fix(redesign): honor title-only requests behind any determiner

.github/workflows/ci.yml

@@ -100,6 +100,8 @@ jobs:
           src/features/dogfood/components/TasteBenchPanel.test.tsx
           src/features/evaluation/data/tasteBenchScenario.test.ts
           scripts/__tests__/releaseWorkflowContracts.test.ts
+          convex/domains/redesign/chatRuns.responseShape.test.ts
+          src/features/redesign/components/UniversalComposer.test.tsx
 
   scratchnode-launch-gates:
     name: ScratchNode launch gates

CHANGELOG/pages/redesign-chat.md

@@ -3,6 +3,49 @@
 Append-only lane for the public redesign chat, reproducible answer receipts, and their
 transition into an authenticated live conversation. Newest entries first.
 
+## 2026-07-17 - Honor title-only requests behind any determiner
+
+`detectRequestedResponseShape` recognized `return only the title` but not `return only
+its title`: the determiner group listed articles (`a`/`the`) and no possessives, so the
+possessive phrasing fell through to the five-section memo and silently overrode an
+explicit user output constraint on a paid run. The determiner set now lives in one
+`TITLE_DETERMINER` constant shared by every title pattern, so the patterns cannot drift
+apart again. User-visible effect: a prompt ending `return only its title` now returns one
+plain title line instead of a Short answer / Why it matters / Evidence / Risks / Next
+action memo.
+
+This was found by re-verifying the 2026-07-16 production audit against `main` rather than
+trusting it. Four of its five P1s were already fixed by #550/#561/#564 within hours of the
+run; only this one survived. The suite stayed green throughout because every title case it
+asserted used the one phrasing that worked - so the regression tests here use the audit's
+**verbatim** production prompts, and both new test files were added to the CI runtime-smoke
+allowlist, which is an explicit file list rather than a full-suite run.
+
+Also corrected the `UniversalComposer` header comment, which still claimed "no provider
+names in UI / provider names appear only in the trace" - the opposite of the disclosure
+contract shipped in #550 - and pinned `DEFAULT_TIERS` to the runtime's `modelForTier` with
+a parity test, since the two were hand-maintained mirrors with no test binding them.
+
+**PR / canonical main commit**: `PENDING #NNN MAIN SHA / FINAL QA`.
+
+**Evidence state**:
+- Source: `pending`
+- Checks: `npx tsc --noEmit --pretty false` -> 0 errors. `npx vitest run
+  convex/domains/redesign/chatRuns.responseShape.test.ts
+  src/features/redesign/components/UniversalComposer.test.tsx` -> 16 passed. Fix reverted
+  in isolation to prove the guard: audit prompt fails `expected { kind: 'memo' } to deeply
+  equal { kind: 'title_only' }` (2 failed / 11 passed), restored -> 16 passed. Redesign
+  sweep `npx vitest run convex/domains/redesign src/features/redesign shared/redesign` ->
+  141 passed, 3 failed in `ScratchnodeEventsSurface.test.tsx`, confirmed pre-existing on
+  clean `main` with all edits stashed and unrelated to this change.
+- Visual proof: `not recorded` - no rendered surface changed; the response body shape is
+  produced by the runtime, and the composer edit is a comment.
+- Preview: `not recorded`
+- Production live: `not recorded`
+
+**Author**: Homen Shum + Claude Opus 4.8.
+**Touches**: this change is chat-runtime only; no other lane.
+
 ## 2026-07-16 - Contract NodeBench to one decision workspace
 
 NodeBench now has one primary product surface: the runtime-backed conversation at

convex/domains/redesign/chatRuns.responseShape.test.ts

@@ -32,6 +32,46 @@ describe("redesign chat runtime response policy", () => {
     expect(detectRequestedResponseShape("Research Acme's market position.")).toEqual({ kind: "memo" });
   });
 
+  // Both prompts below are verbatim from the 2026-07-16 authenticated production
+  // runs. The possessive one silently rendered the five-section memo in prod while
+  // this suite stayed green, because every title case asserted here used "the".
+  it("honors the shape of the verbatim production prompts that regressed", () => {
+    expect(
+      detectRequestedResponseShape(
+        "Production recovery QA run 2026-07-16: From the attached Daily Brief, identify one unresolved claim and return only its title. Do not write, share, approve, or modify any data.",
+      ),
+    ).toEqual({ kind: "title_only" });
+
+    expect(
+      detectRequestedResponseShape(
+        "Production QA run 2026-07-16: Using the attached Daily Brief, return exactly two bullets: (1) the strongest supported claim with its best source, and (2) one concrete review gap. Do not write, share, approve, or modify any data.",
+      ),
+    ).toEqual({ kind: "bullets", count: 2 });
+  });
+
+  it("detects a title request behind any determiner, not just an article", () => {
+    for (const prompt of [
+      "Return only its title.",
+      "Return only their title.",
+      "Just give me his title.",
+      "Return only this title.",
+      "Return only title.",
+      "Provide its title only",
+      "Title-only please.",
+    ]) {
+      expect(detectRequestedResponseShape(prompt)).toEqual({ kind: "title_only" });
+    }
+  });
+
+  it("does not mistake incidental prose about titles for a shape request", () => {
+    expect(
+      detectRequestedResponseShape("Summarize the report and explain why its title is misleading."),
+    ).toEqual({ kind: "memo" });
+    expect(
+      detectRequestedResponseShape("Compare each vendor's job titles across the market."),
+    ).toEqual({ kind: "memo" });
+  });
+
   it("returns one plain title line while omitting memo-only fields", () => {
     const shaped = applyDeterministicResponsePolicy(
       "Provide a title only",

