dependency (b4b6cae)

3.2.2 (2026-07-26)

Dependencies

• deps: bump ai from 6.0.208 to 7.0.37 (#52) (d087fae)
• deps: bump nanoid from 5.1.16 to 6.0.0 (#47) (5c56843)
• update the remaining npm minor and patch releases (c6b8f49)

3.2.1 (2026-07-26)

Dependencies

• deps: bump actions/checkout from 6 to 7 (#34) (fa3bb62)
• deps: bump actions/setup-node from 6 to 7 (#50) (13611a8)

3.2.0 (2026-07-26)

Features

• cli: add -v/--version flag (83bd98d)

3.1.0 (2026-07-25)

Features

• add Claude Opus 5 across Anthropic, Claude SDK, and Copilot providers (7211618)

Bug Fixes

• map Codex reasoning effort onto the model's supported tiers (73de7c6)

Dependencies

• deps: bump the npm-minor-and-patch group with 8 updates (#51) (067f7a7)

Documentation

• add task files for 013 (agy provider), 014 (unified chat tool), 015 (model catalogs) (e25a57c)

3.0.2 (2026-07-17)

Bug Fixes

• restore copilot:codex alias and floor Codex reasoning effort at low (87c04ad)

Dependencies

• deps: bump the npm-minor-and-patch group with 11 updates (#46) (54121bb)

3.0.1 (2026-07-11)

Miscellaneous

• cut 3.0.1 for npm publish (737d403)

3.0.0 (2026-07-11)

⚠ BREAKING CHANGES

• The consensus and conversation MCP tools are removed; their functionality is available as chat modes. The temperature, verbosity, use_websearch, enable_cross_feedback, cross_feedback_prompt, turn_prompt, and singular model parameters are removed from the tool schema.

Features

• merge chat, consensus, and conversation into a unified chat tool with modes (733ce0f)
• modernize provider model catalogs and capabilities (23f0e15)

Bug Fixes

• correct Mistral catalog to match live API (Large is image-capable, 262144 context) (0931754)

Documentation

• describe the unified chat tool and current provider catalogs (7670d28)

2.29.2 (2026-07-10)

Bug Fixes

• require @openai/codex-sdk ^0.144.1 (ae6ef06)

2.29.1 (2026-07-10)

Bug Fixes

• default CODEX_MODEL to gpt-5.6-sol (fc6c4fa)

Documentation

• design: define provider model modernization (3bdf7e7)

2.29.0 (2026-07-09)

Features

• add GPT-5.6 model family (Sol, Terra, Luna) and make it the default (c91c9d4)

Dependencies

• deps: bump the npm-minor-and-patch group with 10 updates (#37) (4bc3576)
• deps: bump the npm-minor-and-patch group with 6 updates (#39) (069b2f0)

2.28.1 (2026-06-22)

Bug Fixes

• copilot: resolve CLI runtime path instead of relying on SDK heuristic (c254dd4)

Dependencies

• deps: bump the npm-minor-and-patch group with 11 updates (#35) (ac9ce43)
• deps: bump the npm-minor-and-patch group with 6 updates (#33) (96afe93)

2.28.0 (2026-06-10)

Features

• replace gemini-cli provider with Antigravity CLI (agy) subprocess provider (a274441)

Miscellaneous

• finish prettier removal in validate script and docs (a72ea67)

2.27.2 (2026-06-10)

Dependencies

• pin claude-agent-sdk 0.3.169 and codex-sdk 0.138.0 for release-age policy (ea59a01)

2.27.1 (2026-06-10)

Dependencies

• deps: bump @github/copilot-sdk to 1.0.0 and update minor dependencies (#28) (6ca4d71)

2.27.0 (2026-06-10)

Features

• add Claude Fable 5 to Anthropic and Claude providers (a677c33)

Dependencies

• deps: bump the npm-minor-and-patch group with 11 updates (#26) (e0e1445)
• deps: bump the npm-minor-and-patch group with 6 updates (#27) (5754541)

2.26.1 (2026-06-01)

Documentation

• correct release automation notes for Trusted Publishing (6b41673)

2.26.0 (2026-06-01)

Features

• add Claude Opus 4.8 to Copilot provider (daa7c97)

2.25.1 (2026-05-31)

Dependencies

• deps: bump @google/genai from 1.51.0 to 2.7.0 (595da47)

2.25.0 (2026-05-31)

Features

• add conversation tool for turn-based multi-model round-table (0b99879)

Refactor

• tidy conversation tool after simplify review (6aa4848)

2.24.0 (2026-05-28)

Features

• add Claude Opus 4.8 support to claude and anthropic providers (fa1365c)

Dependencies

• deps: bump the npm-minor-and-patch group with 9 updates (#19) (ba29d5d)

2.23.0 (2026-05-21)

Features

• add Gemini 3.5 Flash support to google and gemini-cli providers (0d6b25a)

Dependencies

• deps: bump the npm-minor-and-patch group with 7 updates (#14) (def4b12)
• deps: bump the npm-minor-and-patch group with 9 updates (#16) (6621efc)

2.22.8 (2026-05-06)

Dependencies

• deps: bump the npm-minor-and-patch group with 8 updates (#11) (5e0cd5a)

2.22.7 (2026-04-25)

Documentation

• document release-please automation in CLAUDE.md (b67a2e2)

2.22.6 (2026-04-25)

Miscellaneous

• remove bootstrap-sha now that v2.22.5 anchors history (802f545)

2.22.5 (2026-04-25)

Bug Fixes

• release-please: correct tag format and bootstrap-sha (a30b0fb)

[Unreleased]

[2.22.4] - 2026-04-24

Fixed

• Codex Provider: Bumped @openai/codex-sdk 0.123.0→0.124.0 to match the Codex CLI 0.124.0 release and unblock the new gpt-5.5 default model (earlier SDK rejected it with "model does not exist").

Changed

• Dependencies: Updated @anthropic-ai/sdk 0.90.0→0.91.0 and @anthropic-ai/claude-agent-sdk 0.2.118→0.2.119.

[2.22.3] - 2026-04-23

Changed

• Codex Provider: Default to GPT-5.5 via the new CODEX_MODEL env var (default gpt-5.5). The model is now passed explicitly to the Codex SDK rather than relying on the CLI's default. Friendly name, description, and aliases updated to reflect GPT-5.5.
• Dependencies: Updated @openai/codex-sdk 0.118.0→0.123.0, @anthropic-ai/sdk 0.86.1→0.90.0, @anthropic-ai/claude-agent-sdk 0.2.110→0.2.118, @mistralai/mistralai 2.2.0→2.2.1, ai 6.0.164→6.0.168, vite 8.0.8→8.0.10, eslint 10.2.0→10.2.1, vitest / @vitest/coverage-v8 4.1.4→4.1.5

Fixed

• Tests: Repaired tests/unit/async/cache-ttl.test.js after the AsyncJobStore/FileCache API drift — tests now pass explicit jobIds and use writeSnapshot/readSnapshot, and exercise expiry via cleanup(maxAgeMs) instead of LRU's performance.now()-based TTL (which fake timers don't advance reliably)

[2.22.2] - 2026-04-23

Added

• Codex Provider: Pass images through to Codex via the SDK's local_image input (forwarded to the CLI as --image). Enables image-to-image workflows with $imagegen and visual context in chat. Requires @openai/codex-sdk 0.118+.

Changed

• Codex Provider: Flip supportsImages to true for the codex model and rewrite message conversion to emit structured SDK input (string | UserInput[]) when images are attached. Text-only turns continue to send a plain string. Images without an on-disk file path (e.g. raw base64) are skipped with a debug log.

[2.22.1] - 2026-04-23

Added

• Codex Provider: Auto-elevate sandbox to workspace-write when the prompt contains $imagegen so Codex can save generated image files. Higher modes explicitly set by the user (workspace-write, danger-full-access) are left untouched.

[2.22.0] - 2026-04-16

Added

• Anthropic Provider: Added Claude Opus 4.7 model support — most capable model for complex reasoning and agentic coding with adaptive thinking, 1M context, server-side compaction, and 128K output
• Copilot Provider: Added Claude Opus 4.7 model support via Copilot subscription

Changed

• Anthropic Provider: Remapped effort levels to better match Anthropic's xhigh tier — none→low, minimal→low, low→medium, medium→high, high→xhigh, max→max
• Anthropic Provider: Fixed none effort level sending adaptive thinking without an effort parameter; now correctly maps to low
• Claude SDK Provider: Updated default model from Opus 4.6 to Opus 4.7
• Model Aliases: Unversioned opus and claude-opus aliases now resolve to Claude Opus 4.7 across all providers
• Dependencies: Updated @github/copilot-sdk 0.2.1→0.2.2, @google/genai 1.49.0→1.50.1, ai 6.0.154→6.0.164, dotenv 17.4.1→17.4.2, lru-cache 11.3.3→11.3.5, nanoid 5.1.7→5.1.9, vitest 4.1.3→4.1.4

[2.21.1] - 2026-03-24

Changed

• All Providers: Increased execution timeouts to 10 minutes (600s) across all CLI providers — Claude, Codex, Gemini CLI, and all Copilot models — to prevent premature timeout errors on long-running requests
• Dependencies: Updated @anthropic-ai/claude-agent-sdk 0.2.77→0.2.81, @google/genai 1.45.0→1.46.0, ai 6.0.116→6.0.138, openai 6.31.0→6.32.0, vite 8.0.0→8.0.2, vitest 4.1.0→4.1.1, eslint 10.0.3→10.1.0

[2.21.0] - 2026-03-17

Added

• OpenAI Provider: Added GPT-5.4 mini and GPT-5.4 nano model support — fast, efficient models optimized for coding, subagents, and computer use

[2.20.9] - 2026-03-17

Fixed

• Copilot Provider: Fixed responses being duplicated — the assistant.message event (full content) was emitted alongside streaming deltas, doubling output

[2.20.8] - 2026-03-17

Fixed

• Copilot Provider: Replaced disk-patching of vscode-jsonrpc with a node:module resolve hook that rewrites the extensionless vscode-jsonrpc/node import to include .js. Works in all environments (npm, pnpm, global installs) without filesystem writes.

