Feature request: Highly specific AI helper functions

[Joshfindit]

Overall goal: To reduce context window usage

Suggestions:

1. Pull specific line numbers from a file (takes single line or start and end)
2. Pull the whole "section" from a file. This is an ambitious request as it involves analyzing code structure but it can be built up slowly.
   • Markdown: Pull the everything nested under a header
   • JS/TS/Go/Rust: Pull a function definition
3. File identification via file. Helpful to avoid loading binary files masquerading as text files
4. Incremental/Patch-Based File Editing
   • Grants the ability to apply multiple non-contiguous changes in a single tool call
5. Symlink management including resolving symlink targets
6. Modify tree to have a ".treeignore" concept. It's common that the agent trees a folder containing .git, node_modules, or other deep dev folders. It would be nice for the agent to get "path": "/home/josh/project/.git", "type": "ignored by .treeignore. call just this path to tree it anyway"

[towry]

Better use a different approach than doing this kind of thing in a filesystem tool. Use someting like windsurf's fast context tool.

[Joshfindit]

@towry How does that work in an MCP tool?

Windsurf's fast context is embedding search + a small LLM that reviews the results. One could theoretically implement this as an MCP server but it would add additional resource drain (generating and storing embeddings for all code in a way that isn't shared by other processes and is therefore duplicated as well as the resources required to query the LLM searcher).

2 is ambitious as mentioned but it's much lighter-weight than what Windsurf is doing. LSP and a language server would be the resource ceiling.

4 is already being done (to great effect) by Codex