[lawrencecchen]

Summary

The feat-cmux-next tip 79b76ff is red in arch-ratchet: [app_lines] CmuxNextApp Swift lines: +27 (CmuxNextApp +27) over the last green arch tip c5d7e8d. The only commit that grows the App is 79b76ff (#32752): it added the pendingOnly session list to Feed/AgentDialogDesk.swift (+25) and the capabilities field to Feed/AgentDialogLine.swift (+6).

This change fixes forward with no change in behavior. It moves the acpmux side of the dialog read out of CmuxNextApp into CmuxNextAgentPane, the package that already owns AcpmuxPersonLine:

• AgentDialogLine (the protocol, its default empty capabilities, the AcpmuxPersonLine conformance) moves to CmuxNextAgentPane/Approvals/AgentDialogLine.swift, now public.
• The pending read moves to AgentDialogLine.pendingPrompts() in the same file: _acpmux/sessions {pendingOnly: true} once when acpmux serves session-list-pending-v1, else every page (limit: 64, cursor loop with the same duplicate and repeated-cursor rules), then _acpmux/info for each session with a pending prompt.
• The prompt data moves to AgentDialogPrompt (it was AgentDialogDesk.Pending, which stays as a typealias). pendingListCapability is now on AgentDialogPrompt. The question read stays in the App, next to the desk, because it uses CmuxAgentQuestion.

The code is a move: the method bodies are the same, with line. removed.

Testing

• Arch ratchet, as tip-watcher.py runs it (arch-ratchet.py --repo . --base c5d7e8d52824 --head <head>):
  • before (79b76ff): [app_lines] CmuxNextApp Swift lines: +27, push refused
  • after (60f3cde): app_lines -52 (CmuxNextApp -52), no architecture count rises. No other row goes up.
• python3 scripts/cmux-next/prepush-gate.py and --only swiftformat: no new violations.
• Compile at 2678fe9: fleet job ad699fe40000ca23c410a738, compile check CmuxNext: passed (204 s), app: passed, cmux-tui: passed. (60f3cde: job 05a2216aa333bc3a13e8b0ae, all ok.)
• Tests, package-test-lane.sh suite Packages/macOS/CmuxNext AgentDialogControlTests:
  • base aa30aed (step d185ed7c816937c1be51c62c, aws-m4pro-1): Test run with 10 tests in 1 suite failed ... with 1 issue: Agent reads answer on a loaded app: dialog list skips the paced session list, catalog reads skip the read barrier #32752's own listAsksAnAcpmuxWithPendingOnlyForTheChatsWithAnAskOnce counts 3 items, not 2. Its fake acpmux ignored pendingOnly and listed s-tool twice. The second commit makes the fake answer pendingOnly as acpmux does (hub/peers.rs pending_session_summaries: each session with an ask once). Test-only.
  • head 2678fe9, Blacksmith (hq run 41567692653): Test run with 10 tests in 1 suite passed. Then the XCTest phase aborted (signal 6) after CoreData: error: Failed to create NSXPCConnection.
  • head 2678fe9, aws-m4pro-1 (step 7ecb851de11017fd019e334b): Test run with 10 tests in 1 suite passed. Then xctest exited with signal 6 after the CoreData 134020 store error (the known infrastructure abort, STATUS 7919/7933). The step is red only from that abort: no test failed.

AgentDialogControlTests and SubagentLimitTests stay in CmuxNextAppTests (they test the desk and the limit flow, which stay in the App). They now import CmuxNextAgentPane for the protocol; the pendingOnly test reads AgentDialogPrompt.pendingListCapability.

Observed / Not run

• Observed: the arch ratchet before and after (above); the prepush gate; the fleet compile and test results above.
• Not run: no live dogfood of cmux agents dialog list (the behavior is unchanged; this is a move).

Changelog

none

Proof

No capture: no UI change; a code move between two packages.

[github-actions]

cmux status _{(one comment, edited in place; each line is updated by the job that owns it)}

Review 2678fe9: HQ feed card · gallery before/after runs · checks

App of this head (app.zip, kept 1 day)

gh workflow run reload-build.yml -R manaflow-ai/cmux --ref feat-cmux-next -f tag=pr32858 -f ref=base-red-arch-32752 -f platform=macos -f backend_mode=local

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4157787538001 (cc7cdf2066)
• cmux NEXT nightly 1.0.0-nightly.4158055269401 (e5e182cf5b)

[lawrencecchen]

Review (claude-opus-independent): SHIP at 2678fe9

What I checked (read only, no local build):

• Moved logic, compared line by line with the App version at 79b76ff: the pendingOnly path (gated on session-list-pending-v1, one _acpmux/sessions {pendingOnly: true} call, empty when it fails), the paged path (limit: 64, cursor, dedupe on peer/id, a repeated cursor or a failed or malformed page gives nil and then empty), and the prompt read (pendingPermissions > 0, then _acpmux/info, entries with no permissionId or request skipped). They are the same. The only change is that line. became self. inside the protocol extension.
• Access control and isolation: AcpmuxPersonLine already has public isPerson, capabilities, call(_:_:) and close(), so it satisfies the public @MainActor protocol in its own module. The private helpers are in a protocol extension in the same file. Pending is now a typealias of AgentDialogPrompt, and question stays as an App extension. I found no other AgentDialogLine or AgentDialogPrompt in Packages/. The pendingPrompts properties on browser tabs belong to a different type and cannot clash.
• Other conformers: SubagentLimitTests.FakeLine (with the new import) and SubagentLimitHandlers (already imports CmuxNextAgentPane) still use the default capabilities. The test target compiled at this head.
• Test fake: the new pendingOnly reply matches acpmux pending_session_summaries and list_reply.rs::pending, which return only rows with an ask, in one reply with no nextCursor. The paged reply keeps its duplicate row, so the peer/id dedupe test still covers that path. The pendingOnly test still asserts 2 items, exactly one sessions call, pendingOnly == true and no limit. This does not weaken the test.
• Blacksmith run 41567692653 at this sha: all 10 AgentDialogControlTests passed. The step went red only from the CoreData XPC noise in the runner.
• arch-ratchet.py --base c5d7e8d52824 --head HEAD: app_lines -52, no architecture count rises.

[lawrencecchen]

BASE RED XCTEST (STATUS 8187, 8257): the xctest signal 6 after the Swift Testing suites pass is a base red, not a defect of this PR and not a worker fault. Cause: the test-bundle run-loop guard (#25861) holds the main thread of xctest when XCTest start-up turns the main run loop (hung stack from aws-m4pro-6, step 3bd61ca1). The fix is on branch fix/base-red-xctest-contacts (PR follows). For this PR: a Swift step whose only red is xctest ... exited with unexpected signal code 6 after "Test run with N tests ... passed" is this base red. It does not hold the PR.

[lawrencecchen]

BASE RED XCTEST: the fix for the xctest signal 6 base red is PR #33013 (the test-bundle run-loop guard held the main thread of xctest). Red/green on aws-m4pro-6 and -9 run now.