Files
Netcatty/components
陈大猫andClaude Opus 4.7 9bf4aed44f fix(autocomplete): stop prepending theme cwd ("~ ") to completed commands (closes #806) (#814)
* fix(autocomplete): honor typed keystrokes when the prompt parser over-captures

Closes #806.

## Root cause

findPromptBoundary stops at the first "PROMPT_CHAR + space" it sees on
the current line. Themes that render additional content after the
prompt char — most notably oh-my-zsh robbyrussell's "➜  ~ " where "~"
is the cwd — trip it: promptText becomes "➜ ", userInput becomes
"~ sudo id". Every consumer downstream treats the theme's cwd marker
as part of the user's command, so:

  1. recordCommand logs entries like "~ sudo id" into history.
  2. fuzzyQueryHistory later returns those polluted entries as
     suggestions.
  3. When the user hits Tab, insertSuggestion compares
     suggestion.text ("~ ls") against userInput ("~ lo"), falls into
     the Ctrl-U-plus-rewrite path, and the phantom "~ " ends up on
     the real command line.

The reporter hit this right after `sudo` because sudo's password
interaction gave history enough polluted entries to start winning
fuzzy matches; without sudo the popup stays empty so the Ctrl-U
rewrite path never fires and the bug is invisible.

## Fix

Track what the user actually typed in an independent keystroke buffer
(typedInputBufferRef) inside the autocomplete hook:

- Append every printable char / paste chunk.
- Pop on backspace, word-kill on Ctrl+W.
- Clear on Enter, Ctrl+C, Ctrl+U, and any escape sequence / unhandled
  control char (cursor moves we can't follow invalidate the buffer).

Introduce reconcilePromptWithTypedInput: if detectPrompt's userInput
ends with the typed buffer and is longer, the parser over-captured —
move the excess back to promptText so userInput matches what was
actually typed. Apply at every detectPrompt call site
(fetchSuggestions, the stale-result recheck, insertSuggestion).

For Enter-record the typed buffer wins outright when present, but
only after a live detectPrompt confirms we're at a shell prompt —
otherwise a password-entry Enter would log the password as a
command.

insertSuggestion / ghost-text accept update the typed buffer to the
accepted text so a subsequent Enter records the right command.

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

* fix(autocomplete): track keystroke-buffer reliability, skip it after cursor moves

Codex flagged (#814 P1) that clearing typedInputBufferRef on escape /
control sequences and then re-appending printable keys leaves the
buffer holding only the post-navigation suffix of the real line.
A classic Up-arrow-recall workflow — ↑ to pull "git commit -m fix"
out of history, append one char, Enter — would record just that one
char as the command, polluting history and skewing future fuzzy
matches.

Add typedBufferReliableRef as a companion flag:

- Reset (reliable=true) on Enter / Ctrl-C / Ctrl-U (zle wipes the
  line, our buffer is a true view of the empty line again).
- Also reset by insertSuggestion and ghost-text right-arrow accept
  once they write the full accepted text and we re-align the buffer
  to it.
- Cleared (reliable=false) when any escape sequence, unhandled
  control char (Ctrl-P / Ctrl-N / Ctrl-R / Ctrl-A / Ctrl-E / ...)
  arrives — those can move the cursor or swap the zle line in ways
  an append-only buffer can't follow.

All four call sites now gate on the flag:

- reconcilePromptWithTypedInput receives the buffer only when
  reliable, so an unreliable buffer never trims the detector's
  userInput (avoids a symmetric flavor of the original bug where
  the detector is right and the buffer is wrong).
- Enter-record prefers the buffer only when reliable; otherwise it
  falls straight through to detectPrompt.
- The Ctrl+Right (next-word ghost accept) append is skipped when
  unreliable so we don't seed the buffer with just that word.

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

* fix(autocomplete): resync typed buffer when sub-dir select rewrites the line

Codex flagged (#814 P2) that handleSubDirSelect rewrites the command
line via writeToTerminal(Ctrl-U + cmdPrefix + fullPath) but never
touches typedInputBufferRef. After the rewrite the buffer still holds
whatever was typed before, so pressing Enter records that stale partial
input as the executed command — polluting history and steering later
suggestions off course.

Same commit also routes handleSubDirSelect through
reconcilePromptWithTypedInput. The raw detectPrompt would include the
robbyrussell "~ " cwd marker in the command prefix it reconstructs,
which is the original symmetric #806 bug leaking into this path too.

After the rewrite, set the buffer to the newly written command string
and flip reliability back on — the terminal line content now matches
it exactly, so the next Enter-record does the right thing.

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

* fix(autocomplete): reset typed buffer when a paste chunk carries a newline

Codex flagged (#814 P2) that multi-character paste payloads skip the
top-of-handleInput Enter guard (which compares data === "\r" exactly),
so a paste like "cmd\r" goes through the paste branch and the "\r" gets
appended to typedInputBufferRef verbatim. The shell executes "cmd", but
our buffer is left holding "cmd\r...", still marked reliable. The next
Enter then records whatever combined stale string lives there.

Detect line terminators inside multi-char paste chunks: slice from the
last \r or \n onward and keep only that tail as the new buffer content
(and flip reliability back on, since the tail now matches the shell's
zle line). Skip synthesizing recordCommand entries for the flushed
intermediate lines — onCommandExecuted in createXTermRuntime already
tracks pasted multi-line input independently, so duplicating the logic
here would risk double-counting.

