Confirmed: two writable viewers can corrupt Vim input

Two isolated browser contexts viewing one real PTY both answer live terminal queries. In Vim 9.1, a duplicated mode-status reply causes a normal edit to save the wrong text. Independent Claude Opus 5.5 verification confirmed the effect and isolated the harmful response.

Supported contract and owner

The source exposes no elected protocol-reply owner. Each Ghostty handle has a reply callback; each writable web surface forwards it through ordinary terminal.write. The server writes both replies into the same PTY. Multi-view grid ownership is explicit, but reply ownership is absent. The supported application expectation is ordinary Vim editing remaining correct when another writable viewer attaches.

Boundary: renderer callback, per-handle dispatch, PTY write service. This is same-PTY fanout, not cross-terminal callback routing.

Real-client proof

Open a file containing alpha\n using vim -Nu NONE -n -i NONE. Type iHELLO, Escape, then :wq. Both contexts use the actual web app and separate browser storage and sockets, with shared fixture authentication. The second context also executes commands, establishing writable reachability.

StateSaved text
One viewerHELLOalpha\n
Two writable viewers, three trials\nalpha\n
Two viewers, Vim t_u7=HELLOalpha\n
Second viewer closedHELLOalpha\n

A raw receiver emitting one live cursor-position request separately established reply fanout:

One viewerTwo viewersSecond closed1 reply2 replies1 reply

Correction: the harmful duplicate is DECRPM

A transparent PTY tap captured Vim's actual queries and received bytes. Replaying those bytes into bare Vim reproduced the failure with only a duplicated ESC[?12;2$y mode-status response. Duplicating either cursor-position response alone preserved the edit. Disabling t_u7 also removes Vim's DECRQM query, so that control does not isolate CPR. A CPR-only fix or test could miss the failure.

A temporary addition to the existing public Ghostty core test verifies one DECRQM response in a single-core control, then fails with two responses for the same query delivered to two cores sharing one input sink. This models aggregation; the actual two-client proof establishes reachability. Existing single-renderer query controls pass. All test edits were restored.

Late-viewer lifecycle control

After Vim completed, a passive raw PTY receiver emitted no queries. Opening a new isolated viewer produced zero unsolicited bytes. Closing and reopening its page also produced zero bytes. Surrounding live-query controls each received one reply. This establishes no replay-input failure for that flow; it does not establish correctness for every history or recovery path.

Duplicate audit and limits

Both-state eligible issue and PR searches found no owner for duplicate live replies. Trusted merged #24352 owns shared-grid sizing, not reply arbitration. Adjacent fish startup, setup color probes, retained-query stripping and saved-script process ownership do not cover this path. Unvouched PRs were excluded.

Actual-client verifier SHA: 304cf149c8cb5e2c1ab921428773361752d7aff2. Hunter reproduction: 6d3f6423a341. DECRQM owner test and late-viewer control: 449deb635ec0; subsequent relevant core changes concern modified F3 encoding, not generated replies.

Linux web and Vim 9.1 only. Desktop, native mobile, read-only viewers and other applications were not tested. Frequency depends on simultaneous writable views and applications issuing affected queries. Vim's internal key-decoding path was not traced. No performance claim.

Boat VMs bx_ykcqez7c, bx_amu6re98 and follow-up bx_kkyqejr2 were reset before their proofs, used serially per VM, and released. All owned browsers and servers ended. No fixes, filings, commits, pushes or top-level threads.

PostPlan uses fallback colors because the active client theme could not be read.