v2.0.0

🎉 DAIV v2.0.0 – deepagents, Label-Driven Workflows & Platform Tools

DAIV 2.0.0 is a major release that replaces the entire custom agent framework with the deepagents library, introduces label-driven issue workflows, and adds native git platform tools for GitHub and GitLab.

---

⚠️ Breaking Changes

• Issue title prefix (DAIV:) is no longer supported as a trigger. Use labels (daiv, daiv-auto, daiv-max) instead.
• pull_request.branch_name_convention removed from .daiv.yml. Define branch naming in AGENTS.md instead.
• Celery is gone. Background processing is now handled by Django Tasks.
• PlanAndExecuteAgent, CodebaseChatAgent, and PullRequestDescriberAgent classes have been removed in favor of the deepagents-based agent framework.

---

🚀 Highlights

• 🧠 deepagents library integration — The core agent framework has been completely replaced with a unified middleware-based architecture powered by deepagents, removing thousands of lines of custom orchestration code.

• 🏷️ Label-driven agent behavior — Issue labels now control how DAIV operates:

  • daiv / daiv-auto / daiv-max trigger the agent
  • daiv-auto auto-approves the plan and proceeds with implementation
  • daiv-max enables high-performance mode with more capable models and higher thinking levels

• 🛠️ GitHub CLI & GitLab specialized tools — New native gh and GitLab tools replace generic API calls, giving the agent richer platform-aware capabilities.

• 📝 Review Addressor pushes code — The Review Addressor can now push code changes directly to the repository as a draft merge request, closing the loop on review feedback.

• 💬 Inline comments — Merge request reviews now support inline comments for more precise feedback.

• 🔒 Sandbox command policy — Repositories can define allowed/disallowed bash commands via .daiv.yml, giving project owners fine-grained control over what the agent can execute.

• 🌐 web_fetch tool with SSRF protection — Replaces the MCP fetch tool with built-in multicast and IPv6 protection.

• ⚡ Celery → Django Tasks — Background processing migrated to django-tasks database backend, simplifying deployment.

• 🗄️ Postgres → Redis checkpoints — Agent checkpoints moved to Redis to prevent sub-agent hangs during concurrent execution.

---

✨ New Capabilities

• /agents and /clear slash commands
• Code review, security audit, plan, and skill creator builtin skills
• Context7 MCP server integration for live library documentation
• .agents/skills directory support for project-specific skills
• Per-repository model configuration via models section in .daiv.yml
• create_api_key management command
• django-crontask scheduler for periodic tasks (e.g., automatic GitLab webhook setup)
• Sentry AI integrations for full observability across Anthropic, OpenAI, Google GenAI, LangChain, and LangGraph

🤖 New Models

claude-opus-4.6, claude-sonnet-4.6, gpt-5.2, gpt-5.3-codex, glm-5, minimax-m2.5, kimi-k2.5

Default models updated to claude-sonnet-4.6 and gpt-5.3-codex.

🗑️ Removed Models

gpt-5.1, gpt-5.1-codex, deepseek-v3.1-terminus, gemini-2.5-pro, grok-code-fast-1, glm-4.6, qwen3-max, qwen3-coder-plus, kimi-k2-thinking, minimax-m2

---

🔧 Other Changes

• Migrated type checking from mypy to ty
• Migrated pre-commit tooling to prek
• Deferred sandbox session creation until first bash invocation
• Base sandbox image now includes git
• Anthropic prompt caching migrated to automatic cache control for OpenRouter
• Redis configured with persistence (appendonly yes) and memory limits
• Default worker replicas increased to 2
• Quick actions renamed to slash commands and merged with skills system
• Completely rewrote PR metadata generation as the diff_to_metadata module with AGENTS.md context integration
• Improved documentation for Review Addressor, Slash Commands, and Issue Addressor

🐛 Fixes

• Health check endpoints no longer generate irrelevant Sentry transactions
• Empty GitHub repositories handled gracefully when reading config files

🗑️ Removed

• Removed builtin maintaining-changelog skill
• Removed Celery worker configuration and bootstrap scripts
• Removed the quick_actions Django app in favor of the slash_commands module

---

Full Changelog: v1.1.0...v2.0.0

v1.1.0

🎉 DAIV v1.1.0 – Skills, Safer Prompts & Smarter Planning

DAIV v1.1.0 introduces an Agent Skills system, safer prompt templates, and better support for large repos and evaluations.

