Help shape the OpenSpec roadmap! What do you want to see?

[TabishB]

Hey all!
We pushed out 0.7.0 yesterday. Here’s a look at what’s next:

Near term

• Native integration with all major tools that support custom slash commands (Windsurf. Roo Code, Gemini CLI etc.)
• /verify command to check off specs and tasks more reliably
• Q&A Proposal mode where the agent asks clarifying questions to catch missing details

Longer term

• TUI dashboard (looking at opentui)
• Interactive change walkthrough with progressive disclosure, like reviewing a PR step by step
• Roadmapping features: plan multiple changes and add them to a roadmap. Right now we can only see what changes are proposed, but not the priority or the dependencies between them.

Let us know which of these you are most excited for, and drop any suggestions below.

[nicholas-alonzo]

Would you consider adding Continue as another tool?

[TabishB]

@nicholas-alonzo Hey sure, I'll add it to the roadmap. Seems like they do support slash commands: https://docs.continue.dev/customize/deep-dives/prompts

Do keep in mind openspec works with anything that uses https://agents.md/

[nicholas-alonzo]

Awesome, thank you! I'll take a look at agents.md, still learning about all these tools

[TomLucidor]

I want to see that some of the tricks (including ones in MCP) in BMAD, SuperClaude, and/or Claude Flow can be included in OpenSpec. Ideally OpenSpec can self-evolve to create sub-specs that covers soliciting requirements, tracing/documenting dependencies, and doing CI/CD by default.
"Learning from mistakes" and loop-breaking is also a plus. Documents should not just cover specs, but bug tracing + troubleshooting SOPs such that they can spot potentials for improvement, without nuking the whole codebase. https://github.com/PV-Bhat/vibe-check-mcp-server
Ideally, if there are CLI tooling out there that includes intelligent model routing, have SLMs do simple verification and script execution, while using LLMs to write quality code and thinking deeply about errors. Generalizing CCR would be good https://github.com/musistudio/claude-code-router?tab=readme-ov-file#router

[filipeestacio]

I'd like to see support for different backends. I really want to promote OpenSpec in a corporate setting, but it's being challenged on the practice of saving everything in markdown files.

I appreciate that there are advantages to keeping the notes close to the code, but for the purposes of exposing the specs and progress to non-technical staff or non-developers, it would be great if it could save context in tools like Confluence, Notion, Obsidian, etc and keep track of tasks in Jira or other task managers.

[TomLucidor]

Foam/Obsidian for minimalism and robust MCP options (less vendor locks).

[filipeestacio]

I'm experimenting with Clickup. Ideally it would be a solution leaning on MCP but agnostic from the vendor. Maybe configurable?

[AaryanGlobal]

It could be great to think of having a prompt inject to integrate with platforms like Superthread.com that already have an MCP to add context cards and project management.

[feng7605]

You can consider designing the whole process, development-testing-code submission branch.

[TomLucidor]

Check BMAD first if you want a whole process ready

[mai-dev-j]

I would suggest keep it as simple as it is now, no too many docs and no long list of commands

[rendeyuwei]

1. Depending on the complexity of the task, can the process be adjusted accordingly?
2. Should summarizing lessons learned from tasks and applying them to each subsequent iteration also be established as a standard process?

[airtonix]

Same as @filipeestacio in #111 (comment)

I want to see different ways to store the planning and task artifacts.

I don't want these md files living in the repo where the agents work to implement the changes.

perhaps you can provide plugin interface a way to abstract how these artifacts are created?

Personally, i'd like to see options like:

• LocalFileSystemStore
• BasicMemoryStore
• RemoteObsidianRestStore
• ConfluenceStore
• GithubProjectStore
• JiraStore
• SqliteStore
• VibeKanbanStore
• etc

[benjaminalgreen]

+1 especially for VibeKanbanStore. I'm trying to hack something together for that since I discovered it.

[itsklimov]

I have just installed it and it seems like the framework promotes Sequential execution.

Would be cool to have planning with opportunity for parallel development.
& do development paralleled where possible.

and another think is to use subtasks/subagents where possible to save context window of the main session.

[TabishB]

@itsklimov Yep, there's some better parallel support coming. it's a bit tricky because working on things in parallel leads to spec conflict which we would need to define mechanisms to resolve.

[aleksandr1592]

I’m also very interested in sub-agents - not necessarily running in parallel, although that would be nice. What I really want is the ability to break tasks down so that each agent only has the context it actually needs. It would also be great to introduce roles into the workflow, such as developer and QA.

