feat: jump to symbols from assistant outputFeat/reference scoped symbol navigation

[jensenojs]

I really like this feature, lol

Summary

• add reference-scoped symbol navigation for assistant output
• highlight jumpable symbols from files already referenced in the conversation
• make <CR> jump to symbol definitions when file/diff targets do not match
• keep gf limited to file/diff targets
• highlight assistant file references with a lighter path color

Notes

This uses Tree-sitter over the current conversation's Code References. It does not add LSP, workspace-wide symbol search, or a persistent symbol index.

Tests: ./run_tests.sh

[jensenojs]

Design note for review:

This is deliberately not a workspace symbol search. The feature only looks at files already referenced in the current conversation.

Render path:

assistant text
-> file references get OpencodeReference
-> conversation Code References are collected
-> Tree-sitter reads definitions from those referenced files only
-> matching output tokens get OpencodeSymbolReference

Keypress path:

<CR>
-> existing file/path/diff target first
-> if no target, collect current conversation refs
-> read current Tree-sitter definitions from those referenced files
-> jump/pick matching symbol definitions

The render highlight is only a hint. It does not store jump targets. This avoids stale extmarks becoming navigation authority.

No LSP, no workspace-wide search, no persistent index, no edit-event subscriptions. gf remains file/diff-only; only gets the symbol fallback.

[sudo-tee]

I love the feature and find it really useful.

But I noticed a big performance hit in snapshot collection

A semi large session will cause a couple of seconds delay before displaying in the output buffer.

[jensenojs]

Because performance issues exposed some oversights in my vibe-coding process, I took some time to reflect and clean up the code, which is why it took a bit longer. However, we can see where the problems were from the previous comments. A simpler conceptual approach should look something like this:

The GIF might have been compressed a bit too much, so the color differences might not be very noticeable. But generally speaking, the "symbol" and "file" jumps will update according to our changes. You can try commenting them out or uncommenting them to test it.

The durable fact is the current conversation’s reference-file set. Session load rebuilds it from messages; message updates and file lifecycle events refresh it and dirty rendered assistant text parts. Render reads that file set plus current Tree-sitter definitions, then writes visual affordances and RenderState targets. Keypress does not rediscover refs from screen text; it only executes the current RenderState target, resolving symbol targets against that target’s candidate files at press time.

So extmarks are no longer navigation authority. They are just the visual projection of RenderState targets.

After this PR and the other one(#439) are merged, my first instinct when opening opencode might no longer be the TUI, there are (currently) no interaction methods in the TUI that opencode.nvim cannot easily handle; conversely, we can do many things that the TUI simply cannot.

I think having some reasonable interactions in the output window is a good thing. When I have time later, I might tinker with the input window as well. Better delivery methods are all low-hanging fruit, and these can likely be showcased in the README. I doubt the other opencode.nvim plugin has this functionality, hhh.

However, I may not be able to dedicate as much energy to this project in the next month or two as I am now. I will likely still assist with some bug fixes or PR reviews, as I need to spend some time preparing for my next job. But please trust that I have laid a solid foundation for your agent to take over this feature. The ui/AGENTS directory maintains the minimal mental model of this feature; it should be relatively "cheap" for the agent to understand what each part of the code is doing right or wrong (which, in my view, is more important in a sense).

[jensenojs]

One direction I'm interested in for the future is automatic prompt injection/enhancement. My GPT has been too stupid lately, and manual fine-tuning is too exhausting. I want to create an automatic prompt injection/enhancement system to save my time for those high-quality conversations that I don't dare reply to lightly. If it could achieve over 80% accuracy compared to my usual prompt responses, that would be a very good thing.

• [discuss] : CursorTab as the input method for the Agent era? cursortab/cursortab.nvim#90 (comment)

This job isn't very difficult; I've already started the refactoring over there, but there's still quite a bit of work to do lol

[sudo-tee]

I am not sure about updating the AGENTS. As it adds way more tokens than it's worth. Many you could transform this into a skill?

[jensenojs]

I am not sure about updating the AGENTS. As it adds way more tokens than it's worth. Many you could transform this into a skill?

Okay, I'll withdraw them all for now and consider adding them back later when they are worth formalizing as skills

[sudo-tee]

With this PR I lost the ability to gf on a file path like this one