Commit Graph
5 Commits
Author SHA1 Message Date
陈大猫 80d9b33c59 feat(terminal): confirm-to-fill sudo password hint (#1281)
Rework sudo password autofill from auto-fill to a Tabby-style hint + Enter to confirm. When a sudo command is armed and a password prompt appears, show a dimmed inline hint instead of sending the password; Enter pastes the saved password and submits, any other key dismisses it.

Confirmation removes the credential-leak class (nothing is sent without the user pressing Enter at a visible hint), so detection is relaxed to a broad match (Ubuntu/PAM bare "Password:", "[sudo] password for…", localized prompts) and the per-host toggle is removed — always available when the host has a saved password.

Safety guards:
- don't arm when the hint can't render (no overlay) so Enter isn't silently intercepted;
- swallow Escape/Backspace so the byte never reaches the no-echo prompt;
- clear the pending hint once output moves past the prompt (sudo timeout/failure/returns to shell) so a later Enter can't leak the password to the shell.

Implementation ~140 lines; full suite green; manually verified on a real Linux host.
2026-06-07 12:39:06 +08:00
陈大猫andClaude Opus 4.7 f4aa6ddb46 fix #1013: stop ghost text from drawing over untracked echoed input (#1042)
Inline (ghost-text) suggestions render suggestion.substring(trackedInput.length)
after the cursor, where trackedInput is a client-side reconstruction of the
command line (buffer heuristics + keystroke prediction, to mask SSH echo
latency). On hosts with non-standard echo — hardware bastion hosts / network OS
like `ecOS#` (#1013, previously #756 / #906) — that reconstruction drifts and
the ghost gets painted over characters the user already typed (`int` + ghost
`terface` -> `intterface`).

Add a fail-safe consistency check: on each post-echo render, if the real
terminal line before the cursor contains the tracked input followed by more
untracked, non-whitespace characters (reality is AHEAD of what we tracked),
hide the ghost instead of drawing it over real text. SSH echo latency is the
opposite case (the line is a prefix-behind of the tracked input) and is
deliberately not flagged, so the ghost stays responsive on slow links. The
check is ASCII-only (wide-char column mapping is ambiguous) and fail-open, so
it can only ever suppress a ghost that would otherwise corrupt — never change
correct behaviour.

This converts the recurring "ghost shows already-typed characters" bug into
"ghost simply doesn't show" on devices we can't track reliably.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-21 15:40:35 +08:00
bincxzandClaude Opus 4.7 938d1ef48b Fix ghost text duplicating glyphs after Tab completion (#906)
The reliability gate at handleInput's adjustToInput call froze the
ghost at its last show()-time tail in any path where the typed buffer
becomes unreliable (Tab pass-through to shell, history recall, cursor
moves). When the user kept typing into that gap, the next render
advanced the cursor past the ghost's anchor while the ghost text
stayed put — a → -accept then pasted the stale tail on top of the
just-typed glyphs (e.g. "systemctl s" + typing "t" → screen showed
"systemctl sttop firewalld").

Add GhostTextAddon.applyKeystroke so the ghost can evolve its own
currentInput off raw keystrokes (printable / Backspace / Ctrl-W),
seeded by whatever the last show() captured from the live xterm
reading. handleInput now uses the existing adjustToInput on the
reliable path (preserves multi-char paste re-alignment) and routes
single-keystroke events through applyKeystroke on the unreliable
path, fixing the visual misalignment and the duplication-on-accept
in one shot.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-10 23:53:11 +08:00
陈大猫andClaude Opus 4.7 73b27ad7c4 fix(autocomplete): sync ghost text to live input on every keystroke (#815)
* fix(autocomplete): sync ghost text to live input on every keystroke

Ghost text was displayed based on whatever input was passed to
GhostTextAddon.show() at fetch time. Between a user's keystroke and
the next debounced fetchSuggestions firing (~100ms), the on-screen
line had already advanced one character but ghost.getGhostText() still
returned the pre-update tail. Pressing → during that window pasted the
stale tail on top of the new char — e.g. type "do", suggestion shows
"cker ls"; type "c", accept immediately → "doc" + "cker ls" lands as
"doccker ls" instead of the expected "docker ls".

Two-layer fix:

1. New GhostTextAddon.adjustToInput(newInput) that re-renders the ghost
   against a fresh input without waiting for a new fetch: shrinks /
   grows the tail if the suggestion still prefix-matches, hides
   otherwise. Called from handleInput after every buffer mutation
   (printable, backspace, Ctrl-W, paste tail) when the buffer is
   reliable. Unreliable-buffer paths skip the call to avoid making the
   ghost lie.