---

🚀 Highlights

• 🧠 Agent Skills system using Anthropic-style progressive disclosure.

• 🛠 New skills:

  • creating-agents-md-file – generates AGENTS.md following the AGENTS.md spec.
  • maintaining-changelog – maintains or creates changelogs (Keep a Changelog style).
  • creating-daiv-yml-config – helps author .daiv.yml with sandbox / formatting settings.

• 📁 Chunked file reading in read_tool (start_line, max_lines, default 2000 lines) with clear truncation messages.

• 🧪 SWERepoClient for SWE-bench–style runs on public OSS repos (no credentials, temp clones).

• 🤖 Added model support: gpt-5.1, gpt-5.1-codex, gpt-5.1-codex-mini.

• 💸 Anthropic caching middleware now supports OpenRouter.

---

🔧 Agent & Middleware Changes

• PlanAndExecuteAgent:

  • Default fallback planning model: gpt-5.1.
  • Code review model: gpt-5.1-codex-mini.
  • More structured, human-friendly plans with code minimalism guidelines.
  • New EXECUTION_THINKING_LEVEL to enable thinking during execution.
  • Uses ToolStrategy for execution to better handle union types.

• Switched from toolkits to LangChain v1 middlewares:

  • FileNavigationMiddleware
  • FileEditingMiddleware
  • MergeRequestMiddleware
  • WebSearchMiddleware

• Replaced plan_think_tool with TodoListMiddleware for explicit planning todos.

• All prompt templates migrated from Jinja2 → Mustache to mitigate code injection vectors.

• New quick action: /clone-to-topic to clone issues into all repos matching specified topics.

---

🐛 Fixes

• format_code_tool now applies patches even when the format command fails.
• .git is no longer included in sandbox archives, reducing size and repo exposure.
• Fixed InvalidNamespaceError for branch names with periods (e.g. fix/python-version-3.11) via sanitized LangGraph namespaces.

v1.0.0

🎉 DAIV v1.0.0 – Plan & Execute, GitHub Client & 1.0 Readiness

DAIV 1.0.0 turns the Plan & Execute flow into the core automation experience, adds GitHub as a first-class client, and ships a cleaner, more robust toolset for working with real codebases and pipelines.

---

🚀 Highlights

• GitHub as a client

  • You can now use GitHub as the codebase client, integrating DAIV directly into your GitHub workflows.

• Smarter, end-to-end PlanAndExecuteAgent

  • Runs sandbox commands to:

    • Install/update dependencies and keep lockfiles in sync.
    • Generate translations and perform other repo-wide tasks.

  • Understands images from user messages (via InjectImagesMiddleware), so screenshots/diagrams can inform planning.

  • Can fix pipelines using the new pipeline and job_logs tools to inspect pipeline status and job logs while planning.

  • Uses the new review_code_changes tool to check code changes against the planned tasks, ensuring they’re correct and complete before finishing.

  • Reads per-repo instructions from [AGENTS.md](https://agents.md/) (following the AGENTS.md format), so behavior is tailored to each repository.

• Repo-aware safety & control

  • New omit_content_patterns lets you hide file contents from agents (while keeping file names visible), ideal for secrets, noisy files, or large binaries.
  • CodebaseChatAgent now has a configurable RECURSION_LIMIT so you can control how deep it goes on follow-ups.

• Broader model support

  • Added support for new models across:

    • OpenAI: gpt-5, gpt-5-nano, gpt-5-mini, gpt-5-codex
    • Grok: grok-code-fast-1
    • Anthropic: claude-sonnet-4.5, claude-opus-4.1, claude-haiku-4.5
    • DeepSeek: deepseek-v3.1-terminus
    • Z-AI: glm-4.6
    • Qwen: qwen3-max, qwen3-coder-plus
    • MoonshotAI: kimi-k2-thinking

---

💡 Developer Experience & Workflows

• Simplified navigation & edit tools

  • Older read tools have been replaced with:

    • glob, grep, ls, read

  • Older write tools have been replaced with:

    • write, edit, delete, rename

  • Sandbox execution is now unified under a single bash tool.

