M7: Take launch inputs in wrapping multi-line boxes

[Sandstorm831]

Parent

Spec #276; implementation sub-issue of M7 #275.

What to build

Text Launch inputs on Start a Run render as OpenTUI's native multi-line textarea: word wrap, a minimum height of one line, a maximum of max(6, a third of the terminal height), then internal scroll. Enter submits (Continue); Shift+Enter, Ctrl+Enter, Alt+Enter, and Ctrl+J insert a newline, overriding the textarea default. Pasted line endings are normalised, and a paste of three or more lines or over 150 characters shows as a [Pasted ~N lines] placeholder that expands to the full text on submit. File, file-set, choice, and verdict inputs keep their controls. The prompt is rebuilt against OpenCode's, not copied, under ADR 0018, and the vendor provenance record notes what was taken and what differs. Headless multi-line values keep passing through the existing launch input values with no new flag (stories 71–76).

Acceptance criteria

• Renderer tests prove wrapping, the height cap and internal scroll, Enter submit, each newline key, multi-line paste kept intact, and placeholder collapse and expansion on submit.
• A headless test proves a multi-line launch input value is accepted unchanged.
• The vendor provenance record for the rebuilt prompt passes the structural step.
• Prototype Crucible's launch and Run information architecture #23 evidence: keymap and focus, small terminals and resize, large content, interaction tuning, meaning without colour, and renderer and platform evidence, each shown by testRender frames with the fake Renderer Port.
• CI scenario: the Start a Run input renderer tests and the headless launch test are green on Windows x64, macOS arm64, and Linux x64.

Blocked by

• M7: Report structural violations from one rule catalogue in a dedicated gate step #277
• M7: Report guidance-structure violations in the contract and check Module guidance shape #278
• M7: Report vendor-provenance violations in the contract #279
• M7: Report release-workflow violations in the contract #280
• M7: Enforce unused code and clear today's findings #281
• M7: Fail lint on any warning and add two OpenCode style rules #282

Implementation packet

Owner: TUI presentation (Start a Run inputs step), vendor provenance record

Ratchet: renderer tests for textarea submit, newline, paste, and placeholder

Deletes: the single-line launch input that drops pasted newlines

Gate: ADR 0027 three-OS CI and recorded human evidence, with the slice scenario above.

Dependencies: no new runtime dependency

[Sandstorm831]

Starting context

Files are pinned at b976b6c, the main commit that closed #281, the last blocker. OpenCode is cited at 228e9095ba, the snapshot the session-screen research used. UPSTREAM pins 1ead9e3d7f. Between the two commits, component/prompt/index.tsx differs only in three ...→… string edits, so line numbers and behaviour match. config/keybind.ts, keymap.tsx, and prompt/part.ts are identical.

Input	Why it matters	Authority
Spec #276: stories 71–76; Implementation Decisions › Start a Run › Multi-line inputs (Wrapping and height, Keys, Paste, Other input kinds), Build route, Headless; Cross-cutting › OpenCode; Testing Decisions › Seams 4	This is the target. The spec does not settle: Up/Down inside a multi-line box, where the full text lives between paste and Continue, what the box and Review show after back-navigation, whether a summarised paste is trimmed, and how Review renders a long value.	Canonical spec
ADR 0018 › rebuild tiebreak	The prompt is state-coupled, so it is rebuilt. First study the OpenCode prompt, then record on this issue what was taken and what was deliberately changed. No OpenCode source is copied, so no file carries the Vendored from OpenCode marker.	Canonical decision
isTextLike, InputsStep bindings, the single-line <input>, key hint	text, file, and file-set all render the same <input> today. Only text becomes a textarea, so isTextLike splits. The step binds Up/Down to move between inputs and Enter to Continue. The keymap fires these before a focused textarea sees the key. So Enter-to-Continue already works, and Up/Down never move between lines.	Current behavior
Draft values, the per-digest draft reset, InputsStep wiring, and Review's input rows	setValue runs on every edit, and Continue only changes the step. Whatever the draft holds is what Review assesses and what launches. Review prints each value in one <text>, so a multi-line value renders as several lines there.	Current behavior
OpenCode 228e9095ba packages/tui/src/component/prompt/index.tsx: :L1345 (maxHeight = max(6, floor(rows / 3))), :L1369-L1395 (textarea, minHeight={1}, IME double-defer on submit), :L1396-L1420 (onPaste: CRLF/CR→LF, preventDefault), :L1183-L1222 (the summary threshold at :L1206-L1213; the pasted text is trimmed at :L1185), :L1149-L1181 (placeholder as a virtual extmark), :L1026-L1034 with packages/tui/src/prompt/part.ts:L24-L29 (expansion on submit); packages/tui/src/config/keybind.ts:L163-L164; packages/tui/src/keymap.tsx:L136-L173 and :L229-L232 (managed textarea layer)	What to rebuild against. OpenCode keeps the full paste in a side store keyed by extmark and expands it only on submit. It trims the pasted text, so a summarised paste loses leading and trailing whitespace. Both the textarea's extmarks and registerManagedTextareaLayer exist in Secant's pinned OpenTUI 0.4.5 and @opentui/keymap.	Research
Session-screen research › Prompt and its truncation rows	The same facts in prose, including the kitty keyboard note. Its "clear on submit" rule belongs to the Workbench prompt, not this step.	Research
OpenTUI 0.4.5 textarea defaults (defaultTextareaKeyBindings in @opentui/core) and renderer defaults (useKittyKeyboard ?? {})	The defaults are return→newline, meta+return→submit, linefeed→newline. createProductionRenderer leaves useKittyKeyboard unset, and OpenTUI then requests the kitty protocol, as OpenCode does.	Current behavior
Scratch probes at b976b6c under testRender, with the textarea swapped in for text	These are measured, not read from source. The current <input> turns a pasted alpha\nbeta\r\ngamma into alphabetagamma, and Review shows target: alphabetagamma. With a default-bound textarea and kittyKeyboard: true, Shift/Ctrl/Alt+Enter and Ctrl+J insert no newline and do not continue. Ctrl+J arrives as ctrl+j, not linefeed. Without kitty, Shift+Enter and Ctrl+Enter arrive as plain Enter and continue, and only Ctrl+J inserts a newline. Plain Enter continued in both modes. A 300-character value wrapped to 4 rows at width 96. A bracketed paste reached the textarea as three lines.	Current behavior
UPSTREAM header pin, rebuilt-against rules, the #219 prompt-study entry; the release inventory reads the first Commit:	Vendor provenance record touched. The #219 entry already studies component/prompt/index.tsx and is the model for the new Taken/Changed entry. Changing the header commit would change the release inventory's OpenCode version.	Current guidance
checkVendorProvenance's record check and change review › UPSTREAM currency	The structural step only checks that UPSTREAM exists once some file carries the vendor marker. It never reads entries, so the "passes the structural step" criterion holds whatever the entry says. The entry is kept honest only by review.	Current behavior
TUI invariants: keymap versus input, bounded columns, and the Start a Run seam	Any key the keymap binds pre-empts the textarea. The inputs column is overflow="hidden" with flexShrink={0} rows, and a growing box has to live inside it. The step owns its own transient state and bindings.	Current guidance
Start a Run renderer tests: mountFlow, the BETA and GAMMA fixtures, the inputs, Review, and refusal test	Fixture touched. mountFlow passes no kittyKeyboard, so the modified-Enter tests need it. The existing test types per key, asserts target: hi on Review, and asserts a refusal clears the invalidated field. All three must still hold.	Current tests
Headless: --input and parseInputs; the note text-input fixture and its harness	Fixture touched. A scratch probe on this harness launched --input note=line one\nline two\r\nline three with --trust. It exited 0 with the Run succeeded, and runGroup.readRun(id).run.launch held the value byte for byte, CRLF included. The headless criterion needs only a test, no code. The file-set comment saying the CLI "cannot pass newlines yet" is stale in-process.	Current behavior
Slicing rule 7	Lists the #23 items the acceptance criteria copy.	Current guidance

