Cygwin resize/input stalls and split-screen flicker: candidate fixes for review

[chris123git]

We encountered intermittent input stalls during resize and full-screen flicker while changing scrollback margins in a custom TinTin++ Shell on Windows/Cygwin. We have candidate native changes and offline regression results, and would appreciate upstream review before treating them as generally applicable fixes.

Diagnosis and candidate changes

• Temporary phase instrumentation isolated the intermittent stall to the fd0 zero-timeout readiness check on Cygwin. The candidate Cygwin input path instead performs a scoped nonblocking read and yields when no byte is available. Original descriptor flags are restored before scripts or child processes run. Other platforms retain their readiness check.
• Input processing reports whether it consumed a byte; both the regular input loop and daemon caller stop the current drain when no byte is available.
• On Cygwin, the screen-size query uses the terminal-size API rather than requesting a terminal response that can enter the command editor.
• Alternate-buffer state is tracked so margin-only updates do not repeatedly re-enter and clear that buffer.

Verification and limits

Three fresh uninstrumented native runs passed 45 checks each, including 48 total resize cycles. Full Shell tests separately passed 16 resize cycles and intermediate-frame scrolling/flicker checks. Editor/history and packaged communications tests passed for both configured ports using synthetic fixtures.

These were offline Windows Sandbox tests, not live MUD acceptance. They do not establish that every upstream environment has the same problem. The full local patch also contains unrelated protocol/input extensions; those should be separated from a focused upstream submission.

Candidate binary SHA-256: 0b37c76a5bd78c5b8b2b4eb0ea48d074010b6abd55a1b5fc96329bc8019deac9.

Would you prefer focused pull requests for these changes, or minimal reproductions in separate issues first? Is there a maintainer-supported trusted Windows distribution/build route? The local candidate is unsigned; we are not seeking a workaround that disables Windows security.

No player files, credentials, private configuration, or raw personal logs are included here.

[chris123git]

Follow-up/correction — 2026-09-14

Longer quiet-resize testing reproduced the stall in the initially reported 0b37c76a... build and in the isolated subset. The scoped nonblocking read was not a complete fix; the earlier short passing runs should not be treated as acceptance.

Baseline: TinTin++ 2.02.61, Cygwin 3.6.10. An exact upstream Git commit has not been established. Later phase instrumentation observed the same input read outstanding for over four seconds after a resize. Inspection suggests a possible race in Cygwin's console resize/read-discard handling, but we did not capture the internal failing stack; that explanation remains a hypothesis.

The updated Cygwin-console candidate:

• Clears ENABLE_WINDOW_INPUT while the client owns console input, and polls visible console geometry through a readable CONOUT$ handle every 50 ms from the main loop.
• Does not flush or consume the keyboard queue for resize diagnostics. Non-console descriptors retain their existing path.
• Tracks process ownership and nested hold/resume around synchronous #SYSTEM / #SCRIPT, with release on reset/normal exit. Forked children cannot restore the parent's ownership.

Current uninstrumented executable SHA-256:
120df966e2ba91e8f631736c8f8c8b17574d9f26b298e7b705a2a96cdd98af51

Offline verification on that build:

• 10 quiet-resize runs: 160 explicit geometry checkpoints passed.
• 10 traffic/stress runs: 800 resize events and 160 geometry checkpoints passed.
• 37 isolated console lifecycle assertions passed.
• Actual-client SYSTEM/SCRIPT, nested commands, empty/failing commands, input after return, and normal exit passed against an unmodified control, with the initial window-input bit tested both off and on.

These are Windows Sandbox/loopback results, not host-visible or live-MUD acceptance, and not proof of a universal fix. The tested executable still includes other local extensions; these results do not certify a new standalone upstream patch.

Would you review this console-ownership approach, and is there a maintainer-supported no-cost signed/trusted Windows build route, or a free open-source signing route the project could use? An earlier custom candidate was blocked by Smart App Control; this current candidate remains unsigned and has not been attempted on the host. We are not asking to disable protection, add a locally trusted root, or bypass policy.

No player data, credentials, private paths, raw logs, or files are attached.

[scandum]

Maybe ask your llm to keep descriptions shorter and use simpler language.

Best to split PRs up as much as possible. Resizing is a known issue, though improving it is on my roadmap, and the general suggestion is to use #log instead of a large scrollback buffer. If this takes setting #config buffer_size 1000000 to reproduce it's not a genuine issue to most users.

I'm hesitant to mess with the scrollback code, but feel free to send the PR, and I'll take a look.