[2.20.7] - 2026-03-17

Fixed

• Copilot Provider: Fixed ERR_MODULE_NOT_FOUND for vscode-jsonrpc/node under Node.js strict ESM resolution by adding runtime patch for missing exports field in vscode-jsonrpc@8.2.1

[2.20.6] - 2026-03-17

Fixed

• Copilot Provider: Fixed overly broad auth error detection that caused false positives (e.g. errors containing "OAuth" or "authorization" were incorrectly caught as authentication failures)
• Copilot Provider: Fixed error messages referencing non-existent copilot auth login command; now correctly references gh auth login
• Copilot Provider: Auth error now includes original error message for easier debugging

[2.20.5] - 2026-03-14

Changed

• Dependencies: Updated all dependencies to latest versions
  • @anthropic-ai/claude-agent-sdk 0.2.74 → 0.2.76
  • lru-cache 11.2.6 → 11.2.7
  • openai 6.27.0 → 6.29.0

[2.20.4] - 2026-03-13

Changed

• Dependencies: Updated all dependencies to latest versions
  • @anthropic-ai/claude-agent-sdk 0.2.63 → 0.2.74
  • @anthropic-ai/sdk 0.74.0 → 0.78.0
  • @github/copilot-sdk 0.1.29 → 0.1.32
  • @google/genai 1.43.0 → 1.45.0
  • @mistralai/mistralai 1.14.1 → 1.15.1
  • @openai/codex-sdk 0.110.0 → 0.114.0
  • ai 6.0.108 → 6.0.116
  • openai 6.26.0 → 6.27.0
  • vite 7.3.1 → 8.0.0
  • @vitest/coverage-v8 4.0.18 → 4.1.0
  • eslint 10.0.2 → 10.0.3
  • vitest 4.0.18 → 4.1.0

[2.20.3] - 2026-03-07

Changed

• File Range Specifier: Support alternative separator characters in partial file read syntax

[2.20.2] - 2026-03-05

Added

• GPT-5.4 Support: Added gpt-5.4 and gpt-5.4-pro models to OpenAI provider
  • 1M context window, 128K output (272K for Pro)
  • gpt-5 alias now resolves to gpt-5.4
  • gpt-5-pro alias now resolves to gpt-5.4-pro
• Copilot Provider: Added gpt-5.4 model entry, gpt-5 alias updated

Removed

• Deprecated OpenAI models: Removed gpt-5.2, gpt-5-2025-08-07 (GPT-5.0), and gpt-5.2-pro

Changed

• Dependencies: Updated to latest versions
  • @openai/codex-sdk 0.101.0 → 0.110.0
  • openai 6.25.0 → 6.26.0

[2.20.1] - 2026-03-03

Changed

• Dependencies: Updated to latest versions
  • @anthropic-ai/claude-agent-sdk 0.2.50 → 0.2.63
  • @github/copilot-sdk 0.1.25 → 0.1.29
  • @google/genai 1.42.0 → 1.43.0
  • @mistralai/mistralai 1.14.0 → 1.14.1
  • @modelcontextprotocol/sdk 1.26.0 → 1.27.1
  • ai 6.0.97 → 6.0.108
  • openai 6.22.0 → 6.25.0
  • eslint 10.0.1 → 10.0.2

[2.20.0] - 2026-03-03

Fixed

• CWD passthrough for globally installed binary: process.chdir() in bin/converse.js was overwriting the caller's working directory before config.js could capture it, causing relative file paths from MCP clients to resolve against the package root instead of the caller's directory
  • Captures process.cwd() before chdir and exposes it as CLIENT_CWD env var
  • Adds --cwd <path> CLI argument for explicit override
  • Normalizes Git Bash paths (/c/Users/... → C:\Users\...) on Windows so path.resolve() works correctly

[2.19.3] - 2026-02-27

Fixed

• Copilot Provider reasoning_effort crash: Models that don't support reasoning effort (Gemini, Claude, GPT-4.1) no longer cause session.create to fail when reasoning_effort is specified
  • Three-tier capability detection: static flags for known models, SDK listModels() query for unknown models, optimistic retry with fallback for unresolvable defaults
  • Added supportsReasoningEffort: true to GPT-5+ model definitions as a fast-path cache
  • Case-insensitive error matching for retry resilience against SDK wording changes
• Wrong timeouts for Copilot models: invoke() was hardcoded to the base copilot config (120s timeout) instead of resolving the actual model config (e.g., 300s for codex models, 600s for codex-max)

Changed

• Extracted findModelConfig() as a standalone helper, refactored getModelConfig() to delegate to it

[2.19.2] - 2026-02-24

Changed

• Updated chat and consensus tool descriptions to include copilot and copilot:codex model examples
• Simplified reasoning_effort description to remove model-specific qualifier

[2.19.1] - 2026-02-23

Changed

• Simplified continuation_id tool descriptions for chat and consensus tools

[2.19.0] - 2026-02-22

Added

• Custom Continuation IDs: Callers can now provide their own meaningful continuation IDs (e.g., "my-project-analysis") instead of receiving server-generated conv_ IDs
  • Unrecognized IDs start a new conversation under that exact ID rather than being silently replaced
  • Response metadata includes custom_id: true when the ID doesn't match standard conv_ format and is a new conversation (not a resume)
  • Async mode validates custom IDs for filesystem safety — only letters, numbers, hyphens, and underscores allowed (max 128 chars)
  • Consistent behavior across chat and consensus tools in sync, async submission, and async completion paths
  • Aligns implementation with existing README documentation of custom ID support

[2.18.0] - 2026-02-22

Added

• Copilot Provider reasoning_effort Support: The reasoning_effort parameter is now forwarded to the Copilot SDK's reasoningEffort session config instead of being silently ignored
  • Maps tool-level values to Copilot SDK enum: none/minimal → low, low → low, medium → medium, high → high, max → xhigh
  • Applied at session creation via client.createSession({ reasoningEffort }) where the SDK reads it

Fixed

• Copilot Provider: Removed incorrect "reasoning_effort not supported" log message — the @github/copilot-sdk has supported reasoningEffort on SessionConfig since v0.1.25
• Copilot Tests: Updated stale model assertions referencing removed models (gemini-2.5-pro, gemini-3-flash-preview, grok-code-fast-1, raptor-mini, goldeneye) to match current SUPPORTED_MODELS

[2.17.0] - 2026-02-21

Added

• Copilot Model Selection via copilot:modelname Syntax: Choose specific AI models through Copilot provider using a prefix syntax
  • Use copilot:gpt-5.2, copilot:claude-sonnet-4.6, copilot:gemini-2.5-pro, etc. to select specific models
  • Version shortcut aliases: copilot:gpt-5 → gpt-5.2, copilot:codex → gpt-5.3-codex, copilot:sonnet → claude-sonnet-4.6, copilot:opus → claude-opus-4.6, copilot:haiku, copilot:gemini, copilot:flash, copilot:grok
  • 23 Copilot-available models added: OpenAI (gpt-4.1, gpt-5-mini, gpt-5.1, gpt-5.2, codex variants), Anthropic (claude-haiku-4.5, claude-sonnet-4/4.5/4.6, claude-opus-4.5/4.6/4.6-fast), Google (gemini-2.5-pro, gemini-3-flash/pro-preview, gemini-3.1-pro-preview), xAI (grok-code-fast-1), and fine-tuned models (raptor-mini, goldeneye)
  • copilot: prefix takes precedence over keyword routing — copilot:claude-sonnet-4.6 routes to Copilot, not Anthropic
  • COPILOT_MODEL env var supports aliases and prefix stripping (e.g., COPILOT_MODEL=codex)
  • Case-insensitive prefix detection and alias resolution
  • Unknown models passed through to SDK for future compatibility
  • Sync metadata now reports resolved SDK model slug instead of raw input

Changed

• Dependencies: Updated to latest versions
  • @anthropic-ai/claude-agent-sdk 0.2.49 → 0.2.50
  • ai 6.0.94 → 6.0.97

[2.16.0] - 2026-02-20

Added

• MCP Cancellation Support for Sync Tool Calls: Pressing Escape in Claude Code now aborts in-flight provider API calls instead of silently discarding the response
  • Router accepts MCP SDK's extra.signal and forwards it per-request to tools
  • Chat and consensus sync paths pass signal through to provider invoke() calls
  • OpenAI provider passes signal as RequestOptions to SDK .create() calls (both sync and streaming)
  • Consensus Phase 2 (cross-feedback) skipped entirely if signal is aborted after Phase 1
  • Cancelled calls do not persist conversation state to continuation store
  • Abort errors logged at debug level (not error) across router, tool, and provider layers
  • callTool test utility accepts optional signal parameter for test coverage

[2.15.0] - 2026-02-19

Added

• Gemini 3.1 Pro Model Support: Added gemini-3.1-pro-preview to Google provider
  • 1M context window, 64K max output tokens
  • Expanded thinking levels: minimal, low, medium, high (up from 3.0's binary low/high)
  • Thinking level logic updated to pass granular levels when the model supports them
  • Aliases: gemini-3.1, gemini3.1, gemini-3.1-pro, 3.1-pro, pro, gemini-pro, gemini pro

Changed

• Gemini 3.0 Pro Removed: Replaced entirely by Gemini 3.1 Pro; all 3.0 aliases (gemini-3, gemini3, gemini-3-pro, gemini-3-pro-preview, 3-pro) now resolve to 3.1 Pro
• Gemini CLI Provider: Updated SDK model name from gemini-3-pro-preview to gemini-3.1-pro-preview
• Dependencies: Updated to latest versions
  • @google/genai 1.41.0 → 1.42.0
  • @anthropic-ai/claude-agent-sdk 0.2.44 → 0.2.47
  • @github/copilot-sdk 0.1.24 → 0.1.25
  • ai 6.0.87 → 6.0.93

[2.14.0] - 2026-02-17

Added

• Claude Sonnet 4.6 Model Support: Added claude-sonnet-4-6 to Anthropic provider
  • 64K max output tokens, 200K context (1M beta)
  • Adaptive thinking mode (thinking: {type: "adaptive"}) — recommended over legacy budget-based thinking
  • Effort parameter is GA (no beta header required)
  • sonnet / claude-sonnet aliases now resolve to Sonnet 4.6
• Server-Side Context Compaction: Beta support for Opus 4.6 and Sonnet 4.6 via compact-2026-01-12 header
  • Automatically summarizes older context when approaching context window limits