• Better PR descriptions

  • PullRequestDescriberAgent now bases its summaries on diffs instead of commit messages, making PR descriptions more accurate and concise.

• Config & repo UX

  • Repository configuration schema has been refactored to be more flexible and easier to use.
  • DAIV now tracks file changes on the real filesystem with GitPython-backed commits, instead of in-memory tracking.

---

🛠 Reliability & Performance

• More robust async & DB handling

  • Celery tasks now properly handle Django connection pooling, preventing connection pool exhaustion.

• More efficient tooling

  • PlanAndExecuteAgent no longer re-reads the same files twice before applying changes.
  • Sandbox sessions are reused across bash calls via LangGraph’s store, cutting overhead.
  • DuckDuckGo search now uses the modern ddgs package.
  • Date-time format in prompts no longer includes hours/minutes, improving prompt cacheability.

---

⚠️ Breaking Changes (Upgrade Notes)

• Single-repo CodebaseChatAgent

  • Now answers questions about one repository at a time, identified via a repository ID header.

• Repository config schema

  • The repo config file format has changed; you’ll need to update existing configs to the new schema.

• Quick action syntax

  • @daiv help → @daiv /help
  • @daiv plan execute → @daiv /approve-plan
  • @daiv plan revise → @daiv /revise-plan

• Codebase indexation removed

  • The old indexation feature is gone, replaced by the new navigation tools (glob, grep, ls, read, etc.).

• Runtime & infra

  • Default stack updated to Python 3.14 and Postgres 17.6.

---

🔥 Removed / Replaced

• Replaced by Plan & Execute / Review flows

  • IssueAddressorAgent → superseded by PlanAndExecuteAgent.
  • PipelineFixerAgent → superseded by ReviewAddressorAgent + pipeline / job_logs tools.

• Removed agents & tools

  • CodeDescriberAgent, CodebaseSearchAgent, ImageURLExtractorAgent, SnippetReplacerAgent
  • RunSandboxCodeTool
  • All notebooks

• Dropped model support

  • claude-sonnet-4, claude-opus-4, deepseek-chat-v3.1, o4-mini

v0.3.0

🎉 DAIV v0.3.0 – Quick Actions, Cleaner Workflows

We’re thrilled to announce DAIV v0.3.0, bringing powerful new quick action triggers, smarter automation, and key stability fixes for a smoother experience.

⚡ What’s New

• Quick Actions for Issues and Merge Requests:
  You can now perform actions simply by commenting! DAIV listens to specific phrases to trigger automations:

  • @daiv plan revise – triggers plan revision on an issue.
  • @daiv plan execute – approves a plan on an issue.
  • @daiv pipeline repair – repairs a pipeline directly from a merge request.
  • @daiv pipeline repair apply – applies a repair plan to a pipeline directly from a merge request.

🔁 Updated Behavior

• Agents are now quick action–driven:

  • PipelineFixerAgent no longer requires webhooks—just a comment.
  • IssueAddressorAgent can now approve plans via quick action.

• Async-first Architecture:
  RunSandboxCommandsTool and RunSandboxCodeTool are now fully async, improving performance and scalability.

🐛 Bug Fixes

• ReviewAddressorAgent now correctly handles single-line review comments without a range.
• IssueAddressorAgent now properly removes bot labels from issue titles, fixing errors related to title parsing. (#435)
• Fixed handling of bot mentions in review comments, eliminating confusing agent prompts. (#436)

v0.2.1

🎉 DAIV v0.2.1 – Smarter Agents, Better Integration

We’re excited to release DAIV v0.2.1, featuring smarter agent behavior, improved workflows, and enhanced deployment capabilities.

