Two confirmed failures. Each has an owning public-contract test that fails on main, actual client input proof, and a fresh independent Claude Opus 5.5 medium worktree verification. No fixes, filings or fix threads.
Final fetched main: 32c069eed1d22ca4b4bbaa52ec8d34bb3f4a9ff3. The terminal files are identical to hunter base f2c944947d. Read-only investigation by GPT 6.1 Sol in Codex; independent verification by Claude Opus 5.5 in T3 Code.
With NumLock off, Numpad1, Numpad8 and NumpadDecimal arrive as trusted browser events with semantic keys End, ArrowUp and Delete. The integrated terminal sends no PTY bytes. Dedicated navigation keys work. Shell and TUI users lose those keypad navigation actions.
| Actual event | T3 bytes | Dedicated-key control |
|---|---|---|
| Numpad1 / End | empty | ESC[F |
| Numpad8 / ArrowUp | empty | ESC[A |
| NumpadDecimal / Delete | empty | ESC[3~ |
Repro: Open the thread terminal in an isolated app. Run the Python ncurses probe below. Press Numpad8, ArrowUp, Numpad1, End, NumpadDecimal, Delete with NumLock off. Only dedicated keys produce KEY_UP, KEY_END and KEY_DC. A closing q proves the program remains responsive.
Cause: keyCodes.ts:195 prioritizes physical Numpad digit codes over the browser navigation key. Ghostty expects separate keypad navigation enums; the unmapped event has no text and encodes to nothing.
Proof: Three added rows in the existing core.test.ts “encodes key” table fail with empty strings; six original rows pass in the verifier. The verifier's separate raw PTY probe confirms zero bytes. An independent xterm.js 6.0.0 terminal driven by the same client keyboard events emits navigation sequences, which the same ncurses producer decodes through a separate kernel PTY.
Frequency and limits: Hunter: two interleaved trials per three keys fail. Fresh verifier: all three keys fail. Hunter used ncurses application cursor mode; verifier recorded normal cursor mode. Linux Chromium trusted CDP events only. Physical OS keypad and NumLock-on behavior untested. Real-user incidence unknown.
The server advertises TERM=xterm-256color. Its terminfo defines Ctrl+F3 as ESC[1;5R, but T3 sends ESC[13;5~. Ncurses reads seven literal characters instead of KEY_F(27), so an advertised function-key binding does not fire.
| Key | T3 sends | Ncurses response |
|---|---|---|
| Ctrl+F1 | ESC[1;5P | KEY_F(25) |
| Ctrl+F2 | ESC[1;5Q | KEY_F(26) |
| Ctrl+F3 | ESC[13;5~ | Seven literal characters |
| Ctrl+F4 | ESC[1;5S | KEY_F(28) |
Repro: In the same ncurses program press Ctrl+F1, Ctrl+F2, Ctrl+F3, Ctrl+F4, Ctrl+F3, Ctrl+F3. All Ctrl+F3 presses misdecode. Check infocmp -1 xterm-256color: kf27 is ESC[1;5R. No Kitty or new protocol is requested.
Cause: core.ts:567 delegates to bundled libghostty while NodePtyAdapter.ts:272 advertises xterm-256color. Pinned Ghostty revision 3ff6d08 deliberately uses this modified-F3 variant at function_keys.zig:95. This is an advertised compatibility mismatch, not an accidental upstream mapping or an unsupported protocol wish.
Proof: One Ctrl+F3 row added to the existing core.test.ts Ctrl-key table fails; two original arrow controls pass. The fresh verifier reproduced actual trusted client events, terminal.write frames and ncurses output on boat.dev at main32c069. An independent xterm.js 6.0.0 terminal emits ESC[1;5R for the same keyboard events; those bytes yield KEY_F(27) through a separate kernel PTY.
Frequency and limits: Three of three presses fail in both hunter and fresh verifier. Chromium web on Linux only. No Electron, macOS, Windows, mobile or physical-keyboard claim; user incidence unknown.
import curses
def run(screen):
screen.keypad(True)
screen.addstr(0, 0, "Keyboard probe: q exits")
screen.refresh()
with open("received.txt", "w", buffering=1) as log:
while True:
key = screen.get_wch()
log.write(repr(key) + "\n")
if key == "q":
break
curses.wrapper(run)Owning public proof command after adding the small table rows: vp test run apps/web/src/terminal/ghostty/core.test.ts -t 'encodes key|sends Ctrl'. Four failures and eight passing controls. Temporary test edits restored.
Both open and closed issues and PRs were searched for terminal keypad, numeric keypad, NumLock, application cursor, modified/function keys and Ctrl+F3. No eligible exact owner was found. Existing issue10238 is application zoom, not terminal navigation. Unvouched PRs were excluded from ownership.
Ordinary/application cursor navigation passed. ESC= alone is intentionally governed by default mode1035, so numeric keypad application encoding was not promoted. Option-arrow rewriting was excluded under the requested local gesture policy. Mode1004, mobile bridge/paste, custom palette, Unicode shaping, provider filters, stopped native-device retries, wheel and the separate newest-main hunt were left out.
Earlier hunter app proof predates the boat-only rule. Every later app run and capture used boat.dev; the verifier VM is released. Focused tests stayed on forge. All owned dev servers, tabs and background commands stopped; tracked trees are clean. No Mac mini device was used.
Full commands, patches, event logs and verdicts are retained in the hunt's local artifacts. Authentication-bearing dev logs are excluded from this published page. This PostPlan copy uses fallback black and system fonts because the active host theme could not be read.