Changed

• Sonnet 4.5 Deprecated: Marked as legacy; sonnet alias moved to Sonnet 4.6
• Sonnet 4 Removed: Removed from supported models entirely
• Dependencies: Updated all dependencies to latest versions
  • @anthropic-ai/sdk 0.73.0 → 0.74.0
  • @anthropic-ai/claude-agent-sdk 0.2.32 → 0.2.44
  • @google/genai 1.38.0 → 1.41.0
  • @modelcontextprotocol/sdk 1.25.3 → 1.26.0
  • openai 6.17.0 → 6.22.0
  • eslint 9.39.2 → 10.0.0
  • And others (see package.json)

Fixed

• OpenAI Tests: Updated tests for gpt-5-pro → gpt-5.2-pro model rename

[2.13.0] - 2026-02-14

Added

• GitHub Copilot SDK Provider: New provider using @github/copilot-sdk for CLI-authenticated Copilot access
  • Singleton CopilotClient with auto-start/restart; fresh session per request
  • Push-to-pull streaming bridge (SDK events → async generator)
  • Tool permission handler with COPILOT_TOOL_ACCESS (read-only | full)
  • Model passthrough via COPILOT_MODEL env or explicit model param
  • Aliases: copilot, copilot-sdk, github-copilot
  • Registered in all auto-selection paths (chat sync/async, consensus sync/async) with priority after claude, before openai
• ESLint Test Override: Disabled prefer-arrow-callback for test files to prevent eslint --fix from breaking mock constructors

Fixed

• Async Chat Auto-Selection: Used providerOrder priority array instead of Object.keys(providers) for consistent provider selection between sync and async paths
• Async Consensus Streaming: SDK providers (copilot, codex, claude, gemini-cli) now stream via invoke({stream: true}) with async iterator detection, instead of falling through to non-streaming path
• Stream Normalizer Context: Pass provider name in normalizer context from chat and consensus call sites, fixing 'unknown' provider in passthrough normalizer
• Config Usable-Provider Validation: Check SDK package availability via import.meta.resolve instead of always-true array length check
• Copilot SDK Availability: isCopilotSDKAvailable() now uses import.meta.resolve instead of always returning true
• Copilot Config Injection: Provider reads COPILOT_TOOL_ACCESS and COPILOT_MODEL from config object instead of process.env directly
• Copilot TOOL_ACCESS Validation: Invalid values are caught at config load time (matching CODEX_SANDBOX_MODE validation pattern)
• Copilot session.send(): Awaited SDK's session.send() which returns a Promise
• Copilot assistant.message: Handle final-content event as delta fallback when deltas are coalesced

Changed

• Node.js Requirement: Minimum version bumped from 20 to 24 (required by @github/copilot-sdk)
• Dependencies: Added @github/copilot-sdk ^0.1.23

[2.12.0] - 2026-02-07

Fixed

• Codex Provider: reasoning_effort parameter was silently ignored — it was passed to runStreamed() which only accepts outputSchema and signal. Now correctly set as modelReasoningEffort on ThreadOptions (passed to startThread/resumeThread), where the SDK actually reads it. Values are mapped to the SDK's enum: none/minimal → minimal, max → xhigh.

[2.11.0] - 2026-02-05

Added

• Claude Opus 4.6 Model Support: Added claude-opus-4-6 to Anthropic provider
  • 128K max output tokens (doubled from Opus 4.5's 64K)
  • 1M token context window (beta, via context-1m-2025-08-07 header)
  • Adaptive thinking mode (thinking: {type: "adaptive"}) — recommended over legacy budget-based thinking
  • Effort parameter is now GA (no beta header required), with new max effort level
  • opus / claude-opus aliases now resolve to Opus 4.6
• GPT-5.3-Codex Aliases: Added gpt-5.2-codex and gpt-5.3-codex aliases to Codex provider
• API Key Validation Utilities: Added idValidation.js for API key format detection

Changed

• Claude SDK Provider: Default model updated from claude-opus-4-5 to claude-opus-4-6
• Codex Provider: Updated friendly name to reflect GPT-5.3-Codex availability
• Gemini CLI Provider: Improved error handling and configuration
• Updated dependencies
  • @anthropic-ai/sdk 0.72.0 → 0.73.0
  • @anthropic-ai/claude-agent-sdk 0.2.23 → 0.2.32
  • @openai/codex-sdk 0.92.0 → 0.98.0

Removed

• Test Inventory: Removed backlog/test-inventory.md (served its purpose)

[2.10.0] - 2026-01-29

Changed

• Updated all dependencies to latest versions
  • @anthropic-ai/sdk 0.71.2 → 0.72.0 (structured outputs: output_format → output_config)
  • @anthropic-ai/claude-agent-sdk 0.2.9 → 0.2.23
  • @openai/codex-sdk 0.86.0 → 0.92.0 (memory leak fix, dynamic tools injection)
  • @google/genai 1.37.0 → 1.38.0
  • @mistralai/mistralai 1.11.0 → 1.13.0
  • openai 6.16.0 → 6.17.0
  • ai 6.0.38 → 6.0.62
  • @modelcontextprotocol/sdk 1.25.2 → 1.25.3
  • cors 2.8.5 → 2.8.6
  • lru-cache 11.2.4 → 11.2.5
  • vite 7.3.1 (unchanged)
  • vitest 4.0.17 → 4.0.18
  • @vitest/coverage-v8 4.0.17 → 4.0.18
  • prettier 3.8.0 → 3.8.1

[2.9.7] - 2025-12-15

Fixed

• Codex SDK sandboxMode: Fixed incorrect property name passed to Codex SDK
  • Was passing sandbox but SDK expects sandboxMode
  • This caused sandbox restrictions to be silently ignored
  • Now correctly enforces read-only, workspace-write, or danger-full-access modes

[2.9.6] - 2025-12-15

Changed

• GPT-5.2 Pro Model: Updated GPT-5-pro to GPT-5.2-pro (released Dec 11, 2025)
  • gpt-5-pro alias now points to GPT-5.2-pro (latest pro model)
  • Added new alias: gpt-5.2-pro-2025-12-11

[2.9.5] - 2025-12-12

Changed

• GPT-5.2 Model Support: Replaced GPT-5.1 with GPT-5.2 (released Dec 11, 2025)
  • gpt-5 alias now points to GPT-5.2 (latest flagship)
  • Added new aliases: gpt-5.2, gpt5.2, gpt 5.2, gpt-5.2-2025-12-11
  • Removed GPT-5.1 (deprecated, sunset ~March 2026)
  • GPT-5.2 offers 38% fewer errors and 30% fewer hallucinations vs GPT-5.1

Documentation

• CLAUDE.md: Added testing guidance to avoid running full test suite during development

[2.9.4] - 2025-11-30

Changed

• Tool Documentation: Added reminder to use files parameter instead of pasting code
  • Updated chat and consensus tool descriptions to emphasize using files param
  • Updated files parameter descriptions with IMPORTANT note about proper usage
  • Updated help prompt File Context section with guidance

[2.9.3] - 2025-11-26

Changed

• Help Prompt Simplified: Removed topic argument from help prompt
  • Now always returns full comprehensive guide without requiring arguments
  • Improves UX in MCP clients like Claude Code that prompt for optional arguments

[2.9.2] - 2025-11-26

Removed

• CODEX_DEFAULT_MODEL config: Removed unused configuration option
  • Config was defined but never actually used in the codebase
  • Default behavior now uses Codex CLI's latest model (no model override)
• PORT and HOST server configs: Removed redundant configuration options
  • These duplicated HTTP_PORT and HTTP_HOST from transport section
  • Only transport configs were actually used; server configs were dead code
• Deprecated OpenAI models: gpt-4o, gpt-4o-mini, o3-mini
  • Superseded by GPT-5 family (gpt-5, gpt-5-mini, gpt-5-nano) and o4-mini
• Deprecated Google models: gemini-2.0-flash, gemini-2.0-flash-lite
  • Deprecated by Google with shutdown scheduled for Feb 2026
  • Use Gemini 2.5 (flash, pro) or Gemini 3.0 Pro instead
• Deprecated Anthropic models: claude-3-5-sonnet-20241022, claude-3-7-sonnet-20250219, claude-3-5-haiku-20241022
  • Claude 3.5 Sonnet retired in Oct 2025
  • Use Claude 4 series (claude-opus-4-5, claude-sonnet-4-5, claude-haiku-4-5) instead

[2.9.1] - 2025-11-26

Changed

• Schema Descriptions: Improved files parameter descriptions in chat and consensus tools

[2.9.0] - 2025-11-26

Added

• Partial File Line Range Support: Specify line ranges when including files
  • Syntax: file.txt{10:50} extracts lines 10-50 inclusive
  • Start-only: file.txt{100:} extracts from line 100 to end of file
  • End-only: file.txt{:20} extracts first 20 lines
  • Works with both files parameter in chat and consensus tools
  • Context header shows range info: (lines 10-50 of 200)
  • New pathParser.js utility for range parsing and extraction

Removed

• File Size Limits: Removed unused maxTextSize and maxImageSize limits
  • These were internal defaults (1MB text, 10MB images) that were never exposed
  • Files of any size can now be processed (limited only by system memory)

[2.8.5] - 2025-11-24

Fixed

• Anthropic Request Payload: Remove spread of non-API parameters into request
  • Non-API parameters like continuationStore were being spread into the API request
  • Fixes "continuationStore: Extra inputs are not permitted" error

[2.8.4] - 2025-11-24

Fixed

• Anthropic Beta Endpoint: Use anthropic.beta.messages.create() for beta features
  • Effort parameter requires the beta endpoint, not the standard messages endpoint
  • Now correctly routes to beta endpoint when beta features (effort, 1M context, etc.) are enabled
  • Fixes "output_config: Extra inputs are not permitted" error with Opus 4.5

[2.8.2] - 2025-11-24

Fixed

• Claude SDK Image Support: Images now work correctly with the Claude provider

  • Implemented streaming input mode for multimodal content as required by SDK
  • Single message mode (string prompt) doesn't support images per SDK documentation
  • Images are now sent via AsyncGenerator<SDKUserMessage> instead of plain string

• Claude SDK error_max_turns Fix: Resolved error when using files parameter

  • Increased maxTurns from 1 to 10 to allow file reading operations
  • SDK needs additional turns to read files using internal tools

Changed

• Auto Model Selection Priority: Subscription-based providers now prioritized
  • New order: codex → gemini-cli → claude → openai → google → xai → anthropic → mistral → deepseek → openrouter
  • Chat tool: model: "auto" picks first available from priority list
  • Consensus tool: models: ["auto"] expands to first 3 available (ideally codex, gemini-cli, claude)

[2.8.0] - 2025-11-24

Added

• Claude SDK Provider: New claude provider for subscription-based access via Claude Agent SDK

  • Uses @anthropic-ai/claude-agent-sdk for authentication via claude login command
  • No API key required - leverages existing Claude Code CLI authentication
  • Model name: claude with aliases claude-sdk and claude-code
  • Supports both streaming and synchronous execution modes
  • Automatically handles SDK message types (system, assistant, result)
  • Pre-normalized streaming events compatible with ProviderStreamNormalizer

• Model Routing Enhancement: Smart routing distinguishes SDK vs API access

  • claude, claude-sdk, claude-code → routes to new claude provider (SDK-based, subscription)
  • claude-sonnet-*, claude-3-*, opus, haiku, etc. → routes to anthropic provider (API-based)
  • Enables using Claude Pro/Max subscription alongside API access

[2.7.0] - 2025-11-24

Added

• Claude Opus 4.5 Support: Added new claude-opus-4-5-20251101 model

  • 64K max output tokens, 200K context window
  • Full thinking/extended reasoning support (up to 64K thinking tokens)
  • New supportsEffort property for Opus 4.5 exclusive effort parameter
  • Aliases: opus, claude-opus, opus-4.5, opus-4-5, opus4.5, claude-opus-4.5, etc.

