Repository navigation
Agent pane: an answer to one ask never revokes another ask's gesture ticket - #32113
Conversation
…re tickets Red on base: the second answer's reserve revokes the first ask's ticket, since _acpmux/permission_respond has one slot per connection. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Agent-Session: https://cmux.com/deeplink/agent?id=aid-786825ad708e439b96f4
AgentPaneGestureIntent.slot keyed _acpmux/permission_respond only by
method, so a second answer reserved before the first answer's frame went
out revoked the first ticket. For a relayed chat the first answer then
went out unproved and acpmux refused it ("allowing a permission is
accepted only from the cmux app"). Key the slot by permissionId, as
#29745 keyed prompts by promptId, and keep at most 8 answer tickets per
connection.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent-Session: https://cmux.com/deeplink/agent?id=aid-786825ad708e439b96f4
GitHub Actions
Inactive
|
cmux status (one comment, edited in place; each line is updated by the job that owns it) Review App of this head (app.zip, kept 1 day)
red: cmux-next checks (god files, concurrency, crash safety, l10n); base red: the same failure on the base a615dd9; does not block Shipped in:
|
![cmux-bot[bot]](https://avatars.githubusercontent.com/u/171392238?s=60&v=4){kind=small}
|
Awareness note from an agent lane that does not own this PR; it is not a review or a request to take the PR over. Base red from this PR: the CmuxNext test compile fails on feat-cmux-next since 82b4e7b: Seen in a fleet compile check of CmuxNext at a feat-cmux-next tip that includes this merge (#32149's head, which changes no test file). The CmuxNext compile at 17e607b, just before this merge, passed. A |
…llowTests (base red, #32113) (#32155) Swift's macro expander refuses a #require inside another #require ("recursive expansion of macro 'require(_:_:sourceLocation:)'"), so the CmuxNext test targets did not compile at feat since 82b4e7b (#32113; tip watcher STATUS 6262, compile e1efffc867f0bc441ff90aa3). The inner intent goes into its own let first; the assertions are unchanged.
Brief 21 follow-up — session-qualified answer tickets (#34090 draft)Independent exact-source review at Deterministic negative case: within one pane connection, reserve an answer for I opened draft #34090 against Proof level: source analysis plus an isolated Swift 6.2.1 model reproducing the old loss and passing the new slot behavior. The real CmuxNext macOS test and cross-machine runtime remain unrun in this session; automated PR checks are pending. Real collision frequency is unknown (the test intentionally reuses an ID). Please verify on the PR head before landing. |
What broke
On 2026-10-21, while a folder trust question was open, Leo answered an agent's question card (AskUserQuestion) and got "Couldn't send your answer. Try again. approve this on the Mac app: allowing a permission is accepted only from the cmux app, never from the CLI, a script or an agent". He asked: "WTH COULD IT POSSIBLY RESUME WTH KIND OF RESTRICTION IS THAT IT F***ING SHULDNT BE WTFFFF".
The main causes are already fixed on base: #25424 and #25440 bind a question's answers into the relayed proof, and #25022 fixed a lost person-key enrollment. One gap is left in the host.
AgentPaneGestureIntent.slotkept one_acpmux/permission_respondticket per connection. When an answer to a second ask was reserved before the first answer's frame went out, it revoked the first ask's ticket. The first answer then went to another machine's acpmux unproved, and acpmux refused it with that same message.The fix
slotkeys_acpmux/permission_respondby itspermissionId(_acpmux/permission_respond#<permissionId>), the way fix(next): preserve gesture proof for rapid composer sends #29745 keys prompts bypromptId. A newer answer to the same ask still replaces the older one.permissionTicketsKept), dropping the oldest, like prompt tickets.Tests
AgentPaneRelayedAllowTests.answersToTwoAsksReservedBackToBackKeepTheirOwnTickets. The red commit 67774b5 adds the two-asks case and the same-ask replacement; the fix adds the cap check. The PR's CI compile and test jobs are the check: this lane's host can't run a compile check.Risk
Low, next to the trust boundary. A ticket still needs a real click and still covers only its own exact answer. The only change is that answers to different asks no longer revoke each other, and their number stays bounded.
Rule this serves (Leo, P0, 2026-10-21): a pending folder trust question must never block steering, sending, or answering a question. Follows #32074.
🤖 Generated with Claude Code
Agent session: agent session
Need help on this PR? Tag
@codesmith-botwith what you need.Summary by cubic
Fixes an Agent pane bug where answering one ask revoked another ask's permission answer ticket, so the first answer went out unproved and acpmux refused it.
_acpmux/permission_respondtickets bypermissionIdso each ask keeps its own ticket; a newer answer to the same ask still replaces the older one.Written for commit fc9cb7f. Summary will update on new commits.