Reading rule

Start with the spec's Multi-line inputs bullets. Then read InputsStep whole (lines 759–960), then OpenCode's onPaste, pasteInputText, and the submit expansion, then the start-run.test.tsx inputs test.

• Only text inputs change. file, file-set, choice, and verdict keep their controls. A file-set value is newline-separated paths, so the single-line field still cannot enter two paths. The spec keeps that gap, so leave it.
• Enter continues. Shift+Enter, Ctrl+Enter, Alt+Enter, and Ctrl+J insert a newline. Bind Ctrl+J as ctrl+j too, because that is how it arrives under kitty. Two places can own submit and newline: the step's keymap return binding plus textarea keyBindings, or the managed textarea layer. That choice is left to the implementer. Either way, no key may both insert a newline and continue.
• Normalise only pasted line endings. Values from headless stay byte-exact. Never send [Pasted ~N lines] in place of the text: Review and the launch draft must receive the full text.
• The UPSTREAM entry goes under "rebuilt against" in the M6: Expose live Interactive interrupt and preserve its Session #219 entry's Taken/Changed form, and the same summary goes on this issue (ADR 0018). The header Commit: stays 1ead9e3d7f.
• M7: Show a working scanner while a Turn works #292, the scanner, also adds an UPSTREAM entry. The two tickets can run in parallel, and the merge conflict in UPSTREAM is trivial. M7: Highlight Start a Run choices and hold Continue while models are checked #286 edits other steps in start-run-views.tsx. No edge links it to this ticket, so expect a textual merge there too.
• Which modified Enter keys reach Secant in each OS's default terminal is real-terminal evidence (ADR 0027). CI proves the keys through testRender with kittyKeyboard: true.

Open for the human

Choices the spec leaves open, each with the drafting recommendation. They are settled with the human when this context is refreshed for ready-for-agent.

1. Up/Down. Recommendation: move between inputs only when the cursor is on the box's first or last line. Reason: today the step's bindings take Up/Down from any text field.
2. Placeholder. Recommendation: display only. The draft always holds the full text, and after back-navigation the box shows it.
3. Paste trimming. Recommendation: do not trim pastes as OpenCode does. Reason: story 73 keeps a paste intact.
4. Review. Recommendation: Review keeps printing the full value, covered by the large-content evidence.
5. UPSTREAM commit. Recommendation: keep 1ead9e3d7f, because the release inventory reads that header. The cited OpenCode files match 228e9095ba.

Still unknown:

• Which terminals deliver modified Enter keys. This matters for the human evidence.
• How the box's height cap behaves inside the step's overflow="hidden" column. This matters, and it was not probed.

Owner: TUI presentation (Start a Run inputs step), vendor provenance record
Ratchet: renderer tests for textarea submit, newline, paste, and placeholder
Deletes: the single-line launch input that drops pasted newlines
Gate: ADR 0027 three-OS CI and recorded human evidence, with the slice scenario above.
Dependencies: no new runtime dependency