jo-duchan/tapflow

When no decoder can be picked, tapflow does not stream and there is nothing else it can do — so the message explaining that is the entire product experience for whoever lands there. Three things are wrong with it, all within a few lines of each other. The Android message is hardcoded Korean. AndroidViewer.tsx:588 renders "이 환경에서는 스트리밍을 표시할 수 없습니다. / Chrome/Edge 또는 HTTPS 환경에서 다시 시도해 주세요." to everyone. This is the widest of the three: it reaches every person on that path, not a subset, and an English-speaking self-hoster gets two lines they cannot read at the one moment the product needs to explain itself. iOS shows nothing at all. IOSViewer.tsx:125 is onUnsupported: () => { /* iOS has no separate unsupported UI; the hook warns */ }, and the warning is console.warn in useDecoderStream.ts:73. So the same environment gives an Android tester an explanation in the wrong language and an iOS tester a black rectangle with no ex

What happened? On Windows, the full relay suite is not portable: powershell corepack pnpm --filter @tapflowio/relay test Steps to reproduce Use Windows with Node.js and Corepack available. Clone the repository and install dependencies. From the repository root, run: corepack pnpm --filter @tapflowio/relay test Expected behavior The relay test suite should complete successfully on Windows, or explicitly skip only assertions that cannot be meaningful on Windows. ZIP fixtures should not require a system zip executable, and temporary test resources should close cleanly before cleanup. Platform Neither (relay / dashboard / CLI) tapflow version 0.1.0 (source checkout; @tapflowio/relay 0.19.0) Node.js version v24.13.0 macOS version (iOS agent only) No response Mac architecture (the machine running the agent) arm64 (Apple Silicon) Xcode version (iOS agent only) No response Relevant logs Test Files 4 failed | 57 passed | 1 skipped (62) Tests 5 failed | 699 passed | 1 skipped (705

What Two small UI fixes (#586 follow-up branch) put MacResources.tsx and CommentPanel.tsx through the a11y gate, which surfaced a backlog well beyond the diff. The blocking findings were fixed there — the chart's readings are keyboard-reachable through a slider cursor, the composer's controls are named, attachments announce their filename. These were reported alongside them and are not fixed: MacResources has no unconditional h1. The only one is the selected Mac's name, which renders conditionally, so the no-selection, loading and empty states have an h2 (the chart titles, and the sidebar's "Macs") with no h1 above them. The route needs a stable page heading. The Mac list's state lives only in CSS. Selection is a bg-accent class and online/offline is a coloured dot, so a screen-reader user can reach the region but cannot tell which Mac is selected or which are up. Wants aria-current and a visually-hidden status word. The comment file-validation message is not announced. Pick a file

Context #565 added a runtime smoke test to .github/workflows/docker-publish.yml. It boots the built image on both architectures, checks that / serves the dashboard HTML and that /api/v1/auth/status answers, and confirms the JWT secret survives container recreation. That proves the process starts and the HTTP surface works. It does not touch the relay's actual job. RelayServer attaches a WebSocketServer to the same HTTP server (packages/relay/src/RelayServer.ts:249), and every frame between an agent and a browser goes through it. An image whose upgrade path is broken while static files and REST still answer would publish green and be useless to whoever pulls it. Scope Add one step to the smoke, after the existing readiness loop: Open a WebSocket to ws://127.0.0.1:4000/ against the running container and assert the handshake completes (HTTP 101). Assert the relay then closes it with code 1008 (see below). Fail the step, and therefore the merge job, when either does not happen. The ex

changelogEntryOwed in scripts/check-changeset.mjs regexes the raw changeset body with no markdown awareness, so a marker quoted inside a ``` fence silently releases a user-facing change from the CHANGELOG demand. It is reachable rather than theoretical: the natural changeset for a follow-up to the gate itself quotes the marker to document it. Every other near-miss fails in the safe direction — different casing, an indented marker in a list or blockquote, trailing whitespace or CRLF, the marker split across two lines all fail to match and ask for an entry that may already have been written. Only the fence, and a marker inside the --- frontmatter (self-limiting, since changeset version would reject the YAML), go the dangerous way. Strip fenced regions before matching, or require the marker in the last non-empty paragraph. Found by the adversarial review on the backfill-and-gate branch.