🚀 What’s New

• OpenAI o3 model is now supported.
• Docker images are now built and pushed from the main branch—making edge version testing easier than ever.

💡 Key Improvement

• More targeted reviews with ReviewAddressorAgent:
  It now only addresses merge requests that explicitly mention DAIV in the discussion thread—reducing noise and making automated reviews more relevant and intentional.

🔧 Enhancements

• Default model for CodebaseChatAgent set to gpt-4.1.

• PlanAndExecuteAgent now defaults to medium thinking level for more balanced execution.

• Enhanced PipelineFixerAgent:

  • Ensures complete_task is called exactly once per run.
  • Includes richer troubleshooting context.

• Deployment docs updated with MCP proxy setup details.

• MCP server updated to @sentry/mcp-server@0.12.0.

• Base Python image bumped to 3.13.5.

🐛 Fixes

• MCP proxy requests now properly include an Authorization: Bearer header when a token is configured (#419).

🔥 Removed

• Dropped support for Anthropic's claude-3-7-sonnet.

v0.2.0

🚀 DAIV v0.2.0 Released – Now with MCP Tools Support!

We’re thrilled to announce the latest release of DAIV, and it’s a big one:
This is the first version introducing support for MCP tools, enabling agents to interact with external services in a secure, modular and extensible way.

🧠 Key Features:

• 🔌 MCP Tools (Initial Support):
  Agents can now use built-in MCP tools like fetch and sentry to interact with external systems.

  Support for custom MCP tools is on the roadmap for upcoming versions.

• 📊 Improved Agent Reasoning:

  • Rewritten prompts for PlanAndExecuteAgent, ReviewAddressorAgent, and PipelineFixerAgent for better structure, clarity, and workflow guidance.
  • Improved agents thinking workflow, increasing the quality and predictability of execution plans.

• ⚡ Async by Default:
  The entire project has been migrated to async for improved performance and scalability.

Explore the release on GitHub and let us know what you think!

Full change log

v0.1.5

Added

• Added cleanup_indexes command to clean up outdated indexes and inaccessible repositories.

Fixed

• Fixed connection closed or lost on ConnectionPool by using check_connection to verify if the connection is still working.
• Fixed chunks length check to use the correct number of tokens instead of the number of characters.

v0.1.4

Added

• Added support to claude-sonnet-4 and claude-opus-4 models from Anthropic.

Changed

• Optimized PullRequestDescriberAgent prompt to improve the quality of the responses for a 0-shot agent.
• Optimized CodebaseChatAgent prompts to improve the quality of the responses, reduce hallucinations, gatekeeping first and improve the reasoning capabilities of the agent.
• Updated PlanAndExecuteAgent to use claude-sonnet-4 as the default model for planning and execution.

Fixed

• reply_reviewer node of ReviewAddressorAgent was not using the correct tool to reply to the reviewer comments. We completely refactored the agent to turn it more reliable and robust.
• SearchCodeSnippetsTool was being called with repository parameter even when repo_id was being provided, leading to errors. Now we support conditionally add the repository parameter to the signature of the tool.
• Sometimes Document.id was being defined as an uuid when retrieving the document from the database, leading to errors.

v0.1.3

Added

• Added mapping for yaml language for .yaml and .yml extensions.

Fixed

• Temperature is being sent on o4-mini model, which is not supported.
• Large chunks were being indexed, causing errors on the embedding process. Now it will skip chunks that are too large (more than 2x the chunk size). #378

v0.1.2

Added

• Added support to 04-mini model from OpenAI.
• Added support to define LANGSMITH_API_KEY as docker secrets.

Changed

• Improved plan comment template readability by adding a separator between the each step of the plan.
• Normalized WEB_SEARCH_API_KEY to be AUTOMATION_WEB_SEARCH_API_KEY and followed the same pattern for other keys.

Fixed

• Fixed ImportError when LanguageParser try to parse a files with tree-sitter-languages, which is not installed.
• Fixed system checks to verify if required environment variables or docker secrets are set up.
• Fixed PushCallback to consider only merge requests created by DAIV to avoid indexing every merge request on the project.