feat(cursor): collapse the picker into umbrella rows routed through the catalog resolver

devlog/_plan/260828_cursor_ndjson_backlog_train/050_merge_train.md

@@ -0,0 +1,50 @@
+# 050 — cursor PR merge train (wp map for the merge-round loop)
+
+User instruction (2026-08-28): merge the cursor rounds one at a time; the
+instruction is the maintainer approval for these session-authored PRs.
+
+## Rounds (dependency-first)
+
+| R | PR | head | gate |
+|---|---|---|---|
+| R1 | #2774 backlog coalesce | codex/runturn-backlog-coalesce 286a1e5a5 | checks 25 SUCCESS + 1 SKIPPED — green; sol-medium pre-merge review |
+| R2 | #2795 midstream echo | codex/cursor-midstream-echo | retarget to dev post-R1; CI re-run green |
+| R3 | #2769 failed_precondition | codex/claude-classified-error-status 16cb875b8 | checks green; review |
+| R4 | #2801 umbrella core | codex/cursor-umbrella-core 54965ef03 | CI FAIL: test 1/4 update-stop-first launcher-recovery timeout (46.8s, waitForProxy false) — UNRELATED to catalog diff (no update/launcher files touched); same infra-flaky class dev itself shows (dev run 33134096643 fails a different macos test). Gate: causal fix or evidence-backed unrelated-flake disposition + fresh green run; never rerun-until-green without a cause |
+| R5 | #2802 umbrella wire | codex/cursor-umbrella-wire | retarget to dev post-R4; CI green |
+
+## Per-round procedure
+
+1. Exact head SHA + full check rollup via gh.
+2. sol-medium reviewer: independent diff review, VERDICT line.
+3. Blockers folded or rebutted with rationale; repairs get focused tests.
+4. gh pr merge --squash --delete-branch; record merge SHA.
+5. Child retarget (gh pr edit --base dev) + verify checks restart.
+6. Post-merge: origin/dev log + no new cursor-test failures.
+
+## Round log
+
+- R1 (#2774): reviewer PASS (Tesla, sol-tier; coalescing phase-safe, consumers
+  checked). MERGED squash 5511a424c via --admin (user merge instruction =
+  maintainer approval; branch policy requires review). Head branch deleted.
+  SIDE EFFECT: base deletion auto-closed stacked #2795, which GitHub cannot
+  reopen (base ref gone). Recovery: cherry-picked 58ee805/a652f0d/e167311
+  onto origin/dev (990a83f5e; 17 tests + tsc green on rebased head),
+  force-pushed the branch, opened successor PR #2803 vs dev.
+  LESSON for R4/R5: retarget the child to dev BEFORE merging the parent with
+  --delete-branch, or merge parent without branch deletion.
+- R2 (#2803, successor of #2795): CI 23 ok / 0 fail (CodeRabbit status
+  marker non-required); prior audits stand (cherry-pick clean). MERGED
+  squash via --admin, branch deleted.
+- R3 (#2769): reviewer PASS (Avicenna; precedence + claude derivation +
+  72 focused tests + clean merge simulation). MERGED squash via --admin,
+  branch deleted.
+- R4 (#2801): CI failure root-caused by investigator (Zeno): update-stop-first
+  45s readiness deadline exhausted on loaded runners (46-47s failures on 4+
+  unrelated PRs; catalog diff has no launcher imports, isolated shard).
+  Causal fix 22c073e03 raised the deadline to 90s (derived budget + pinned
+  arithmetic keep it honest). Fresh CI fully green (0F, macos SUCCESS).
+  MERGED squash 7232a60a7. #2802 retargeted to dev BEFORE branch deletion
+  (R1 lesson applied) — but the parent squash still made the old chain
+  CONFLICTING; wire branch cherry-picked onto dev (874f59734, 116 tests +
+  tsc green) and force-pushed; #2802 stayed OPEN base=dev.

