agent-session: reliable tracking system + Codex picker GUI + debug trace

[azooz2003-bit]

Reliable agent-session tracking for the iOS coding-agent GUI, Codex interactive-picker support, and an end-to-end debug trace. Built and dogfooded iteratively this session across Claude + Codex.

Reliability system (two-pillar: observe floor + hooks)

• Tree-aware end backstop: liveness is judged from the surface's process tree, not a single recorded pid, so a launcher/intermediate dying (subrouter, node shim) re-binds to the real agent instead of falsely ending a live session.
• Observe-floor detection: a throttled off-main process-tree scan finds untracked codex/claude under a cmux surface and binds it, resolving identity without hooks (codex via the rollout fd it holds open; claude via --session-id/--resume argv).
• Reducer/match hardening: list reducer ignores the unversioned stateChanged (versioned descriptorChanged is authoritative); the live send/list gate is deterministic pid-liveness, not the terminal-title heuristic; resume re-bind keyed on the real terminalPanel.id, fired whenever a restored surface carries a resumable binding.
• Transcript resolution honors CLAUDE_CONFIG_DIR / CODEX_HOME; accurate "can't find transcript" copy for the $HOME-cwd case.
• Codex wrapper Settings toggle (mirrors Claude); script-file hooks (#!/bin/sh in ~/.cmux/hooks) for both wrapper-injected and persistent ~/.codex/hooks.json commands so non-shell runtimes (subrouters) stop erroring with os error 2.

Codex interactive pickers in the GUI

• request_user_input parses into the same tappable ChatQuestion the GUI renders for Claude's AskUserQuestion.
• mobile.chat.answer is agent-aware (Codex needs Enter after the digit).
• Answered pickers resolve to the chosen option (non-interactive) via the function_call_output; multi-question codex calls resolve each card by question id.

Debuggability

• DEBUG-only agentChat.* trace across the whole pipeline (hook ingest, process-tree detection, state transitions, transcript resolve, transcript batch). One grep 'agentChat\.' /tmp/cmux-debug-<tag>.log shows the full flow.

Verification

• Shared package (CmuxAgentChat) tests: 129/129 pass.
• macOS app builds clean (tagged codexwrap).
• Dogfooded live: codex pickers render tappable + answer + collapse to selection; observe-floor detection resolves real rollout fds; tree-aware liveness; script-file hooks direct-exec returns {} 0.

Known follow-ups (not regressions)

• Persistent-hooks cleanup: a re-install over old inline SessionStart/UserPromptSubmit leaves a stale inline duplicate (fresh installs are clean); the fire-and-forget format dodges the ownership matcher.
• A $HOME-cwd Claude session that Claude itself doesn't persist a transcript for; cmux reports it honestly. An FSEvents transcript watcher (capture every file the agent writes) is the planned robust fix.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

---

Summary by cubic

Rebuilt agent-session tracking for iOS with a versioned single source of truth, process‑tree detection, and full Codex support (wrapper, resume, GUI pickers) so sessions stay reliably live and editable. Adds a DEBUG trace and a host-side scripts/cmux-chat-debug.py inspector, restores the data‑plane boundary for mobile.chat.sessions, and fixes a Swift‑6 warning.

• New Features

  • Reliable SoT: session version, version‑gated descriptorChanged upserts, and mobile.chat.session snapshot pull; iOS re‑pulls on foreground.
  • Liveness: per‑pid exit watcher + tree‑aware end backstop rebinds the real agent; observe‑floor scan binds untracked sessions (Codex via open transcript fd, Claude via --session-id/--resume).
  • Codex: bundled cmux-codex-wrapper with fire‑and‑forget hooks; sibling PATH shim exports CMUX_CODEX_WRAPPER_SHIM; wrapper fires session-start on resume; persistent hooks emit #!/bin/sh scripts; Settings toggle; honors CODEX_HOME.
  • GUI: Codex request_user_input renders as tappable ChatQuestion; mobile.chat.answer appends Enter for Codex; multi‑question resolution by questionID.
  • API/UI: mobile.chat.sessions scopes by the surface’s current workspace, retains ended sessions, and re‑pins to a reopened live session; data‑plane only.
  • Debug: agentChat.* DEBUG logs across hook ingest, detection, state, and transcript; DEBUG‑only tracing inside mobile.chat.sessions; host‑side scripts/cmux-chat-debug.py (now fails loudly on unset CMUX_TAG/CLI errors and uses ANSI clear); the resume‑initiated log is gated under #if DEBUG.
  • Build: fixed a Swift‑6 captured‑var self warning in the process‑exit handler.

• Migration

  • No action required; the Codex wrapper is bundled and per‑surface PATH shims install automatically.
  • Optional: to cover Codex launched outside cmux, run cmux hooks setup --agent codex to install persistent hooks.
  • You can disable the Codex integration in Settings → Automation (exports CMUX_CODEX_HOOKS_DISABLED=1).
  • CI: refreshed Swift file‑length budget; no behavior change.

^{Written for commit fe9df5e. Summary will update on new commits.}

Summary by CodeRabbit

• New Features
  • Added a Codex Integration toggle, including per-surface Codex shims, hook injection, and wrapper-routed Codex resume flows.
  • Improved agent chat questions with questionID to support multi-question Codex and id-based answer completion.
  • Added per-session versioning and mobile “authoritative snapshot” fetching using live process discovery.
• Bug Fixes
  • Improved session reconciliation to prevent stale/out-of-order updates and improved chat recovery when returning from background and when reopening ended sessions.
  • Improved Codex hook injection to prefer stable script-based commands with safe fallbacks.
• Documentation
  • Added/updated Codex integration and agent session tracking documentation.
• Tests
  • Updated Codex wrapper/command rendering coverage and session version gating tests; removed obsolete transcript resolver tests.

[greptile-apps]

New Codex settings strings missing translations for 17+ locales

The four new settings.automation.codex* strings (settings.automation.codex, settings.automation.codex.subtitleOn, settings.automation.codex.subtitleOff, settings.automation.codex.note) only ship with en and ja entries. The sibling settings.automation.claudeCode strings they mirror carry translations for en, ja, zh-Hans, zh-Hant, ko, de, es, fr, it, da, pl, ru, bs, ar, nb, pt-BR, th, tr, and uk. Every user running cmux in one of those 17 locales will see raw English in the new Codex Integration settings card.

Rule Used: Flag production user-facing text that is not fully... (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[greptile-apps]

Static mutable state missing actor isolation annotation

pendingResumeIntents and liveInstance are static stored properties on a non-@MainActor class, yet every caller (recordResumeIntent from Workspace.swift, start() from the app delegate) runs on the main actor. Without an explicit @MainActor annotation on these statics, Swift 6 strict concurrency cannot verify the isolation is safe — a future caller from a background context would be a silent data race. Marking the statics and the static method @MainActor fixes this and makes the invariant compiler-enforced rather than convention-enforced.

Rule Used: Flag new or materially worsened Swift 6 actor isol... (source)

[greptile-apps]

Implicit singleton side-channel for initialization ordering

The liveInstance / pendingResumeIntents pattern is a static side-channel that routes session restore events through a global buffer instead of a direct call graph. The ordering invariant ("restore fires before start(); start() drains the buffer") is correct today but nowhere enforced structurally. Consider making the restore path pass its intents into start(intents:) as parameters, eliminating both static vars entirely.