• Effort Parameter (Beta): New effort parameter support for Opus 4.5

  • Maps reasoning_effort to Anthropic's effort parameter (low, medium, high)
  • Automatically adds effort-2025-11-24 beta header when using Opus 4.5
  • Controls response thoroughness vs token efficiency
  • Mapping: minimal/low → "low", medium → "medium", high/max → "high"

[2.6.1] - 2025-11-24

Fixed

• Consensus Tool Export: Enhanced consensus export to include actual model responses
  • Consensus conversations now export full response content from each model
  • Added formatted sections for "Initial Responses" and "Refined Responses"
  • Each response is clearly labeled with model name and response type
  • Responses are properly formatted with markdown headers and separators
  • Fixed issue where exports only showed generic summary instead of actual content
  • Preserves both initial and refined responses when cross-feedback is enabled

[2.6.0] - 2025-11-24

Added

• Conversation Export Feature: New export parameter for chat tool enables saving conversations to disk
  • Automatically creates organized folder structure with continuation_id as folder name
  • Saves each request/response pair as numbered text files (1_request.txt, 1_response.txt, etc.)
  • Generates comprehensive metadata.json with conversation settings and parameters
  • Implements incremental export with "write-if-missing" optimization for performance
  • Atomic metadata updates ensure file always reflects current conversation state
  • Sanitizes continuation_id to prevent path traversal security issues
  • Supports both synchronous and asynchronous chat execution modes
  • Graceful error handling ensures export failures don't interrupt conversations
  • Cross-platform support for Windows and Unix file systems
  • Respects CLIENT_CWD for user-friendly file locations

[2.5.3] - 2025-11-24

Fixed

• Gemini CLI Provider: Fixed AI SDK v5 ModelMessage format validation
  • Changed image format to use image property instead of data
  • AI SDK validates ModelMessage format before passing to provider
  • Images now use { type: 'image', image: base64 } per AI SDK v5 spec
  • Fixes "messages must be ModelMessage[]" validation error
  • Renamed conversion function to convertToModelMessages for clarity

[2.5.2] - 2025-11-24

Fixed

• Gemini CLI Provider: Fixed message format conversion for images and files
  • Added convertToGeminiCliMessages() function to properly convert messages
  • Images now use { type: 'image', data: base64 } format per Gemini CLI SDK guide
  • Fixes "Invalid prompt: messages must be ModelMessage[]" error when using files/images
  • Conversion only affects Gemini CLI provider, other providers remain unchanged

[2.5.1] - 2025-11-24

Fixed

• Gemini CLI Provider: Fixed streaming event format for async mode compatibility
  • Changed delta events to use data.textDelta instead of content field
  • Ensures proper integration with ProviderStreamNormalizer
  • Enables real-time progress updates during async execution

Changed

• Auto Mode Priority: Updated model selection priority for model: "auto"

  • New priority order: codex > gemini-cli > openai
  • Prioritizes subscription-based and CLI providers over pay-per-use APIs
  • Applies to both chat and consensus tools

• Model Routing: Enhanced routing to prevent provider conflicts

  • Added gemini-cli as explicit alias for Gemini CLI provider
  • Removed gemini from keyword matching to avoid Google API conflicts
  • Positioned gemini-cli before google in provider registry for proper priority

[2.5.0] - 2025-11-24

Added

• Gemini CLI Provider: OAuth-based access to Gemini models via subscription
  • New gemini-cli provider using ai-sdk-provider-gemini-cli package
  • OAuth authentication via Gemini CLI (credentials stored in ~/.gemini/oauth_creds.json)
  • Access Gemini 3.0 Pro Preview through Google subscription (Google One AI Premium or Gemini Advanced)
  • No API key required - uses Google account login instead of pay-per-use API
  • Model name gemini now routes to CLI provider (for API access, use specific names like gemini-2.5-pro)
  • Full support in both chat and consensus tools
  • Streaming support with ProviderStreamNormalizer integration
  • Setup: npm install -g @google/gemini-cli && gemini to authenticate
  • Enhanced agentic features available through CLI that aren't in standard API

Changed

• Model Routing: Updated gemini alias behavior
  • gemini → routes to Gemini CLI provider (OAuth-based subscription access)
  • For Google API access, use specific model names: gemini-2.5-pro, gemini-2.0-flash, etc.
  • Aliases pro, flash, pro 2.5 remain unchanged for API-based access

Fixed

• Consensus Tool: Fixed model routing for CLI-based providers
  • Codex now correctly routes to codex provider instead of openai in consensus mode
  • Gemini CLI correctly routes to gemini-cli provider in consensus mode
  • Both CLI providers (Codex and Gemini CLI) now work properly in consensus tool
  • Added exact match routing checks in src/tools/consensus.js to match src/tools/chat.js

Documentation

• Updated docs/PROVIDERS.md with comprehensive Gemini CLI setup guide
• Updated docs/API.md with Gemini CLI model information and usage examples
• Added authentication setup instructions and best practices
• Documented differences between Gemini CLI and Google API providers

[2.4.0] - 2025-11-19

Added

• Google Provider: Added Gemini 3.0 Pro model support
  • gemini-3-pro-preview: Google's newest model with enhanced reasoning capabilities
  • 1M context window, 64K max output tokens
  • Level-based thinking mode (low/high) instead of token-based budget
  • New media_resolution parameter for controlling image/PDF/video processing quality
  • Values: MEDIA_RESOLUTION_LOW, MEDIA_RESOLUTION_MEDIUM, MEDIA_RESOLUTION_HIGH, MEDIA_RESOLUTION_UNSPECIFIED
  • Defaults to MEDIA_RESOLUTION_HIGH for Gemini 3.0 models (automatically applied)
  • Supports streaming, web search (grounding), and multimodal inputs
  • Aliases: gemini-3, gemini3, gemini-3-pro, 3-pro

Changed

• Google Provider: Updated default model aliases to point to Gemini 3.0 Pro
  • gemini, pro, gemini-pro, and gemini pro now resolve to gemini-3-pro-preview (latest version)
  • Previous Gemini 2.5 Pro now accessible via explicit aliases: pro 2.5, gemini pro 2.5, gemini-2.5-pro-latest
  • reasoning_effort parameter mapping for Gemini 3.0:
    • minimal, low → thinking level: low
    • medium, high, max → thinking level: high
  • Gemini 2.5 models continue using token-based thinking budget (backward compatible)
• Chat Tool: Added optional media_resolution parameter for Gemini 3.0 models

[2.3.1] - 2025-11-17

Added

• OpenAI Provider: Added "none" reasoning effort support for GPT-5.1
  • GPT-5.1 now supports reasoning_effort: "none" for faster responses with increased steerability
  • Optimized for use cases requiring quick responses without extended reasoning
  • Only available on GPT-5.1 and later models (not available on GPT-5.0)
• Chat & Consensus Tools: Updated reasoning_effort parameter to include "none" option
  • New enum values: none, minimal, low, medium, high, max
  • "none" provides fastest responses for GPT-5.1+ models

[2.3.0] - 2025-11-17

Added

• OpenAI Provider: Added GPT-5.1 model support
  • gpt-5.1: Latest flagship model with same capabilities as GPT-5 family
  • gpt-5.1-2025-11-13: Fully qualified model name
  • Aliases: gpt-5, gpt5, gpt 5, gpt5.1, gpt 5.1
  • Same parameters and capabilities: 400K context, 128K output, web search, reasoning effort, verbosity control

Changed

• OpenAI Provider: Updated simple model aliases to point to latest GPT-5 version
  • gpt-5, gpt5, and gpt 5 now resolve to GPT-5.1 (latest version)
  • Previous GPT-5 (2025-08-07) now accessible as gpt-5-2025-08-07 (fully qualified name) or via gpt-5.0 aliases
  • Aliases for old model: gpt-5.0, gpt5.0, gpt 5.0, gpt-5-2025-08-07

[2.2.0] - 2025-10-15

Added

• Anthropic Provider: Added Claude Haiku 4.5 support
  • claude-haiku-4-5-20251001: Fast and intelligent model with extended thinking
  • 200K context window, 64K max output tokens (8x increase from Haiku 3.5)
  • Full thinking support with configurable reasoning effort (1024-64K thinking tokens)
  • Image/vision support and streaming capabilities
  • Aliases: haiku-4.5, haiku-4-5, claude-haiku-4.5, haiku4.5, haiku, claude-haiku

Changed

• Anthropic Provider: Updated simple model aliases to point to latest versions
  • haiku and claude-haiku now resolve to Haiku 4.5 (previously Haiku 3.5)
  • sonnet and claude-sonnet now resolve to Sonnet 4.5 (previously Sonnet 4.0)
  • opus and claude-opus already pointed to Opus 4.1 (no change)
  • Users can now use simple aliases to always get the latest model in each family

Documentation

• Updated PROVIDERS.md with Claude Haiku 4.5 model specifications
• Updated image support documentation to include Claude 4 series models

[2.1.0] - 2025-10-07

Fixed

• Codex Provider: Fixed SDK hang issue with cleaner workaround approach
  • Removed pnpm patch dependency (not published with npm packages)
  • Always use thread.runStreamed() internally, bypassing buggy thread.run()
  • Consume stream synchronously when stream: false is requested
  • Explicitly break after turn.completed event in our own code
  • Works for all users regardless of package manager (npm, pnpm, yarn)
  • No more 5-minute hangs - responses return in ~10-15 seconds

