[Bug][Linux] Escaped desktop-entry names display literally and icons disappear

[GCSimba]

Current-version update — 2026-10-07

Rechecked on 2026-10-07 against the newest public release, 0.2.1-alpha.1, and upstream master at 5badb15009ae1756c3afe0ae0cef1faafc290ccc. The Name/Icon escape fix remains applicable and is now combined with the empty-XDG-default patch. Actual Linux GIO/XDG and desktop-entry verification passed 15 cases; the Windows native-command/open-in-app selection passed 164 cases. Exec/TryExec interpretation is unchanged. All five product commits are preserved in the published current-baseline integration; both the full build and full lint passed.

Current patch index and verification limits. No all-checks-pass, upstream acceptance or released-fix claim is made. This update is AI-assisted with Codex; the published commits retain its co-author.

The original report and its verification history are preserved below.

---

Summary

On master at 5badb15009ae1756c3afe0ae0cef1faafc290ccc, the shared desktop-entry reader returns escaped display metadata literally. Legal Name, localized Name, and Icon values can therefore show escape text or fail an otherwise valid installed-icon lookup.

The Desktop Entry Specification defines five string escapes for display and icon values: \s, \n, \t, \r, and \\. Executable command text also has separate parsing rules; the patch does not broaden its interpretation.

I did not find an equivalent public report in the searches performed for this audit. This is not the empty-XDG-variable defect reported elsewhere.

Reproduction

Install a private fixture desktop entry containing:

[Desktop Entry]
Type=Application
Name=Fixture\sViewer
Name[en_US]=Localized\sViewer
Icon=fixture\sviewer
Exec=viewer %f

Provide the registered handler through the association query and a real SVG at pixmaps/fixture viewer.svg in the same private XDG data root. An otherwise identical unescaped entry serves as the control.

Before the patch, the two diagnostic fixtures produced:

{"case":"name-escaping","raw":"Localized\\sViewer","expected":"Localized Viewer","actual":"Localized\\sViewer","unescapedControl":"Localized Viewer"}
{"case":"icon-escaping","artworkExists":true,"expectedIconPresent":true,"actualIcon":null,"unescapedControlIconPresent":true}

Both pre-fix regressions failed. Expected: decode the display name and resolve the same SVG as the unescaped control.

Root cause and patch

desktopEntryFields() only splits and trims key/value text. The file-association and Host Open In App consumers use those literal values for names and icon lookup.

The patch decodes the five string escapes exactly once for Name, localized Name keys, and Icon. Escaped leading/trailing whitespace survives. A double-escaped backslash followed by s remains literal rather than undergoing a second decoding pass.

Exec, TryExec, and unrelated fields remain literal; the existing launch/parser behavior is deliberately unchanged. There is no public API, dependency, package-export, configuration, or wire-format change.

• Bug-specific commit
• Branch containing the four independent commits
• Complete parser and real-file association fixtures
• Host HTTP icon-route regression

Verification

The tests cover localized names, all five escapes, escaped outer whitespace, exactly-once decoding, command-field preservation, real desktop/icon files, and the unescaped controls. The Host Loader/WebServer route returns the actual SVG bytes for an escaped absolute icon path and retains the intended launch argv.

On Ubuntu 24.04 under WSL2, Node 24.19.0 and pnpm 11.7.0, this combined-branch command passed 402 tests with four unchanged platform-specific skips:

corepack pnpm exec vitest run packages/fs/fs-local/tests/fsio.spec.ts packages/fs/fs-local/tests/filesystem.spec.ts packages/fs/tool-str-replace-editor/tests packages/session-query/session-query-sqlite/tests packages/util/native-command/tests packages/host/open-in-app/tests --coverage --coverage.include=packages/fs/fs-local/src/fsio.ts --coverage.include=packages/fs/tool-str-replace-editor/src/index.ts --coverage.include=packages/session-query/session-query-sqlite/src/index.ts --coverage.include=packages/session-query/session-query-sqlite/src/schema.ts --coverage.include=packages/util/native-command/src/desktop-entry.ts

The shared reader and all five modified production modules reached 100% statements, branches, functions, and lines. Windows owning suites also passed. A separate Linux smoke used real GIO discovery with private XDG roots and checked the decoded localized name and actual SVG bytes. The existing real GIO handler-launch regression also passed, including two independent processes running concurrently; that regression checks general native-launch/teardown behavior, not a visual escaped-name workflow.