[WarlockSyno]

I would love for the workflow of https://github.com/GWUDCAP/cc-sessions and this be combined. Sessions is very good at keeping Claude on task and preventing it from making changes that are not approved or planned, but the workflow is lacking without the spec based tasks.

It would be amazing if some of the hooks and scripts that cc-sessions uses were integrated to keep Claude on task. I much prefer to use OpenSpec but Claude likes to do whatever it wants sometimes and forgets what and why we are working on something.

Part of the sessions task creation process has some benefits over the OpenSpec one, but overall OpenSpec is a lot nicer to use.

Oh and the Sessions status line is really nice to use too.

[Karnak19]

What I'd like to see is a ticketing system linking/integration using MCP or CLI during the init command. In my case we are using Linear with the MCP but it should works with any other.

I've modified the commands to ensure:

• the proposal create a ticket
• the apply moves the ticket to in progress, creates a branch, then start to work.
• the archive is moving the ticket to in review, commit, push the branch and opens a PR.

Why ? I fear to use openspec update and have to re-update the part for my Linear integration.

We could go further to also sync our specs in our code and in the cloud to have it shared with non-tech people in the team

[RodrigoGP-RSM]

Have you considered implementing writing specifications as a skill rather than a custom command, or extending the tool using the Open Skill Standard?

[TabishB]

@RodrigoGP-RSM Yes, we'll add official skill support soon. FWIW I don't think skills will drastically change OpenSpec. The way we use AGENTS.md / CLAUDE.md pretty much work like a skill at the moment.

[ShuBuShao]

希望能够支持多语言

[xihangzhou]

I would like to know how to split my spec into small parts that can describe some domain knowledge so that AI can produce more accurate code

[TabishB]

@xihangzhou This is being looked into, I have this problem myself. Will most likely come as a skill for "breakdown"

[harikrishnan83]

• A Constitution (as in Spec-Kit) or Steering Docs (like Kiro) to establish the rules for how the application should be built. The project.md file seems to be a mix of both tech and non-tech and does not seem to be enforced strictly.
• Since OpenSpec is brownfield friendly it will be good to have a separate command like /openspec:baseline? this will clearly articulate that I am currently establishing the source of truth spec and anything here on out will be delta specs. Right now when I introduce OpenSpec to an existing codebase, I am using the project.md to establish baseline and it seems to mix up tech, domain and other aspects.
• Right now design as well as implementation related artifacts are all generated during openspec:apply. If we have an /openspec:plan phase where artifacts such as schema design etc. are created as standards based documents, then it will be easy to review those to make sure those are aligned before we proceed to tasking.

I have not yet not yet fully thought through below point.

• Since many of us may be moving to OpenSpec from other tools, does it make sense provide any optional pathways that help with transitioning without context loss when spec format changes? or is it enough to just ignore those spec formats and let OpenSpec create a baseline spec based on the implementation alone? (I do realise that in other tools the specs may be out of sync with the implementation and that can cause confusion)

[jakubvedral]

I completely agree with this.

[clay-good]

I recently made a proposal / cli tool for trying to establish baseline in brownfield environments: #634

[GITHUBWYW]

I have a question about the large language model. Does the openspec support Gemini very well? It seems like the response from Gemini isn't same with Claude.

[TabishB]

@GITHUBWYW OpenSpec works with most models it's just the quality of the results vary depending on the model used. Usually models with high reasoning perform better in general at planning tasks. We recommend using Opus 4.5 and GPT 5.2 at the moment for OpenSpec.

It can still work relatively well with Gemini 3. It's hard to provide instruction sets that are different per model, so it becomes a matter of how well can a model follow instructions.

Either ways in grand scheme of things don't sweat over details like these - just give it a try and see if it produces good results for you :)

[mdlmarkham]

What about using Beads to manage the backlog?

[j2h4u]

Having the Beads support out of the box would be great!

[harikrishnan83]

In general backlog management integration will be helpful. Here is a quick walkthrough of how I was able to integrate OpenSpec with Linear MCP for product backlog - https://www.youtube.com/watch?v=u7tlYs2WgQU

[devdnn]

In general backlog management integration will be helpful. Here is a quick walkthrough of how I was able to integrate OpenSpec with Linear MCP for product backlog - https://www.youtube.com/watch?v=u7tlYs2WgQU

It was good to see an integration and a good separation of what from the how.

Any plans to propose to merge to OpenSpec?

[harikrishnan83]

Thanks @devdnn . Glad you liked it. I have shared the same on the discussions forum - #487 to get more feedback. Based on that we can decide on what is the right way to package this capability. Can you please comment on that discussion so that we can continue the conversation on that thread?. Regards.