Changed

• Codex Provider: Removed all debug logging added during investigation
• Codex Provider: Simplified code by removing unreachable legacy thread.run() path
• Build: Removed SDK patch file (workaround makes it unnecessary)

[2.0.1] - 2025-10-07

Fixed

• Codex Provider: Fixed 5-minute timeout issue when Codex completes without emitting turn.completed event
  • Applied permanent pnpm patch to @openai/codex-sdk@0.45.0 to fix SDK's thread.run() hanging bug
  • Updated stream normalizer to handle natural stream closure when CLI process exits
  • Force streaming mode for all Codex calls (workaround until OpenAI fixes SDK)
  • See CODEX_SDK_PATCH.md for detailed patch documentation
• Codex Provider: Normalized Windows extended-length paths (\\?\C:\... → C:\...)
  • Added normalizeExtendedPath() utility to strip \\?\ prefix
  • Codex responses now show clean paths instead of extended-length notation

Documentation

• Added CODEX_SDK_PATCH.md documenting SDK bug and permanent patch solution
• Documented pnpm patch workflow for team sharing and future updates

[2.0.0] - 2025-10-07

Added

• Codex Provider: OpenAI Codex integration for agentic coding assistance
  • Thread-based conversation sessions with persistent context via continuation_id
  • Local file system access with configurable sandbox modes
  • Support for model: 'codex' in Chat tool
  • Configuration options:
    • CODEX_SANDBOX_MODE: read-only (default), workspace-write, danger-full-access
    • CODEX_WORKING_DIRECTORY: Optional working directory (defaults to CLIENT_CWD)
    • CODEX_SKIP_GIT_CHECK: Skip Git repository validation (default: true)
    • CODEX_APPROVAL_POLICY: Command approval behavior (default: never)
    • CODEX_DEFAULT_MODEL: Default Codex model (default: gpt-5-codex)
  • Streaming support with full event handling (item.completed, turn.completed, errors)
  • Requires ChatGPT login or CODEX_API_KEY environment variable
  • See .env.example for configuration details

Changed

• BREAKING: Tool Parameter Order: Moved prompt parameter to last position in both chat and consensus tools
  • Makes it easier to parse tool calls in Claude Code when prompts are very long
  • All other parameters maintain their relative order
  • No functional changes, only schema ordering

Fixed

• Codex Provider: Fixed API key handling to use SDK's native apiKey option instead of environment manipulation
• Codex Provider: Added comprehensive event handlers for turn.failed, error, item.started, item.updated events
• Codex Provider: Enhanced debug logging with working directory type detection and execution timing

[1.18.0] - 2025-10-07

Added

• Codex Provider: OpenAI Codex integration for agentic coding assistance
  • Thread-based conversation sessions with persistent context
  • Local file system access with configurable sandbox modes
  • Support for model: 'codex' in Chat tool
  • Configuration options:
    • CODEX_SANDBOX_MODE: read-only (default), workspace-write, danger-full-access
    • CODEX_SKIP_GIT_CHECK: Skip Git repository validation (default: true)
    • CODEX_APPROVAL_POLICY: Command approval behavior (default: never)
    • CODEX_DEFAULT_MODEL: Default Codex model (default: gpt-5-codex)
  • Requires ChatGPT login or CODEX_API_KEY environment variable
  • See .env.example for configuration details

[1.17.2] - 2025-10-06

Changed

• OpenAI Provider: Removed unused reasoningEffort property from GPT-5 Pro model config

[1.17.1] - 2025-10-06

Fixed

• OpenAI Provider: Fixed GPT-5 Pro reasoning effort handling
  • Automatically enforces reasoning_effort: 'high' for GPT-5 Pro (only supported value)
  • Prevents API errors when users specify 'medium' or other unsupported values

[1.17.0] - 2025-10-06

Added

• OpenAI Provider: Added GPT-5 Pro support
  • gpt-5-pro: Most advanced reasoning model (400K context, 272K output)
  • Designed for the hardest problems requiring extended compute time
  • Supports web search, images, and reasoning summaries
  • Responses API only (no streaming support)
  • Defaults to high reasoning effort
  • Aliases: gpt5-pro, gpt-5pro, gpt 5 pro, gpt-5 pro, gpt-5-pro-2025-10-06

[1.16.0] - 2025-10-03

Added

• Anthropic Provider: Added Claude Sonnet 4.5 support

  • claude-sonnet-4-5-20250929: Latest Sonnet with enhanced intelligence
  • 200K standard context / 1M beta context window support
  • 64K max output tokens with extended thinking capabilities
  • Aliases: claude-4.5-sonnet, sonnet-4.5, claude-sonnet-4.5, sonnet4.5

• Anthropic Provider: Beta 1M context window support

  • Enabled for Claude Sonnet 4.5 and Claude Sonnet 4 models
  • Automatic context-1m-2025-08-07 beta header when using supported models
  • Uses modern SDK betas parameter API instead of deprecated defaultHeaders

Changed

• Dependencies: Updated to latest versions

  • @anthropic-ai/sdk: 0.57.0 → 0.65.0 (8 versions)
  • openai: 5.11.0 → 6.1.0 (major version upgrade)
  • @google/genai: 1.12.0 → 1.22.0 (10 versions)
  • @mistralai/mistralai: 1.7.5 → 1.10.0
  • @modelcontextprotocol/sdk: 1.17.1 → 1.19.1
  • Various dev dependencies updated (eslint, cross-env, vitest)

• Anthropic Provider: Migrated to modern beta features API

  • Now uses betas parameter in messages.create() calls
  • Removed deprecated defaultHeaders approach for beta features
  • Improved compatibility with latest Anthropic SDK

Fixed

• Code Quality: Fixed linting errors
  • Changed let to const for non-reassigned variables
  • Fixed string quote consistency
  • Removed trailing whitespace

[1.15.1] - 2025-10-01

Changed

• Tool Descriptions: Shortened chat and consensus tool descriptions for better clarity
  • Removed implementation details (e.g., "handles partial failures gracefully")
  • Focus on tool behavior, use cases, and parameter usage
  • Chat tool: Explicitly mentions continuation_id parameter for multi-turn conversations
  • Consensus tool: Reduced from 334 to 225 characters while maintaining essential information
  • Optimized for LLM understanding rather than human marketing

[1.15.0] - 2025-10-01

Added

• Configuration: New DISABLE_ASYNC_TOOLS environment variable to disable async execution features

  • When enabled, removes check_status and cancel_job tools completely
  • Removes async parameter from chat and consensus tool schemas
  • Help documentation automatically reflects filtered tools based on configuration
  • Useful for deployments that don't need background execution capabilities

• Google Provider: Added Gemini 2.5 Flash Lite model

  • gemini-2.5-flash-lite: Lightweight fast model with 1M context window
  • Supports images, thinking mode, and web search with grounding
  • Efficient for quick responses with lower resource usage

• OpenRouter Provider: Added Z.AI GLM 4.6 model

  • z-ai/glm-4.6: 200K context window model with improved coding and reasoning
  • Better performance in agentic tasks and tool usage
  • Enhanced writing quality and role-playing capabilities

Changed

• Google Provider: Updated Gemini Flash model references
  • gemini-2.5-flash now uses gemini-flash-latest endpoint
  • Added aliases for gemini-2.5-flash-preview-09-2025 and gemini-2.5-flash-latest
  • Ensures automatic access to latest Flash model improvements

Fixed

• Consensus Tool: Fixed display showing incorrect model counts (e.g., "3/1" instead of "3/3")
  • Now correctly shows successful models vs total models in completion status
  • Applies to both synchronous and asynchronous consensus executions
  • Uses providerCalls.length instead of models.length for accurate counting

[1.14.4] - 2025-09-21

Added

• XAI Provider: Added support for new Grok 4 Fast models
  • grok-4-fast-reasoning: Cost-efficient reasoning model with 2M token context window
  • grok-4-fast-non-reasoning: Fast non-reasoning variant for quick responses
  • Both models support function calling, structured outputs, and web search
  • Pricing: $0.20 input / $0.50 output per 1M tokens

Changed

• Mistral Provider: Updated all models to latest versions
  • magistral-medium-2509 (v1.2): Now includes vision support and 128K context window
  • magistral-small-2509 (v1.2): Now includes vision support and 128K context window
  • mistral-medium-2508 (v3.1): Improved tone and performance with 128K context window
  • All models now support image inputs after the September 2025 update

[1.14.3] - 2025-09-12

Changed

• XAI Provider: Removed discontinued grok-3 and grok-3-fast models
  • These models are no longer available from X.AI
  • Added grok-4 as an alias for grok-4-0709 for cleaner model names
  • Updated all documentation and examples to use grok-4 instead of grok-4-0709
  • grok-code-fast-1 remains available for fast, economical coding tasks

Fixed

• Updated consensus tool examples to use simplified grok-4 model name instead of versioned grok-4-0709
• Cleaned up test files to remove references to discontinued models

[1.14.2] - 2025-09-10

Changed

• OpenAI Provider: Increased timeouts for GPT-5 and O3 models to handle longer processing times
  • gpt-5: 5 minutes → 1 hour
  • gpt-5-mini: 3 minutes → 30 minutes
  • gpt-5-nano: 2 minutes → 10 minutes
  • o3: 5 minutes → 10 minutes
  • o3-pro: 30 minutes → 60 minutes
  • o3-deep-research-2025-06-26: 90 minutes → 120 minutes

[1.14.1] - 2025-09-02

Added

• XAI Model Support: Added support for grok-code-fast-1 model
  • 256K context window, optimized for agentic coding tasks
  • Economical pricing ($0.20 input / $1.50 output per 1M tokens)
  • Full streaming support and OpenAI-compatible features
  • Includes aliases: grok-code-fast, grok-code-fast-1-0825

[1.14.0] - 2025-08-26

Added

• AI-Powered Summarization: Intelligent title generation and content summarization for async operations

  • Automatic title generation (up to 60 chars) from user prompts at request initiation
  • Status check returns an up-to-date summary of the progress based on the partially streamed response
  • Final summaries (1-2 sentences) generated for completed responses
  • Smart summaries in check_status tool for better context understanding
  • Configurable via ENABLE_RESPONSE_SUMMARIZATION and SUMMARIZATION_MODEL (default: gpt-5-nano) environment variables

