Feature/composer additions

[VaterVato]

Reasonix for VS Code brings the local Reasonix coding agent into the editor. It implements the ACP v1 client surface used by Reasonix 1.0 (main-v2) and keeps the extension focused on IDE integration: chat, native sessions, editor resources, tool review, approvals and questions, terminals, models, modes, and plans.

This repository is the standalone VS Code extension package. It does not include the Reasonix CLI or upstream Reasonix source code.

The extension does not bundle a Reasonix binary. It uses reasonix.binaryPath when configured, then falls back to resolving reasonix from PATH.

Custom build: Add to Chat, Ctrl+L, and drag & drop

This fork adds three composer conveniences on top of the upstream extension:

• Right-click → "Reasonix: Add to Chat" on a file/folder in the Explorer (or in the editor) inserts a workspace-relative @ mention at the composer caret. Selected code in the editor is inserted together with its file reference.
• Ctrl+L (Cmd+L on macOS) with text selected does the same from the keyboard.
• Drag & drop files or folders from the Explorer into the chat input. Hold Shift while dragging (a VS Code 1.91+ requirement for webviews). Images become attachments.

[Francesco146]

hi, since this repository doesn't seem to be maintained, can you reproduce the issue #17? if you can and consider to fix it, you could publish your forked extension while this one is not updated

[SivanCola]

Thanks for this, and sorry for the wait — there's a lot in the branch and reviewing it properly took some time.

Short version: the core feature is something I want in the extension, and the mention/protocol layer is well built. But I can't merge the branch as it stands: it carries fork-specific artifacts that shouldn't exist upstream, one product decision that needs its own discussion, and two defects I found while reading. I'd like to take it in pieces.

Verified locally on 7d8eae1: npm run lint and npm test both pass (80 tests), and it merges cleanly onto current main — the "conflicting" state GitHub shows is stale.

What I want to keep

• reasonix.addToChat with the explorer/editor context menus and Ctrl+L, plus drag & drop. This fits the product and the three entry points converging on one mention model is the right shape.
• The fileDrop validation in webviewProtocol.ts — scheme allowlist, URI count and length caps, offset bounds. Exactly the right level of paranoia for a message crossing the webview boundary.
• The acknowledged mentionsPicked / mentionsApplied queue with debounce and retry on stateSnapshot, so chips queued before the webview is ready aren't lost.
• The quoted @ mention grammar in resourceMentions.ts. This is my favourite part of the PR: keeping CJK paths readable instead of %E6%98%8E…, the comment explaining why encodeURIComponent isn't sufficient (it leaves ' and ) unencoded, and those still break the token grammar), and the round-trip tests.

Two defects

1. media/styles.css line 1395 — the bytes are E2 80 3F. The third byte of › (U+203A = E2 80 BA) was overwritten with ?:

content: "<invalid>;

That is invalid UTF-8, and it leaves the CSS string unterminated, so the whole content declaration is dropped and the composer's chevron disappears. Looks like a cp936 round-trip somewhere in the Windows packaging flow — worth checking whether anything else in the branch took the same damage.

2. syncAttachmentsFromDom() in src/webview.ts — the early-return guard is a tautology:

const indexes = chips.map((chip) => Number(chip.dataset.attachmentIndex)).filter(...);
if (indexes.length === chips.length && indexes.every((index, i) => chips[i].dataset.attachmentIndex === String(index))) {
  return;
}

indexes[i] is read from chips[i].dataset.attachmentIndex, so the comparison holds whenever every chip index is still in range — the function is a no-op in the normal case. Add chips A(0) and B(1), then delete A: chips = [B(dataset=1)], indexes = [1], lengths match, every passes, early return, and pendingAttachments stays [A, B]. The file the user removed is still sent with the next prompt, and the "attachment order follows the text flow" behaviour doesn't hold either. I think the check wants indexes.every((index, i) => index === i).

What can't go upstream

• dist/reasonix-vscode.vsix tracked in git (125 KB binary) and the .gitignore un-ignore that keeps it there.
• update-reasonix-fork.ps1 — hardcoded D:\Geeksoft\Node.js and the 99.0.0 version trick. Perfectly sensible in your fork, not something this repo can carry.
• The "Custom build" sections in README.md and README.zh-CN.md telling readers to install the fork's VSIX. Same reason.

Needs a separate decision

The USD→CNY conversion, the reasonix.cnyPerUsd / reasonix.cnyPerUsdAuto settings, and the daily open.er-api.com fetch on activation. That adds an outbound network call to the extension and makes RMB the default cost display for every user. I'm not saying no, but it's a privacy and product decision that shouldn't ride along with a composer feature.

The contenteditable rewrite

This is the part I'm most hesitant about. It replaces a working <textarea> with ~200 lines of hand-written offset↔DOM position mapping, and IME composition, undo/redo, paste and multi-line editing are all in the blast radius. webview.ts can't be unit tested (it calls acquireVsCodeApi() at module scope), so CI covers none of it. I'd want this on its own branch with a manual QA pass before it goes in.

Suggested split

1. resourceMentions quoted token grammar + tests — independent, self-tested, lowest risk. I'd merge this first.
2. addToChat + menus + Ctrl+L, with chips rendered in the existing attachment tray. Keeps the feature, avoids the rewrite.
3. Drag & drop + the fileDrop protocol.
4. The contenteditable composer, separately.
5. Usage card collapse — happy to discuss, but separately from the CNY conversion.
6. Dropped: the VSIX, the ps1 script, the README fork sections, the .gitignore change.

Two notes. The reasoning-block commits (01cf997, bf509a3) are already covered — I fixed #17 in #18, including the transcript scroll lock that this branch doesn't touch, and credited you as co-author on the commit since that behaviour is adapted from your work. And CI on this branch currently fails at the npm audit step for reasons unrelated to your changes (new advisories on fast-uri / qs, both via @vscode/vsce); the lockfile fix is in #18, so rebasing once that lands will give you a real CI signal.

Happy to review each piece as it comes.

---

摘要（中文）：核心功能（Add to Chat / Ctrl+L / 拖放）我是想要的，mention 与协议层写得也扎实——fileDrop 的入参校验、带 ack 的重发队列、以及 resourceMentions 里保持中文路径可读的引号 token 语法（含 round-trip 测试）都很好。但整个分支没法直接合：一是夹带了 fork 专属产物（tracked VSIX、update-reasonix-fork.ps1、README 的 Custom build 章节、.gitignore 反选）；二是人民币换算 + 启动时请求汇率 API 属于需要单独讨论的产品/隐私决策；三是有两个实际缺陷（styles.css:1395 的 › 第三字节被 ? 覆盖，导致 CSS 字符串未闭合、声明被丢弃；syncAttachmentsFromDom() 的早退判断恒真，删掉 chip 后附件仍会被发送）；四是 contenteditable 重写风险集中且 CI 完全覆盖不到，建议单独提。建议按上面 1–6 的顺序拆分，第 1 项可以先合。思考块那两个 commit 已在 #18 中处理，并已把你署为 co-author。