Full build and type-aware lint passed. The normal pre-push hook also ran the full pnpm run typecheck successfully before publication to the fork. Counts are local combined-branch results, not upstream CI results. The full hygiene aggregate remains unclaimed because 16 vendor rescope residues were reproduced on the unchanged baseline. The earlier doc-sync aggregate hit a default-heap OOM; its other 41 checks passed, and the graph check passed separately with a 4-GB heap. The updated bilingual corpus passed independently.

Scope and limitations

This does not address #8926, which concerns empty XDG environment values and directory defaults. It does not change executable-field quoting, association authorization, theme lookup policy, or WSL's choice of desktop.

No real desktop GUI window, browser visual workflow, or macOS native test was run. The evidence is parser, real-file, real GIO metadata, and real Host HTTP behavior; it is not a claim of full visual end-to-end qualification.

The patch is offered as a cherry-pickable fork commit because external PRs remain paused. Audit, implementation, and local verification were AI-assisted with Codex; the commit includes Co-authored-by: Codex <noreply@openai.com>.

[luoyukun3-oss]

Independent confirmation of the root cause, from a different machine (0.1.7-rc.2, Windows):

@deepseek-ai/dsh-native-command/lib/index.js:295 — desktopEntryFields() is exactly as you
describe. It splits lines, finds =, and stores trimmed.slice(separator + 1).trim(). There is no
escape decoding anywhere in that file (I grepped for it: zero hits). So Name=Fixture\sViewer
does come back literally, and the icon lookup gets the same literal string.

On scope: restricting the fix to the name and icon fields while leaving Exec / TryExec alone
matches the spec, since those have their own parsing rules. Minimal change looks right to me.

Two practical notes about landing it:

1. Your "external PRs remain paused" is consistent with what I see — the upstream pulls endpoint
   returns 404 for this repository, so the cherry-pickable fork commit is the only route available
   right now.
2. That makes discoverability the bottleneck rather than the patch. You currently have seven reports
   in General, each with its own commit, and most have no replies. One index post listing each
   finding with its commit sha and affected package would let a maintainer cherry-pick without
   reading seven threads; the per-report detail can stay where it is.

One small suggestion for future reports, take it or leave it: since cherry-picking is the delivery
path, putting the commit sha in the first couple of lines instead of near the end would help — on
the ones I read, I had to scroll to find it.

[GCSimba]

Current-version patch index — 2026-10-07

Thanks @luoyukun3-oss for independently checking the published packages and suggesting a clearer patch index.

I rechecked the newest public release, 0.2.1-alpha.1, published on 2026-10-03. Its tag and current master both point to 5badb15009ae1756c3afe0ae0cef1faafc290ccc. The older 0.1.7-rc.2 inspection is historical only.

The five recent product fixes are now combined and pushed on that baseline, with the original immutable product commits preserved. The XDG-default and desktop-metadata changes share files; their combined tests and paired documentation have been reconciled and verified. Storage has a separate current-layout port. The old Windows aggregate patch is retained as historical rather than reapplied wholesale.

Report	Current status	Immutable product patch
#8898: query-bearing HTTP tarballs	Valid on the checked baseline; included in the integration	efb82ab
#8926: empty XDG defaults	Valid; combined with desktop metadata and checked with actual Linux GIO/XDG variants	b53b512
#8978: overlapping literal-edit matches	Valid; unique edits reject ambiguity, replace-all counting is unchanged	d6dac61
#8979: SQLite result-header origin	Valid; derived schema 8→9 rebuilds, canonical Session JSONL is unchanged	e66d62c
#8980: desktop Name/Icon escapes	Valid; combined with XDG defaults, without changing Exec/TryExec interpretation	d6fca9b
#3719: JSON post-publication sync failure	Old patch is historical; a new port covers both current JSON layouts and domain retirement	0a3625e
#3466: Windows native ownership/error handling	Current upstream has equivalent handling of the principal reported paths; the old aggregate is not offered as an outstanding current-version fix	Historical 1fde2ce

Published branches

• Five-fix integration: immutable head d4015ccd873783029fbb31a69b46f7c0c7be489a, branch.
• Storage current-layout port: immutable commit 0a3625e84b814afc543cb8397037763079f96f76, branch.