2. Defense-in-depth at both ghost-accept sites (→ and Ctrl-→):
   recompute the tail against the live typed buffer instead of trusting
   getGhostText's show()-time state. If the suggestion no longer
   prefixes the live buffer, hide without writing. Ctrl-→ additionally
   resyncs ghost.show() to the live buffer before picking the next word
   so getNextWord operates on an up-to-date tail.

* fix(autocomplete): defer ghost text updates to the next xterm render

The previous pass made adjustToInput re-show the ghost synchronously on
every keystroke, but xterm hasn't echoed the triggering char yet at
that moment — cursorX is still the pre-keystroke position. Painting
the shrunken tail there left it visibly overlapping with the char
xterm was about to draw, and the ghost only snapped to the right
column on the next onRender tick. That one-frame overlap is the
"jitter" the reporter still saw.

Switch adjustToInput to a defer-and-reapply pattern:

- On every keystroke that should re-align the ghost, stash the desired
  input in pendingInput and hide the element immediately. The
  transient blank frame is preferable to an overlap glyph.
- The existing term.onRender listener now checks for a pending update
  first: by that tick xterm has processed the echo, cursorX has
  advanced, and we can paint the new tail at the correct column via
  applyInputUpdate.
- New isActive() exposes "has a live suggestion even if hidden waiting
  for render" so a fast "type + →" / "type + Ctrl-→" sequence in the
  hide-until-render gap still hits the accept branch and grabs the
  recomputed tail from the live buffer.

show() and hide() clear pendingInput so an explicit state change
supersedes any queued adjust.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): restore ghost text, predict-anchor-shift on each keystroke

The previous refactor broke inline completion entirely:

1. useTerminalAutocomplete force-disabled showGhostText whenever
   showPopupMenu was on — and both are true by default, so ghost
   never rendered.
2. GhostTextAddon put its overlay container *under* xterm's screen
   via insertBefore + no z-index. xterm's default renderer paints
   theme.background across every cell including empty ones, so the
   ghost was fully occluded by the canvas even when the hook *did*
   call show().

Fixes both issues and lands the correct per-keystroke strategy the
jitter report was asking for:

- Drop the showGhostText-vs-showPopupMenu gate; respect user settings.
- Put the ghost container back on top of the screen (appendChild +
  z-index 1).
- Track anchorInputLength at show() time. adjustToInput now advances
  the ghost's left by (newInput.length - anchorInputLength) cells
  *synchronously* — i.e. it predicts where xterm's cursor will land
  once the echo arrives, instead of re-reading the live cursorX that
  hasn't advanced yet. textContent is trimmed in the same call, so
  ghost + real-input stay aligned across SSH echo latency with no
  one-frame overlap or blank gap.
- Updated GhostTextAddon.test.ts expectations for the new behavior
  (and cast the fake-document through unknown to fix the pre-existing
  TS error).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): address ghost text review feedback

Follow-ups on the predict-anchor-shift from the previous commit,
based on a code-reviewer pass:

- Backspace / Ctrl-W de-sync: updatePosition's Math.max(0, ...) was
  clamping the delta to zero when newInput shrank below the show-time
  input length. The ghost then stayed pinned at the original anchor
  column while the real cursor walked back left, leaving a gap
  between the cursor and the ghost. Let the delta go negative so the
  ghost tracks the cursor backwards; clamp the resulting left at 0
  instead of clamping the delta.
- Resize staleness: onResize now also resets lastLeft/lastTop and
  re-renders, so the dedup cache in updatePosition doesn't hide a
  now-stale pixel coordinate after xterm recomputes cell dims.
- Added a regression test for the backspace path covering both the
  step-back-below-anchor case and the clamp-at-0-on-overshoot case.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): don't accept whole suggestion when buffer is unreliable

Codex flagged (#815 P1 ×2) that the live-buffer recompute on → and
Ctrl-→ falls into a degenerate path when typedBufferReliableRef is
false. My previous cut used live = "" as the fallback, but
fullSuggestion.startsWith("") is always true — so:

- → would write the entire suggestion over whatever is on the line
  (post history-recall ↑, Ctrl-R reverse search, etc.).
- Ctrl-→ would reanchor the ghost at the start and getNextWord would
  hand back the first token, duplicating leading content on top of
  the recalled command.