Rule Used: Flag Swift fixes that patch symptoms while leaving... (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[coderabbitai]

Actionable comments posted: 12

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (2)

Sources/TerminalController+MobileChat.swift (1)

365-369: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Validate refreshed bindings with the refreshed workspace ID.

After refreshSessionBindings, this still resolves with the stale workspaceID captured from the old record. If the refresh repaired the workspace binding, the retry is discarded and mobile send/interrupt/answer still fail.

Proposed fix

         if let refreshed = await service.refreshSessionBindings(sessionID: sessionID),
-           let surfaceID = refreshed.surfaceID,
-           mobileChatBindingResolves(workspaceID: workspaceID, surfaceID: surfaceID),
-           mobileChatBindingIsCurrentAgent(refreshed) {
-            return ["workspace_id": workspaceID, "surface_id": surfaceID]
+           let surfaceID = refreshed.surfaceID {
+            let refreshedWorkspaceID = refreshed.workspaceID ?? workspaceID
+            if mobileChatBindingResolves(workspaceID: refreshedWorkspaceID, surfaceID: surfaceID),
+               mobileChatBindingIsCurrentAgent(refreshed) {
+                return ["workspace_id": refreshedWorkspaceID, "surface_id": surfaceID]
+            }
         }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Sources/TerminalController`+MobileChat.swift around lines 365 - 369, The
refreshed binding check in the mobile chat flow is still using the old workspace
value, so the retry can be rejected even after a successful repair. Update the
logic in TerminalController+MobileChat’s session refresh path to use the
workspace ID from the refreshed binding returned by refreshSessionBindings, and
pass that refreshed workspace value into mobileChatBindingResolves alongside
refreshed.surfaceID and mobileChatBindingIsCurrentAgent.

Sources/SessionIndexModels.swift (1)

340-364: 🔒 Security & Privacy | 🟠 Major | ⚡ Quick win

Quote sessionId before passing the command to /bin/sh -c.

Line 351 embeds sessionId directly into the inner shell command. If the Codex session id ever contains shell metacharacters, the wrapped /bin/sh -c will interpret them; quote it like the other dynamic Codex arguments.

Proposed fix

-            var parts = ["\(AgentResumeArgv.codexWrapperShellExecutableToken) resume \(sessionId)"]
+            var parts = ["\(AgentResumeArgv.codexWrapperShellExecutableToken) resume \(Self.shellQuote(sessionId))"]

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Sources/SessionIndexModels.swift` around lines 340 - 364, The Codex resume
command built in SessionIndexModels.swift embeds sessionId directly into the
inner /bin/sh -c command, which can be interpreted as shell syntax if it
contains metacharacters. Update the resume command assembly in the codex wrapper
path so sessionId is quoted the same way as the other dynamic Codex arguments
(for example alongside model and effort) before it is concatenated into parts
and passed through AgentResumeArgv.portableCodexResumeShellCommand.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@CLI/CMUXCLI`+CodexFireAndForgetHooks.swift:
- Around line 10-17: The wrapper currently routes all entries in
codexWrapperInjectionEvents through codexFireAndForgetAgentHookShellCommand,
which breaks the blocking behavior for PreToolUse and PermissionRequest. Update
CMUXCLI+CodexFireAndForgetHooks so those two events do not use the
fire-and-forget wrapper, either by removing them from
codexWrapperInjectionEvents or by adding a blocking path in the wrapper that
matches codexHookCanRunFireAndForget’s behavior. Keep the existing native
injection behavior aligned with the persistent path for those event types.

In
`@Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift`:
- Around line 228-231: The chat refresh flow still restarts when the app
backgrounds because `chatRefreshKey` changes and `refreshChatSessions()` has no
early exit. Add a guard at the start of `refreshChatSessions()` in
`WorkspaceDetailView` to return immediately when `scenePhase` is `.background`,
before calling `store.makeChatEventSource()` or starting the stream. Keep the
existing fallback logic for the non-background path so the task only runs while
the app is active.

In
`@Packages/Shared/CmuxAgentChat/Sources/CmuxAgentChat/Model/ChatSessionDescriptor.swift`:
- Around line 41-46: The synthesized Equatable for ChatSessionDescriptor now
includes version, so descriptor diffs can fire on activity-only registry writes.
Update AgentChatTranscriptService.descriptorChangedMeaningfully to normalize or
ignore version alongside lastActivityAt, and compare a sanitized descriptor so
pure hook/pre/post-tool activity bumps do not trigger descriptorChanged; use
ChatSessionDescriptor and descriptorChangedMeaningfully as the main touchpoints.

In
`@Packages/Shared/CmuxAgentChat/Sources/CmuxAgentChat/Parsing/CodexTranscriptParser.swift`:
- Around line 215-232: All questions in a multi-question card are not being
registered under the same pending key, so only the first one gets resolved.
Update CodexTranscriptParser’s question appending loop so every ChatMessage
created in the enumerated questions block uses pendingKey: callID, while keeping
the existing id scheme and seq/timestamp/role/kind behavior. This ensures
TranscriptBatchAssembler.resolve can match the single function_call_output
against all related questions and clear them together.

In `@Resources/bin/cmux-codex-wrapper`:
- Line 307: The resume payload built in cmux_codex_resume_payload is
hand-assembling JSON and interpolates $PWD raw, so quotes or backslashes can
corrupt the payload. Update the session resume path in cmux-codex-wrapper to
serialize the payload through a JSON-safe encoder or the cmux CLI instead of
string concatenation, keeping cmux_codex_resume_sid and cwd as properly escaped
fields. Ensure the generated payload remains valid JSON for paths containing "
or \ so the session-start hook can parse it reliably.

In `@Sources/Mobile/AgentChat/AgentChatSessionRegistry.swift`:
- Around line 276-277: The rollout path detection in AgentChatSessionRegistry
should not hard-code the “.codex” root, because it misses sessions created under
a custom CODEX_HOME. Update openCodexRolloutPath to पहचान the rollout file by
its sessions-directory shape and .jsonl suffix, or resolve the root from process
environment/config instead of checking for "/.codex/sessions/". Keep the
existing path matching logic in that method but make it root-agnostic so custom
session roots are included.
- Around line 85-92: The early return in syncProcessExitWatch(for:) is leaving
an existing exit watcher active when the same session transitions to .ended, so
adjust the watcher synchronization logic to always cancel and clear any existing
watcher before checking the new record state. Keep the existing pid/sessionID
reuse guard for still-active sessions, but ensure that when record.state becomes
.ended the exitWatchers entry is removed and its source is cancelled rather than
returning early.
- Around line 197-203: The observed-session path in
AgentChatSessionRegistry.update(sessionID:session:) is still calling update and
bumping the version even when rec.surfaceID, rec.workspaceID,
rec.transcriptPath, and rec.pid are already set. Add a guard in that branch so
the closure only runs when the incoming session can fill at least one missing
binding, and skip the update entirely otherwise to avoid unnecessary descriptor
pushes.

In `@Sources/Mobile/AgentChat/AgentChatTranscriptResolver.swift`:
- Around line 30-42: The config-root fallback in AgentChatTranscriptResolver
currently uses the app process environment, so it misses
agent-launcher/subrouter overrides for CLAUDE_CONFIG_DIR and CODEX_HOME. Update
the transcript resolution flow to use the per-session agent environment or
observed config root when resolving missing transcriptPath values, by threading
that information through AgentChatTranscriptResolver’s init/configRoot path or
by recording the resolved transcript path during hook/process detection.