devlog/_plan/260828_cursor_umbrella_catalog/030_closure.md

@@ -10,3 +10,40 @@
 3. Back-compat proof: legacy-id wire table test green (every 69 id routes
   to the same wire id as before, or documented intentional change).
 4. Stack finalization: PR A -> dev, PR B stacked; retarget checks.
+
+## Closure results (2026-08-28)
+
+### Cleanliness comparison
+
+| Measure | Before (opencodex) | After | senpi | omo-ai@beta |
+|---|---|---|---|---|
+| Picker rows (cursor, non-router) | 65 | 47 seed rows / 31 umbrella identities | ~raw roster + grouped-with-fast-splits | none (no cursor map at all) |
+| Variant duplicate rows | 22 (13 thinking + 7 fast + 2 x 1m) | 1 (claude-4-sonnet-1m real wire id) + composer-2.5-fast (no effort base) | thinking split retained for Claude; fast groups separate | n/a |
+| Capability truth surfaces | 2 (effort-map tables + discovery seed annotations) | 1 (catalog.ts CURSOR_CAPABILITIES) | 3 (static TS table + 336-row generated alias JSON + display-name regexes) | 0 |
+| Thinking handling | 13 separate picker rows | dimension; merged into base for ALL families | Claude-only thinkingMode split | delegated |
+| Fast handling | 7 separate rows | dimension; aliases only | separate catalog groups; parameter fast always "false" | delegated |
+| 1M/Max-Mode | single synthetic kimi-k3-1m row | window metadata generalized (claude/gemini/kimi/gpt-5.6 1M) + evidence-gated ultra->maxMode (static kimi-k3 + live maxModeModels union) | window/maxWindow fields; maxMode from name regex + family pattern (window-size inference we rejected as unsupported) | delegated |
+| Back-compat | n/a | every legacy slug byte-identical (oracle + pinned-session tests) | variant-id fallback silently degrades to representative id | n/a |
+
+LOC: effort-map.ts (229) still present as the test oracle only — zero src/
+consumers remain (request-builder/discovery now import catalog.ts; discovery
+keeps two legacy helpers for the transition). catalog.ts is 541 lines
+INCLUDING the full capability table that previously lived across two files
+plus prose. Deletion of effort-map.ts is queued for the post-merge cleanup
+once the oracle freezes to literal fixtures.
+
+### Picker proof
+
+cursorUmbrellaRows(): 31 identities. Excerpt: kimi-k3 {efforts:[low,high,max],
+window:1000000, maxModeVerified:true} — the old kimi-k3-1m row is gone and its
+capability rides the base. Seed: 51 rows (4 router + 47).
+
+### Stack
+
+| PR | base | head | state |
+|---|---|---|---|
+| #2801 core | dev | codex/cursor-umbrella-core 54965ef03 | open |
+| #2802 wire | codex/cursor-umbrella-core | codex/cursor-umbrella-wire 075c5705a | open, retarget to dev after #2801 |
+
+Verification totals: 1137 tests / 56 cursor+catalog files pass on the wire
+head; tsc 0; privacy scan pass. CI is the wide gate per user instruction.

src/adapters/cursor/catalog.ts

@@ -275,10 +275,21 @@ export interface ParsedCursorVariantId {
 }
 
 function stripLevelSuffix(id: string): { stem: string; level?: string } {
+  // Prefer the parse whose stem is a KNOWN capability, and among known stems
+  // the most specific (longest) one: "gpt-5.5-extra-high" must parse as
+  // gpt-5.5-extra + high (its real single-rung wire id), not gpt-5.5 +
+  // extra-high (A-gate blocker 2 family).
+  let fallback: { stem: string; level?: string } | undefined;
+  let best: { stem: string; level?: string } | undefined;
   for (const token of LEVEL_TOKENS) {
-    if (id.endsWith(`-${token}`)) return { stem: id.slice(0, -(token.length + 1)), level: token };
+    if (!id.endsWith(`-${token}`)) continue;
+    const candidate = { stem: id.slice(0, -(token.length + 1)), level: token };
+    fallback ??= candidate;
+    if (CURSOR_CAPABILITIES[candidate.stem] && (best === undefined || candidate.stem.length > best.stem.length)) {
+      best = candidate;
+    }
   }
-  return { stem: id };
+  return best ?? fallback ?? { stem: id };
 }
 
