[PattedyrAI]

Draft review copy in the existing fork: the connected GitHub account could not create the upstream PR (must be a collaborator). The base is pinned to upstream commit 2a738273e74e2bc4637bb0977fa7818970760621, so this PR contains only the 14 files for this request.

Selecting CSV and JSON and then typing a qualification currently clears the selections and sends only the text. This keeps the choices and additional text together in the same question response on web/desktop and mobile. Refs pingdotgg/t3code#26401 (TKT-9903).

Customizable single-choice questions now wait for explicit submission so there is time to add details. Selecting an option preserves typed text, and selecting it again permits a text-only answer. Exact option values, thread-reference projection, editable prefilled answers, attachments, and explicitly choice-only questions retain their handling. Pi's string-only dialog response now serializes a combined answer instead of cancelling it.

Earlier merged work was checked: upstream #12577 parks displaced text as a later draft; upstream #17637 delays question takeover while typing; upstream #22019 preserves native choice-only restrictions. This changes the first behavior deliberately. The reporter requested implementation and this PR; maintainer approval of the product direction is still pending, so this is a draft.

Verification

At a8e4b33a19049fa296e9f4f4f8da876ca85041d0, Windows, Node 26.12.1:

• Six web/mobile regressions failed before the client fix; the Pi regression failed before its adapter fix.
• 66 focused client tests and all 90 Pi adapter tests pass. Browser cases cover persistent checkmarks, both interaction orders, explicit submission, and the combined answer payload.
• Changed-file checks: format, lint, both knip passes, GraphQL generation, and typechecks for nine affected packages pass.
• The full affected test run has 387 passes / 13 failures. The same 13 Claude-adapter failures reproduce with the unchanged upstream test file: 12 attachment/path cases and one tool-search settings case. Their failure names match exactly. These remain an owned verification limitation; Linux CI is needed.
• Windows Git follows dependency junctions during the check wrapper's cache fingerprint scan. Only that cache fingerprint was temporarily disabled to execute all checks; the original script was restored and is absent from this diff.

Reported installation: T3 Code Nightly 0.1.3-nightly.20261027.3754, packaged commit d2c4403, read from installed application metadata.

Remaining review and evidence

R2 independent Opus review was attempted but could not start because the configured Claude provider is unauthenticated. Author source review is not independent approval.

Full-client verification and before/after captures remain missing: automatic policy review rejected launching the isolated Chrome profile. The browser-test evidence above is component/state coverage, not a claim of live provider delivery or native-mobile verification. No user screenshot was uploaded. Production and the live install were not changed.

Implemented with GPT-6.1-Sol through the Codex harness in T3 Code.

[PattedyrAI]

The fork's Agent context budget check failed because the workflow explicitly compares against the fork's stale main, rather than this PR's pinned upstream base branch.

This change does not modify apps/server/src/provider/agentContextBudget.json: the blob is identical at base 2a738273e74e2bc4637bb0977fa7818970760621 and head a8e4b33a19049fa296e9f4f4f8da876ca85041d0. The reported increases (278 system-context bytes and 3533 MCP schema bytes) are already in upstream and are outside this diff.

I am retaining the failure as a fork-baseline limitation rather than changing the budget or applying an increase-approval label. The actual code CI is queued/running. Upstream submission/access and the review/evidence gaps in the PR description remain unresolved.

Follow-up on the failure wake: the budget blob is fdc39bc78bc0adec275edb2fae22739d909a7a5f at both PR head and its pinned base, while the fork main uses a54335979eb4ceb71a3d871d26c9f23aa59db606. This confirms a baseline comparison failure, not an increase introduced by this PR.

The queued Changes job requests blacksmith-2vcpu-ubuntu-2404, has no assigned runner, and the fork runner API currently reports zero registered runners. The remaining CI is unverified while waiting for runner availability; no runner installation, workflow change, or budget exception has been applied.

[PattedyrAI]

Superseded by the actual upstream draft PR: pingdotgg#26505. The implementation head remains a8e4b33. Upstream submission succeeded through the authenticated GitHub CLI; the earlier connector rejection did not establish an upstream repository restriction.