In `@Sources/Mobile/AgentChat/AgentChatTranscriptService.swift`:
- Around line 272-276: The DEBUG trace in AgentChatTranscriptService’s
transcript resolution path is logging the raw working directory via the
cmuxDebugLog call, which can expose sensitive path details. Update the log in
this unresolved-session branch to avoid printing record.workingDirectory
directly and instead emit only a presence indicator or a redacted placeholder.
Keep the existing sessionID and agentKind context, but ensure the cwd portion is
sanitized while preserving usefulness for debugging.

In `@Sources/TerminalController`+MobileChat.swift:
- Around line 153-160: The error response in TerminalController+MobileChat’s
session-not-found path is leaking the session identifier by including it in the
err payload’s data. Remove the session_id entry from the .err construction in
this MobileChat terminal handler and keep only the user-facing message/code so
the caller does not receive session IDs in API error bodies.
- Around line 219-224: Resolve the terminal binding a single time in the send
flow instead of calling both mobileChatTerminalParams(sessionID:) and
mobileChatTerminalPanel(sessionID:) separately. In
TerminalController+MobileChat.swift, update the send path to reuse one resolved
terminal target (or a shared lookup helper) so attachments, separators, and text
all go to the same panel/params instance even if the registry changes between
awaits. Keep the existing not_found error handling, but base it on that one
resolution.

---

Outside diff comments:
In `@Sources/SessionIndexModels.swift`:
- Around line 340-364: The Codex resume command built in
SessionIndexModels.swift embeds sessionId directly into the inner /bin/sh -c
command, which can be interpreted as shell syntax if it contains metacharacters.
Update the resume command assembly in the codex wrapper path so sessionId is
quoted the same way as the other dynamic Codex arguments (for example alongside
model and effort) before it is concatenated into parts and passed through
AgentResumeArgv.portableCodexResumeShellCommand.

In `@Sources/TerminalController`+MobileChat.swift:
- Around line 365-369: The refreshed binding check in the mobile chat flow is
still using the old workspace value, so the retry can be rejected even after a
successful repair. Update the logic in TerminalController+MobileChat’s session
refresh path to use the workspace ID from the refreshed binding returned by
refreshSessionBindings, and pass that refreshed workspace value into
mobileChatBindingResolves alongside refreshed.surfaceID and
mobileChatBindingIsCurrentAgent.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 826345d6-03e3-4ff9-8c75-d7e6f76c0c9a

📥 Commits

Reviewing files that changed from the base of the PR and between 261b950 and 8889850.

📒 Files selected for processing (49)

• CLI/CMUXCLI+AgentHookDefinitions.swift
• CLI/CMUXCLI+CodexFireAndForgetHooks.swift
• CLI/cmux.swift
• Packages/Shared/CmuxAgentChat/Sources/CmuxAgentChat/Model/ChatQuestion.swift
• Packages/Shared/CmuxAgentChat/Sources/CmuxAgentChat/Model/ChatSessionDescriptor.swift
• Packages/Shared/CmuxAgentChat/Sources/CmuxAgentChat/Parsing/CodexTranscriptParser.swift
• Packages/Shared/CmuxAgentChat/Sources/CmuxAgentChat/Parsing/TranscriptToolCompletion.swift
• Packages/Shared/CmuxAgentChat/Sources/CmuxAgentChat/Store/ChatSessionListReducer.swift
• Packages/Shared/CmuxAgentChat/Tests/CmuxAgentChatTests/ChatSessionListReducerTests.swift
• Packages/iOS/CmuxMobileShell/Sources/CmuxMobileShell/MobileChatEventSource.swift
• Packages/iOS/CmuxMobileShell/Sources/CmuxMobileShell/MobileChatSessionResponse.swift
• Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift
• Packages/macOS/CMUXAgentLaunch/Sources/CMUXAgentLaunch/AgentResumeArgv.swift
• Packages/macOS/CMUXAgentLaunch/Sources/CMUXAgentLaunch/Workstream/WorkstreamEvent.swift
• Packages/macOS/CMUXAgentLaunch/Tests/CMUXAgentLaunchTests/AgentResumeArgvTests.swift
• Packages/macOS/CmuxControlSocket/Sources/CmuxControlSocket/Coordinator/MobileHost/ControlCommandCoordinator+MobileHost.swift
• Packages/macOS/CmuxControlSocket/Sources/CmuxControlSocket/Coordinator/MobileHost/ControlMobileHostContext.swift
• Packages/macOS/CmuxSettings/Sources/CmuxSettings/Keys/IntegrationsCatalogSection.swift
• Packages/macOS/CmuxSettings/Sources/CmuxSettings/Stores/AgentIntegrationSettingsReading.swift
• Packages/macOS/CmuxSettings/Sources/CmuxSettings/Stores/AgentIntegrationSettingsStore.swift
• Packages/macOS/CmuxSettingsUI/Sources/CmuxSettingsUI/Sections/AutomationSection.swift
• Packages/macOS/CmuxTerminal/Sources/CmuxTerminal/Spawn/TerminalSurface+StartupEnvironment.swift
• Packages/macOS/CmuxTerminal/Sources/CmuxTerminal/Spawn/TerminalSurfaceSpawnPolicy.swift
• Packages/macOS/CmuxTerminal/Sources/CmuxTerminal/Surface/TerminalSurface+RuntimeSurfaceCreation.swift
• Packages/macOS/CmuxTerminal/Sources/CmuxTerminal/Surface/TerminalSurface.swift
• Packages/macOS/CmuxTerminalCore/Sources/CmuxTerminalCore/SurfaceValues/TerminalSurfaceClaudeCommandShim.swift
• Packages/macOS/CmuxTerminalCore/Sources/CmuxTerminalCore/SurfaceValues/TerminalSurfaceCodexCommandShim.swift
• Resources/Localizable.xcstrings
• Resources/bin/cmux-codex-wrapper
• Sources/AppDelegate.swift
• Sources/Mobile/AgentChat/AgentChatSessionRecord.swift
• Sources/Mobile/AgentChat/AgentChatSessionRegistry.swift
• Sources/Mobile/AgentChat/AgentChatTranscriptResolver.swift
• Sources/Mobile/AgentChat/AgentChatTranscriptService+TitleDetection.swift
• Sources/Mobile/AgentChat/AgentChatTranscriptService.swift
• Sources/RestorableAgentSession.swift
• Sources/SessionIndexModels.swift
• Sources/SurfaceResumeCommandCanonicalizer+PortableAgentExecutable.swift
• Sources/TerminalController+ControlMobileHostContext.swift
• Sources/TerminalController+MobileChat.swift
• Sources/TerminalController+MobileWorkspaceList.swift
• Sources/TerminalSurfaceRuntimeWiring.swift
• Sources/Workspace.swift
• cmux.xcodeproj/project.pbxproj
• cmuxTests/AgentChatTranscriptResolverTests.swift
• cmuxTests/RestorableAgentSessionIndexTests.swift
• cmuxTests/SessionIndexViewTests.swift
• docs/agent-session-tracking-spec.md
• docs/codex-agent-detection-plan.md

💤 Files with no reviewable changes (3)

• Sources/Mobile/AgentChat/AgentChatTranscriptService+TitleDetection.swift
• Sources/TerminalController+MobileWorkspaceList.swift
• cmuxTests/AgentChatTranscriptResolverTests.swift

[coderabbitai]

🚀 Performance & Scalability | 🟡 Minor

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
# Verify the task keyed by chatRefreshKey calls refreshChatSessions without a scenePhase guard.
rg -n -C4 'task\s*\(id:\s*chatRefreshKey|refreshChatSessions\(\)|chatRefreshKey|scenePhase' Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift

Repository: manaflow-ai/cmux

Length of output: 3048

---

🏁 Script executed:

sed -n '244,265p' Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift

Repository: manaflow-ai/cmux

Length of output: 1207

---

Do not restart the chat stream while backgrounded

The chatRefreshKey sets foreground = 0 when scenePhase == .background (lines 228‑231). Because .task(id:) cancels the old task and immediately spawns a new one when the key changes, the transition to background currently launches refreshChatSessions() again. The function lacks a guard, so it will attempt to establish the event source and stream while suspended, contradicting the comment that the stream should "tear down" in the background.

