Pheron LogoPheron Agent
Download

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.

[1.0.6] - 2026-08-01

Added

  • Lark Suite Integration — Bridges Lark/Feishu's official MCP server for messaging, chats, calendar, Base, docs, and tasks via a bring-your-own-app connection.
  • LemonSqueezy Integration — Direct REST bridge for managing orders, customers, subscriptions, discounts, and license keys.
  • Kit (ConvertKit) Integration — Direct REST bridge for managing subscribers, broadcasts, sequences, and tags.
  • MCP Hub Disconnect Support — Every connected service now has a one-click way to clear its credential and disconnect.
  • Audacity Integration — Control a running Audacity instance directly via its own scripting protocol for audio editing and effects.
  • Apple Notes, Reminders & Office Export — Native Notes/Reminders support, real Excel/PowerPoint/Word export from Numbers/Keynote/Pages, and on-device background removal.

Fixed

  • Web Search & Citation Reliability — Fixed a batch of research-answer issues, including dropped search results, false "incomplete answer" rejections, and wrongly-flagged citations.
  • New Tool Discoverability — Notes, Reminders, background removal, and the Office-document tools are now correctly reachable for simple requests.
  • Multi-Step Task Completion — Fixed cases where the agent could describe a next step without executing it, or mark a report as done without writing it.
  • Memory & Recall Reliability — "Remember this" requests now reliably save instead of silently doing nothing.
  • Safer Tool Retry Logic — A tool disabled after repeated failures now recovers automatically, and permanent restrictions are reported immediately instead of retried.
  • Calendar Date Accuracy — Fixed a bug where events could silently save with an incorrect date.
  • Stripe & Git Integration Fixes — Corrected Stripe's action set after an upstream change and stabilized the Git MCP integration.
  • File & Folder Operation Safety — Fixed edge cases in path-permission checks and folder move/copy behavior.
  • General Reliability — Smaller fixes to telemetry delivery, browser automation, timeout budgets, and system-load detection.

[1.0.5] - 2026-07-24

Fixed

  • Source Citation Safety — Fixed an issue where the agent could cite non-existent source URLs, dates, or version numbers by moving citation safety checks to the active execution path.
  • Compound Request Execution — Fixed an issue where compound multi-part requests (e.g. asking for telemetry and OS version together) could return with only half the answer by enforcing missing tool calls.
  • Shell Output Redirect Safety — Prevented bare single-file shell redirects (command > file) from bypassing binary protection and write safety checks.
  • Bot Detection & CAPTCHA Filtering — Web search now detects and filters out CAPTCHA/bot-challenge pages from search engines to prevent reasoning from challenge text.
  • Google Search JS Stabilization — Improved Google search result fetching by waiting for client-side JavaScript rendering to complete.
  • Safari Fallback Resilience — Genuinely opens visible Safari tabs for search fallback with clear permission guidance when required.
  • Biometric & Keychain Concurrency — Fixed Touch ID timeout handling and unblocked background Keychain reads from blocking tool-availability checks.
  • Daemon Retry Limits — Prevented failing background daemon connections from retrying indefinitely.

Added

  • Authoritative Source Researching — Agent now prioritizes official project data, structured specs, and direct documentation over third-party search snippets.
  • Expanded GitHub Toolsets — Added access for GitHub Actions, code security, Dependabot, discussions, advisories, gists, projects, labels, and notifications.

[1.0.4] - 2026-07-06

Added

  • MCP Tool Bridges — Pheron Agent now integrates with external Model Context Protocol (MCP) servers including Git, Playwright browser automation, Perplexity web search, Stripe, GitHub, Notion, Unreal Engine, and Zapier.
  • MCP Hub & Connections — Added a dedicated card-grid wizard in Settings > Connections to easily configure, save, and test credentials for external tool connections.
  • Context-Aware Recommendations — The agent now suggests connecting missing integrations in Settings > Connections when a task requires a credential-gated tool.
  • Unified Screen & Accessibility Reasoning — Chained screen-capture descriptions, OCR, and AX tree analysis for more cohesive and reliable browser/screen-related actions.
  • Local User Profile — Preferences discovered by the agent and user identity information are now saved in a readable Markdown profile (UserProfile.md).
  • Performance Tab Redesign — Merged Health and Analytics tabs under Settings into a single tab featuring real-time CPU, memory, and speed usage trend charts.
  • Disk Telemetry Support — Telemetry reports now include boot volume free space alongside CPU and memory stats.