 /**
@@ -287,12 +298,22 @@ function stripLevelSuffix(id: string): { stem: string; level?: string } {
  * `gpt-5.1-codex-max` and `gpt-5.5-extra` — whose tails collide with effort
  * tokens — never mis-parse (A-gate round-1 blocker 2).
  */
+/**
+ * Real wire ids that merely END in "-1m" — they are distinct catalog rows the
+ * wire serves verbatim, never the synthetic ultra marker (A-gate blocker 2:
+ * a real legacy wire identity must not parse as `<base>-1m`).
+ */
+const REAL_1M_WIRE_IDS: ReadonlySet<string> = new Set(["claude-4-sonnet-1m"]);
+
 export function parseCursorVariantId(rawId: string): ParsedCursorVariantId {
   const id = rawId.trim();
   // 1. Exact base identity.
   if (CURSOR_CAPABILITIES[id]) {
     return { baseId: id, kind: defaultKindFor(id), ultra: false, known: true };
   }
+  if (REAL_1M_WIRE_IDS.has(id)) {
+    return { baseId: id, kind: "regular", ultra: false, known: false };
+  }
   // 2. cursor- wire prefix (regular grok wire forms).
   if (id.startsWith("cursor-")) {
     const inner = parseCursorVariantId(id.slice("cursor-".length));
@@ -444,10 +465,32 @@ export function resolveCursorSelection(
     ? `${capability.wirePrefix}${canonicalId}`
     : canonicalId;
   const ultraRequested = parsed.ultra || reasoning?.toLowerCase() === "ultra";
-  const maxModeArmed = capability.maxModeVerified === true || liveMaxModeIds?.has(parsed.baseId) === true;
+  const evidence = liveMaxModeIds ?? liveCursorMaxModeBases;
+  const maxModeArmed = capability.maxModeVerified === true || evidence.has(parsed.baseId);
   return { wireId, canonicalId, maxMode: ultraRequested && maxModeArmed, known: true };
 }
 
+/**
+ * Live Max-Mode evidence (GetUsableModels maxModeModels). Provider discovery
+ * records the BASES the live roster flags; the resolver unions this with the
+ * static `maxModeVerified` gate so ultra generalizes automatically as evidence
+ * arrives — never from window size (devlog 260828 blocker-4 fold).
+ */
+let liveCursorMaxModeBases: ReadonlySet<string> = new Set();
+
+export function recordLiveCursorMaxModeModels(liveIds: readonly string[]): void {
+  const bases = new Set<string>();
+  for (const id of liveIds) {
+    const parsed = parseCursorVariantId(id);
+    if (parsed.known) bases.add(parsed.baseId);
+  }
+  liveCursorMaxModeBases = bases;
+}
+
+export function liveCursorMaxModeBasesForTests(): ReadonlySet<string> {
+  return liveCursorMaxModeBases;
+}
+
 export interface CursorUmbrellaRow {
   readonly id: string;
   readonly efforts: readonly string[];
@@ -456,6 +499,27 @@ export interface CursorUmbrellaRow {
   readonly maxModeVerified: boolean;
 }
 
+/**
+ * Grok Fast keeps the parameterized wire shape (base id + effort/fast
+ * parameters) rather than a flattened -fast id — current Cursor clients send
+ * it that way and the flat form is rejected. Returns undefined for every
+ * other id.
+ */
+export function cursorGrokFastSelection(
+  pickedId: string,
+  reasoning: string | undefined,
+): { wireBaseId: string; effort: string } | undefined {
+  const parsed = parseCursorVariantId(pickedId);
+  if (!parsed.known || parsed.kind !== "fast") return undefined;
+  const capability = CURSOR_CAPABILITIES[parsed.baseId];
+  if (capability?.wirePrefix !== "cursor-") return undefined;
+  const spec = capability.variants.fast;
+  if (!spec) return undefined;
+  const effort = cursorVariantEffort(spec, parsed.level ?? reasoning);
+  if (effort === undefined) return undefined;
+  return { wireBaseId: parsed.baseId, effort };
+}
+
 /**
  * The umbrella picker rows: one per base whose default variant is selectable.
  * Thinking merges into the base row; fast/thinking-fast/legacy slugs stay