Add an explicit guard at the entry of refreshChatSessions() to exit early if the app is in the background:

private func refreshChatSessions() async {
    guard scenePhase != .background else { return }
    guard let source = store.makeChatEventSource() else {
        chatSessions = []
        applyChatModeFallback()
        return
    }
    // ... rest of the function

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In
`@Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift`
around lines 228 - 231, The chat refresh flow still restarts when the app
backgrounds because `chatRefreshKey` changes and `refreshChatSessions()` has no
early exit. Add a guard at the start of `refreshChatSessions()` in
`WorkspaceDetailView` to return immediately when `scenePhase` is `.background`,
before calling `store.makeChatEventSource()` or starting the stream. Keep the
existing fallback logic for the non-background path so the task only runs while
the app is active.

[coderabbitai]

🚀 Performance & Scalability | 🟠 Major | ⚡ Quick win

Don't let version force activity-only descriptor pushes.

version now participates in synthesized Equatable, but AgentChatTranscriptService.descriptorChangedMeaningfully only normalizes lastActivityAt. Since registry writes stamp new versions for hook activity, pre/post-tool activity can emit descriptorChanged despite the “pure activity bumps” guard.

Proposed fix

 private static func descriptorChangedMeaningfully(
     previous: AgentChatSessionRecord?,
     current: AgentChatSessionRecord
 ) -> Bool {
     guard var normalizedPrevious = previous else { return true }
     normalizedPrevious.lastActivityAt = current.lastActivityAt
+    normalizedPrevious.version = current.version
     return normalizedPrevious.descriptor != current.descriptor
 }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	/// Monotonic per-session revision, bumped by the host on every change to
	/// this session. The client reconciles best-effort pushes against
	/// authoritative pulls by this number: apply a push only when its version
	/// is strictly greater than the last applied, and replace wholesale from a
	/// snapshot pull. A missed or duplicated push self-heals on the next pull.
	public var version: Int = 0
	private static func descriptorChangedMeaningfully(
	previous: AgentChatSessionRecord?,
	current: AgentChatSessionRecord
	) -> Bool {
	guard var normalizedPrevious = previous else { return true }
	normalizedPrevious.lastActivityAt = current.lastActivityAt
	normalizedPrevious.version = current.version
	return normalizedPrevious.descriptor != current.descriptor
	}

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In
`@Packages/Shared/CmuxAgentChat/Sources/CmuxAgentChat/Model/ChatSessionDescriptor.swift`
around lines 41 - 46, The synthesized Equatable for ChatSessionDescriptor now
includes version, so descriptor diffs can fire on activity-only registry writes.
Update AgentChatTranscriptService.descriptorChangedMeaningfully to normalize or
ignore version alongside lastActivityAt, and compare a sanitized descriptor so
pure hook/pre/post-tool activity bumps do not trigger descriptorChanged; use
ChatSessionDescriptor and descriptorChangedMeaningfully as the main touchpoints.

[coderabbitai]

🎯 Functional Correctness | 🟠 Major

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
# Inspect how the assembler registers pendingKey messages and resolves them,
# to confirm a single call_id output resolves at most one message.
fd -t f 'TranscriptBatchAssembler.swift' --exec cat -n {}

Repository: manaflow-ai/cmux

Length of output: 5202

---

All multi-question cards must register the same pendingKey.

The TranscriptBatchAssembler.resolve method correctly iterates over all messages grouped under a pendingKey, but the parser only registers the first question (index == 0) with pendingKey: callID. Subsequent questions (index > 0) are appended with pendingKey: nil, so they are never added to the pending dictionary and remain unresolved even after the user answers.

Register every question under the same pendingKey: callID so the single function_call_output resolves all cards:

for (index, question) in questions.enumerated() {
    let baseID = callID ?? "line-\(seq)"
    assembler.append(
        ChatMessage(
            id: index == 0 ? baseID : "\(baseID)-q\(index)",
            seq: seq,
            role: .agent,
            timestamp: timestamp,
            kind: .question(question)
        ),
        // Pair ALL questions with the same call id so the output resolves every card.
        pendingKey: callID
    )
}

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In
`@Packages/Shared/CmuxAgentChat/Sources/CmuxAgentChat/Parsing/CodexTranscriptParser.swift`
around lines 215 - 232, All questions in a multi-question card are not being
registered under the same pending key, so only the first one gets resolved.
Update CodexTranscriptParser’s question appending loop so every ChatMessage
created in the enumerated questions block uses pendingKey: callID, while keeping
the existing id scheme and seq/timestamp/role/kind behavior. This ensures
TranscriptBatchAssembler.resolve can match the single function_call_output
against all related questions and clear them together.

[coderabbitai]

🗄️ Data Integrity & Integration | 🟡 Minor | ⚡ Quick win

Resume payload JSON is not escaped — a "/\ in $PWD yields malformed JSON.

cmux_codex_resume_sid is UUID-validated, but $PWD is interpolated raw. On macOS, directory names may legally contain " and \, which break the hand-built JSON. The session-start hook then fails to parse and the resumed session never re-binds to its live pid — i.e. it stays .ended/read-only, the exact failure this change targets. Consider serializing via the cmux CLI or a JSON-safe encoder rather than string interpolation.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Resources/bin/cmux-codex-wrapper` at line 307, The resume payload built in
cmux_codex_resume_payload is hand-assembling JSON and interpolates $PWD raw, so
quotes or backslashes can corrupt the payload. Update the session resume path in
cmux-codex-wrapper to serialize the payload through a JSON-safe encoder or the
cmux CLI instead of string concatenation, keeping cmux_codex_resume_sid and cwd
as properly escaped fields. Ensure the generated payload remains valid JSON for
paths containing " or \ so the session-start hook can parse it reliably.

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Don't hard-code .codex for rollout detection.

openCodexRolloutPath misses Codex sessions using CODEX_HOME because custom roots won’t contain /.codex/sessions/. Match the rollout file shape under a sessions directory, or use the process environment/config root.