Fixed

  • Multi-Turn Conversation Context — Fixed context loss between consecutive turns in the same conversation thread, ensuring the agent remembers the immediate context.
  • Model Context Limits — Corrected local model context budget scaling issues that artificially restricted usable token windows on higher-RAM systems.
  • Memory Loop Fixes — Resolved a tool-calling loop trigger when looking up user-recall details (e.g., "do you remember my name?").
  • Keychain & Authentication Stability — Fixed OAuth flow callbacks for Notion/Zapier and restored deleted Keychain entries during local test executions.
  • Task Switch Performance — Stopped background processes and command executions immediately on timeout or task cancellation to prevent CPU leakages.
  • Turkish Command Preposition Routing — Fixed a misrouting bug where Turkish prompts containing "üzerinden" (via) were incorrectly sent to the math/calculation path.

[1.0.3] - 2026-06-19

Added

  • Personal memory & recall — the agent now reliably remembers and surfaces facts you've explicitly shared (background, CV, preferences) when you ask about them; closed a deep retrieval gap where saved facts could become effectively unsearchable
  • Multi-language file/folder commands — "organize this folder" style requests now recognized in 13 languages (added ES, FR, DE, PT, IT, RU, ZH, JA, KO, AR alongside TR/EN), not just Turkish/English
  • MusicDNA report actions — analysis results now include "Open Report" and "Show in Finder" buttons to jump straight to the generated .dna.md / .report.plist files
  • Telemetry — Supabase integration: all telemetry events now flow through telemetry_events with authenticated requests, retry logic, and a synchronous flush on quit
  • Energy tracking — IOKit-based: real CPU+GPU+ANE joule measurements via powermetrics, shown live in the menu bar effort indicator
  • Analytics default-on: analytics now defaults to enabled when no explicit preference is set

Fixed

  • Lost context after a clarifying question — answering the agent's follow-up question (e.g. "which date format?") could previously derail the conversation into unrelated results (a stray "ram" substring match was misrouting these replies); the agent now stays on the original task after you answer
  • Faster personal-recall responses — eliminated a wasted reasoning turn when the agent looks up something you previously told it
  • Apple Music playback and volume control: confirmation now reflects the actual player state, fixing silent failures when Music wasn't already running
  • Telemetry: RAM/inference metrics and authentication no longer report stale or zero values; failed analytics batches no longer fail silently
  • Debug builds now sign with the correct development team, fixing missing entitlements

[1.0.2] - 2026-06-03

Added

  • Background task processing — start a new conversation while a task is still running; the old conversation stays in the sidebar with a ⟳ indicator and continues in the background
  • Task interruption — Stop button (and Escape key) cancels a running task mid-execution
  • Model Hub — full model catalogue: 30+ local MLX models (Qwen3, Llama 4, Gemma 3/4, Mistral, Devstral, Phi-4, DeepSeek) in a 3-column grid; hardware-adaptive display
  • VLM (Vision) support expanded: Qwen2.5-VL 7B added for 48 GB+ systems
  • Help → Model Catalog documentation section with full file lists and RAM requirements
  • License deep link — pheron://activate?key=... URL scheme for one-click activation
  • Qwen3 Dense: 0.6B · 1.7B · 4B · 8B · 14B · 32B
  • Qwen3 MoE: 30B-A3B · Coder-30B-A3B · Next-80B-A3B · 235B-A22B · Coder-480B-A35B
  • Llama 4 Maverick (512 GB)
  • Mistral Small 3.2 24B · Devstral Small 24B · Mistral Large 123B · Devstral 2 123B
  • DeepSeek V4 Flash (192 GB)
  • VLM: Qwen2.5-VL 7B (48 GB+)

Changed

  • Session titles now use the first message instead of the model name
  • Model Hub VLM section shown separately
  • Settings → AI tab now contains the Configuration section
  • WebSearchTool reliability improvements

Fixed

  • License activation window recreates correctly when opened with a pre-filled key

[1.0.1] - 2026-06-01

Changed

  • Minimum RAM updated to 16 GB across all docs and Info.plist

Fixed

  • Profile pane private relay Apple ID display (shows "Apple Account" + Apple logo)
  • Settings window resize for Profile and Analytics tabs
  • Help menu missing Refund Policy item
  • In-app Help bundle path (documents were not loading)
  • Documentation UI navigation paths corrected throughout

[1.0.0] - 2026-06-01

Public Release

Added

  • Apple Sign In with Supabase authentication
  • License activation via Lemon Squeezy

Fixed

  • Settings window auto-resizes per tab content
  • Analytics tab window sizing fix (async data load)
  • Profile pane window sizing fix