discover_project_configs labelled every project by basename alone, so
~/work/app/.mcp.json and ~/personal/app/.mcp.json both read as
"Project: app". Paths dedupe correctly, so both profiles existed -- they
were just indistinguishable in the picker, and picking the wrong one meant
editing, backing up, and writing the wrong repo's config. api/web/app/
server/client as repo names make this common.
- disambiguate_project_labels(dirs): pure, testable. Labels stay
"Project: <name>" until a basename collides, then only the colliding
ones widen toward the root one component at a time
("Project: work/app" vs "Project: personal/app"), widening further if the
parent also collides. The common no-collision case is unchanged.
- Full path goes in the combo item's ToolTipRole, so a profile is always
verifiable by hover regardless of label.
- Tightened the "is this a project config?" check: the docstring claimed a
top-level object but the code only checked is_file(), so a .mcp.json that
was a JSON array or garbage still became a profile and only failed on
open. It now must parse (strict) to a dict, else it's skipped.
Tests: +6 (no-collision basename, colliding widen, deeper widen, and
discover_project_configs disambiguation + skipping array/garbage/missing).
463 passed, ruff clean.
Closes#74
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EKwBecy6N83jnqQmw8ezwE
Phase 1 of cross-client support: introduce the ClientSpec adapter and route Claude Desktop + Claude Code through it with no behavior change. Parameterizes servers_key/disabled_key/discovery and the entry translation seam; extract_servers/apply_servers/_server_sections/external_change_summary default to Claude's layout so no-spec calls are byte-identical. Verified: 458 tests + a synthetic non-mcpServers client proving the seam generalizes, CI green on all platforms incl. Windows, and the live GUI confirmed loading/saving identically. requirements.txt now declares cryptography as the runtime dep it always was.
Refs #5
bcc_core imports cryptography at module load (catalog signature
verification), so it's required to even start the app -- but
requirements.txt listed only PySide6, so a from-source run died with
ModuleNotFoundError: No module named 'cryptography'. The frozen release
builds were unaffected because PyInstaller follows the import, which is
why this never surfaced until someone ran the GUI from source to review
this branch.
Move cryptography into requirements.txt (runtime) and re-label it in
requirements-dev.txt as runtime rather than test-only; the version pin is
unchanged (>=42.0).
Refs #5
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EKwBecy6N83jnqQmw8ezwE
Cross-client support (Cursor / Windsurf / VS Code) was blocked on Claude's
layout being hard-coded throughout the code: servers always under the literal
"mcpServers" key, a fixed set of Claude file locations, and "which client is
this?" answered by sniffing a filename. Adding a client that differs on any of
those axes meant chasing those assumptions through a dozen sites.
This is phase 1 of #5: the keystone refactor, with NO behaviour change. It adds
a ClientSpec adapter that captures the three things that vary across clients --
the top-level servers key, the config discovery paths, and the per-server value
shape -- plus the capability flags that were previously computed inline from a
filename (does the client expand ${VAR}? can we offer Restart?).
- ClientSpec (frozen dataclass): servers_key, disabled_key, config_filename,
expands_env_refs, supports_restart, and entry_to_internal/entry_from_internal
-- the per-server translation seam, identity for any mcpServers-shaped client,
the single point a differently-shaped client (VS Code's type/inputs form)
overrides.
- CLAUDE_DESKTOP and CLAUDE_CODE specs; both use mcpServers + the existing
parking key, so their translation is the identity and nothing changes for
today's users. resolve_client(path) reproduces the old filename rule exactly;
each Profile now carries its resolved .client.
- extract_servers / apply_servers / _server_sections / external_change_summary
take an optional spec and default to Claude's layout, so every existing call
site and test that omits a spec is byte-for-byte unchanged. The cardinal rule
now generalises: apply_servers only ever writes the client's own two keys,
parameterised rather than hard-coded.
- profile_targets_claude_desktop and client_expands_env_refs are now thin reads
off the profile's spec -- one source of truth for client identity instead of
scattered filename checks -- with identical answers.
- GUI: the load, Copy-to, save and stale-merge paths pass the profile's spec
into the core calls. Mechanical; no logic moved into bcc.py (which CI can't
test -- no PySide6).
Tests: +14. Existing suite unchanged and green (behaviour preservation). A
synthetic non-mcpServers spec ("servers" key, a different disabled key, a
per-server `type` field) exercises the whole pipeline -- extract, apply,
masking, external-change diff -- proving the seam actually generalises before
any real client depends on it. 458 passed, 1 skipped; ruff clean.