Proposed fix

-            if path.hasSuffix(".jsonl"), path.contains("/.codex/sessions/") {
+            let fileName = (path as NSString).lastPathComponent
+            if fileName.hasPrefix("rollout-"),
+               fileName.hasSuffix(".jsonl"),
+               path.contains("/sessions/") {
                 return path
             }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	if path.hasSuffix(".jsonl"), path.contains("/.codex/sessions/") {
	return path
	let fileName = (path as NSString).lastPathComponent
	if fileName.hasPrefix("rollout-"),
	fileName.hasSuffix(".jsonl"),
	path.contains("/sessions/") {
	return path
	}

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Sources/Mobile/AgentChat/AgentChatSessionRegistry.swift` around lines 276 -
277, The rollout path detection in AgentChatSessionRegistry should not hard-code
the “.codex” root, because it misses sessions created under a custom CODEX_HOME.
Update openCodexRolloutPath to पहचान the rollout file by its sessions-directory
shape and .jsonl suffix, or resolve the root from process environment/config
instead of checking for "/.codex/sessions/". Keep the existing path matching
logic in that method but make it root-agnostic so custom session roots are
included.

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | 🏗️ Heavy lift

Use the agent environment for config-root fallbacks.

This captures only the cmux app process environment. If CLAUDE_CONFIG_DIR or CODEX_HOME is set by the agent launcher/subrouter, fallback resolution still searches the app’s default roots for observed/resumed sessions without a recorded transcriptPath. Plumb the per-session config root from hook/process observation, or record the resolved transcript path during detection.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Sources/Mobile/AgentChat/AgentChatTranscriptResolver.swift` around lines 30 -
42, The config-root fallback in AgentChatTranscriptResolver currently uses the
app process environment, so it misses agent-launcher/subrouter overrides for
CLAUDE_CONFIG_DIR and CODEX_HOME. Update the transcript resolution flow to use
the per-session agent environment or observed config root when resolving missing
transcriptPath values, by threading that information through
AgentChatTranscriptResolver’s init/configRoot path or by recording the resolved
transcript path during hook/process detection.

[coderabbitai]

🔒 Security & Privacy | 🟠 Major | ⚡ Quick win

Don't log raw working directories in the DEBUG trace.

cwd can include usernames, repo names, or customer/project paths. Log presence or a redacted value instead.

Proposed fix

             cmuxDebugLog(
                 "agentChat.transcript.resolve session=\(record.sessionID.prefix(8)) "
-                + "kind=\(record.agentKind.sourceName) cwd=\(record.workingDirectory ?? "nil") UNRESOLVED"
+                + "kind=\(record.agentKind.sourceName) cwd=\(record.workingDirectory == nil ? "nil" : "set") UNRESOLVED"
             )

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	#if DEBUG
	cmuxDebugLog(
	"agentChat.transcript.resolve session=\(record.sessionID.prefix(8)) "
	+ "kind=\(record.agentKind.sourceName) cwd=\(record.workingDirectory ?? "nil") UNRESOLVED"
	)
	`#if` DEBUG
	cmuxDebugLog(
	"agentChat.transcript.resolve session=\(record.sessionID.prefix(8)) "
	"kind=\(record.agentKind.sourceName) cwd=\(record.workingDirectory == nil ? "nil" : "set") UNRESOLVED"
	)

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Sources/Mobile/AgentChat/AgentChatTranscriptService.swift` around lines 272 -
276, The DEBUG trace in AgentChatTranscriptService’s transcript resolution path
is logging the raw working directory via the cmuxDebugLog call, which can expose
sensitive path details. Update the log in this unresolved-session branch to
avoid printing record.workingDirectory directly and instead emit only a presence
indicator or a redacted placeholder. Keep the existing sessionID and agentKind
context, but ensure the cwd portion is sanitized while preserving usefulness for
debugging.

Source: Coding guidelines

[coderabbitai]

🔒 Security & Privacy | 🟠 Major | ⚡ Quick win

Do not echo session IDs in API error bodies.

The caller already has sessionID; returning it in data exposes a session identifier in an error payload.

As per coding guidelines, “user-facing errors, alerts, command output, API error bodies … must not expose … session IDs.”

Proposed fix

-                data: ["session_id": sessionID]
+                data: nil

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	return .err(
	code: "not_found",
	message: String(
	localized: "mobile.chat.error.sessionNotFound",
	defaultValue: "That agent session is no longer available."
	),
	data: ["session_id": sessionID]
	)
	return .err(
	code: "not_found",
	message: String(
	localized: "mobile.chat.error.sessionNotFound",
	defaultValue: "That agent session is no longer available."
	),
	data: nil
	)

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Sources/TerminalController`+MobileChat.swift around lines 153 - 160, The
error response in TerminalController+MobileChat’s session-not-found path is
leaking the session identifier by including it in the err payload’s data. Remove
the session_id entry from the .err construction in this MobileChat terminal
handler and keep only the user-facing message/code so the caller does not
receive session IDs in API error bodies.

Source: Coding guidelines

[coderabbitai]

🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win

Resolve the target terminal only once per send.

mobileChatTerminalParams(...) and mobileChatTerminalPanel(...) each re-resolve/refresh the binding. If the registry changes between awaits, attachments can paste using one terminalParams while separators/text are sent to another panel.

Proposed fix

         guard let terminalParams = await mobileChatTerminalParams(sessionID: sessionID) else {
             return .err(code: "not_found", message: Self.chatTerminalBindingErrorMessage, data: [
                 "session_id": sessionID
             ])
         }
-        guard let terminalPanel = await mobileChatTerminalPanel(sessionID: sessionID) else {
+        guard let resolved = mobileResolveWorkspaceAndSurface(params: terminalParams, requireTerminal: true),
+              let surfaceId = resolved.surfaceId,
+              let terminalPanel = resolved.workspace.terminalPanel(for: surfaceId) else {
             return .err(code: "not_found", message: Self.chatTerminalBindingErrorMessage, data: [
                 "session_id": sessionID
             ])
         }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	guard let terminalParams = await mobileChatTerminalParams(sessionID: sessionID) else {
	return .err(code: "not_found", message: Self.chatTerminalBindingErrorMessage, data: [
	"session_id": sessionID
	])
	}
	guard let terminalPanel = mobileChatTerminalPanel(sessionID: sessionID) else {
	guard let terminalPanel = await mobileChatTerminalPanel(sessionID: sessionID) else {
	guard let terminalParams = await mobileChatTerminalParams(sessionID: sessionID) else {
	return .err(code: "not_found", message: Self.chatTerminalBindingErrorMessage, data: [
	"session_id": sessionID
	])
	}
	guard let resolved = mobileResolveWorkspaceAndSurface(params: terminalParams, requireTerminal: true),
	let surfaceId = resolved.surfaceId,
	let terminalPanel = resolved.workspace.terminalPanel(for: surfaceId) else {
	return .err(code: "not_found", message: Self.chatTerminalBindingErrorMessage, data: [
	"session_id": sessionID
	])
	}

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Sources/TerminalController`+MobileChat.swift around lines 219 - 224, Resolve
the terminal binding a single time in the send flow instead of calling both
mobileChatTerminalParams(sessionID:) and mobileChatTerminalPanel(sessionID:)
separately. In TerminalController+MobileChat.swift, update the send path to
reuse one resolved terminal target (or a shared lookup helper) so attachments,
separators, and text all go to the same panel/params instance even if the
registry changes between awaits. Keep the existing not_found error handling, but
base it on that one resolution.

[greptile-apps]

The cmuxDebugLog call in noteResumeInitiated is not wrapped in #if DEBUG, but cmuxDebugLog is only defined inside #if DEBUG in Sources/App/DebugLogging.swift. This unconditional call will produce a compile error in release builds. All three other cmuxDebugLog calls added in this file are correctly guarded — this one was missed.