convex/domains/redesign/chatRuns.ts

@@ -842,13 +842,29 @@ function parseRequestedCount(value: string): number | null {
   return Number.isInteger(count) && count >= 1 && count <= 12 ? count : null;
 }
 
+// Determiners that may precede "title". Articles alone drop possessive phrasings
+// such as "return only its title", which then fall back to the five-section memo
+// and silently override the user's explicit shape request. Kept in one place so
+// the patterns below cannot drift apart.
+const TITLE_DETERMINER = String.raw`(?:a |an |the |its |their |his |her |our |your |this |that )?`;
+
+const TITLE_ONLY_PATTERNS: RegExp[] = [
+  /\btitle[- ]only\b/,
+  new RegExp(
+    String.raw`\b(?:only|just) (?:give|return|output|provide|write|respond with)?\s*(?:me )?`
+      + TITLE_DETERMINER
+      + String.raw`title\b`,
+  ),
+  new RegExp(
+    String.raw`\b(?:give|return|output|provide|write|respond with) (?:me )?`
+      + TITLE_DETERMINER
+      + String.raw`title only\b`,
+  ),
+];
+
 export function detectRequestedResponseShape(prompt: string): RequestedResponseShape {
   const normalized = prompt.replace(/\s+/g, " ").trim().toLowerCase();
-  if (
-    /\btitle[- ]only\b/.test(normalized)
-    || /\b(?:only|just) (?:give|return|output|provide|write|respond with)?\s*(?:me )?(?:a |the )?title\b/.test(normalized)
-    || /\b(?:give|return|output|provide|write|respond with) (?:me )?(?:a |the )?title only\b/.test(normalized)
-  ) {
+  if (TITLE_ONLY_PATTERNS.some((pattern) => pattern.test(normalized))) {
     return { kind: "title_only" };
   }
 

src/features/redesign/components/UniversalComposer.test.tsx

@@ -1,6 +1,7 @@
 import { act, fireEvent, render, screen } from "@testing-library/react";
 import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
 
+import { modelForTier } from "../../../../convex/domains/redesign/chatRuns";
 import {
   CANCEL_ARM_DELAY_MS,
   DEFAULT_TIERS,
@@ -28,6 +29,20 @@ describe("UniversalComposer runtime controls", () => {
     vi.clearAllMocks();
   });
 
+  // The preflight names a provider and model before a paid submit, but DEFAULT_TIERS
+  // is a hand-maintained mirror of the runtime's modelForTier. Drift between them
+  // makes the disclosure advertise a model the runtime never runs — and never prices.
+  it("discloses the exact model the chat runtime resolves for every offered tier", () => {
+    expect(DEFAULT_TIERS.length).toBeGreaterThan(0);
+
+    for (const tier of DEFAULT_TIERS) {
+      expect({ id: tier.id, model: tier.model }).toEqual({
+        id: tier.id,
+        model: modelForTier(tier.id),
+      });
+    }
+  });
+
   it("does not let the second click of submit immediately cancel the new run", () => {
     const onSubmit = vi.fn();
     const onStop = vi.fn();

src/features/redesign/components/UniversalComposer.tsx

@@ -7,8 +7,11 @@
  *   - Bottom row: `+` tools menu (left) · mic (left) · keyboard hint · circular send (right).
  *   - Tools / paid-tier hints / estimates live behind the dropdown popover, not as a 4-pill row.
  *
- * Spec: "no provider names in UI" — the dropdown shows tiers (Auto / Quick / Deep / Compare),
- * not model names. Provider names appear only in the trace.
+ * Disclosure contract: a paid run must reveal its resolved provider and model
+ * BEFORE submit, not only afterwards in the trace. `DEFAULT_TIERS` therefore
+ * mirrors `modelForTier` in convex/domains/redesign/chatRuns.ts, and a parity
+ * test pins the two together — if they drift, the preflight advertises a model
+ * the runtime does not charge for.
  *
  * Submit: Enter sends. Shift+Enter inserts newline. ⌘↵ also sends.
  */