Repository navigation
Enable Claude Code integration by default again - #247
Merged
Merged
Conversation
Contributor
Preview Videos and ScreenshotsOpen Workspace (1 hr expiry) · Open Dev Browser (1 hr expiry) · Open Diff Heatmap Failed to capture screenshots for commit
Generated by manaflow preview system |
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
BillionClaw
added a commit
to BillionClaw/cmux
that referenced
this pull request
Mar 24, 2026
The pre-0.60 "Claude Code hooks" section is no longer relevant after PR manaflow-ai#247. Fixes manaflow-ai#2009
BillionClaw
added a commit
to BillionClaw/cmux
that referenced
this pull request
Mar 24, 2026
The Claude Code hooks section documented manual JSON configuration for notification hooks. Since PR manaflow-ai#247 (released in 0.60.0), Claude Code integration is enabled by default — cmux automatically injects notification hooks when the claude command is invoked. This manual configuration is no longer needed and is misleading for users on 0.60+.
BillionClaw
added a commit
to BillionClaw/cmux
that referenced
this pull request
Mar 24, 2026
Claude Code integration is now enabled by default (PR manaflow-ai#247, 0.60.0), so the manual hooks configuration in settings.json is no longer needed. Kept OpenAI Codex and OpenCode integration examples which are still relevant. Fixes manaflow-ai#2009.
bn-l
pushed a commit
to bn-l/cmux
that referenced
this pull request
Apr 3, 2026
hhsw2015
pushed a commit
to hhsw2015/cmux
that referenced
this pull request
May 22, 2026
Drag-storm log analysis (bd23f95 build, captured via os_log) shows the freeze symptom is per-RPC SSH subprocess cost, not main-thread saturation: - 27 outbound pane.resize during a ~10s cmux divider drag - each oneshot.ok durMs=600-1000ms over the SSH master - ~19s of cumulative SSH-master time spent on pane.resize alone - applyDividers.end durMs=0 across 294 calls (main is fast) - pump.stats kept reporting (PTY consumer not starved) Each pane.resize today opens a fresh ssh subprocess + remote api-bridge + bincode handshake — that's the ~700ms baseline. Multiplied by 27 fires per drag, the SSH master gets jammed and the raw-pty-attach channel multiplexed on the same master can't land output bytes — user sees "panel input invisible" and right-click menu hangs while the master is in flight. The 80ms debounce was too short for bursty mouse motion: a single panel observed 7-10 fires per drag because brief mouse pauses (>= 80ms) re-fired the timer instead of coalescing. With 4 panels in the workspace and ~10 fires each, you get the 35-40 outbound pane.resize calls the log shows. Bumping to 400ms keeps drag bursts coalesced into one fire per panel at drag end. Real fix is the persistent api-bridge (task manaflow-ai#247): collapses the per-RPC ~700ms baseline to <50ms and makes the debounce length immaterial.
hhsw2015
pushed a commit
to hhsw2015/cmux
that referenced
this pull request
May 22, 2026
Routes every non-streaming RPC through a per-host long-lived HerdrApiClient connection cached by HerdrPersistentClientRegistry. With herdr daemon >= 0.6.0-cmux9 the API socket loops on the same connection (see daemon commit f91ffdf...cmux9 ssh side "feat(api): persistent socket connections"); cmux now reuses one ssh subprocess + one remote api-bridge for all RPCs instead of paying ~700ms ssh+bincode handshake per call. Why this matters in practice: drag-storm os_log capture from build c25cf0d (debounce 80→400ms) showed 31 outbound pane.resize in a 10-second cmux divider drag, each oneshot.ok durMs in the 644-2130ms range, totalling ~21s of cumulative SSH-master time. The raw-pty-attach output channel multiplexed on the same SSH master got starved for the duration — user observed "panel input invisible / right-click menu hangs" while bytes still flowed TUI-side. The 400ms debounce mitigated nothing because the per-RPC cost dominates; reducing the per-RPC cost is what unblocks the drag freeze. Architecture: - HerdrPersistentApiClient (actor) per host. Holds optional HerdrApiClient + reconnect lock. ensureConnected() coalesces concurrent connect attempts so a burst of RPCs after a disconnect doesn't spawn 30 parallel ssh subprocesses. - request(method, params) does up to one transparent reconnect on a transport-level error. Application errors (HerdrApiError with daemon codes like pane_not_found) propagate as-is — we don't retry refusals. - HerdrPersistentClientRegistry caches by HerdrHost.id and exposes forget/forgetAll. EventPump.tearDownHost calls forget so a removed host doesn't leak its ssh subprocess. HerdrOneShotRPC.send/request now thin facades that route through the registry. All existing call sites (Workspace, Sidebar, DividerSync, FocusSync, ZoomSync, CloseHandler, KillCommands, WorkspaceSync, EventPump's pane.set_split_ratio path) pick up the new transport without any change. Backward compat: legacy daemons (pre-cmux9) close the socket after one response. The persistent client sees that as a transport disconnect and lazy-reconnects on the next call — visibly identical to the old per-call connection model. No client side change is needed for the user to keep working against an old remote daemon; they just don't see the latency win until the daemon gets upgraded. Tests: - HerdrPersistentApiClientTests.testThreeSequentialRequestsReuseOneConnection spins up an in-test UDS daemon mock, drives three echo RPCs through HerdrPersistentApiClient, and asserts the daemon saw EXACTLY ONE inbound connection. Without manaflow-ai#247 this would be 3. - HerdrPersistentApiClientTests.testRegistryReusesClientForSameHost locks the registry's identity-by-host-id contract and the forget()-then-fresh contract. Daemon side already shipped as v0.6.0-cmux9 with multiple_requests_share_one_socket_connection integration test in herdr/tests/api_ping.rs.