Suggested change

	let sessionID = Self.normalizedSessionID(rawSessionID, source: source)
	let now = Date()
	cmuxDebugLog(
	"agentChat.resumeInitiated session=\(sessionID.prefix(8)) source=\(source) "
	+ "surface=\((surfaceID ?? "nil").prefix(8)) existed=\(records[sessionID] != nil)"
	)
	let sessionID = Self.normalizedSessionID(rawSessionID, source: source)
	let now = Date()
	#if DEBUG
	cmuxDebugLog(
	"agentChat.resumeInitiated session=\(sessionID.prefix(8)) source=\(source) "
	+ "surface=\((surfaceID ?? "nil").prefix(8)) existed=\(records[sessionID] != nil)"
	)
	#endif

[coderabbitai]

Actionable comments posted: 1

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (2)

Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift (1)

260-278: 🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

Fail closed instead of repinning by terminal recency.

repinToReopenedSession() repins an ended pinned chat to the most recently active newer non-ended session for the same terminal. Terminal identity + recency is not an authoritative session lineage signal, so this can show the wrong conversation when a terminal has multiple newer sessions. Only repin when the descriptor carries a structured resume/reopen relationship to the pinned session; otherwise leave chosenChatSession nil and let the existing fallback exit chat mode. As per path instructions, correctness-critical agent/session identity must use one authoritative structured source and fail closed instead of using “better than nothing” fallbacks.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In
`@Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift`
around lines 260 - 278, The repin logic in repinToReopenedSession currently uses
terminalID plus lastActivityAt to choose a newer live session, which can attach
the chat to the wrong conversation. Update WorkspaceDetailView so the repin only
happens when there is an explicit structured resume/reopen relationship from the
descriptor to the pinned session, and do not infer it from recency or terminal
matching alone. If no authoritative link exists, leave chosenChatSession nil and
let the existing fallback exit chat mode.

Source: Path instructions

CLI/CMUXCLI+AgentHookDefinitions.swift (1)

386-406: 🩺 Stability & Availability | 🔵 Trivial

Separate the script-path computation from the file write.

codexPersistentHookScriptCommand invokes writeCodexHookScript whenever a hook command is generated, including during the equality check inside isCmuxOwnedHookCommand. Although writeCodexHookScript is idempotent (it skips writing if the file contents match), invoking a filesystem-write operation within a predicate violates separation of concerns and introduces unnecessary I/O on hot paths.

Refactor into two distinct functions:

1. codexPersistentHookScriptPath(subcommand:_:) — computes the deterministic file path without touching the disk.
2. installCodexPersistentHookScript(...) — performs the atomic write and is called only from the explicit install or refresh flow.

Update isCmuxOwnedHookCommand to compare the command against the computed path (or inline fallback) without triggering any write operations.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@CLI/CMUXCLI`+AgentHookDefinitions.swift around lines 386 - 406,
`codexPersistentHookScriptCommand` is doing filesystem writes while merely
generating/comparing hook commands, which affects hot-path predicates like
`isCmuxOwnedHookCommand`. Split the logic so
`codexPersistentHookScriptPath(subcommand:_:)` only computes the deterministic
script path, and `installCodexPersistentHookScript(...)` performs the atomic
write only in the explicit install/refresh flow. Then update
`isCmuxOwnedHookCommand` to use the computed path or inline fallback without
calling `writeCodexHookScript`, keeping the existing
`codexPersistentHookScriptCommand` behavior as a thin wrapper if needed.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@CLI/cmux.swift`:
- Around line 33353-33371: The PostToolUse scalar conversion in
boundedPostToolUseScalarValue is treating NSNumber values 0/1 as Bool, which
corrupts exitCode metadata. Update the type checks in
boundedPostToolUseScalarValue to distinguish real CFBoolean/Bool values from
numeric NSNumber instances, using the same CFGetTypeID approach already used in
safeV2DetailValue, and keep numeric exitcode values flowing through as numbers.

---

Outside diff comments:
In `@CLI/CMUXCLI`+AgentHookDefinitions.swift:
- Around line 386-406: `codexPersistentHookScriptCommand` is doing filesystem
writes while merely generating/comparing hook commands, which affects hot-path
predicates like `isCmuxOwnedHookCommand`. Split the logic so
`codexPersistentHookScriptPath(subcommand:_:)` only computes the deterministic
script path, and `installCodexPersistentHookScript(...)` performs the atomic
write only in the explicit install/refresh flow. Then update
`isCmuxOwnedHookCommand` to use the computed path or inline fallback without
calling `writeCodexHookScript`, keeping the existing
`codexPersistentHookScriptCommand` behavior as a thin wrapper if needed.

In
`@Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift`:
- Around line 260-278: The repin logic in repinToReopenedSession currently uses
terminalID plus lastActivityAt to choose a newer live session, which can attach
the chat to the wrong conversation. Update WorkspaceDetailView so the repin only
happens when there is an explicit structured resume/reopen relationship from the
descriptor to the pinned session, and do not infer it from recency or terminal
matching alone. If no authoritative link exists, leave chosenChatSession nil and
let the existing fallback exit chat mode.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: ccb79efd-36c5-4be8-b926-7fd39058a4cf

📥 Commits

Reviewing files that changed from the base of the PR and between 8889850 and 0c8cad8.

📒 Files selected for processing (5)

• CLI/CMUXCLI+AgentHookDefinitions.swift
• CLI/cmux.swift
• Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift
• Packages/macOS/CMUXAgentLaunch/Sources/CMUXAgentLaunch/Workstream/WorkstreamEvent.swift
• Resources/Localizable.xcstrings

💤 Files with no reviewable changes (2)

• Resources/Localizable.xcstrings
• Packages/macOS/CMUXAgentLaunch/Sources/CMUXAgentLaunch/Workstream/WorkstreamEvent.swift

[coderabbitai]

Caution

Inline review comments failed to post. This is likely due to GitHub's internal server error or limits when posting large numbers of comments. If you are seeing this consistently it is likely a permissions issue. Please check "Moderation" -> "Code review limits" under your organization settings.

Actionable comments posted: 1

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (2)

Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift (1)

260-278: 🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

Fail closed instead of repinning by terminal recency.

repinToReopenedSession() repins an ended pinned chat to the most recently active newer non-ended session for the same terminal. Terminal identity + recency is not an authoritative session lineage signal, so this can show the wrong conversation when a terminal has multiple newer sessions. Only repin when the descriptor carries a structured resume/reopen relationship to the pinned session; otherwise leave chosenChatSession nil and let the existing fallback exit chat mode. As per path instructions, correctness-critical agent/session identity must use one authoritative structured source and fail closed instead of using “better than nothing” fallbacks.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In
`@Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift`
around lines 260 - 278, The repin logic in repinToReopenedSession currently uses
terminalID plus lastActivityAt to choose a newer live session, which can attach
the chat to the wrong conversation. Update WorkspaceDetailView so the repin only
happens when there is an explicit structured resume/reopen relationship from the
descriptor to the pinned session, and do not infer it from recency or terminal
matching alone. If no authoritative link exists, leave chosenChatSession nil and
let the existing fallback exit chat mode.

