Dragging a non-image file leaves a shell-escaped path in the composer that no file tool can open

[aphelion31]

Summary

Dragging a non-image file into the TUI leaves the composer holding a shell-escaped path, which no file tool can open. Drop /tmp/a b/c d.txt and the composer gets /tmp/a\ b/c\ d.txt. That is correct shell quoting but it is not a path: Read/Edit resolve it literally, look for a directory component containing a backslash, and fail.

So the user drags a file in and the agent cannot read it. Images are unaffected (they become [image N] attachments); this only bites .txt/.rs/.pdf/etc.

Environment

• macOS 15, aarch64
• jcode v0.84.0 and current master

Reproduction

$ mkdir -p "/tmp/jcode drop repro" && echo content > "/tmp/jcode drop repro/my notes.txt"

Drag that file from VS Code into the composer. It types:

/tmp/jcode\ drop\ repro/my\ notes.txt

$ [ -f '/tmp/jcode\ drop\ repro/my\ notes.txt' ] && echo opens || echo "does not open"
does not open
$ [ -f '/tmp/jcode drop repro/my notes.txt' ] && echo opens
opens

Root cause

parse_dropped_paths already does the hard part: it resolves shell escapes, '...', "...", and file:// URLs, and verifies each token is an existing file. Those resolved paths are just never written back for non-image drops.

promote_dropped_images (crates/jcode-tui/src/tui/app/input.rs) returns early unless every dropped path is an image:

pub(super) fn promote_dropped_images(app: &mut App) -> bool {
    let Some(images) = dropped_image_files(&app.input) else {
        return false;   // <- a dropped .txt lands here, composer left escaped
    };

dropped_image_files maps over the paths and ?-returns on the first non-image, so one text file discards the whole batch.

Suggested fix

When the composer is exactly dropped paths, rewrite it to the verified paths parse_dropped_paths already produced. Since it confirmed each file exists, the rewrite is lossless. Guards that matter:

• only fire when the whole composer parses as dropped paths, so prose merely containing a path is never rewritten
• no-op when the path is already clean, so a typed path does not burn an undo entry
• keep multi-file drops quoted so they stay separable

Branch with a fix and tests, if useful as a reference: https://github.com/aphelion31/jcode/tree/fix/macos-clipboard-image-applescript

The tests assert Path::is_file() on the resulting composer rather than a string shape, so they pin the actual requirement (openable). Verified against real files whose directory and filename both contain spaces:

[PASS] R1 vscode shell-escaped   -> /tmp/.../my notes.txt
[PASS] R2 single-quoted          [PASS] R3 double-quoted
[PASS] R4 file:// URL            [PASS] R5 clean path untouched
[PASS] R6 prose not rewritten    [PASS] R7 image drop still attaches
[PASS] R8 multi-file separable   [PASS] R9 bracketed-paste route

Neutering the fix flips exactly R1-R4 to FAIL while the guard cases R5-R9 stay PASS, so the tests bind to the fix rather than passing incidentally. cargo test --test e2e is unaffected: 55 pass, and the 3 failures are pre-existing (target/selfdev/jcode missing in my checkout) and reproduce identically on the parent commit.

Notes / non-goals

Deliberately not handled, since neither can be distinguished from a legitimate filename:

• ~/... — the shell expands it before the terminal sends it, so a real drop never produces it
• percent-encoded paths without a file:// scheme — indistinguishable from a filename that genuinely contains %20

Filing as an issue rather than a PR because pull_request_creation_policy is collaborators_only, so POST /repos/1jehuang/jcode/pulls returns aphelion31 does not have the correct permissions to execute CreatePullRequest. Happy to open one if that ever changes, or you're welcome to take the diff and rewrite it.

[1jehuang]

Thanks for the report. Completed non-image file drops are normalized before dispatch, without rewriting partial key streams. Quoting, undo and image handling remain covered.

Proposed fix: #1218, commit 4e71f87. Seven targeted tests passed after reproducing four baseline failures. Broader drop and input cohorts passed 26 and 15 tests with overlap. Follow-up acceptance on the exact published binary: Enter normalized an escaped filename in the actual TUI transcript, and the real daemon read tool opened that exact path successfully.

The PR remains a draft pending clean-branch CI and review. It documents the local build limitations. This issue remains open until the fix is merged and verified.

---

— Jcode agent (automated triage), on behalf of @1jehuang