Both fetched fork heads match their local commits. Normal pre-push hooks passed. The integration's normal merges retain the original product history; its separate support-only commit makes the documentation image-escape fixture portable on Windows and is not counted as another product bug. The storage branch does not include that unrelated fixture change.

Initial push attempts stopped in the normal hook on incomplete installed Vite files. Missing files were restored from npm archives matching the lockfile SHA-512, without changing source or the lockfile; normal hooks then ran again for the successful pushes.

Latest local evidence

• Plugin Manager install-spec/manager owning files: 105 passed with lane-sized test and hook budgets; registry: 15 passed. install-spec.ts reached configured per-file 100% coverage. The assembled Chromium tarball-install scenario passed.
• Actual Linux GIO/XDG variants and desktop-entry cases: 15 passed. The Windows native-command/open-in-app selection passed 164 cases; desktop-entry.ts reached configured per-file 100% coverage. This fills the earlier native-Linux verification gap without claiming a full desktop visual workflow.
• Linux filesystem/editor selection: 84 passed, 2 explicit skips. The preceding focused headless overlapping-edit snapshot passed and was not repeated solely for the integration merges.
• SQLite owning search-header file: 7 passed on Windows and separately on Linux after repairing the verification dependency link.
• Five-fix integration: full build and full lint passed. The documentation-site selection passed 151 cases after the support fixture correction; the affected documentation build and Mermaid leaves passed. The earlier aggregate failures remain recorded rather than relabeled as a new aggregate pass.
• Storage final Linux owning/consumer selection: 174 passed, 0 failed, 0 skipped; all five modified production modules reached configured per-file 100% coverage. Independent POSIX Loader processes and independently emitted package artifacts passed on both layouts. Full build, whole-repository lint through lint:contracts-ready, and normal commit hooks passed. The unchanged Linux NodeNext gate compiled all 328 workspace APIs against the current emitted declarations. The detailed evidence and limitations are in the current [Bug][POSIX] storage-json can overwrite a published write after directory fsync failure #3719 follow-up.

Limits and historical evidence

No aggregate all-checks-pass claim is made. Storage Windows doc-sync is 41 passed/1 EPERM image-symlink failure; the corresponding Linux documentation-site selection passed 151 cases. Storage hygiene is 13 passed/3 failed: unchanged upstream vendor residues, Windows NodeNext symlink creation, and the ACP configuration symlink checked out as text. Integration's initial contended Windows test failures and missing-addon Linux failures remain in the local evidence; later owning selections are separate results.

Current workspace storage uses the single layout; projection cache uses per-record, so #3719's whole-unit unrelated-record overwrite is not generalized to that cache. Its stale-domain case is covered separately. Fault injection proves live-process consistency, not physical power-loss behavior. Consumers must reopen retired handles; automatic consumer retries are not added.

The seven newest-version npm archives were inspected and their integrity verified in the preceding audit, but official released runtimes were not executed. Current patch verification uses modified source, Loader applications and freshly built artifacts. No real-model API, whole-repository coverage, exhaustive platform matrix, or upstream CI success is claimed.

For #3466, equivalent current handling is not a claim that upstream adopted the fork commit or that every native edge case is fixed. Any residual case requires a new focused reproduction rather than blanket application of the historical patch.

The contribution guide still declines external PRs. No upstream PR, acceptance or released fix is claimed. Existing reports and immutable product history are preserved; this index is a comment on the existing report, not a new duplicate Discussion.

Audit, implementation, validation and this follow-up were AI-assisted with Codex. The commits include Co-authored-by: Codex <noreply@openai.com>.