Problem

With a custom-answer question open, selecting terminal output and choosing Add to chat left the visible answer unchanged while overwriting the hidden ordinary message draft. Dismissing the question exposed the unexpected terminal reference.

Change

Reject terminal-context insertion while a question is visible and show an informational notice explaining that the question must be answered or dismissed first. The answer and ordinary draft remain intact. Normal terminal-context insertion is unchanged.

Scope and approval

Fixes https://github.com/pingdotgg/t3code/issues/24745

GitCafe: https://git.cafe/pingdotgg/t3code/issues/24745

Triaged bug via the automated issue flow. This focused correction implements the report's stated fallback for unsupported terminal context in question answers. No new state, contracts, or provider behavior.

Verification

• Reproduced in the real web app on current main with Codex request_user_input_async, an empty custom answer, and terminal text hello; confirmed the ordinary draft gained the reference after dismissal. Control without a question inserted visibly.
• Targeted vp lint apps/web/src/components/chat/ChatComposer.tsx, vp exec tsc --noEmit -p apps/web, and git diff --check origin/main passed after merging main.
• Capture-proof fallback: no existing component test mounts this composer handle; draft-store tests do not exercise the wrong caller routing. No implementation-mirroring test was added.
• Two isolated reviewers returned MERGE. Independent proof audit returned SUFFICIENT and rebuilt the affected custom-answer capture: only the blinking terminal cursor differed, with zero differences outside its disclosed mask at zero tolerance.
• Evidence index, replay/fixture bundle, and replay instructions. Scripts and fixtures are uploaded, not committed.

Evidence compares c79ef8c454 before with 4e080eb004 after. Final head merges main without changing this handler. Linux real web app, 1440×900 and 390×844 CSS viewports, DPR 2; desktop saved PNGs are downsampled to 1280×800, responsive PNGs are 780×1688. Responsive web is not native mobile. Electron shares this handler; its shell and the reporter's Windows environment were not separately exercised. No provider/backend lifecycle or relay/tunnel claim.

Interaction sequence: begin with draft-before in the ordinary draft and the specified question input; select terminal hello, right-click Add to chat, inspect the visible answer, dismiss questions, inspect the ordinary draft before retry, then retry once. C/Q/L now show a notice without changing input or draft. Q/L previously returned silently. R shows the corrected hidden-draft behavior. N verifies unchanged normal insertion.

The continuous after recording covers question transitions, Q, L, question closure, inspection before retry, and normal retry. Separate C clips show the rejected custom-answer interaction. Q uses multi-select because single-select auto-submits after 200ms.

Matched capture evidence — corrected raw diffs

Before commit: c79ef8c454. After commit: 4e080eb004. Existing independently audited captures reused; no new runtime verification.

Diff0 and Diff8 are full-frame, unmasked RGB absolute differences. Both use 8× channel amplification, clipped to 255. Diff0 retains every pixel whose maximum channel difference is >0; Diff8 retains pixels whose maximum channel difference is >8/255 (threshold applied before amplification). Black means no retained difference. No mask, crop, registration, or background overlay is applied. SHA-256 values identify the actual PNG files.

Corrected interpretation

The earlier narrow-view zero-change wording must not be read as pixel identity. For C390, outside the analytic toast/shadow and caret/cursor exclusions, 32,950 pixels differ at zero threshold; every such difference is ≤2/255. At >8/255, zero pixels remain outside those exclusions. The raw Diff0 image retains these low-level differences; Diff8 suppresses them by its stated threshold.

For analytic comparison only, the existing exclusions are PNG-coordinate rectangles: desktop C: toast/shadow (900,50)–(1280,220), answer caret (525,417)–(529,437), terminal cursor (528,570)–(538,589); C390: toast/shadow (0,110)–(780,390), answer caret (239,940)–(243,980), terminal cursor (678,1190)–(695,1210); Q: desktop toast/shadow only; L: desktop toast/shadow and terminal cursor. These exclusions are not applied to any image in the table. The prior analysis found zero >8 pixels outside those exclusions for C/C390/Q/L; it does not establish zero differences at threshold 0.

All 12 newly generated diffs were inspected individually with the image tool. C/C390 include answer-caret and terminal-cursor changes alongside toast/shadow changes. Q/L show toast-region changes; L also has the previously identified cursor exclusion. N is unmasked: 675 pixels differ at >0, and 125 at >8 (0.012207% of 1,024,000); visible differences include the age label, composer caret, and thin sidebar raster. R is unmasked: 48,750 pixels differ at >0 and 19,104 at >8, including toast, age-label, and ordinary draft/terminal-context changes. These captures do not support a claim that nothing else changed.

Desktop PNGs are 1280×800 saved images of the stated 1440×900 CSS viewport; C390 is 780×1688 at DPR2. Responsive web is not native mobile verification.

FRICTION

None during evidence correction. No repo, server, browser, or forge actions were taken; no screenshots were retaken and no replay was run. Existing independent-auditor acceptance was relied on as requested. Prior immutable index wording remains superseded by this correction; paste this table and interpretation into the lead PR body. All 12 diffs were uploaded directly using the ghdrop skill.

What else changed: the intended notice appears for custom, choice-only, and literal questions. Recordings include real focus/caret/cursor changes. Normal-control stills differ by 0.012207% of pixels from the age label, caret, and thin rasterized border. Narrow C has 32,950 outside-mask pixels differing by at most 2/255; its threshold comparison is zero, not its zero-tolerance comparison. No claim that nothing else changed.

Models: Claude Opus 5.5 (implementation/review) and GPT-6.1-Sol (lead/review/proof), through T3 Code's Codex and Claude harnesses.