Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

Dexter changelog: what's new each month

Every stable Dexter release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 8 months covered; the current month updates daily.

September 2026

2 releases: 1.0.6 → 1.1.0

September brought model updates across all major providers and introduced new session management features. Dexter now defaults to GPT 6 Astra and upgraded all saved sessions to the latest available models.

Highlights

  • New session notes, archived history, and context-window reset commands (/notes and /compact) for better conversation management.
  • All major AI models updated: GPT 6 Astra, Claude Sonnet 5, Gemini 3.8 Flash, and others across OpenAI, Anthropic, Google, Xai, Alibaba, and DeepSeek.
  • Queued messages now display under the working indicator so you can see what Dexter will process next.
  • Automatic upgrade of saved older model IDs to current versions for all sessions.

New

  • Session notes feature for storing context and metadata.
  • Archived history for managing conversation records.
  • Context-window reset command (/compact) to clear history within a session.
  • /notes command for session note management.
  • Queued message visibility under the working indicator.

Improved

  • Model picker updated to include GPT 6 Astra, Sol, and Luna options.
  • Automatic model upgrade applied to all saved sessions on first run.
  • Tab key now fills slash commands instead of executing them, allowing editing before running.

Fixed

  • Tab completion on slash hints now fills the command for editing instead of immediately running it.
  • Claude 5 structured output now works correctly when thinking is enabled.

August 2026

1 release: 1.0.5

August brought improvements to Dexter's finance data handling, ensuring complete results are returned even when data spans multiple pages.

Highlights

  • Finance API responses now automatically paginate to return all results instead of truncating at page boundaries.
  • Finance operations now have a longer timeout to handle larger paginated datasets.

Improved

  • Automatic pagination for finance API responses including prices and financials.
  • Finance sub-tool timeout increased from default to 60 seconds to support paginated fetches.

July 2026

2 releases: 1.0.2 → 1.0.4

July brought security controls for bash execution, new financial analysis capabilities, and significant evaluation framework improvements. The month focused on safety, compliance tooling, and test reliability.

Highlights

  • Bash permission rules engine lets you control which commands Dexter can run with allow/ask/deny policies.
  • New SEC insider ownership and beneficial ownership tools for tracking stock holdings and activist stakes.
  • Evaluation framework now supports partial-credit scoring, contradiction detection, and detailed performance metrics.
  • Upgraded to GPT 5.6 model family (Sol, Terra, Luna) with automatic migration from earlier versions.

New

  • Bash permission rules engine with allow/ask/deny rules stored in .dexter/settings.json
  • Always allow approval option for simple bash commands
  • SEC Form 3/5 insider ownership tool for positions, options, and RSUs
  • SEC 13D/13G beneficial ownership tool for 5%+ stakeholders
  • Rubric-based partial-credit scoring for evaluations
  • Contradiction detection in evaluation responses
  • Reproducible sampling mode for consistent test runs
  • Configurable models, concurrency controls, and timeouts for evaluations

Improved

  • Default OpenAI models upgraded to GPT 5.6 family with auto-upgrade from 5.5/5.4/5.2
  • Evaluation reporting now includes metadata, latency, and failure metrics
  • Finance evaluation answers corrected for improved accuracy
  • Rubric criteria refined for better scoring alignment

Fixed

  • Inaccurate finance evaluation answers
  • Rubric criteria misalignment in scoring

June 2026

3 releases: 2026.6.3 → 1.0.0

June brought major workflow enhancements with parallel task delegation, improved web research capabilities, and new interactive features for gathering user input mid-task. The product also transitioned to semantic versioning.

Action needed
  • Product versioning switched from CalVer to SemVer; current version is 1.0.0.

Highlights

  • Spawn focused sub-agents to run independent research, analysis, and general-purpose tasks in parallel within a single turn.
  • Agent can now pause and ask users 1–4 multiple-choice questions inline during execution, with optional free-text responses.
  • Web fetching now converts HTML to markdown, caches requests, and extracts content faster using lightweight models.
  • New CLI subagent progress UI shows per-delegate status and tool-use counts for transparent task monitoring.
  • Claude Fable 5 added to model selector.

New

  • spawnsubagent tool to delegate focused sub-tasks to isolated workers with parallel execution.
  • askuserquestion tool to pause execution and ask 1–4 multiple-choice questions with optional free-text "Other" and notes.
  • Inline question widget in chat with tabbed options and multi-select support.
  • CLI subagent progress UI displaying per-delegate status and tool-use activity.
  • Claude Fable 5 model option.
  • HTML-to-markdown conversion for web-fetched content using Readability and Turndown.