When the buffer is unreliable, empty buffer ≠ empty line — the line
has content we're not tracking. Fall back to the ghost's own cached
state instead of recomputing:

- → reliable: recompute tail vs live buffer, flip buffer to the
  accepted suggestion, reliability back on.
- → unreliable: use ghost.getGhostText() (shown-at-show-time tail)
  and don't touch the buffer/reliability flag.
- Ctrl-→ reliable: resync ghost to live, then proceed as before.
- Ctrl-→ unreliable: skip the resync, derive the shrink baseline from
  fullSuggestion - current-ghost-tail so the next-word logic still
  works off whatever the ghost was actually showing.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): hide ghost on single-byte cursor/recall control chars

Reviewer caught that Ctrl-P / Ctrl-N / Ctrl-R / Ctrl-A / Ctrl-E and
friends flip typedBufferReliableRef to false but don't hide the
ghost — leaving it rendering a tail tied to the pre-recall line. The
previous commit's unreliable-→ fallback then reads that stale tail
via ghost.getGhostText() and writes it onto the recalled line,
reproducing the very duplication class the fallback was meant to
prevent (just triggered by Ctrl-P instead of ↑).

Mirror what the escape-sequence branch already does: clearState() +
return. Once the ghost is hidden, ghost.isActive() is false at the →
and Ctrl-→ gates, so the accept-path doesn't fire at all until a
fresh fetchSuggestions re-anchors it.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): drop accepted-command cache on cursor/recall keys

Reviewer pointed out that the early returns in the single-byte
ctrl-char and escape-sequence branches leave lastAcceptedCommandRef
untouched. If the user accepts a suggestion via → and then immediately
hits Ctrl-R or ↑ to pick a different command, the fast Enter path
(lines ~611-612) still reads the cached accepted command and records
it — logging the old suggestion instead of whichever command the
reverse-search or history-recall actually ran.

Null lastAcceptedCommandRef at the top of both branches (same place
we hide the ghost and flip reliability off) so accept + recall + Enter
records the recalled command, not the stale accept.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): also null accepted-command cache on Ctrl-C / Ctrl-U

Reviewer flagged this class of bug is still reachable via Ctrl-C /
Ctrl-U. The branch handling those kills the zle line, but the early
return leaves lastAcceptedCommandRef pointing at a command that is
no longer on the line: accept "git status" via → → Ctrl-C to abandon
→ type "ls" → Enter logs "git status" via the fast path instead of
"ls".

Same one-liner as the other early-return branches: null the cache
alongside clearState(). Now the cache's lifetime truly ends at any
event that invalidates the accept.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): null accepted-command cache on bracketed paste too

Fifth-pass reviewer caught the last symmetric gap: the bracketed-paste
branch appends pasted bytes to the buffer but leaves lastAcceptedCommandRef
set. Accept "git status" via → then bracketed-paste " --short" (no
embedded newline), press Enter — the fast path at line 611 still reads
"git status" and logs that instead of "git status --short".

Mirror the non-bracketed paste branch: null the cache before clearState()
returns. All handleInput paths that extend or invalidate the line now
consistently end the cache's lifetime.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): predict ghost column by cell width + wrap at EOL

Review caught two geometry bugs in GhostTextAddon.updatePosition that
only surfaced outside the ASCII happy path:

- CJK / fullwidth / emoji glyphs occupy two xterm cells but the
  predictor advanced by one char-length per code unit, so ghost
  drifted one cell left for every wide char typed and visibly
  overlapped the user's glyph.
- When the predicted column crossed term.cols the real cursor wrapped
  to the next row, but the predictor just piled more pixels onto
  `left` — ghost walked off the right edge instead of following
  onto the next line.

Fix both by switching from code-unit count to a small EAW-style
width classifier, then applying row wrapping via
  col = (anchorX + cellDelta) % cols
  rowOffset = Math.floor((anchorX + cellDelta) / cols)
against the current term.cols. Fake terminal in the test suite now
exposes cols/rows so the unit tests can exercise both invariants:

- "advances the anchor by two cells when a CJK glyph is typed"
- "wraps the ghost to the next row when the predicted column crosses cols"

Known limitation the review already flagged: on backspace-after-wide
we don't have per-grapheme widths to reverse exactly, so the negative
delta falls back to code-unit width on the deleted slice. The slice
is `currentSuggestion[currentInput.length..anchorInputLength]` which
is the same text the user would have typed, so it's correct when
only ASCII edits; wide-char backspace can still drift by one cell.
Fixing this cleanly needs a per-grapheme buffer and is out of scope.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): honor showGhostText toggle while a ghost is on screen

Codex flagged (#815 P2) that fetchSuggestions gates new ghost shows
on settingsRef.current.showGhostText, but handleInput's adjustToInput
call had no such guard. A ghost that was already active at the moment
the user turned showGhostText off would keep tracking the typed
buffer via adjustToInput on every keystroke, so the "disabled" setting
only took hold after some unrelated path called clearState().

