A real ncurses program using established A_INVIS makes hidden text partially visible in T3's integrated terminal. A block cursor exposes one character; combining conceal with underline exposes the hidden word's width. Fresh independent Claude Opus 5.5 verification confirmed both against native Ghostty 1.3.1.
| Producer state | T3 web terminal | Native Ghostty |
|---|---|---|
| Conceal, cursor on first character | Visible S inside block cursor | Empty block cursor |
| Conceal + underline | Visible S and underline beneath hidden word | No text or underline |
Crops from actual captures, without compositing or altered pixels. Same Python curses program and TERM=xterm-256color.
T3 integrated terminal, concealed + underline:
Native Ghostty 1.3.1, block cursor with blinking disabled:
T3 advertises xterm-256color; its terminfo defines invis as SGR 8. The existing public core test preserves invisible=true and underline, and the normal text renderer suppresses invisible glyphs. Inverse, faint and ordinary underline are existing controls, not new features.
The owning renderer has two paths that bypass the flag: apps/web/src/terminal/ghostty/renderer.ts:386 redraws cell.text over a block cursor without checking invisible, and the decoration loop at line 353 draws underline without that check. Core decoding already retains the correct state. These are two manifestations of one conceal-rendering failure.
python3 producer.py from the accompanying hunt artifacts. No model turn is required.For future reruns, hold one VM with fleet-boat-bench hold --label terminal-conceal --ttl 7200 HEAD, reset it before proof, then run:
fleet-boat-bench exec --vm <id> --record --in .local/hunt/in --out .local/hunt/out pingdotgg/main -- 'bash "$FLEET_IN/run.sh"; bash "$FLEET_IN/reference.sh"'
The verifier's reference.sh fixes native window targeting, disables cursor blink and sets the same TERM. Release the VM before reporting. Harness files remain in the local notes.
Both hunter and fresh verifier reproduced both states, once per run. Population frequency is unmeasured and likely low: it requires a program emitting conceal and either a block cursor over that cell or concealed underline. Normal password prompts usually disable echo instead. This does not establish a password prompt leak. Strikethrough and overline share the unchecked decoration path but were not independently exercised and are not additional confirmed symptoms.
Linux web client only. No native mobile, Electron, macOS or Windows claim; no device or mini used. No new font/theme/zoom, border37080 or unproved29050 finding. No performance claim. Twenty-four adjacent public core tests passed; visual failure is proved by actual screens, not a synthetic snapshot test.
Hunter runtime: 50a2f5b9f415665f48b4d71df4ff68c858578199. Independent verifier runtime: 45649628dcb7d07cf6b91d94f94d327037a2967f. Final fetched main: c8a1b5ff0b457860a7b8daad5963026cef96ac63. Web and server terminal directories are identical across the verifier and final main; no runtime claim on unrelated final-main changes.
Both-state issue and eligible PR searches for conceal, invisible, SGR 8, invis, hidden text, cursor and ncurses found no exact covering owner. The historical Grok partial-screen report concerns backgrounded long jobs, not this producer or conceal. Ineligible submissions were excluded from coverage.
Hunter held VM bx_vwz9fxk8 explicitly released before verifier. Verifier's one fresh exec began before the new held-VM instruction and finished unchanged, then auto-deleted its VM. Both report no running servers or devices. No fixes, filings, commits, pushes or new top-level threads. Ledger updated. Hunt by GPT 6.1 Sol through Codex; verification by Claude Opus 5.5 at medium effort through Claude Agent.
PostPlan uses true-black fallback colors because the active host theme could not be read.