Refs #5
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EKwBecy6N83jnqQmw8ezwE
Adds the authoring + safety layer for ${VAR} references: expand_env_refs preview, per-client gating via profile_targets_claude_desktop (Claude Code expands natively, Desktop does not), and fixes two backwards behaviours (masking hid placeholders; args_secret_warning fired on the recommended fix). BCC never expands on write. 444 tests, ruff clean.
Closes#76
The blocker on this issue was whether BCC or the client does the expanding.
Answer, from Anthropic's docs: Claude Code expands ${VAR} and
${VAR:-default} itself, in command, args, env, url and headers, for both
project .mcp.json and user-scope ~/.claude.json. Claude Desktop has no
documented support.
So this is a per-client capability, not a global one, and BCC does NOT
expand on write: resolving a reference into the file would put the secret
back on disk -- the whole thing the user is avoiding -- and would defeat a
feature the client already implements correctly. BCC authors, validates and
warns; expand_env_refs exists to preview what the client will do.
Semantics mirror the documented ones exactly, including the unusual bit:
an unset variable with no default is left as literal ${VAR} text rather
than blanked, because that is what Claude Code passes through.
Gating uses the existing profile_targets_claude_desktop(), so a config that
is correct under Claude Code and broken under Desktop is reported against
whichever profile is actually loaded. The two warnings are worded
differently on purpose -- 'this client will never expand these' is a
different problem from 'this variable looks unset here'.
Two existing behaviours were backwards for this feature and are fixed:
- Secret masking hid placeholders. is_secret_key('API_KEY') is true, so
${API_KEY} rendered as dots -- making a reference indistinguishable from
a stored credential, which is the one distinction that makes the feature
worth adopting. should_mask_value() now skips references, in the table
delegate, _redact_server_data and redact_args alike.
- args_secret_warning fired on placeholders. Moving a token into ${VAR} is
the recommended fix for that warning; continuing to warn punished the
fix. It now skips references while still flagging a real secret that
follows one.
Real secrets are still masked everywhere they were before -- asserted, not
assumed.
Refs #76
Three conflicts, two of them semantic rather than textual:
- bcc.py QSS: this branch added the noticeBanner rules using the old
module-level constants ({MUTED}, {ACCENT}); main had since moved the
stylesheet onto palette slots ({p.muted}). Took main's form and
translated the notice rules into it -- picking either side wholesale
would have either dropped the banner styling or reintroduced globals
that test_stylesheet_builder_has_no_hardcoded_colours now forbids.
- bcc.py methods: both sides appended to MainWindow (update-notice
handlers vs theme handlers). Additive, kept both.
- tests/test_core.py: the usual EOF append. Kept both blocks.
_build_menu_bar auto-merged cleanly (View menu above, Help menu below);
verified both are present with their menu roles intact.
Verified: 272 test functions = 265 (main) + 7 (this branch), no
duplicates; 421 passed, ruff clean.
Union conflict at the end of tests/test_core.py -- both branches appended a
test block. Kept both, main's #72/#73 block first. Verified: 265 test
functions = 238 baseline + 15 (#77) + 12 (theming), no duplicates.
Reported from the field: running v1.2 against a repo with v1.3.0 published
gave no prompt, and there appeared to be no way to check manually. The
checker itself works; it was invisible, for two reasons.
#78 -- the notice was written to the shared status label, which 21 other
call sites rewrite. The check runs off-thread and lands a second or two
after launch, right as the user starts clicking, so the next selection or
refresh wiped it. Exactly the bug fixed for the MSIX warning in #35, which
got a persistent banner; that fix was never carried to the update notice.
Adds NoticeBanner: a persistent, dismissible notice carrying its own action
button. It's a shared widget rather than a second bespoke banner, so the
next thing needing the user's attention doesn't reach for the status bar
again. (The MSIX banner still uses its own QLabel -- migrating it is a
follow-up, deliberately not bundled with a bug fix.)
#79 -- the only 'Check for updates' affordance was a button inside the
About dialog, which is not where anyone looks. Worse, the About action was
created without a menu role, and Qt auto-assigns AboutRole to actions whose
text begins with 'About', relocating it into the macOS application menu --
so the notice's own hint, 'Help > About to view it', pointed at a menu that
on macOS doesn't contain the item.
Help now has its own 'Check for updates...' item with an explicit
ApplicationSpecificRole, and the About action states its AboutRole rather
than inheriting it invisibly. The menu-driven check is never throttled and
always reports back -- the user asked, so silence would read as broken.