Two-part fix:

- Add a useEffect watching settings.showGhostText. When it flips false,
  hide the active ghost immediately so the disabled setting applies to
  whatever was already on screen.
- Gate the adjustToInput call in handleInput behind
  settingsRef.current.showGhostText too, so subsequent keystrokes under
  the disabled setting don't try to move or re-show a ghost.

Codex's earlier P2 about wrap-at-EOL on line 236 is already resolved
by e61f0e8b (predict-column-with-wrap + CJK width); that comment is
against an older commit.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): self-heal stale anchor + handle backward-wrap on delete

Codex flagged two real geometry gaps in the predict-anchor-shift math:

1. Stale anchor on high-latency shells. show() captures cursorX from
   xterm at debounce-fire time, but under SSH round-trip latency the
   user's latest keystroke may not have echoed yet — cursorX is still
   the pre-echo column. With updatePosition now purely anchor-based
   (no longer reading live cursorX on every render), that stale anchor
   becomes frozen; the ghost stays one-plus cells off for the whole
   suggestion session until another show() rebuilds it.
2. Backspace crossing a wrapped row boundary. Math.max(0, ...) clamped
   targetCol at zero, so deletions past column 0 stayed pinned to the
   current row instead of wrapping back to the previous row — exactly
   the symmetric case the forward wrap added in e61f0e8b handles.

Fixes:

- Self-heal in updatePosition: while no adjustToInput has moved us
  from the show-time baseline (currentInput.length === anchorInputLength),
  re-read live cursorX/Y each render tick. Once the user starts typing
  the anchor is frozen and delta math takes over.
- Normalize the wrap for negative targetCol: `col = targetCol % cols`
  plus `if (col < 0) col += cols`, `rowOffset = Math.floor(targetCol/cols)`
  naturally yielding -1 on underflow. Clamp `top` at row 0 so a
  runaway negative doesn't render above the terminal.

Two new tests cover both invariants:
- "self-heals a stale anchor on render while no adjustToInput has fired"
- "wraps the ghost to the previous row when deletion crosses a row boundary"

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(autocomplete): restore ghost/popup mutual-exclusivity guard in hook

Codex flagged (#815 P2) that dropping the popup-wins-over-ghost
normalization inside useTerminalAutocomplete weakens the hook's own
defensive invariant. The repo enforces mutual exclusivity in two
places already — SettingsTerminalTab toggles one off when the other
turns on, and domain/models.ts normalizes stored settings so
autocompletePopupMenu === true forces autocompleteGhostText to false
— so on the normal Terminal.tsx → store path only one of the two
arrives as true. But the hook's own defaults (DEFAULT_AUTOCOMPLETE_SETTINGS)
have both flags true, and any caller that builds settings directly
from those defaults (tests, future embedders) would end up rendering
popup + inline ghost simultaneously against the repo-wide contract.

Restore the guard, comment it as defensive rather than load-bearing
so future readers don't mistake it for the hiding-invisible-ghost
bug I was fixing last time (that was really the insertBefore /
z-index issue in GhostTextAddon.ts, not this normalization).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-23 02:06:26 +08:00
陈大猫andClaude Opus 4.6 9a7d4decff feat: 终端命令自动补全系统 (#527)
* feat: 终端命令自动补全系统

实现类似 WindTerm/Fish 的终端命令自动补全功能,不依赖机器学习:

- 历史命令持久化存储:按主机分组,频率+时间衰减排序,跨会话共享
- 前缀匹配引擎:支持精确前缀匹配和模糊匹配(首字符+连续字符+词边界加权)
- Prompt 检测器:识别 bash/$、zsh/%、fish/> 等常见 prompt 模式,排除 vim/less 等程序
- Ghost Text 插件:xterm.js 自定义 addon,光标后灰色行内建议,→ 接受全部,Ctrl+→ 接受一词
- 弹出补全菜单:浮动列表 UI,↑↓ 导航,Tab/Enter 选中,Esc 关闭,来源标记(h/c/s/o/a)
- @withfig/autocomplete 集成:600+ 命令规范的子命令、选项、参数补全
- 上下文感知:解析命令行 token,根据当前位置提供对应类型的补全
- 用户配置:启用/禁用、Ghost Text、弹出菜单、防抖延迟、最小字符数等
- 快速打字防误触:检测打字速度,快速输入时抑制建议
- 输入防抖 100ms,异步匹配不阻塞 UI

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: 补全菜单混合展示历史命令和 spec 子命令

- 输入已知命令名(如 docker)时即使没有空格也预览子命令
- 历史命令条数从 8 降为 5,留空间给 spec 建议
- 修复 wordIndex === 0 时 spec 补全被跳过的问题