[emreakay]

Hi — I’d love to see an IDE plugin for IntelliJ and VS Code, similar to what Kiro built. It would be great if the plugin could display specs and tasks inside a Kanban-style board so you can visualize and manage changes directly from the editor.

[j2h4u]

Just search VSC marketplace for "opencode", you'd be surprised

[jasonKRR]

"More than anything, I hope to see improvements in the functional aspects.

Currently, the biggest problems are:

1. The proposal design task content does not match up every time.
2. Even when following the tasks, things are often missed or done incorrectly.

To address this, we need the following features:

1. A way to verify that all documents share the same content and follow the right standards when writing a proposal.
2. A function to validate and correct the work at each task phase to ensure it is being done properly."

[TabishB]

Yeah I hear you, reliability of tasks has been a challenge, especially across multiple models (some models are worse at following instructions than the others) and across varying context sessions (if the context session window is already full this will lead to a higher risk of incorrectly formatted files).

We're trying to fix this by giving users who care about reliability a way to step through file creations instead of creating them all at once. We've documented this new workflow under the experimental workflow docs: https://github.com/Fission-AI/OpenSpec/blob/main/docs/experimental-workflow.md

For users that do want to create all documents at once we've preserved an equivalent /opsx:ff command, but we can't guarantee proper formatting all the time when using the fast forward mode.

[j2h4u]

some models are worse at following instructions than the others

Hey! Quick question--do you guys have any "evals" you could share? Not talking about anything super formal or detailed, just a sentence or two. If you could name a couple models you think are better or worse, that'd be awesome

[TabishB]

@j2h4u Nothing official at the moment unfortunately. Part of the re-factoring we've done is to make room for this down the line so we can more reliably track how OpenSpec works and performs with different models.

That being said we consistently hear that Opus 4.5 and GPT-5.2 are very good at planning using OpenSpec (I personally use Opus 4.5 all the time).

[amardeep]

I would love if there can be a section in docs how to best set it up for monorepos with multiple sub projects and tech stacks.

[viteinfinite]

Hi, sorry for the bittersweet feedback. I've been a openspec fan for a long time and I also spent some time recording a podcast praising OpenSpec for its effectiveness and simplicity. Sadly, I believe this version goes against what made OpenSpec great: straightforwardness.

I personally dislike:

• having to get back to the doc to recall how this is supposed to work
• needing to complete one more step to achieve the same results I achieved before
• having to fill the proposal document: is it necessary? Why wasn't it needed before? What's the purpose of filling it and having to edit it again after ff?
• :proposal, :apply, :archive was great naming. openspec was a great name too. Why changing it? What user problem does renaming to opsx even solve?
• ❌ last but not least: 10 skills?! Welp! As much as I loved this, this is simply a bad practice, and a sadly very efficient way to burn your "skill budget"

I personally feel like OpenSpec wants to grow into a BMAD-like framework, which verbosity and complexity is something users like me have always tried to avoid.

(On a more positive note, thank you for the work that you put in this. I just personally believe it's not the right direction).

[clay-good]

I recently made a proposal: #634

I've built spec-gen, an open-source CLI tool that reverse-engineers OpenSpec-compatible specifications from existing codebases. It addresses the "cold start" problem for teams wanting to adopt OpenSpec on established projects (brownfield environments). It included a Claude Code skill, OpenSpec skill, Agents.md, and a cli tool. I'd love to contribute this to OpenSpec.

[randybias]

Would really love to see support for Agent Teams in the latest Claude Code. You can kind of make it work right now, but it's kind of janky. My flow is to use planning mode and go back and forth until there is a comprehensive plan. Then clear and bypass permissions, but you kind of have to stop it immediately and then give it a prompt something like this:

Deploy a NEW team of agents who will use the openspec skill to manage the implementation, where you have a Sr Product Mgr who plans and sets up changes, a Sr Architect who helps with the implementation planning and details, a Sr Developer who implements the code, a Code Reviewer that reviews the code for security, gaps, completeness, and taste (clean, well thought out code), a Sr QA engineer that creates tests, expands test coverage as necessary and makes sure the tests pass, and a Sr Tech Writer that updates documentation and the tentacular skill.  ALL agents on the team should spawn as many subagents as necessary to speed themselves up.

The Sr Developer and Sr QA Engineer should use the latest Sonnet model.  The rest of the team should all use the latest Opus model.

This works pretty well in that the different agents employ the different skills, but obviously it could be improved. Clear alignment between roles and openspec skills would be good as well as pre-canned sets of agent teams. It enhances the usability of openspec if the skills are aligned with those roles.