• Enhanced Async Job Storage: Improved job state tracking with new fields

  • accumulated_content: Full streaming content instead of limited preview
  • title: AI-generated descriptive title for each job
  • final_summary: Concise summary of completed job results
  • Removed streaming_preview field (now generated on-demand from accumulated content)

• SummarizationService: New centralized service for all summarization operations

  • Uses fast models (gpt-5-nano, gemini-2.5-flash) for minimal latency
  • Graceful fallback to text snippets when disabled or on errors
  • Non-blocking implementation ensures main flow continues even if summarization fails
  • Temperature set to 0.3 for consistent, focused summaries

• FileCache Integration: Persistent storage for async job results

  • Wire up FileCache to persist job state across server restarts
  • Comprehensive integration tests for cache recovery and TTL management
  • Improved memory management with proper cleanup

Changed

• Check Status Tool: Enhanced display with AI-generated summaries

  • Running jobs show AI-generated summaries based on accumulated content when checked
  • Job listings include titles for quick identification
  • Completed jobs display final summaries in listings
  • Async formatting functions for on-demand summary generation

• Chat Tool Integration: Title and summary generation during streaming

  • Generates title from user prompt at request start
  • Accumulates full content during streaming (replacing 200-char preview)
  • Creates final summary for responses over 100 characters

• Consensus Tool Integration: Multi-provider summary aggregation

  • Combined content accumulation from all providers
  • Handles both single-phase and two-phase (cross-feedback) flows
  • Provider-specific previews maintained alongside combined summaries

Technical

• New formatStatus.js utility for async status formatting
• Configuration schema extended with summarization settings
• Updated AsyncJobStore to accept arbitrary job fields
• Comprehensive test coverage for summarization features
• Fixed integration tests for async workflow scenarios

[1.13.0] - 2025-01-20

Added

• Async Execution Support: Run chat and consensus tools in background mode

  • Use async: true parameter for non-blocking execution
  • Monitor progress with check_status tool
  • Cancel running jobs with cancel_job tool
  • Persistent conversation state across async operations

• Job Management System: Complete async job lifecycle management

  • AsyncJobStore with LRU cache for memory management
  • EventBus for real-time progress updates
  • JobRunner for concurrent task execution
  • Automatic cleanup of completed jobs

• Progress Tracking: Real-time status updates for async operations

  • Streaming progress for individual providers
  • Combined progress for consensus operations
  • Detailed error reporting and recovery

[1.12.0] - 2025-01-19

Added

• Execution Time Display: Added smart execution time formatting to tool responses

  • Shows time in seconds with appropriate precision (0.05s, 1.2s, 15.3s, 1m6s)
  • Displays in metadata header for both chat and consensus tools
  • Time measurement accounts for actual LLM response durations

• Enhanced Metadata Display: New comprehensive metadata shown at start of responses

  • Chat: [⏱️ 2.3s | 🤖 openai | 📱 gpt-5 | 🔗 conv_abc123]
  • Consensus: [⏱️ 8.7s | ✅ 2/3 models | 🔗 conv_xyz789]
  • Environment-aware (automatically disabled in test environments)

• Detailed Failure Reporting: Specific model failure information for consensus tool

  • Shows which models failed and in which phase (initial vs refinement)
  • Example: "• gemini-2.5-pro (refinement failed)" and "• grok-4 (initial failed)"
  • Helps users understand exactly what went wrong during consensus gathering

Changed

• BREAKING: Shorter Continuation IDs: Switched from UUID to nanoid format

  • Before: conv_f47ac10b-58cc-4372-a567-0e02b2c3d479 (41 characters)
  • After: conv_nTC5QoA-ml (15 characters) - 63% shorter
  • Uses cryptographically secure nanoid with URL-safe alphabet
  • Backward Compatible: Old UUID format still accepted and validated
  • Zero collision risk tested with 100,000+ generated IDs

• Improved Consensus Success Counting: More accurate model success tracking

  • When cross-feedback enabled: counts only models succeeding in both phases
  • When cross-feedback disabled: counts initial phase successes
  • Properly accounts for refinement phase failures

Technical

• Migrated from Node.js crypto.randomUUID() to nanoid for ID generation
• Enhanced validation regex to accept both UUID and nanoid formats
• Updated all test patterns to match new continuation ID format
• Added comprehensive failure detail collection and formatting

[1.11.2] - 2025-01-19

Fixed

• Relative Path Support: Fixed relative path handling in chat and consensus tools
  • File validation now uses the same working directory as context processing
  • Relative paths like "./file.txt" and "file.txt" now work correctly
  • Both tools now consistently use auto-detected client working directory for path resolution
  • Fixed issue where file validation would fail but context processing would succeed with relative paths

[1.11.1] - 2025-01-19

Fixed

• File Extension Support: Removed arbitrary file type restrictions
  • All file types now supported for text processing (previously limited to specific extensions)
  • Fixed .cshtml, .razor, .php, .jsp and other web development files being blocked
  • Only images (.jpg, .png, etc.) are treated specially (base64 encoded)
  • Removed unused getSupportedExtensions() and isFileTypeSupported() functions

[1.11.0] - 2025-01-15

Added

• Automatic Client Working Directory Detection: The server now automatically detects where it was invoked from
  • Uses INIT_CWD, PWD, or npm_config_local_prefix environment variables
  • Enables proper relative path resolution from the client's directory
  • Works seamlessly with npx and npm execution

Changed

• File Access Security: Made file path security restrictions optional (disabled by default)
  • Removed mandatory directory restrictions that prevented access to files outside the server directory
  • Security checks can be re-enabled with enforceSecurityCheck: true option
  • Files can now be accessed from any location on the system
• Relative Path Resolution: Fixed to resolve from client's working directory instead of server's directory
  • Relative paths like ./file.txt now work correctly from where the command was invoked
  • Both absolute and relative paths are fully supported

Fixed

• Fixed file access issues where both absolute and relative paths were incorrectly rejected
• Fixed "File access denied" errors when trying to access files outside the Converse directory

Changed (Previous)

• BREAKING CHANGE: Consensus Tool Models Parameter: Simplified model specification from object array to string array
  • Old format: [{"model": "gpt-5"}, {"model": "gemini-2.5-pro"}]
  • New format: ["gpt-5", "gemini-2.5-pro"]
  • Affects all consensus tool calls in client code
  • Input schema updated to accept items: { type: "string" } instead of object structure
  • All tests, documentation, and examples updated to reflect new format

Migration Guide

• Update all consensus tool calls to use string arrays:

  // Before
  models: [{ model: "gpt-5" }, { model: "gemini-2.5-pro" }];
  
  // After
  models: ["gpt-5", "gemini-2.5-pro"];

[1.10.1] - 2025-08-09

Changed

• Consensus Tool Auto Model Selection: Enhanced "auto" model behavior for consensus tool
  • Now expands to first 3 available providers instead of just one
  • Provider priority order: OpenAI → Google → XAI → Anthropic → Mistral → DeepSeek → OpenRouter
  • Automatically selects providers based on configured API keys
  • Enables multi-model consensus without manual model specification
• Default Model Updates: Changed OpenAI default model from o3 to gpt-5 for both chat and consensus tools
• Documentation: Updated README with comprehensive auto model selection behavior for both tools

Technical Details

• Consensus tool with ["auto"] intelligently expands to multiple providers
• Chat tool continues to use single provider selection for efficiency
• Each provider uses its optimal default model when selected via auto

[1.10.0] - 2025-08-08

Added

• Google Provider: Added comprehensive Google API configuration options
  • GEMINI_API_KEY support: Primary API key for Google Gemini models (recommended)
  • GOOGLE_API_KEY fallback: Still supported, but GEMINI_API_KEY takes priority
  • Google Vertex AI support: Full enterprise-grade Vertex AI integration
    • GOOGLE_GENAI_USE_VERTEXAI: Enable Vertex AI mode
    • GOOGLE_CLOUD_PROJECT: Google Cloud project ID
    • GOOGLE_CLOUD_LOCATION: Deployment region (e.g., us-central1)
    • GOOGLE_API_VERSION: API version selection (v1, v1beta, v1alpha)
  • Automatic detection of configuration mode (API Key vs Vertex AI)
  • Support for both Gemini Developer API and Vertex AI API endpoints

Changed

• Environment Configuration: Updated .env files to use GEMINI_API_KEY for clarity
• Documentation: Enhanced README with Google/Gemini API options and Vertex AI setup

Technical Details

• Google provider now supports three initialization modes:
  1. Gemini Developer API with GEMINI_API_KEY (simplest)
  2. Gemini Developer API with GOOGLE_API_KEY (backward compatible)
  3. Google Vertex AI with project/location configuration (enterprise)
• API version can be configured for both Gemini and Vertex AI modes
• Improved validation to handle both API key and Vertex AI configurations

[1.9.0] - 2025-08-07

Added

• OpenAI Provider: Added support for GPT-5 family models, OpenAI's latest flagship series
  • GPT-5: Latest flagship model with 400K context window, 128K max output tokens
    • Superior reasoning, code generation, and analysis capabilities
    • Full support for streaming, function calling, structured outputs, web search, and MCP
    • Aliases: gpt5, gpt 5, gpt-5-2025-08-07
  • GPT-5-mini: Faster, cost-efficient version for well-defined tasks
    • Same 400K context and 128K output as GPT-5
    • Optimized for speed and cost ($0.25 input, $2 output per 1M tokens)
    • Aliases: gpt5-mini, gpt-5mini, gpt 5 mini, gpt-5-mini-2025-08-07
  • GPT-5-nano: Fastest, most cost-efficient version
    • Same 400K context and 128K output capabilities
    • Best for summarization and classification ($0.05 input, $0.40 output per 1M tokens)
    • No web search support
    • Aliases: gpt5-nano, gpt-5nano, gpt 5 nano, gpt-5-nano-2025-08-07
  • All GPT-5 models don't support temperature parameter
  • Updated model recommendations to prefer GPT-5 family over O3 for various use cases
• New API Features for GPT-5:
  • Minimal reasoning effort: New minimal option for fastest responses with few reasoning tokens
  • Verbosity control: New verbosity parameter (low/medium/high) to control output length
    • Low: Concise answers, minimal code commentary
    • Medium: Balanced responses (default)
    • High: Thorough explanations and detailed code
  • Both features supported across entire GPT-5 family (GPT-5, GPT-5-mini, GPT-5-nano)
  • Enhanced chat tool to support these new parameters with proper defaults