hhsw2015
pushed a commit
to hhsw2015/cmux
that referenced
this pull request
May 22, 2026
CRIT-1 — test compile break.
testMalformedPayloadSurfacesAsApiError used `[String: Any?]` which
the JSONSerialization API never returns and which can't be
subscripted via the `??` flatten — the cmuxTests target failed to
build. Switch the helper return type to `[String: Any]` and
extract the error body via a clean `as? [String: Any]` cast.
MED-4 — close()-during-connect leaked a fresh HerdrApiClient.
The previous design: A is awaiting `api.start()` at the top of
the connect path. close() runs (stopped = true; awaiters drained;
client is still nil so close exits without doing anything). A's
start() then succeeds and writes `client = api` against an
already-stopped instance, leaking the api forever (no caller will
release it). Now: after start() succeeds, check `stopped` and if
flipped, close the just-spawned api and throw `disconnected`
instead of publishing it. close() therefore guarantees no surviving
HerdrApiClient on a stopped instance.
HIGH-2 — direct end-to-end test for daemon error surfacing.
Replaced the raw-socket "exercise the mock fixture" approach with
testDaemonErrorEnvelopeSurfacesAsHerdrApiError, which drives a
request through HerdrPersistentApiClient against a mock daemon
configured to emit a HerdrApiError envelope (`force_error_*`
method prefix). Asserts:
- persistent.request throws HerdrApiError(code: "invalid_request")
- the connection stays alive afterwards (a follow-up ping on
the same socket succeeds)
- daemon.connectionCount stays at 1 (error envelope did not
cause a reconnect)
This locks the contract that manaflow-ai#247 requires: per-request errors
are surfaced as application errors, not as transport drops that
would void the persistent connection.
Mock daemon's `makeResponse` now emits the error envelope for any
method starting with `force_error_` so future tests can drive
arbitrary daemon error flows without changing the mock plumbing.
lawrencecchen
added a commit
that referenced
this pull request
Oct 2, 2026
Bump the ghostty submodule from b1a49b601 to 257a40956, the merge of manaflow-ai/ghostty#252. That PR adds the offscreen embedded platform (GHOSTTY_PLATFORM_OFFSCREEN = 6, CPU frame and Linux dmabuf callbacks, app-thread draw) behind -Dembedded-offscreen, so the cross-platform GPUI and cmux-browser clients can render daemon terminals offscreen on Linux and Windows from the same fork. The default macOS GhosttyKit build does not change, and cmux-next does not use the new platform. feat-cmux-next now tracks fork main (coordinator decision), so the bump also brings the fork-main merges between the old pin and the #252 base dcb1d6965: #242 (PTY tee after the parser, under the renderer mutex), #243, #245, #247, #250 (prompt line and replay fixes), #246, #248, #249 (VT replay blank cells keep the default style) and #251 (config API window padding, whose fix commit was the old pin). scripts/ghosttykit-checksums.txt pins the published 257a409 GhosttyKit archive (build-ghosttykit.yml run 36989929300), which every cmux-next CI job downloads. docs/ghostty-fork.md documents the offscreen platform, #242 and a map of every merge in the range, with conflict notes. The cmux main pin's local-only binding flag (324c02738) is not on fork main; cmux-next does not use it.
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
ClaudeCodeIntegrationSettings.defaultHooksEnabledback totrueValidation
xcodebuild -project GhosttyTabs.xcodeproj -scheme cmux -configuration Debug -destination 'platform=macOS' build✅xcodebuild -project GhosttyTabs.xcodeproj -scheme cmux-unit -configuration Debug -destination 'platform=macOS' -only-testing:cmuxTests/GhosttyConfigTests test❌ (blocked by existing compile errors incmuxTests/CmuxWebViewKeyEquivalentTests.swiftreferencingUpdateChannelSettings)./scripts/reload.sh --tag claude-default-on✅