Improved

  • Web fetching now includes HTTP caching for faster repeated requests.
  • Web fetching extracts content from pages using fast models.
  • Question widget degrades gracefully on non-CLI channels (WhatsApp/gateway/subagents) when no interactive user is available.

May 2026

8 releases: 2026.5.1 → 2026.5.29

May brought major improvements to streaming and UI responsiveness, expanded financial and research capabilities, and added multiple web search providers. The month focused on richer data access, better real-time feedback, and smoother performance.

Action needed
  • getsegmentedrevenues tool retired in favor of getfinancialsegments with richer output.
  • Default model changed to GPT 5.5; fast model changed to gpt-5.4-mini.

Highlights

  • Live working indicator now shows turn duration and token estimates in real time, with improved streaming phase visibility.
  • New SEC institutional holdings tools let you look up who holds a ticker and what institutional filers own.
  • Web search now supports multiple providers (Exa, Perplexity, Tavily, LangSearch) with automatic fallback.
  • New write-memo skill generates investment memos with Bear/Base/Bull scenarios and DCF anchors.
  • Updated default models to GPT 5.5 and Opus 4.8 for better reasoning and analysis.

New

  • Live working indicator stats showing turn duration and estimated token count from streaming events.
  • Streaming lifecycle now distinguishes request, thinking, responding, tool-input, and tool-use phases.
  • getfinancialsegments tool providing income statement, balance sheet, and cash flow breakdowns by axis.
  • /search command to choose preferred web search provider with automatic fallback chain.
  • SEC 13F institutional holdings tools: look up ticker holders, filer portfolios, and resolve filer names to CIKs.
  • write-memo skill to draft investment memos with scenario analysis and optional DCF anchors.
  • Market-wide earnings feed with optional ticker filtering and table formatting.
  • LangSearch as a web search provider option.

Improved

  • Streaming lifecycle handling allows UI to show detailed request phases during tool use.
  • CLI query rendering stability improved by keying header insertion to query id.
  • Turn duration now appears in working indicator immediately (removed 3-second delay).
  • Chat performance footer simplified to elapsed time only.
  • Tool event spinner cadence kept consistent with improved timing.
  • Research rules handling improved to prevent LLM confusion and unnecessary tool cascades.
  • Institutional holdings output routed through getmarketdata with formatted CLI tables.
  • Financial search routing updated so queries like "latest earnings" trigger getearnings without requiring a ticker.

Fixed

  • UI stutter from per-chunk re-render storms eliminated by accumulating stream stats.
  • API key save flow no longer shows misleading restart message.
  • Approval prompt now displays when Dexter asks to write research rules.
  • CLI tool status now shows "Done" instead of misleading source count when market data tools return no data.
  • Memory indexer watch loop that re-triggered on SQLite writes and caused keystroke jitters.
  • Cron schedule schema updated to use z.union for Gemini compatibility.

April 2026

4 releases: 2026.4.1 → 2026.4.29

April brought financial data improvements, better user interface responsiveness, and support for DeepSeek V4 models. The month focused on refining existing capabilities and fixing edge cases in data handling.

Highlights

  • Added DeepSeek V4 model support with thinking mode enabled.
  • Implemented local caching for financial API responses to reduce network calls.
  • Improved financial search with 15-second timeouts and markdown table summaries.
  • Added hint bar with slash-command matching and better CLI rendering.
  • Added optional user research rules file (.dexter/RULES.md) for customizing system behavior.

New

  • Support for DeepSeek V4 models including thinking mode, with deepseek-v4-flash as the fast model option.
  • Optional .dexter/RULES.md file for loading custom user research rules into the system prompt.
  • Hint bar with slash-command matching and Escape key behavior.
  • Optional name filter for insider trades lookups.
  • Local caching of financial API responses with per-endpoint TTLs.

Improved

  • Financial search now uses 15-second timeouts on nested sub-tool calls and returns results as markdown table summaries.
  • Stock screener modeling updated to use 'in' for membership checks and range queries with gte + lte operators.
  • Hint bar styling enhancements and CLI rendering efficiency with reduced root re-renders.
  • Earnings output formatting now displays clearer filing metadata and core figures.
  • Centralized slash commands and matching for autocomplete (model, rules, clear, memory, heartbeat, history, help).
  • Better documentation of decimal vs percent field conventions in screener data.

Fixed

  • Embedding and index sync failures no longer disable keyword search functionality.
  • Embedding calls now time out gracefully instead of hanging during indexing or query operations.
  • Earnings tool parsing corrected to handle the API's array response shape properly.
  • Over-eager scheduled job messages no longer appear on the gateway heartbeat path.
  • Gateway heartbeat suppression behavior corrected.