Source: Path instructions

CLI/CMUXCLI+AgentHookDefinitions.swift (1)

386-406: 🩺 Stability & Availability | 🔵 Trivial

Separate the script-path computation from the file write.

codexPersistentHookScriptCommand invokes writeCodexHookScript whenever a hook command is generated, including during the equality check inside isCmuxOwnedHookCommand. Although writeCodexHookScript is idempotent (it skips writing if the file contents match), invoking a filesystem-write operation within a predicate violates separation of concerns and introduces unnecessary I/O on hot paths.

Refactor into two distinct functions:

1. codexPersistentHookScriptPath(subcommand:_:) — computes the deterministic file path without touching the disk.
2. installCodexPersistentHookScript(...) — performs the atomic write and is called only from the explicit install or refresh flow.

Update isCmuxOwnedHookCommand to compare the command against the computed path (or inline fallback) without triggering any write operations.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@CLI/CMUXCLI`+AgentHookDefinitions.swift around lines 386 - 406,
`codexPersistentHookScriptCommand` is doing filesystem writes while merely
generating/comparing hook commands, which affects hot-path predicates like
`isCmuxOwnedHookCommand`. Split the logic so
`codexPersistentHookScriptPath(subcommand:_:)` only computes the deterministic
script path, and `installCodexPersistentHookScript(...)` performs the atomic
write only in the explicit install/refresh flow. Then update
`isCmuxOwnedHookCommand` to use the computed path or inline fallback without
calling `writeCodexHookScript`, keeping the existing
`codexPersistentHookScriptCommand` behavior as a thin wrapper if needed.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@CLI/cmux.swift`:
- Around line 33353-33371: The PostToolUse scalar conversion in
boundedPostToolUseScalarValue is treating NSNumber values 0/1 as Bool, which
corrupts exitCode metadata. Update the type checks in
boundedPostToolUseScalarValue to distinguish real CFBoolean/Bool values from
numeric NSNumber instances, using the same CFGetTypeID approach already used in
safeV2DetailValue, and keep numeric exitcode values flowing through as numbers.

---

Outside diff comments:
In `@CLI/CMUXCLI`+AgentHookDefinitions.swift:
- Around line 386-406: `codexPersistentHookScriptCommand` is doing filesystem
writes while merely generating/comparing hook commands, which affects hot-path
predicates like `isCmuxOwnedHookCommand`. Split the logic so
`codexPersistentHookScriptPath(subcommand:_:)` only computes the deterministic
script path, and `installCodexPersistentHookScript(...)` performs the atomic
write only in the explicit install/refresh flow. Then update
`isCmuxOwnedHookCommand` to use the computed path or inline fallback without
calling `writeCodexHookScript`, keeping the existing
`codexPersistentHookScriptCommand` behavior as a thin wrapper if needed.

In
`@Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift`:
- Around line 260-278: The repin logic in repinToReopenedSession currently uses
terminalID plus lastActivityAt to choose a newer live session, which can attach
the chat to the wrong conversation. Update WorkspaceDetailView so the repin only
happens when there is an explicit structured resume/reopen relationship from the
descriptor to the pinned session, and do not infer it from recency or terminal
matching alone. If no authoritative link exists, leave chosenChatSession nil and
let the existing fallback exit chat mode.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: ccb79efd-36c5-4be8-b926-7fd39058a4cf

📥 Commits

Reviewing files that changed from the base of the PR and between 8889850 and 0c8cad8.

📒 Files selected for processing (5)

• CLI/CMUXCLI+AgentHookDefinitions.swift
• CLI/cmux.swift
• Packages/iOS/CmuxMobileShellUI/Sources/CmuxMobileShellUI/WorkspaceDetailView.swift
• Packages/macOS/CMUXAgentLaunch/Sources/CMUXAgentLaunch/Workstream/WorkstreamEvent.swift
• Resources/Localizable.xcstrings

💤 Files with no reviewable changes (2)

• Resources/Localizable.xcstrings
• Packages/macOS/CMUXAgentLaunch/Sources/CMUXAgentLaunch/Workstream/WorkstreamEvent.swift

🛑 Comments failed to post (1)

CLI/cmux.swift (1)

33353-33371: 🎯 Functional Correctness | 🟠 Major

🧩 Analysis chain

🌐 Web query:

In Swift, does (NSNumber(value: 1) as? Bool) and (NSNumber(value: 0) as? Bool) succeed and return true/false?

💡 Result:

In Swift, casting NSNumber(value: 1) as? Bool succeeds and returns true, and casting NSNumber(value: 0) as? Bool succeeds and returns false [1]. This behavior is governed by the conditional bridging mechanism between NSNumber and Bool [1]. The standard library implementation specifically checks for these values: 1. NSNumber(value: 1) or kCFBooleanTrue is bridged to true [1]. 2. NSNumber(value: 0) or kCFBooleanFalse is bridged to false [1]. If an NSNumber contains a value other than 0 or 1 (such as 0.25 or 5), the conditional cast as? Bool will fail and return nil [1][2]. This behavior ensures that the bridge is restrictive and does not interpret arbitrary non-zero numbers as true, which differs from the behavior of the boolValue property (which treats any non-zero value as true) [1][3].

Citations:

• 1: https://github.com/apple/swift/blob/a28c9d62138e5386ba9341332af218cd232c455e/stdlib/public/SDK/Foundation/NSNumber.swift
• 2: swiftlang/swift-corelibs-foundation#3632
• 3: https://developer.apple.com/documentation/foundation/nsnumber/boolvalue

---

exitCode 0/1 rendered incorrectly as false/true in PostToolUse feed summary

The boundedPostToolUseScalarValue function checks rawValue as? Bool before as? NSNumber. In Swift, NSNumber instances wrapping 0 or 1 successfully cast to Bool via conditional bridging, returning false and true respectively. This causes numeric exitcode values in feedPostToolUseMetadataKeys to be coerced into booleans, corrupting telemetry data.

Adopt the CFGetTypeID check used in safeV2DetailValue to distinguish CFBoolean references from numeric NSNumber types:

🐛 Proposed fix

     private static func boundedPostToolUseScalarValue(_ rawValue: Any) -> (value: Any, truncated: Bool)? {
         if rawValue is NSNull {
             return (NSNull(), false)
         }
-        if let bool = rawValue as? Bool {
-            return (bool, false)
-        }
         if let number = rawValue as? NSNumber {
+            if CFGetTypeID(number) == CFBooleanGetTypeID() {
+                return (number.boolValue, false)
+            }
             return (number, false)
         }
+        if let bool = rawValue as? Bool {
+            return (bool, false)
+        }
         if let string = rawValue as? String {

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

    private static func boundedPostToolUseScalarValue(_ rawValue: Any) -> (value: Any, truncated: Bool)? {
        if rawValue is NSNull {
            return (NSNull(), false)
        }
        if let number = rawValue as? NSNumber {
            if CFGetTypeID(number) == CFBooleanGetTypeID() {
                return (number.boolValue, false)
            }
            return (number, false)
        }
        if let bool = rawValue as? Bool {
            return (bool, false)
        }
        if let string = rawValue as? String {
            let preview = feedUTF8Prefix(
                string,
                maxBytes: feedPostToolUseScalarStringLimitBytes
            )
            return (preview.value, preview.truncated)
        }
        return nil
    }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@CLI/cmux.swift` around lines 33353 - 33371, The PostToolUse scalar conversion
in boundedPostToolUseScalarValue is treating NSNumber values 0/1 as Bool, which
corrupts exitCode metadata. Update the type checks in
boundedPostToolUseScalarValue to distinguish real CFBoolean/Bool values from
numeric NSNumber instances, using the same CFGetTypeID approach already used in
safeV2DetailValue, and keep numeric exitcode values flowing through as numbers.