[1.8.3] - 2025-08-07

Changed

• Anthropic Provider: Updated Claude Opus 4 to the new Opus 4.1 model
  • Model ID changed from claude-opus-4-20250514 to claude-opus-4-1-20250805
  • Added new aliases: claude-opus-4-1, opus-4.1, opus4.1, claude-opus-4.1
  • Maintains all existing aliases for backward compatibility
  • Same capabilities: 200K context, 32K output tokens, extended thinking, image support

[1.8.0] - 2025-08-04

Added

• OpenAI Deep Research Models: Added support for OpenAI's deep research models
  • Added o3-deep-research-2025-06-26 model with 90-minute timeout for comprehensive research
  • Added o4-mini-deep-research-2025-06-26 model with 60-minute timeout for faster research
  • Both models support web search via web_search_preview tool
  • Models can run 30-90 minutes for in-depth analysis and multi-source synthesis
  • Requires setting MCP_TOOL_TIMEOUT environment variable (e.g., 5400000 for 90 minutes)

Changed

• Web Search Implementation: Simplified web search to use only web_search_preview tool type
  • Removed unused web_search tool type references
  • All OpenAI models now consistently use web_search_preview when web search is enabled
  • Removed support for always-search models (gpt-4o-search-preview, gpt-4o-mini-search-preview)

Technical Details

• Deep research models work with existing chat tool - no separate research tool needed
• Models are integrated into the standard OpenAI provider implementation
• Supports all standard features: streaming, images, context, continuation
• Progress notifications and cancellation infrastructure ready for future Claude Code UI support

[1.7.3] - 2025-08-02

Fixed

• Test Suite: Fixed numerous test failures across the codebase
  • Fixed syntax errors in fixture files (duplicate __dirname declarations, invalid JSON)
  • Fixed JSON parsing errors in edge-cases.json (sparse arrays, JavaScript expressions, hex escape sequences)
  • Fixed performance test reliability by using consistent model selection
  • Fixed consensus tool cross-feedback by ensuring proper message alternation for Anthropic API
  • Fixed image validation to handle base64 data URLs properly
  • Fixed mock provider implementations to properly track method calls
• Image Processing: Enhanced image quality settings
  • Updated OpenAI provider to use detail: 'high' for better image analysis
  • Updated XAI provider to use detail: 'high' for better image analysis
• Path Utilities: Removed shebang line from pathUtils.js module

Added

• Test Images: Added test images (fruits.png, tulips.png, baboon.png) for image processing tests
• Dependencies: Added vite as a dependency (was missing)

Improved

• Test Coverage: Enhanced test reliability and coverage
  • Updated image tests to use real images instead of invalid base64 strings
  • Added proper base64 encoding helper for XAI image tests
  • Fixed mock provider tests to properly handle call tracking
  • Improved error message matching in provider tests

[1.7.2] - 2025-01-27

Added

• Cross-Platform Support: Comprehensive cross-platform compatibility improvements
  • Created src/utils/pathUtils.js utility module for platform-agnostic operations
  • Added cross-env and rimraf dependencies for cross-platform npm scripts
  • Platform-specific path handling for Windows, Linux, and macOS
  • Cross-platform timeout commands and process spawning

Changed

• npm Scripts: Updated all scripts to use cross-platform commands
  • Replaced Unix-specific rm -rf with rimraf package
  • All environment variable assignments now use cross-env
  • Scripts now work correctly on Windows, Linux, and macOS
• Path Handling: Improved path operations throughout codebase
  • Fixed path comparisons to use proper URL methods instead of string replacement
  • Line counting now handles both CRLF (Windows) and LF (Unix/Mac) line endings
  • Test files use platform-agnostic path helpers instead of hardcoded paths
• Process Spawning: Updated to use Node.js executable path
  • Tests now use process.execPath instead of hardcoded 'node' command
  • Proper spawn options for Windows compatibility

Fixed

• Windows Compatibility: Fixed multiple Windows-specific issues
  • Path separator handling in file operations
  • Process spawning in test files
  • Timeout commands in validation script
• Test Reliability: Fixed hardcoded paths in tests
  • Replaced Windows-specific paths (C:) with platform helpers
  • Replaced Unix-specific paths (/tmp) with OS temp directory
• JSON Import: Fixed ES module JSON import syntax for better compatibility
  • Tests now use readFileSync and JSON.parse instead of import assertions

Improved

• Code Quality: Enhanced linting and code standards
  • Changed no-unused-vars to warning level for better DX
  • Added missing global variables to ESLint config
  • Fixed numerous linting issues across the codebase

[1.7.1] - 2025-07-28

Changed

• Test Organization: Reorganized integration tests into provider-specific structure
  • Provider tests now in tests/integration/providers/{provider}/ directories
  • Each provider has separate API, features, and image test files
  • Removed archived test files that were replaced
• Test Commands: Updated test command naming for consistency
  • npm run test:e2e now preferred over test:real-api (both still work)
  • Added provider subcategory commands for granular testing
• Documentation: Updated all test documentation to reflect new structure
  • Updated tests/README.md with new test organization
  • Updated main README.md with current test commands
  • Added tests/integration/providers/README.md for provider test guidance

Added

• Provider Image Tests: Added dedicated image processing tests
  • xai/xai-image.test.js - XAI Grok-4 image processing
  • google/google-image.test.js - Google Gemini image processing
• Error Handling Tests: Added comprehensive error handling tests
  • anthropic/anthropic-error.test.js - Rate limiting and edge cases
  • multi-provider-error.test.js - Cross-provider error handling
• Advanced Tests: Added advanced multi-provider scenarios
  • multi-provider-advanced.test.js - Consensus with files, consistency tests
  • debug-tests.test.js - Message format debugging

Fixed

• Test Configuration: Updated suites.config.js to use new test paths
  • Fixed real-api suite to use glob patterns for new structure
  • Properly excludes archived directory from test runs

[1.7.0] - 2025-07-28

Changed

• Configuration: Server name and version are now automatically read from package.json
  • Removed MCP_SERVER_NAME and MCP_SERVER_VERSION environment variables
  • Ensures version consistency across all parts of the application
• OpenRouter Provider: Requires OPENROUTER_DYNAMIC_MODELS=true to use models in provider/model format
  • Previously allowed dynamic models without explicit configuration
  • Now properly enforces the environment variable requirement
• Tool Descriptions: Updated parameter descriptions for better clarity
  • Model examples now show o3, gemini-2.5-pro, grok-4-0709
  • File and image paths show both absolute (Windows) and relative path examples
  • Reasoning effort examples updated to low, medium, high
  • Simplified use_websearch description
• Help System: Help prompt now dynamically generates tool documentation from metadata
  • Ensures consistency between implementation and documentation
  • No more manual updates needed when tool parameters change

Added

• Environment Variables Documentation: Added all missing environment variables to help prompt
  • All API keys (ANTHROPIC, MISTRAL, DEEPSEEK, OPENROUTER)
  • OpenRouter configuration options
  • HTTP server configuration options

Removed

• Unused Build Script: Removed unused build.js script and related npm scripts
  • Project doesn't require build step as it's pure Node.js
  • Removed build and build:fast npm scripts
• Obsolete Environment Variables: Cleaned up documentation
  • Removed references to GOOGLE_LOCATION (already unused)
  • Removed references to XAI_BASE_URL (not configurable via env)

Fixed

• OpenRouter Dynamic Models: Fixed behavior to require explicit enablement
  • Models with "/" format now properly require OPENROUTER_DYNAMIC_MODELS=true
  • Returns clear error message when dynamic models are disabled

[1.6.0] - 2025-07-27

Added

• OpenRouter Provider: Dynamic model discovery support
  • Enable with OPENROUTER_DYNAMIC_MODELS=true environment variable
  • Automatically fetches model capabilities from OpenRouter's endpoints API
  • Supports any model available on OpenRouter without manual configuration
  • Model capabilities are cached for 24 hours to improve performance
  • Added support for openrouter/auto model for automatic model selection
• Model Routing: Enhanced model routing logic
  • Models with "/" format check native providers first before routing to OpenRouter
  • Allows using models like anthropic/claude-3.5-sonnet via OpenRouter when not available natively
  • Maintains backward compatibility with keyword-based routing

Changed

• OpenRouter Provider: Added static configurations for Qwen3 and Kimi models
  • qwen/qwen3-235b-a22b-thinking-2507 - 235B model with thinking capabilities
  • qwen/qwen3-coder - Specialized for coding tasks
  • moonshotai/kimi-k2 - 200K context window

[1.5.5] - 2025-07-26

Fixed

• Anthropic Provider: Increased SDK timeout to 20 minutes for thinking models
  • Prevents "Streaming is strongly recommended" errors for long-running requests
  • Claude 4 series models now work properly with thinking mode enabled
• Tests: Updated test expectations for max_tokens being required by API

[1.5.4] - 2025-07-26

Fixed

• Anthropic Provider: Removed non-existent 'thinking-2025-01-27' beta header
  • Thinking mode is controlled through model selection, not beta headers

[1.5.3] - 2025-07-26

Fixed

• Google Provider: Fixed gemini-2.0-flash configuration - model does not support thinking mode
• Anthropic Provider: Fixed Claude 4 series models token handling
  • No longer set max_tokens for opus-4 and sonnet-4 models, letting SDK use defaults (32k/64k)
  • Prevents "context length exceeded" errors that were actually SDK warnings about streaming
• Tests: Updated test expectations to match new error message formats

[1.5.1] - 2025-07-26

Fixed

• Mistral Provider: Fixed image handling by correcting the image URL field name from image_url to imageUrl to match Mistral API expectations
  • Models supporting images (mistral-medium-3) now properly process image content
  • Resolved validation errors when sending images to Mistral API

[1.4.0] - 2025-07-26

Changed

• BREAKING: Transport Default: Changed default transport from HTTP to stdio for standard MCP compliance
  • Stdio transport is now the default (launched automatically by Claude)
  • HTTP transport available via --transport=http or MCP_TRANSPORT=http for development/debugging
  • Updated CLI help and documentation to reflect new defaults
  • No functionality lost - all transport methods still available

Fixed

• Test Stability: Fixed timeout issue in file context processing test
• Test Environment: Added explicit MCP_TRANSPORT=http to test environment to maintain HTTP testing

Documentation

• Updated README.md to show stdio as default transport
• Updated help text and examples to reflect new transport defaults
• Clarified when to use HTTP transport (development/debugging scenarios)

