Codex P2 on #2671: if only one ProgrammaticAccessOnly value is written
during disable, delete the partial key so disable never leaves a single
hidden verb while the other remains visible.
Codex P2s on #2671: remove treats any leftover runnable verb as still
enabled, and install restores ProgrammaticAccessOnly when enabling fails
after clearing a prior per-user suppression.
Codex P2 on #2671: when no preference exists but a single leftover shell
verb remains, still run install/repair so the Settings toggle cannot stay
off while Explorer still shows one Open-in-Netcatty entry.
Codex P2 on #2671: unelevated refreshes must not report success merely
because non-empty HKLM commands exist. Require writes to succeed or both
verbs to already match the expected executable/command contract.
Codex P2 on #2671: treat the integration as user-suppressed only when
both folder and background keys have ProgrammaticAccessOnly, so a partial
suppress write cannot hide one menu while claiming the feature is off.
Address Codex P2s on #2671: disabled preferences re-suppress on every
startup so NSIS updates cannot resurrect the menu, and registration is
only considered complete when both folder and background commands exist.
Codex P2 on #2671: a partially written registry verb (key exists, no
command value) must not count as an active menu registration, so install
and Settings toggle state stay honest and repair can still run.
Persist the last-applied executablePath in the preference record and
re-apply shell verbs when it differs from the current launcher path so
renamed/moved portable builds keep Open-in-Netcatty working.
Codex P2 on #2671: electron-builder portable apps expose the stable
launcher via PORTABLE_EXECUTABLE_FILE while process.execPath points at a
temp unpack that is deleted on exit. Resolve that path for shell verbs
and bump the preference schema so existing installs re-register once.
Address Codex P2 on #2671: the Settings toggle is per-user, so disable
only clears HKCU and suppresses machine verbs via ProgrammaticAccessOnly
instead of deleting HKLM (even when elevated). Also persist a schemaVersion
so healthy warm starts skip registry work after the first migration.
Codex P2 on #2671: warm starts no longer spawn a full reg.exe write pass
when installer/user shell verbs already match the expected command, label,
and icon. Still rewrites stale --open-terminal-path entries on upgrade.
Address Codex review on #2671: when all-users (HKLM) shell verbs cannot be
deleted unelevated, hide them for the current user via ProgrammaticAccessOnly
instead of failing the Settings toggle. Also stop mirroring machine verbs into
HKCU so uninstall does not leave a broken Open-in-Netcatty entry.
Add a Settings > System switch to enable/disable the Explorer folder
context menu, and fix shell launch so Chromium does not swallow
--open-terminal-path (args after --, =path form, trailing-dot roots).
Closes#2651Closes#2655
Bound the main-process TCC probe to 3s, only treat real private IPv4/IPv6
literals as LAN, prefer proxy as first hop when set, and annotate
unreachability using either the target or the first hop. Closes the
remaining merge blockers for #2663.
Show manual compact for any Catty session with a configured model, not
only when enough history exists for a large head. External agents still
do not get the command.
Gate /compact behind canCompact so the composer is not cleared as a
silent no-op, hide the usage ring until a real context snapshot arrives,
and summarize only the UI head that the persisted compaction boundary
will drop. Drop unused clearSessionMessages surface and add slash-command
regression tests.
Emit plan items as { text, completed } with top-level running|completed status so sdkAgentAdapter and the activity UI can render labels and completion instead of empty rows stuck on running.
Remove the async rpc.request hop that let promptDone win Promise.race on session/prompt JSON-RPC errors. Gate emitDone on turnCompleted, treat any incomplete process close (including exit 0) as failure, and align streaming-json toolResult emission with ACP when rawOutput is present.
Permission replies must select an optionId offered in params.options instead of inventing allow-once. Process close(null, signal) before turnCompleted is a failure for both ACP and streaming-json drivers.
Attach a stdin error listener so async EPIPE after peer exit is not an
unhandled stream error. Treat EPIPE/ERR_STREAM_DESTROYED as benign
(processErrorGuards/terminalBridge); fail the turn only for other mid-turn
write errors.
Only turnCompleted (headless end / ACP prompt success) may ignore teardown
exit codes. streamedAssistantText must not suppress errors, or mid-response
crashes look successful.
Root cause of Token ~1: onPromptComplete resolved promptDone in the same
tick as session/prompt, so Promise.race often discarded the result and the
UI fell back to estimated usage (prompt length / 4). Emit usage when the
prompt response is handled, and map live Grok _meta.usage fields including
cachedReadTokens.
Drop the inactive acp/streaming-json in-memory session when switching
runtimes so switch-back cannot resume a pre-switch thread without
intervening turns. Forward optional session/prompt usage to emitter.usage.
Tool-only turns leave no assistant text; Windows teardown often exits with
code 1. Mark turnCompleted when session/prompt succeeds so close does not
emitError or skip emitDone.
Windows .cmd/.bat installer shims fail with spawn EINVAL when launched
without shell. Route ACP, headless, and models spawns through the same
helper other managed agents use.
When a stored session id cannot resume/load, still inject renderer history
into the new session prompt so multi-turn context is not dropped. Successful
resume/load paths remain unseeded to avoid stacked prior replies.
Prefer ACP session/resume (then load/new) without replaying history into the
current turn, authenticate when Grok advertises methods, isolate acp vs
streaming-json session identities, and apply MCP local-tool lockdown plus
--no-auto-update on both transports.
Suppress session/update emissions until session/prompt starts so
session/load history replay does not fill the current assistant bubble.
Also treat Grok like CodeBuddy: when a resume session id exists, do not
also inject renderer conversation history into the prompt.
Expose a Settings toggle on the Grok Build card (like Codex App Server)
to choose ACP agent stdio or the original headless streaming-json path.
Persist grokRuntime on ExternalAgentConfig and pass it into the driver
via NETCATTY_GROK_RUNTIME on the agent env.
Grok agent stdio session/new rejects plain object env (Invalid params /
McpServer enum). Emit type:stdio servers with env as [{name,value},...]
matching the live protocol, and assert the shipped shape in unit tests.
In MCP tool-integration mode, pass --disallowed-tools for Grok built-in
shell/edit/write so remote session work goes through injected Netcatty MCP,
matching Claude-style policy. Skills mode leaves local shell available for
the Netcatty CLI skill path.
Curate GROK_MODEL_PRESETS and wire getAgentModelPresets so the chat model
picker still has a non-empty list when live grok models discovery fails,
matching the Claude/Codex fallback pattern.