The decision and the wording live in core.update_notice() because the test
suite has no PySide6 (CI installs pytest + cryptography only), so anything
in bcc.py is untestable. A test asserts the notice text names no menu path,
which is what went stale here in the first place.
Closes#78Closes#79
BCC has always been dark-only -- BG #1b1d23, hardcoded at import time, with
no light option and no awareness of the desktop's appearance. On a light
desktop it matches nothing else on screen and there was no way to change it.
Adds a Palette value type in bcc_core with DARK (byte-identical to the
colours v1.3.0 shipped) and a new LIGHT, plus resolve_theme(setting,
system_is_dark) so the decision is testable without a Qt app. View > Theme
offers Match system / Light / Dark, persisted in QSettings under ui/theme,
defaulting to following the system.
The light palette's semantic colours are deliberately not the dark ones
lightened: #4ade80 sits near 1.7:1 against white. They are darkened to clear
WCAG AA, and a contrast test enforces >= 4.5:1 for every text colour against
its surface in both palettes so nobody harmonises them back later.
Three near-black literals were baked into the stylesheet (#1a1205 on-accent
text, #202229 disabled table, #16181d diagnostics pane). Fine with one theme,
invisible breakage with two -- each now has a palette slot, and a test
asserts build_stylesheet contains no hex literals at all.
The ~20 inline setStyleSheet(f"color: {MUTED}") call sites are left alone:
apply_palette rebinds the module-level colour names, and an f-string resolves
its names when it runs, so each call site picks up the new colour on its next
render. Switching theme reapplies the global QSS and re-renders the
inline-styled widgets, so nothing is left dark-on-light.
Refs #75
Two silent-failure bugs found auditing the v1.3.0 features.
#72 -- extract_servers called dict() on every server value, so a config
that was valid JSON but held a non-object server ("foo": "oops", a
number, a list, null) raised on load. The strict parse succeeded, so the
repair path never saw it, and the call sat outside the load try/except:
an unhandled traceback with the window half-swapped to the new profile.
It also made the #54 schema lint unreachable for the most likely
hand-edit mistake -- the load died before the linter ran.
Malformed values are now preserved verbatim on ServerEntry.raw (behind a
NO_RAW sentinel, since a literal JSON null is itself a malformed entry
worth keeping) and written back untouched on Save, so nothing is silently
deleted. lint_servers names the offending entry instead.
#73 -- the stale-file "Merge & save" path reloaded the file from disk and
re-applied the user's servers, but apply_servers only writes mcpServers
and _disabledMcpServers. Named server sets live under _bccServerSets in
the same file, so a set saved that session was dropped from disk and then
from memory, with no warning, on the path the user picks because it
sounds like the safe one.
BCC-owned keys are now declared in BCC_OWNED_KEYS and carried across by
carry_owned_keys, which reports genuinely contested keys so the status
line can say so. Deliberately one-directional: a key absent locally is
left alone on disk, because 'user deleted their last set' and 'another
machine just added sets' are indistinguishable and deleting someone
else's data is the worse failure.
Closes#72Closes#73
Finding 1: can_sign() returned True on an empty changeset, and _on_sign
compared the reviewed blob against a hardcoded "main" rather than the ref
actually reviewed — so the PR path could never sign, and the main-vs-main
path unlocked Sign with zero entries acknowledged. That is how commit b08cf21
signed 19 entries nobody reviewed. can_sign now refuses an empty diff, checks
has_blocking_risk itself instead of trusting a GUI checkbox, and resolves the
TOCTOU pin from the reviewed ref via a pure, testable sign_precondition().
Finding 5: the catalog key and the release key are now separate. The release
key lives in CI and signs checksums; the catalog key stays offline and signs
what users execute. A CI compromise gets the former, not the latter.
Findings 2/3/4/6/7. config.env now gets the deny-list, ASCII, secret and
empty-or-placeholder checks that args always had; version pinning is enforced
at runtime, not only in the maintainer tool; the CI gate pins the expected
pubkey instead of trusting the one in the PR it is reviewing; resolve_catalog
anchors its cap to the bundled version and prefers bundled on ties; catalog
ids are constrained.
Tests rewritten: the old fixtures asserted the unpinned form validates clean
and that config.env passes through verbatim — they enshrined two of the bugs.
Maintainer-only PySide6 tool: semantic per-entry diff of data/catalog.json,
blocking risk predicates (non-empty env values, command allowlist, non-ASCII
homoglyphs, unpinned packages, URL domain changes), automatic off-thread
registry lookups (publisher/age/downloads/near-neighbour), acknowledge-gated
signing with a pinned git blob SHA (refuses to sign bytes that changed since
review), and payload+signature emitted in a single commit.