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: 补全菜单在终端底部时向上展开

当光标在终端下方、空间不足时,弹出菜单向上展开(底边对齐光标行),
避免溢出终端区域。列表顺序和选中逻辑不变——最可能的选项始终在顶部,
用户初始向下选择。参考 Termius 的做法。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: 补全菜单跟随终端主题 + Enter 直接执行命令

1. 补全菜单颜色从终端主题动态派生(color-mix),不再硬编码色值,
   确保与任何主题视觉一致
2. 在弹出菜单中按 Enter 选择命令时,直接插入并发送 \r 执行,
   无需用户再按一次回车
3. Tab/鼠标点击仍然只插入不执行(保留选择后编辑的能力)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: 修复 PR review 发现的全部 20 个问题

功能修复:
- #1 修复 selectAndExecute 导致命令双重录入历史:用 suppressNextEnterRecordRef
  标志位让 handleInput 的 Enter 分支跳过已经录入过的命令
- #2 修复 Prompt 末尾 $ 误判:重写 findPromptBoundary 为从左到右逐字符扫描,
  排除 $HOME/$PATH 等变量引用(检查 $ 前是否有空格、是否在 token 内部)
- #6 快速打字检测实际生效:快速打字时 debounce 延迟翻倍(200ms),等用户停顿
- #8 resolveSpecContext 处理带参数的 option(如 --name value):
  识别 option 的 args 字段,自动跳过下一个 token
- #9 Ghost text 位置随终端滚动/渲染更新:注册 term.onRender 回调
- #13 Escape 键不再拦截 vi-mode:仅在 popup 可见时消费 Escape,
  ghost text 显示时不拦截(ghost text 是被动的,不应阻止 shell 交互)
- #14 所有 setState 统一使用 EMPTY_STATE 常量,不再遗漏 expandUpward 字段

架构修复:
- #3 消除 CustomEvent 通信:改为 onAcceptText 回调注入,
  Terminal.tsx 直接传 writeToSession 回调给 hook,
  删除 createXTermRuntime 中约 20 行 listener 代码和 cleanupAutocompleteListener 字段
- #7 xterm 私有 API 访问集中到 xtermUtils.ts:getCellDimensions 统一入口,
  带缓存机制,仅在首次访问或 terminal 切换时触发 DOM 测量
- #16 删除 getCommandNameSuggestions 中多余的动态自导入 await import("./figSpecLoader")

性能修复:
- #5 合并 ghost text 和 popup 的查询路径:删除独立的 getInlineSuggestion,
  fetchSuggestions 只调一次 getCompletions,ghost text 取 completions[0]
- #10 preloadCommonSpecs 分批加载(每批 8 个,requestIdleCallback 间隔),
  延迟 200ms 启动,且检查 enabled 才执行
- #11 scoreEntry 改为 scoreEntryAt(entry, now),now 在查询开始时缓存一次
- #15 scrollIntoView 从 smooth 改为 instant,消除快速导航动画排队
- #19 loadSpec 添加 in-flight 去重(inFlightLoads Map),同一 spec 并发加载只触发一次 import
- #20 存储满时淘汰改为按 score 排序后保留前半,而非按插入顺序

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: 修复二审发现的全部 10 个问题

功能修复:
- #1(高) insertSuggestion 改用实时 detectPrompt 而非过时的 lastPromptRef,
  修复用户继续打字后 Tab 选择建议导致字符重复插入的 bug
- #2(中) handleInput Enter 录入历史优先用实时 detectPrompt,
  修复快速打字场景下 recordCommand 记录不完整命令
- #9 suppressNextEnterRecordRef 添加 100ms 安全超时清除,防止 flag 残留
- #10 getNextWord 从 index 1 开始搜索分隔符,修复 ghost text 以 / 开头时
  一次接受全部而非逐段的问题

性能修复:
- #3(中) GhostTextAddon 注册 term.onResize 调用 invalidateCellDimensionCache,
  确保 resize/字体变化后 cell 尺寸缓存正确失效
- #4 updatePosition 缓存 lastLeft/lastTop,位置无变化时跳过 DOM 写入;
  字体属性移到 show() 中只设置一次,不再每帧写 6 个 style
- #5 统一 clearState() 函数替代所有 setState({...EMPTY_STATE}),
  带 popupVisible 守卫避免无效 re-render
- #6 hasSpec 中 specs.includes() 改为 Set.has(),O(1) 查找

架构修复:
- #7 Terminal.tsx 中 autocompleteAcceptTextRef 去掉多余的 useCallback 包装
- #8 删除 AutocompletePopup 的 onClose 死代码 prop

