feat(desktop): merge MCP command and args into one command-line field

[GabrielDrapor]

Extracted from #2653 per the review recommendation to split it into independently reviewable units. This is the command-line slice, standalone and independent of the OAuth work.

What

Merges the MCP editor's separate command / args inputs into one command-line field, with shell-style tokenization (quotes, escapes) parsed in mcp-command-line.ts and validated by mcp-editor-validation.ts. The dialog shows one field that round-trips to the stored {command, args} shape.

Tests

apps/desktop/src/main/__tests__/mcp-command-line.test.ts covers tokenization (quoting, escaping, unterminated quotes) and round-tripping; editor-validation tests updated for the merged field.

Co-Authored-By: Claude noreply@anthropic.com

https://claude.ai/code/session_01TMwYxgNEbz2RFmuK6AXGcj

[Astro-Han]

This slice looks very good overall. The split is clean and independently reviewable, the command-line field remains a tokenizer rather than a shell, the stored { command, args } contract is preserved, and the formatter/parser provide one clear round-trip authority. The focused tests and CI evidence also look solid. I did not find any P0–P3 issues.

The only thing missing for me is a small before/after screenshot of the stdio editor dialog. This changes the visible form from separate command and arguments fields to one command-line field, and changes the primary grid from equal columns to a 1:2 ratio. A screenshot at the normal dialog width should be enough to confirm that the command help text, environment field, and working-directory field remain visually balanced.

Once that screenshot is present, I expect this slice to be ready to approve. Thanks for extracting this into such a focused first step.

AI-assisted review disclosure: Codex reviewed the parsing, formatting, validation, renderer wiring, tests, and CI evidence. Astro-Han verified the architectural boundary, found no P0–P3 code issues, and made the final decision to request the remaining visual evidence.

[GabrielDrapor]

Here is the before/after of the stdio editor dialog at the normal dialog width:

Before — separate 命令 field and 参数 textarea, id/command in equal columns:

After — one command-line field with the shell-quoting hint; the id shrinks to slug width and the command line takes the remaining 2/3, with 环境变量 and 工作目录 unchanged below:

[GabrielDrapor]

CI note: all jobs green except two known-flaky/inherited reds — windows_recovery died in npm ci with a registry socket reset (UND_ERR_SOCKET; same job passes on the sibling stack PRs), and e2e's single failure is quote-selection.spec.ts, which fails identically on current main (run 31675601561). A rerun should clear windows_recovery; I don't have rerun permission here. Screenshots for the review ask are in the comment above.

[Astro-Han]

Thanks for the split — the boundary with #2653 is clean (only the MCP editor's 7 files, no OAuth content), the storage contract is untouched (McpStdioServerConfig and the runtime consumption path are unchanged, mcp.json schema identical), old configs round-trip losslessly through formatCommandLine, and the hand-written tokenizer is deliberately non-shell (no expansion/globs/operators), matching the old "not parsed by a shell" semantics. CI is green.

Conclusion: PASS — no blocking findings (no P0/P1/P2).

All findings are P3 (nice-to-have, non-blocking):

• F1 — configFromDraft's defensive throw is half-done (mcp-page.tsx:995-997 only checks !parsed.ok; a '""' input silently saves command: ''). The only caller validates first so it's unreachable — either add the !parsed.command check or drop the throw.
• F2 — empty-string args behavior change is undocumented: the old UI filtered blank lines, the new UI can save args: [''] via "", which spawns with an empty arg. More correct, but worth a line in the description.
• F3 — Windows trailing-backslash paths ("C:\Program Files\") report unbalanced-quote since \" is treated as an escape (standard shell behavior, but commandHelp could mention the \ rule).
• F4 — the CSS 1fr/2fr grid also widens the remote form's URL field (visual only, no correctness impact).
• F5 — test gap (deferrable): no formatCommandLine round-trip test with single/mixed quotes (logic verified correct), no dialog component-level test.
• F6 — mcp-command-line.ts:2 mixes a Chinese word into an English comment.

F1/F2 are one-liners if you touch the file anyway; F5 is a fine follow-up.

---

AI-assisted review disclosure: this review was produced with AI assistance (pi review subagent on ollama-cloud/deepseek-v4-flash). The subagent extracted the tokenizer and ran 40+ edge cases locally (including property-style round-trip tests over 30 tricky tokens — all pass) and grepped the PR branch for residual old-structure references (zero). The P3 findings are static analysis; please weigh them with your own judgment.

中文摘要（AI 辅助审查）

结论：PASS，无 P0/P1/P2，全部 P3。与 #2653 边界干净（只动 MCP editor 7 个文件，无 OAuth 内容）；存储契约零变化（McpStdioServerConfig 与运行时消费路径未动、mcp.json schema 不变）；旧配置经 formatCommandLine 无损往返；手写 tokenizer 刻意非 shell 语义（无展开/通配/操作符），与旧"不经过 shell 解析"一致。解析器经 40+ 边界用例实测正确（含 30 个刁钻 token 的属性式往返测试全过），PR 分支内旧结构零残留。P3 项：F1 configFromDraft 防御性 throw 只防一半（'""' 静默存 command:''，但唯一调用方先校验不可达）；F2 空字符串参数行为变更未声明（旧 UI 过滤空行，新 UI 可经 "" 存 args:['']）；F3 Windows 尾反斜杠路径报 unbalanced-quote（标准 shell 行为，文案可补充）；F4 CSS 1fr/2fr 同时影响 remote 表单 URL 字段（纯视觉）；F5 测试缺口可延后（无单引号/混合引号往返测试、无组件级测试）；F6 注释混入中文。