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

* fix(autocomplete): clear lastAcceptedCommandRef on paste-with-newline early return

Codex flagged (#814 P2) that the multi-line-paste branch clears the
keystroke buffer and bails out before the rest of handleInput runs —
including the line that resets lastAcceptedCommandRef. If the user had
just accepted a suggestion (Tab / → / popup click), the embedded
newline still flushes it in the shell, but our fast-path cache keeps
holding it. The next Enter then takes the lastAcceptedCommandRef
shortcut and logs that old suggestion as the executed command,
polluting history with something the user didn't actually run.

Null lastAcceptedCommandRef.current at the same point we reset the
typed buffer so the fast path stays aligned with the shell.

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

* fix(autocomplete): require typed buffer to align with live line before recording

Codex flagged (#814 P1) that paste paths which bypass handleInput —
the createXTermRuntime hotkey / context-menu / middle-click handlers
all call writeToSession(...) directly — leave typedInputBufferRef
stale while still marked reliable. A "type prefix → paste remainder →
Enter" flow would then record just the keyboard-typed prefix, feeding
garbage back into autocomplete ranking.

Require alignment: livePrompt.userInput must end with the typed buffer
before we trust it. reconcilePromptWithTypedInput already snaps the two
together when they *are* aligned — if its endsWith check fails, the
buffer is stale (or mid-navigation) and we fall back to
livePrompt.userInput instead. That drops the #806 fix for this one
paste-bypass case, but the same flow would have hit the same pollution
before this PR, so it's a no-regression fallback.

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

* fix(autocomplete): route out-of-band paste writes through handleInput

Codex flagged (#814 P1) that the reconcile path in fetchSuggestions
has the same stale-buffer failure mode the Enter-record path now
guards against: snippet / keyboard-paste / selection-paste /
middle-click-paste handlers in createXTermRuntime call
writeToSession directly, so typedInputBufferRef only holds whatever
was typed *after* the paste. reconcilePromptWithTypedInput then
treats the pasted prefix as prompt text and trims it, completions
fetch on the truncated input, and accepting a suggestion rewrites
the command incorrectly.

Fix at the source: notify the autocomplete hook with the raw
(pre-bracket-wrap) bytes at every paste site so its keystroke
buffer absorbs them through the same handleInput path keyboard
input uses. handleInput's multi-char paste branch already resets /
aligns the buffer (and invalidates on embedded escape sequences),
so this single extra call per paste site is enough — no new hook
API needed. The existing onData-driven notification at line 684
already covers the non-paste keyboard path, and the snippet /
paste / pasteSelection / middle-click handlers are the only
remaining paths that bypass it.

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

* fix(autocomplete): preserve inner newlines of bracketed-paste input

Codex flagged (#814 P2) that the multi-char-paste branch in
handleInput drops everything before the last newline, but when
bracketed paste is active those newlines are literal input staying on
the zle line — not command terminators. A multi-line paste like
"cmd1\ncmd2" then left only "cmd2" in typedInputBufferRef and the
next Enter recorded / trusted just the tail.

Teach handleInput to recognize the bracketed-paste wrapper
"\x1b[200~...\x1b[201~" and append the enclosed content verbatim
(reliability flag stays on — we know exactly what was added).

Matching change in createXTermRuntime: pass the final (possibly
bracket-wrapped) bytes to ctx.onAutocompleteInput instead of the raw
pre-wrap text so the handle sees the markers when applicable.
Non-bracketed pastes still hit the existing newline-split branch so
each "\n" resets the buffer to the post-terminator tail.

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

* refactor(autocomplete): route every prompt consumer through getAlignedPrompt

Each Codex round on #814 surfaced one more code path that needed the
"consume the keystroke buffer only when it's aligned with the live
line" gate: Enter-record, fetchSuggestions (×2), insertSuggestion,
handleSubDirSelect, fetchSubDirForIndex. The fixes were correct but
the guard ended up spelled three different ways across the file:

  reconcilePromptWithTypedInput(detectPrompt(term), reliable ? buf : "")

plus a separate `userInput.endsWith(buf)` check in the Enter branch.
That scatter is exactly how the next out-of-band writer gets missed
and regresses #806.

Collapse all six sites onto one helper:

  getAlignedPrompt(term, buffer, reliable) → { prompt, alignedTyped }

The helper owns the policy — reliability + endsWith alignment — in one
place. Non-aligned buffers fall through as raw detector output (same
pre-PR behavior, so the worst case for any future forgotten path is
a degrade, not a pollution). Enter-record additionally consumes
alignedTyped, which is only non-null when the buffer truly matches
the tail, so it can record the clean typed command directly without
redoing the endsWith check.

No behavior change from the previous commit; this is purely
deduplication of the alignment guard.

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

* fix(autocomplete): inherit reliability on bracketed paste instead of resetting

Codex flagged (#814 P1 follow-up) that the bracketed-paste branch
unconditionally flipped typedBufferReliableRef back to true. A
history-recall-then-paste flow (↑ marks the buffer unreliable, then
bracketed paste arrives) would then set reliable=true even though
the buffer only contains the pasted tail, not the recalled head.
getAlignedPrompt's endsWith check can pass trivially for a short
paste tail that happens to equal the last N chars of the recalled
line, and Enter would record just the pasted fragment.

Reliability is now inherited across a bracketed paste rather than
reset: if the buffer was already aligned, appending the paste keeps
it aligned; if the buffer was unreliable (post-recall / post-cursor-
move), it stays unreliable and the alignment guard in getAlignedPrompt
falls through to the raw detector result the way it should.

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 00:40:29 +08:00
..