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: popup 默认不选中任何项,用户按 ↑/↓ 后才选中

修复输入 ls 等简单命令时回车误执行联想结果的问题:
- selectedIndex 初始为 -1(无选中),Enter 直接执行用户输入的命令
- 用户按 ↑/↓ 导航后 selectedIndex >= 0,此时 Enter 才执行选中的建议
- Tab 仍然可以直接接受第一条建议(主动接受行为)
- Enter 无选中时关闭 popup 并让按键透传到终端

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: fig spec 改为从静态资源 fetch 加载,修复生产构建中补全不工作

根因:@vite-ignore 动态 import 在 Electron 生产构建中无法解析
node_modules 路径(app:// 协议只能访问 dist/ 目录)。

修复方案(与 Monaco 编辑器相同的模式):
- 新增 scripts/copy-fig-specs.cjs,prebuild 时将全部 739 个 fig spec
  从 node_modules/@withfig/autocomplete/build/ 复制到 public/fig-specs/
- Vite 自动将 public/ 内容复制到 dist/,app:// 协议可以正常访问
- figSpecLoader.ts 改用 fetch + Blob URL + dynamic import 加载 spec,
  同时保留 @vite-ignore import 作为 fallback(兼容 dev 模式)
- public/fig-specs 加入 .gitignore(构建时生成,不进版本控制)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: ESLint 忽略 public/fig-specs 目录(第三方生成代码)

与 public/monaco 相同的处理方式——这些是从 node_modules 复制的
第三方构建产物,不应被项目 ESLint 规则检查。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: 输入完整子命令名时展示其选项(如 git commit 显示 --message 等)

当 currentToken 完全匹配一个子命令时(如 "git commit" 中的 "commit"),
导航进入该子命令并展示其 options 和 sub-subcommands 作为预览。

之前的逻辑因为 name !== currentToken 过滤掉了完全匹配的项,
且 resolveSpecContext 的 consumedTokens 不包含当前 token,
导致停留在父级而看不到子级的选项。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: 修复 fig spec index.js 解析失败导致补全不工作

根因:index.js 格式为 var e=[...],diffVersionedCompletions=[...];
正则 /var\s+\w+\s*=\s*(\[[\s\S]*?\]);/ 要求 ] 后紧跟 ;,
但第一个数组后面是 , 不是 ;,导致非贪婪匹配跳到第二个 ];,
捕获了两个数组拼在一起,JSON.parse 失败,spec 列表为空。

修复:改用 indexOf 找第一个 [ 和对应的 ],直接截取子串解析。
fig spec 的 index 是简单的字符串平坦数组,无嵌套括号。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: fig spec 改用 URL 直接 dynamic import,移除 fetch+Blob 方案

fetch + Blob URL + import() 方案可能被 Electron CSP 策略阻止。
改为直接用完整 URL 做 dynamic import:
- dev: import("http://localhost:5173/fig-specs/git.js")
- prod: import("app://./fig-specs/git.js")

两种环境下动态 import 都能正常解析模块,无需 fetch 中间步骤。
同时简化 getAvailableSpecs 也用同样方式,移除 fetch+正则解析。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: fig spec 改为通过 Electron IPC 加载,彻底解决 dev/prod 加载问题

之前的方案(静态文件 + dynamic import / fetch + Blob URL)都因为
Vite dev server 对 .js 文件的模块转换和 Electron CSP 限制而失败。

新方案:通过 main process 的 Node.js require() 加载 fig spec,
通过 IPC 传给 renderer:
- main.cjs: 添加 netcatty:figspec:list 和 netcatty:figspec:load handler
- preload.cjs: 暴露 listFigSpecs() 和 loadFigSpec() API
- figSpecLoader.ts: 通过 window.netcatty bridge 调用 IPC

优势:
- main process 直接访问 node_modules,dev 和 production 都可靠
- 无需复制文件到 public/、无需 @vite-ignore hack
- spec 数据通过 IPC 序列化传输,无 CSP 限制
- 删除了 scripts/copy-fig-specs.cjs 和 public/fig-specs/ 相关代码

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: main process fig spec 加载改用 import() 替代 require()

@withfig/autocomplete 是 ESM 包("type": "module"),
CommonJS 的 require() 无法加载 ESM 模块会抛 ERR_REQUIRE_ESM。
改用 dynamic import() 在 async handler 中加载。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: fig spec 加载用 pathToFileURL 绕过 package.json exports 限制

@withfig/autocomplete 的 exports 字段只允许 import "." 和 "./dynamic",
Node.js 严格遵守 exports map 拒绝解析 build/git.js 等子路径。