[1.3.4] - 2025-07-26

Added

• Anthropic Prompt Caching: Implemented automatic prompt caching with 1-hour TTL for system prompts
  • Reduces latency and API costs for repeated requests
  • Minimum 1024 tokens required (2048 for Haiku models)
  • Cache metrics available in response metadata
• Provider Documentation: Added comprehensive documentation for all new providers (Anthropic, DeepSeek, Mistral, OpenRouter)

Fixed

• Anthropic Provider: Fixed thinking budget calculation to properly account for token limits
• Anthropic Provider: Force temperature to 1 when thinking is enabled (API requirement)
• Anthropic Provider: Fixed context length issues with Claude Sonnet 4

Improved

• Test Coverage: Added comprehensive integration tests for all new providers
• Error Handling: Better error messages for model availability and context limits

[1.3.3] - 2025-07-26

Fixed

• Anthropic Provider: Fixed context length calculation for thinking models
• Mistral Provider: Fixed SDK import order to resolve constructor errors

[1.3.2] - 2025-07-26

Fixed

• OpenRouter Provider: Fixed HTTP-Referer header configuration issue by correcting config key casing
• Missing Dependencies: Added @anthropic-ai/sdk and @mistralai/mistralai as dependencies to fix provider initialization errors

[1.3.1] - 2025-07-26

Improved

• Help System: Updated help documentation and resources to display models from all 7 providers (previously only showed 3)
• Auto Model Selection: Enhanced "auto" model selector to support all providers with intelligent defaults:
  • OpenAI: o3 (powerful reasoning model)
  • Google: gemini-2.5-pro (advanced capabilities)
  • Anthropic: claude-sonnet-4-20250514 (Sonnet 4)
  • Mistral: magistral-medium-2506 (frontier-class model)
  • DeepSeek: deepseek-reasoner (reasoning model)
  • XAI: grok-4-0709 (unchanged)
  • OpenRouter: qwen/qwen3-coder (unchanged)
• Model Aliases: Added comprehensive aliases for all models across all providers for easier access
• Provider Detection: Updated mapModelToProvider function to recognize models from all 7 providers

Fixed

• Help Command: Fixed issue where help command only displayed models from original 3 providers
• Model Resolution: Fixed model name resolution to work with all provider models and their aliases

[1.3.0] - 2025-07-26

Added

• New Providers: Added support for 5 new AI providers, expanding model options:
  • Anthropic: Support for Claude models including Opus 4, Sonnet 3.5, and Haiku 3.5
  • Mistral AI: Support for Magistral Medium, Magistral Small, and Mistral Medium 3
  • DeepSeek: Support for DeepSeek Chat (V3) and DeepSeek Reasoner (R1) models
  • OpenRouter: Gateway to access Qwen3 235B Thinking, Qwen3 Coder, and Kimi K2 models
  • OpenAI-Compatible Base Module: Reusable factory for creating providers with OpenAI-compatible APIs

Features

• Unified Provider Interface: All providers implement consistent interface with error handling
• Advanced Model Capabilities:
  • Thinking/reasoning models with configurable effort levels (Anthropic, OpenRouter)
  • Multimodal support for images (Anthropic, Mistral Medium 3)
  • Extended context windows (up to 200K tokens for Claude, 200K for Kimi K2)
• Enhanced Error Handling: Provider-specific error mapping to unified error codes
• Comprehensive Test Coverage: Added extensive unit tests for all new providers
• Dynamic SDK Loading: Lazy loading of provider SDKs for better performance

Improved

• Provider Architecture: Refactored to use base modules for code reuse
• Model Configuration: Rich metadata for each model including capabilities and limits
• Temperature Handling: Fixed temperature parameter conflicts in OpenAI-compatible providers
• Image Validation: Added proper validation for models that don't support images
• Integration Tests: Fixed MCP server initialization with required capabilities

Fixed

• OpenAI-Compatible Providers: Fixed temperature default parameter override issue
• Error Re-throwing: Fixed error handling in Anthropic provider to avoid double-wrapping
• Mock Setup: Fixed dynamic import mocking patterns in provider tests
• API Key Validation: Added proper validation for provider-specific key formats

[1.2.1] - 2025-07-26

Changed

• Dependencies: Updated dotenv from v16.4.7 to v17.2.1
• Dependencies: Updated eslint to latest version (9.17.0)
• Configuration: Added quiet: true option to dotenv configuration to suppress verbose logging output

Fixed

• Tests: Fixed test failures caused by dotenv v17's verbose logging interfering with JSON parsing in MCP protocol tests
• Tests: Updated tests to properly handle MCP protocol error responses instead of expecting thrown errors
• Tests: Added missing prompts and resources capabilities to test server instances

Improved

• Code Quality: All code now passes latest eslint rules and formatting standards

[1.2.0] - 2025-07-26

Added

• Help Prompt: Added comprehensive help prompt (/converse:help) that provides detailed documentation about all tools, parameters, providers, and models
  • Supports topic-specific help queries (tools, models, providers, parameters, examples)
  • Dynamically pulls real-time model information from provider files
  • Explicitly instructs LLMs to share the information with users
• Help Resource: Added MCP resource (converse://help) that exposes the same help documentation plus server version information
  • Accessible via MCP resource protocol for programmatic access
  • Includes current server version from package.json
• MCP Capabilities: Extended server capabilities to support both prompts and resources in addition to tools

Improved

• Documentation: Help content automatically stays up-to-date by fetching model details directly from provider implementations
• User Experience: Both prompt and resource provide comprehensive guidance including model selection tips, configuration advice, and best practices

[1.1.2] - 2025-07-26

Fixed

• Binary Entry Point: Fixed "startServer is not a function" error when running via npx/npm by properly exporting main function from index.js
• Module Structure: Improved module architecture to support both CLI and programmatic usage
• Stdio Transport: Removed console output from bin file to prevent JSON-RPC protocol corruption

Changed

• Entry Point Pattern: index.js now exports main function and only auto-executes when run directly, following Node.js best practices

[1.1.1] - 2025-07-26

Improved

• Consensus Tool Output: Optimized output format by removing redundant rawResponse fields, reducing output size by ~70-80% while maintaining all essential information
• Performance: Significantly reduced memory usage and network payload for consensus tool responses

Changed

• Output Structure: Removed rawResponse from both initial and refined consensus responses while maintaining backward compatibility

[1.1.0] - 2025-07-26

Fixed

• Image Processing: Fixed image handling in chat and consensus tools where images were being sent in a separate message from the prompt, causing XAI (Grok) and Google (Gemini) providers to not receive images correctly
• Message Structure: Both tools now properly merge context (including images) and prompt into a single user message with complex content array
• Provider Compatibility: All three providers (OpenAI, XAI, Google) now correctly process images with their respective format requirements

Added

• Integration Tests: Added comprehensive image processing tests for consensus tool to verify all providers handle images correctly

Improved

• Image Format Validation: Enhanced image format conversion for XAI and Google providers with proper debugging output
• File Validation: Added file existence validation before processing context to prevent errors

[1.0.3] - 2025-07-26

Fixed

• Stdio Transport: Fixed configuration loading error (Cannot convert undefined or null to object) that prevented stdio transport from starting
• Console Suppression: Fixed logger to properly suppress console output in stdio transport mode from startup
• Transport Detection: Moved transport type detection to very early in startup process to prevent any console output interference

Improved

• JSON-RPC Protocol: Enhanced stdio transport reliability by eliminating all console output that could corrupt the protocol stream
• Logger Configuration: Improved logger reconfiguration timing to respect transport mode from the beginning

Changed

• Default Port: Changed default HTTP server port from 3000 to 3157 to avoid common port conflicts

[1.0.2] - 2025-07-26

Fixed

• Console Logging: Replaced remaining console.log and console.error calls with proper structured logger to prevent stdio transport corruption
• Configuration: Fixed console output in config loading that could interfere with MCP JSON-RPC protocol

Changed

• Documentation: Updated model examples to use latest intelligent models (o3, grok-4, gemini-2.5-pro) and fast models (gemini-2.5-flash, o4-mini, gpt-4.1)
• File Paths: Updated example file paths in documentation to use git-bash compatible paths (/c/Users/username/...)

Removed

• Unused Configuration: Removed unused GOOGLE_LOCATION and XAI_BASE_URL environment variables from configuration files
• Legacy Config: Cleaned up unused Docker, DIAL, and OpenRouter configuration remnants from environment files

Improved

• Logger Integration: Enhanced error logging consistency across chat and consensus tools
• Transport Safety: Strengthened stdio transport protection against console output interference

[1.0.1] - 2025-07-25

Fixed

• Binary Script: Fixed Windows compatibility for bin script import path

[1.0.0] - 2025-07-25

Added

• Initial Release: Complete Node.js implementation with functional architecture
• Chat Tool: Single-provider conversational AI with context and continuation support
• Consensus Tool: Multi-provider parallel execution with cross-model feedback
• Provider Support: OpenAI, Google/Gemini, and X.AI/Grok providers
• Token Limiting: Configurable response size limits (default: 25,000, max: 200,000 tokens)
• System Prompts: Dedicated prompts for chat and consensus tools
• Context Processing: File and image support with security validation
• Continuation System: Persistent conversation management
• Configuration Management: Environment-driven configuration system
• Comprehensive Documentation: API reference, architecture guide, and examples
• Test Suite: Unit, integration, and end-to-end tests
• NPX Support: Direct execution via npx FallDownTheSystem/converse
• MCP Compliance: Full Model Context Protocol implementation
• Error Handling: Robust error handling with graceful degradation
• Logging System: Structured logging with configurable levels

Features

• Parallel Consensus: Simultaneous model execution for faster responses
• Cross-Model Feedback: Models can refine responses based on other models' insights
• Auto Model Selection: Intelligent model selection when using "auto" parameter
• Multiple Response Formats: Support for text, JSON, and structured responses
• File Context Processing: Support for multiple file formats with line numbering
• Image Analysis: Base64 image processing for visual context
• Flexible Configuration: Environment variables with sensible defaults
• Provider Abstraction: Unified interface across different AI providers
• Request Validation: Comprehensive input validation and sanitization

Notes

This is a simplified Node.js implementation of an MCP Server focused on providing just the essential Chat and Consensus tools for a streamlined experience. The parallel consensus workflow represents a major architectural improvement, providing faster and more nuanced multi-model analysis.