[greptile-apps]

Unconditional cmuxDebugLog call in noteResumeInitiated will fail to compile in release builds

cmuxDebugLog is defined only inside #if DEBUG (in Sources/App/DebugLogging.swift), but the call inside noteResumeInitiated is not wrapped in a matching #if DEBUG guard. Every other cmuxDebugLog call added in this file (in applyObservedSessions, update, and noteHookEvent) is correctly guarded — this one was missed. The release build will produce a "use of unresolved identifier 'cmuxDebugLog'" error.

[coderabbitai]

Actionable comments posted: 2

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (2)

Sources/TerminalController+MobileChat.swift (2)

374-392: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Use refreshed workspace bindings after hook-store refresh.

After refreshSessionBindings, the code still resolves surfaceID against the old workspaceID. If the refresh fixed a stale workspace binding, this still reports “terminal moved” and prevents sending.

Proposed fix

-        guard let record = service.sessionRecord(sessionID: sessionID),
-              let workspaceID = record.workspaceID else {
+        guard let record = service.sessionRecord(sessionID: sessionID) else {
             return nil
         }
-        if let surfaceID = record.surfaceID,
+        if let workspaceID = record.workspaceID,
+           let surfaceID = record.surfaceID,
            mobileChatBindingResolves(workspaceID: workspaceID, surfaceID: surfaceID),
            mobileChatBindingIsCurrentAgent(record) {
             return ["workspace_id": workspaceID, "surface_id": surfaceID]
         }
@@
         if let refreshed = await service.refreshSessionBindings(sessionID: sessionID),
+           let workspaceID = refreshed.workspaceID,
            let surfaceID = refreshed.surfaceID,
            mobileChatBindingResolves(workspaceID: workspaceID, surfaceID: surfaceID),
            mobileChatBindingIsCurrentAgent(refreshed) {
             return ["workspace_id": workspaceID, "surface_id": surfaceID]

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Sources/TerminalController`+MobileChat.swift around lines 374 - 392, The
mobile chat terminal parameter lookup is still validating the refreshed surface
against the stale workspace value, which can leave “terminal moved” handling
incorrect. In mobileChatTerminalParams(sessionID:), after refreshSessionBindings
returns a refreshed record, use the refreshed workspaceID (falling back to the
original only if needed) when calling mobileChatBindingResolves and when
building the returned workspace_id/surface_id pair. Keep the existing checks
around agent ownership and the refresh path in
TerminalController+MobileChat.swift.

---

34-86: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

Await the observe-floor scan before reading sessions.

mobile.chat.sessions returns the current registry while observeAgentProcesses() runs later in an unstructured Task, so the first pull can still return stale/empty results for an unhooked live Codex/Claude session. Since the dispatch path is already async, make v2MobileChatSessions async and await the throttled off-main refresh before enumerating records.

As per coding guidelines and path instructions, avoid fire-and-forget Task work with meaningful lifecycle, and do not introduce visible staleness windows for correctness-critical session/liveness tracking.

Proposed fix

         switch method {
         case "mobile.chat.sessions":
-            return v2MobileChatSessions(params: params)
+            return await v2MobileChatSessions(params: params)
@@
-    func v2MobileChatSessions(params: [String: Any]) -> V2CallResult {
+    func v2MobileChatSessions(params: [String: Any]) async -> V2CallResult {
@@
-        Task { await service.observeAgentProcesses() }
+        await service.observeAgentProcesses()

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@Sources/TerminalController`+MobileChat.swift around lines 34 - 86,
`v2MobileChatSessions` is launching
`agentChatTranscriptService.observeAgentProcesses()` in a fire-and-forget
`Task`, so the session list can be read before the refresh completes. Make
`v2MobileChatSessions(params:)` async, await the observe-floor scan directly
before collecting sessions, and keep the existing `mobile.chat.sessions`
dispatch path aligned with the async call chain. Use
`agentChatTranscriptService` and `observeAgentProcesses()` as the key symbols to
update so the registry is refreshed before returning results.

Sources: Coding guidelines, Path instructions

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@scripts/cmux-chat-debug.py`:
- Around line 32-35: The cli helper currently hides failures by returning only
stdout from subprocess.run, which can turn missing CMUX_TAG or a failing
cmux-debug-cli.sh into an empty result. Update cli in cmux-chat-debug.py to
check the subprocess result and raise or exit with a clear error when the
command fails or required env like CMUX_TAG is missing, so callers don’t render
“0 records” on failure.
- Around line 124-130: The live refresh loop in cmux-chat-debug.py shells out
with os.system("clear"), which should be removed to avoid shell-execution
checks. Update the --live branch inside the while True loop to clear the
terminal using ANSI escape sequences instead, while keeping the existing
render(show_all) and refresh behavior intact.

---

Outside diff comments:
In `@Sources/TerminalController`+MobileChat.swift:
- Around line 374-392: The mobile chat terminal parameter lookup is still
validating the refreshed surface against the stale workspace value, which can
leave “terminal moved” handling incorrect. In
mobileChatTerminalParams(sessionID:), after refreshSessionBindings returns a
refreshed record, use the refreshed workspaceID (falling back to the original
only if needed) when calling mobileChatBindingResolves and when building the
returned workspace_id/surface_id pair. Keep the existing checks around agent
ownership and the refresh path in TerminalController+MobileChat.swift.
- Around line 34-86: `v2MobileChatSessions` is launching
`agentChatTranscriptService.observeAgentProcesses()` in a fire-and-forget
`Task`, so the session list can be read before the refresh completes. Make
`v2MobileChatSessions(params:)` async, await the observe-floor scan directly
before collecting sessions, and keep the existing `mobile.chat.sessions`
dispatch path aligned with the async call chain. Use
`agentChatTranscriptService` and `observeAgentProcesses()` as the key symbols to
update so the registry is refreshed before returning results.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 09fc7097-444e-4a91-9ec2-5edd0c635544

📥 Commits

Reviewing files that changed from the base of the PR and between 29b4502 and 030db3a.

📒 Files selected for processing (2)

• Sources/TerminalController+MobileChat.swift
• scripts/cmux-chat-debug.py

[azooz2003-bit]

Review triage:

• Fixed (P0, Greptile): the cmuxDebugLog in noteResumeInitiated was unconditional; cmuxDebugLog only exists under #if DEBUG (no release stub), so the bare call would have failed the release/beta build. Now gated, and I swept all changed Swift files for other unconditional calls (none).
• Fixed (script nits): cmux-chat-debug.py now fails loudly on a missing CMUX_TAG / nonzero CLI exit, and uses an ANSI clear instead of os.system("clear").
• Not changing — session_id echoed in not_found error data: pre-existing pattern across the mobile.chat error paths; the client already owns that id (it sent it), so this is correlation metadata, not a leak. Out of scope for this PR.
• Not changing — cwd in the transcript.resolve trace: that trace is #if DEBUG only and writes to a local /tmp/cmux-debug-<tag>.log; the working directory is the single most useful field for diagnosing transcript-path resolution (resolution maps cwd to ~/.claude/projects/<encoded-cwd>/). Redacting it would defeat the diagnostic.