改为手动拼接文件绝对路径 + pathToFileURL 转换为 file:// URL 后 import,
完全绕过 Node.js 的 package exports 限制。

同时修复 promptDetector 不再 trim 尾部空格(用 cursorX 确定实际输入长度),
确保 "git commit " 的尾部空格被保留,触发空 token 显示选项列表。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: 补全菜单添加详情面板 + 清理调试日志

- 选中或悬停补全项时,右侧显示详情面板(类似 VS Code IntelliSense)
  - 显示完整命令名、来源类型标签(Option/Subcommand/History 等)
  - 显示完整的描述文本(不再截断)
- source 标记移到左侧,与描述分离,更易读
- 悬停和键盘选中都能触发详情面板
- 向上展开时详情面板也正确对齐
- 清理所有临时调试 console.log

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: 清理全部调试日志

移除 autocomplete 模块中所有临时 console.log 调试语句,
仅保留 figSpecLoader 中的 console.warn 用于真实错误报告。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: 三审问题修复 — 移除多余 prop、过滤子路径 spec、防路径遍历

1. 移除 Terminal.tsx 传给 AutocompletePopup 的多余 onClose prop
2. getCommandNameSuggestions 过滤含 / 的 spec 名(aws/s3 等不是直接命令)
3. figspec:load IPC handler 添加 .. 路径遍历检查

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 5 个问题全部修复

1. [P1] fuzzy 匹配建议不以 userInput 开头时,用 Ctrl+U 清行再写入完整命令,
   避免 substring 截断产生损坏的命令行
2. [P2] Ghost addon 初始化改用 polling 等待 termRef,解决首次挂载时
   termRef.current 为 null 导致 ghost text 永远不激活的问题
3. [P2] popup overlay 改为 pointer-events-none 透传,仅 popup 自身设
   pointer-events: auto,不再阻止终端区域的鼠标交互
4. [P2] getCompletions 异步返回后重新 detectPrompt 校验输入是否已变,
   丢弃过时的补全结果避免覆盖新状态
5. [P2] prompt 检测支持折行:当 line.isWrapped 时向上回溯查找 prompt 行,
   拼接多行内容作为完整 userInput

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第二轮 3 个问题修复

1. [P2] broadcast 模式下 autocomplete 插入也触发广播 —
   onAcceptText 回调中调用 onBroadcastInputRef 通知其他 session
2. [P2] 支持无尾随空格的 prompt(如 cmd.exe C:\path>)—
   prompt 字符后允许直接是行尾,boundary 为 i+1
3. [P2] 光标移动 escape 序列(Left/Home/End)清除过时建议 —
   不再静默忽略,改为 clearState() 清除 popup 和 ghost text

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第三轮 3 个问题修复

1. [P2] commandBufferRef 处理 Ctrl+U 清行 — fuzzy 匹配发送 \x15 时
   重置 buffer,避免 onCommandExecuted 记录错误的拼接命令
2. [P2] fetchVersionRef 递增计数器废弃过时异步结果 — clearState/Escape
   关闭 popup 时 bump version,getCompletions 返回后检查 version 匹配,
   防止已关闭的 popup 被旧请求重新打开
3. [P2] prompt scanLimit 从 80 提高到 200 — 支持包含 git branch、
   kube context、长路径的 prompt,超过 80 列不再失效

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第四轮 3 个问题修复

1. [P1] 拒绝绝对路径 — figspec:load IPC handler 检查 commandName
   不以 / 或 \ 开头,防止 path.join 丢弃前缀导致任意 JS 执行
2. [P1] cmd.exe prompt > 后不要求空格 — 对 > ❯ ➜ › 等 prompt 字符
   不强制要求后跟空格,支持 C:\src>dir 格式
3. [P2] serial line mode 下 autocomplete 走 serialLineBufferRef —
   在串口 lineMode 时不直接 writeToSession,而是缓冲到 line buffer
   并处理 local echo,与正常按键输入行为一致

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第五轮 — translateToString(false) 保留尾部空格

translateToString(true) 会 trim 行尾空格,导致 cursorX 截取的
userInput 与实际行内容不一致。改为 translateToString(false) 保留
原始空格,确保 "git commit " 的尾部空格被正确保留用于触发选项补全。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: 设置页添加自动补全开关(启用/Ghost Text/弹出菜单)

在终端设置页末尾新增「自动补全」区域,包含三个开关:
- 启用自动补全:总开关
- 行内建议(Ghost Text):光标后灰色建议文本
- 弹出菜单:浮动补全列表