Never shipped to users — excluded from bcc.spec, with a test asserting it.
Publish SHA-256 checksums for every release artifact and sign them with a
domain-separated Ed25519 signature. Skips signing (loudly) if the key secret
is absent rather than failing the release. README documents verification and
states the limit plainly: this proves the file is the one we published; it
does not remove Gatekeeper/SmartScreen warnings.
All network calls are monkeypatched (urllib.request.urlopen) — no live
network in tests. Covers older/newer/equal comparisons, v-prefix,
differing-length tuples, malformed input on both sides, and
fetch_latest_release success/timeout/network-failure/malformed-response
paths.
- Add a real QMenuBar (Help -> About...).
- AboutDialog: app icon, name, version (core.__version__), and links to the
repo/issues/license opened via QDesktopServices.openUrl (system browser,
never in-app navigation). macOS unsigned-app note.
- "Check for updates" button + UpdateCheckWorker (QThread) runs
core.fetch_latest_release() off the UI thread; shows "up to date" or
"vX.Y.Z available" with a button to open the releases page. No binary
download, ever.
- Optional quiet startup auto-check, throttled to once/day via a QSettings
timestamp, off-thread, silent on failure, toggle lives in the About dialog.
Mocks subprocess.run/Popen and patches sys.platform per-OS branch (darwin,
win32, linux) -- never actually kills or launches anything. Covers: correct
command sequence per platform, pkill/taskkill exiting non-zero (nothing to
kill) is NOT treated as failure, a failed relaunch IS reported as failure,
and profile_targets_claude_desktop() correctly distinguishes Claude Desktop
configs from Claude Code / legacy settings.json profiles.
Includes the required regression case: rewrite a file with different
content, force the original mtime back via os.utime, and assert the
fingerprint still differs (because size changed).
MainWindow._loaded_mtime -> _loaded_stat (core.ConfigStat), populated
via core.config_fingerprint() at load/save time. The save-path stale
check now compares the full fingerprint (mtime AND size) instead of
bare mtime equality.
After a successful save, shows a 'Restart Claude Desktop' button next to
the status line. Only shown when the saved profile targets Claude Desktop
(core.profile_targets_claude_desktop) -- never for Claude Code, which has
no GUI process to bounce. Clicking it calls core.restart_claude_desktop()
and reports success/failure. The button hides again on further edits or
when switching/reloading profiles.
Two prior attempts to fix this line via a literal or -escaped
non-breaking space both silently reverted back to a no-op regular-space
replace during transcription. Using chr(0xA0) instead removes any
non-ASCII or backslash-escape character from the source line entirely.
Platform-abstracted restart: macOS uses pkill -x Claude + open -a Claude,
Windows uses taskkill + relaunch via the Start-menu shortcut, Linux uses
pkill claude + a detached Popen relaunch. Not finding a running process is
not an error -- only a failed relaunch is reported as failure.
Also adds profile_targets_claude_desktop() to distinguish Claude Desktop
profiles (claude_desktop_config.json) from Claude Code profiles, used to
gate the restart button in the GUI.
- Add `__version__ = "1.1.0"` as the single source of truth (matches pyproject.toml).
- Add parse_version()/is_newer_version() for numeric (never lexical) version
comparison, handling v-prefix, pre-release suffixes, and malformed input.
- Add fetch_latest_release(): reads the public Gitea releases API
(anonymous, no token) and returns {version, url} or None on any failure.
Never downloads or touches a binary — metadata only.
The previous commit accidentally normalized the non-breaking space
(U+00A0) in _normalize_unicode's replace() call to a regular space
during a copy/paste, turning that replace() into a no-op. Use an
explicit escape instead of the literal character so it can't
be silently corrupted again.
Bare mtime equality can miss a concurrent external write that lands
within the filesystem's mtime resolution (same-second writes), or
where the writer restores the original mtime. Add config_fingerprint()
returning a (mtime, size) ConfigStat pair; the stale-file check in
bcc.py now compares both fields instead of mtime alone. config_mtime()
is kept as-is (still used/tested independently).
The crashed/exited/stderr spawn tests used 0.3-0.5s timeouts. On a slow CI
runner, interpreter startup can exceed that, so the process is still starting
when the timeout fires, gets killed, and is misclassified "ok" (still running)
instead of "crashed"/"exited". Bump those three to 2.0s. The two "ok" paths
(sleep-60, stdin-block) stay at 0.3s since timing out there means success.
the_ogandClaude Opus 4.8