risa-labs-inc/BossConsole

Follow-ups from Claude's review of #1382 (merged as 94ea7ff). Medium - no test checks that the current attempt completes. Hard-wiring isCurrent() to false in PasskeyAuthViewModel.launchAuthentication breaks login entirely, yet all five tests in PasskeyAuthViewModelSessionLifecycleTest still pass. In test 1, complete the second gate with Result.success(Unit), then assert secondSuccess == 1 and isLoading == false while firstSuccess == 0. Low - positional arguments are unpinned. passkeyAuthentication(email, credentialId) takes two Strings, so a transposition would compile. Record the arguments in the fake and assert that credentialId passes through (and is null on the email path). Low - registerPasskey (PasskeyAuthService.kt ~:124/:137) still converts a cancel into Result.failure. Give it the same treatment, or scope the propagateCancellation KDoc to the authentication path. Low - an abandoned attempt can still establish a session if the cancel lands after polling has returned success (C

From the merge review on #741 (none blocked the merge; finding 1 is the one to take first): saveOwner is a non-reactive read captured at composition (WorkspaceButton.kt:99). SplitViewStateRegistry.getState reads a MutableStateFlow outside Compose snapshot state, and registration happens in a LaunchedEffect after first composition, so the dialog's rebind fence can silently drop a save for a live window and log "deregistered" for one that never registered. Resolve the owner at press time inside onSave, fence against that local, key the latch on windowId alone, and make the miss log distinguish never-registered from deregistered. The call-site wiring is the untested half, and every defect in this series lived there. Extracting the menu effect's save wiring into an internal function would let ordinary unit tests pin: snapshot built from the invoking window's own id, a second event coalescing into exactly one re-run on the latest layout, re-run dropped after deregistration, dialog ignoring

Leftovers from the four review rounds on #822 (all non-blocking, none a disclosure or authorization defect; full derivations in the review comments on #822): McpJsonDepth.kt's KDoc enumerates three guarded parses on the invoke path; there are now four (the pre-pass's parseElement joined them in #822). One clause keeps the drift-prevention file truthful. The {{secret:…}} marker is case-sensitive: {{SECRET:…}} reaches a handler as literal text. Decide whether that is the contract (and document it) or normalize. The ledger sanitizer's value class [^\s&,;}]+ stops at the first }, so a substituted value containing } leaves a trailing plaintext tail in redacted output (TOKEN=[REDACTED]}}tail). Agent-authored text only. The residual lookbehind hole: {{secret:<plaintext>}} with no keyword prefix is not redacted in the malformed-refusal snippet. Refused before any prompt; hygiene. parseMcpToolArgs's failure log interpolates t.toString(), and kotlinx appends the offending document. Unreacha

Follow-up to #1455 (central log redaction). The append path wraps throwables so rendered channels (SLF4J line, file writer) sanitize messages and stack traces, but the wrapped throwable's stack frames are copied verbatim, and listeners receive that throwable. A listener that calls stackTraceToString() - a Console panel, a crash reporter - renders raw frame filenames, and a hand-built or native frame can embed a path like /Users/ci/keys.pem. The file path is sanitized at render; the listener path is not. Options: sanitize frame filenames when wrapping (frames are reconstructable via StackTraceElement(String, String, String, int)), or document that listeners must render through LogSanitizer.

#1485 fixed runInTerminal argument quoting; the fix is verified correct (quotePath escapes embedded single quotes on POSIX and PowerShell), but the new pinning test has gaps flagged in review: No embedded-single-quote test case - the one input class that decides whether the escaping is real. Add table-driven coverage: ', ", backslash, empty string, $(), backticks, newline, ;, |, &&. The test installs GitTerminalEventBus.ipcBridge and nulls it in finally instead of saving/restoring the previous bridge. The stub bridge drops other producers' events while installed - cross-test interference if the suite runs classes in parallel. The retry loop races the concurrency it tolerates and publishes duplicate bus events; a deterministic injection seam would be better. quotePath is only exercised through two benign inputs despite being load-bearing for a security property. Good first issue for whoever wants to pick it up.