子开关在总开关关闭时 disabled。中英文 i18n 翻译齐全。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第六轮 3 个问题修复

1. [P1] 光标不在行尾时禁止补全 — 检测 cursorX 后方是否有字符,
   有则 clearState 不显示建议,避免 mid-line 插入导致文本重复
2. [P2] Enter 录入历史改为先尝试实时 detectPrompt,失败则 fallback
   到 lastPromptRef 缓存,应对高延迟 SSH 下 buffer 未回显的情况
3. [P2] fuzzy 替换在 Windows host 上用退格清行而非 Ctrl+U —
   cmd.exe/PowerShell 不支持 Ctrl+U,改为发送 \b 退格序列

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第七轮 — commandBuffer 退格处理 + 接受后历史记录

1. [P2] commandBufferRef 处理 \b 退格 — Windows fuzzy 替换用退格
   清行时正确移除 buffer 末尾字符,避免记录拼接错误的命令
2. [P3] lastAcceptedCommandRef 追踪接受的补全文本 — Tab/→ 接受后
   立即 Enter 时用追踪值录入历史,不依赖可能未回显的 buffer

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第八轮 — 历史记录准确性 + 设置同步

1. [P2] 用户继续编辑后清除 lastAcceptedCommandRef — Tab 接受
   "git status" 后追加 " --short" 再 Enter 时记录完整编辑后的命令
2. [P2] Ghost text →/Tab 接受路径也设置 lastAcceptedCommandRef —
   确保所有接受路径在快速 Enter 时都能准确记录命令
3. [P2] autocomplete 设置加入 SYNCABLE_TERMINAL_KEYS —
   跨设备同步时保留自动补全偏好

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第九轮 — REPL 误识别 + 本地终端 OS 检测

1. [P1] local terminal 的 hostOs 改用 navigator.platform 检测实际 OS,
   避免 Windows 上 fallback 到 "linux" 导致 Ctrl+U 清行失败
2. [P2] 回退 > 无条件接受改动,恢复要求 > 后跟空格或行尾 —
   避免 python >>>、mysql>、sqlite> 等 REPL 被误识别为 shell prompt
3. 新增 REPL NON_PROMPT_PATTERNS:>>>(python)和 word>(mysql/redis)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第十轮 4 个问题修复

1. [P1] cmd.exe prompt C:\path> — 对 > 特判:前面是 \ 或 / 时允许无空格,
   避免误匹配 REPL(python>>>、mysql>)的同时支持 Windows cmd prompt
2. [P2] serial lineMode autocomplete 不再 early return — fall through 到
   共享的 commandBuffer/broadcast 更新逻辑
3. [P2] serial 字符模式 + localEcho 时 autocomplete 插入文本也本地回显
4. [P3] 运行时关闭 autocomplete 时调用 clearState() 清除已显示的 popup

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第十一轮 — option args、PS2 误识别、bridge 缓存

1. [P2] resolveSpecContext 返回 option 的 args — 当光标在 option 参数
   位置时(如 git archive --format |),返回该 option 的 args 而非
   subcommand 的 args,使 tar/zip 等枚举值能正确补全
2. [P2] 排除 bare > 作为 shell prompt — bash PS2 续行提示 > 加入
   NON_PROMPT_PATTERNS,避免在多行命令续行和 REPL 中误触发补全
3. [P3] bridge 不存在时不缓存 null — preload 时 bridge 可能未就绪,
   缓存 null 会永久禁用该命令的 spec 补全

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第十二轮 — prompt 检测取最后一个分隔符

Starship/Powerlevel10k 等 prompt 包含多个 prompt 字符
(如 ➜  repo git:(main) $),之前在第一个 ➜ 就停了,
把后续 prompt 文本当成用户输入。

改为收集所有候选 prompt 边界,返回最后一个。确保
"➜  repo git:(main) $ ls" 中 userInput 正确为 "ls"。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Codex review 第十三轮 — prompt 搜索范围限制 + cmd.exe 路径

1. [P2] prompt 扫描限制在行前 60% — 避免 "echo foo > bar" 中的
   重定向符 > 被当作 prompt 结束(prompt 不会出现在行尾部分)
2. [P3] cmd.exe 路径检测扩展 — 除了 \ / 前缀,也检测行首是否有
   驱动器号 (X:) 模式,支持 C:\Users\me> 等标准 Windows prompt

P1 (高延迟 SSH buffer 滞后) 和 P2 (Enter 时 stale prompt) 属于
prompt 检测方案的固有局限,根本解决需要 OSC 133 Shell Integration,
不在本 PR 范围内。已有 lastAcceptedCommandRef fallback 缓解。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-26 19:45:34 +08:00