March 2026

6 releases: 2026.3.5 → 2026.3.31

March brought major improvements to Dexter's financial capabilities, memory system, and multi-platform support. The month focused on real-time data integration, WhatsApp group chat functionality, and performance optimizations for longer research sessions.

Action needed
  • "Dexter" prefix removed from gateway messages.
  • Hidden files location moved to project root.
  • LangSmith tracing turned off by default.
  • Gemini 3 Pro removed; Gemini 3.1 Pro added instead.

Highlights

  • Real-time financial data support including earnings, market data, and stock screening with sector filters.
  • Complete rewrite of the memory system with better search, temporal decay, and session awareness.
  • WhatsApp gateway now supports group chats with mention-based activation and includes a setup wizard.
  • Node.js runtime support via better-sqlite3, removing the Bun requirement.
  • Concurrent tool execution and mid-run message queueing to reduce wait times during multi-tool operations.

New

  • Heartbeat tool for periodic market checks during trading hours.
  • Persistent markdown memory with hybrid SQLite recall and agent flush integration.
  • xsearch tool and x-research skill for public sentiment research.
  • WhatsApp gateway setup wizard with self-chat and bot phone modes.
  • Real-time earnings sub-tool via Financial Datasets API showing beat/miss results.
  • Stock screener with sector and industry filters and API support.
  • Cron scheduling system for running heartbeat and other tasks.
  • Concurrent execution for safe read-only tools and mid-run message queueing.

Improved

  • Memory search now uses OR queries with lower thresholds for better recall.
  • Memory tools simplified with sensible defaults (append to long-term memory by default).
  • Financial search and metrics consolidated into a single getfinancials tool.
  • Insider trades and news routed through new getmarketdata tool.
  • WhatsApp gateway utils normalization and test coverage.
  • Agent context handling with microcompaction plus full LLM-based compaction for long research sessions.
  • Model invocation improved with message-array streaming and fallback behavior for providers without stable streaming.
  • CLI and chat UX enhanced with queued-message indicators, compaction events, and cleaner tool spinner lifecycle.

Fixed

  • Gateway now uses model and provider from user settings instead of hardcoded values.
  • Memory search returning no results for multi-term queries.
  • Gemini tool schema compatibility issues with exclusiveMinimum errors.
  • Memory search now single source of truth for recall with proper FTS queries.
  • Heartbeat behavior issues.
  • Model/provider sync after runtime model switch so subsequent turns use selected config.

February 2026

9 releases: 2026.2.5 → 2026.2.25

February brought major expansions to Dexter's capabilities, including browser automation, multiple new LLM providers, and a WhatsApp gateway for direct messaging. The month focused on improving efficiency through caching and adding financial tools with better web search integration.

Action needed
  • Read Filings tool parameter changed: filingtype is now an array to support multiple filing types per request

Highlights

  • Browser automation via Playwright now enables web scraping and page interaction.
  • WhatsApp gateway lets users chat with Dexter directly through WhatsApp with QR linking and session management.
  • Added DeepSeek and Moonshot AI as LLM providers, plus Sonnet 4.6 and Claude Opus 4.6 model options.
  • Prompt caching via Anthropic reduces input token costs by approximately 90% on repeated calls.
  • Simplified financial tools with improved web search integration and stock price lookups including market cap.

New

  • Browser implementation with Playwright for web scraping and page interaction
  • WhatsApp gateway with QR code linking and session management
  • DeepSeek (V3 and R1) and Moonshot AI (Kimi K2.5) as LLM providers
  • Perplexity as an optional web search provider
  • Webfetch tool for fetching and parsing web pages with caching
  • Stock price tool with market cap support
  • Performance metrics display showing token usage, duration, and tokens per second
  • Local file cache for financial API responses to eliminate redundant network calls

Improved

  • Context compaction for better token efficiency
  • Granular browser events shown in the UI
  • Model selector now displays model names and added Sonnet 4.6
  • Read Filings tool refactored with structured planning and parallel metadata fetching
  • Financial tools consolidated into financialsearch and websearch with better routing
  • Chat history context improved with better recent turn handling
  • Error messages enhanced with API source information
  • WhatsApp reply formatting cleaned up for markdown bold text

Fixed

  • Exa search client version incompatibility between exa-js 2.x and @langchain/exa resolved
  • WhatsApp gateway now properly handles history catch-up messages and marks them as read
  • Read Filings tool improved error handling and query escaping in prompts
  • LLM error handling improved for better reliability

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Dexter: pricing, features and alternatives · All changelogs