The desktop app is now jcode-desktop2. Drop the legacy crate from the
workspace, collapse the selfdev Desktop build target into Desktop2,
delete legacy-only scripts/packaging, refresh budget baselines, and mark
the legacy desktop design docs as historical.
* style: cargo fmt --all to unblock CI on master
* fix(fuzzy): drop redundant trim before split_whitespace to satisfy clippy
* chore(ci): unblock guardrails - explicit output_style re-exports and synced ratchet baselines (refs #547)
- Replace the cross-crate glob re-export in jcode-base/src/output_style.rs
with explicit items so the wildcard ratchet passes without a baseline bump.
- Re-sync the code-size, test-size, panic, and swallowed-error baselines to
current master, matching the approach in #550.
- test-size and swallowed-error baselines drifted behind already-merged
master commits, leaving master CI red and blocking PR #493
- panic and swallowed-error budget scripts now treat #[cfg(all(test, ...))]
modules as test code, which drops the panic baseline from 56 to 43
Harden Windows installation, uninstallation, hotkey, Copilot-key, power-inhibition, terminal mouse, and onboarding behavior.
Includes strict installer verification, lifecycle cleanup, Windows-specific CI coverage, and the reviewed safety fixes added before merge.
Closes#488.
Move provider_e2e.rs, live_provider_probes.rs, and lifecycle_driver.rs
(~6600 lines) out of jcode-base into a new downstream crate
jcode-provider-doctor, so doctor edits no longer rebuild the
base -> app-core -> tui spine and base compiles faster with lower peak
memory.
- lifecycle_driver had no base-internal consumers outside its own
#[cfg(test)] module, so it moves too (as a test-only module).
- native_doctor_supports_provider stays in base (auth/doctor.rs) because
live_tests' roster annotation calls it; the doctor crate re-exports it
and a sync test asserts the roster matches NativeProviderKind's specs.
- crate:: paths in the moved files rewritten to jcode_base::.
- src/cli/provider_doctor.rs now imports jcode_provider_doctor::*.
- Budget baselines relocated to the new paths (provider_e2e +29 LOC for
the new sync test).
* style: cargo fmt --all to clear formatting drift
Master's CI Format/Quality-Guardrails checks were red from accumulated
rustfmt drift across ~18 files (unrelated to any single change). Apply
`cargo fmt --all` so the formatting gate passes again.
* test(e2e): fix flaky transport parity test (memory sidecar race)
The websocket transport-parity test failed in CI: the WS resume snapshot
saw 1 message while unix saw 2. Root cause: the memory sidecar's
background extraction made its own provider `complete()` call that
consumed the single queued mock response, so the actual user turn streamed
an empty reply and never persisted an assistant message. The unix scenario
only avoided it by incidental timing.
Fixes:
- Disable the memory sidecar/extraction in the e2e test env
(JCODE_MEMORY_ENABLED=0, JCODE_MEMORY_SIDECAR_ENABLED=0) so background
LLM calls can't steal queued mock responses.
- Deterministically wait for the assistant message to persist (poll live
history until 2 messages) before resuming in both transport scenarios,
instead of relying on incidental latency.
Full e2e suite: 49 passed / 0 failed (was 48/1).
* fix(clippy): clear clippy -D warnings debt across workspace
Master's Quality Guardrails clippy step (`clippy --all-targets
--all-features -- -D warnings`) had accumulated failures (it never ran in
CI because the earlier fmt step failed first). Fix them:
- Idiomatic auto-fixes: needless borrows, needless into_iter, collapsible
ifs, sort_by -> sort_by_key, unnecessary casts, redundant clones, etc.
(jcode-base, jcode-import-core, jcode-tui*, jcode-tui-render/-mermaid).
- Gate platform-only launch-hotkey/menubar helpers used solely by macOS
code (+ tests) with `#[cfg(any(test, target_os = "macos"))]` and a
crate-level allow for the non-macOS test build, so they aren't reported
as dead code on Linux.
- memory_recall_bench (dev-bins): factor complex tuple types into
ExtractionJob/RecallResultRow aliases, allow intentionally-unused corpus
fields and the dynamic_gate wrapper, fix doc-list indentation.
`cargo clippy --all-targets --all-features -- -D warnings` is now clean.
* chore(ci): refresh quality-guardrail budget baselines
The code-size, test-size, panic-prone, and swallowed-error baselines were
last refreshed Jun 13-20 and had drifted behind master as PRs merged
(commands.rs/dispatch.rs growth, etc.), so the ratchet steps failed on
every CI run regardless of the change under test. Refresh all four
baselines to the current tree (which also accounts for the e2e
test_support growth from the transport-parity fix) so the guardrails
reflect reality again.
* chore(ios): untrack committed Swift build artifacts (.build/.build-ios)
2284 iOS Swift build outputs under ios/.build and ios/.build-ios were
committed before ios/.gitignore existed. They include symlinks
(ios/.build/debug -> arm64-apple-macosx/debug) that resolve to
directories on checkout, which broke CI's security-preflight secret scan
('grep: ... : Is a directory' -> exit 2 -> 'secret scan failed to
execute'), failing Build & Test (ubuntu).
Remove them from the index (they stay on disk and are already covered by
ios/.gitignore) so the tree no longer ships build artifacts and the scan
runs cleanly.
* fix(security): unblock CI security preflight (quinn-proto bump + lopdf ignore)
The security-preflight cargo-audit step flagged 2 vulnerabilities:
- RUSTSEC-2026-0185 (quinn-proto remote memory exhaustion): fixed by
bumping quinn-proto 0.11.14 -> 0.11.15 (Cargo.lock only).
- RUSTSEC-2026-0187 (lopdf stack overflow on deeply nested PDFs): lopdf
0.34 is pinned by pdf-extract 0.8.2 and cannot be upgraded to the fixed
>=0.42 without an upstream pdf-extract release. Only reachable when
extracting text from a user-opened PDF, not in the auth/provider/network
path. Added to the documented audit-ignore list in
scripts/security_preflight.sh with a tracking row in
docs/SECURITY_DEPENDENCIES.md, matching the existing pattern.
Combined with untracking the ios build artifacts, this makes the
Build & Test (ubuntu) security-preflight step pass.
- scan.rs: replace poison-prone Mutex<u64> counters with AtomicU64
(removes 2 .lock().unwrap() calls; also lighter for the rayon loop)
- aggregate.rs: replace .last().unwrap() with a let-else guard
- memory_recall_bench.rs: replace .unwrap() on Path::parent / Option /
partial_cmp with explicit handling
- refresh panic budget baseline: productivity-core fully cleared;
register the previously-untracked dev bench (expect-on-bad-CLI-arg only)
These four Quality Guardrails baselines were last refreshed on 2026-05-14,
before the Phase A/B crate extraction (4dd91a9c, 4aec863e) moved most modules
from src/ to crates/jcode-app-core/ and crates/jcode-base/. The baselines still
keyed the old src/ paths, so every relocated file registered as a brand-new
oversized/panic-prone/swallowed-error file and the guardrails failed on every
commit. CI only surfaced it now because the clippy step ran first and aborted
the job before these checks.
Regenerated from the current tree (no source changes in this commit). Verified
no new entries originate from this session's edits: the deltas are exactly the
src/ -> crates/ relocations from the earlier refactor.
jeremy