src/adapters/cursor/discovery.ts

@@ -5,6 +5,7 @@ import {
   cursorWireModelIdWithEffort,
   CURSOR_THINKING_MODEL_IDS,
 } from "./effort-map";
+import { parseCursorVariantId } from "./catalog";
 
 export interface CursorModelInfo {
   id: string;
@@ -76,9 +77,16 @@ function stripCursorWirePrefix(id: string): string {
  * ordinary `{base}-{effort}` form, or Cursor's current `{base-without-fast}-{effort}-fast` form.
  */
 export function isCursorModelAvailableForAccount(modelId: string, liveIds: readonly string[]): boolean {
+  // Umbrella matching (devlog 260828_cursor_umbrella_catalog): a live suffix
+  // id counts toward its BASE — any variant dimension (thinking/fast/effort)
+  // proves the account can reach the umbrella. Unknown ids fall back to the
+  // legacy exact/suffix comparison so non-cataloged rows keep matching.
+  const parsedTarget = parseCursorVariantId(modelId);
   return liveIds.some(raw => {
     const id = stripCursorWirePrefix(raw);
     if (id === modelId) return true;
+    const parsedLive = parseCursorVariantId(id);
+    if (parsedLive.known && parsedTarget.known && parsedLive.baseId === parsedTarget.baseId) return true;
     for (const effort of CANONICAL_EFFORT_SUFFIXES) {
       if (
         id === `${modelId}-${effort}` ||
@@ -244,15 +252,13 @@ export function filterCursorConfiguredModelsByLiveDiscovery<T extends { id: stri
 
 /**
  * Models GetUsableModels advertises but whose every Run returns not_found (catalog honesty,
- * devlog 260826_cursor_responses_gap 060). Live probes 2026-08-26: cursor/claude-opus-5 failed
- * 100% ("Cursor Connect error not_found") while its -fast and -thinking siblings — separate
- * wire families — succeed. Quarantined here, in the shared filter, so live, cached, stale, and
- * static serving paths all agree. Custom user provider overrides are not routed through this
- * canonical seed and stay untouched.
+ * devlog 260826_cursor_responses_gap 060). The claude-opus-5 REGULAR wire family is the known
+ * case (probes 2026-08-26: 100% not_found while -fast/-thinking succeed) — under the umbrella
+ * catalog (devlog 260828) that quarantine moved to the RESOLVER level: the capability marks the
+ * regular VARIANT quarantined, the bare slug routes the healthy thinking variant, and the base
+ * row stays in the seed. This row-level set stays for future whole-base quarantines.
  */
-export const CURSOR_KNOWN_UNCALLABLE_MODEL_IDS: ReadonlySet<string> = new Set([
-  "claude-opus-5",
-]);
+export const CURSOR_KNOWN_UNCALLABLE_MODEL_IDS: ReadonlySet<string> = new Set([]);
 
 export const CURSOR_STATIC_MODELS: readonly CursorModelInfo[] = normalizeCursorModels([
   // Context windows and the model lineup mirror Cursor's public models/pricing docs plus the jawcode
@@ -265,26 +271,26 @@ export const CURSOR_STATIC_MODELS: readonly CursorModelInfo[] = normalizeCursorM
   // gemini/grok/kimi-k2.7/gpt-5-mini are reasoning models in the SOT but are sent bare (no tier picker).
   ...CURSOR_ROUTER_MODEL_IDS.map(id => ({ id, contextWindow: CONTEXT_200K, supportsReasoningEffort: false })),
 
-  { id: "claude-sonnet-5", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
+  // Umbrella seed (devlog 260828_cursor_umbrella_catalog): one row per BASE
+  // model. Thinking merges into the base (the resolver routes the thinking
+  // variant); fast / thinking-fast / -1m stay routable as aliases but add no
+  // rows. Windows follow CURSOR_CAPABILITIES where the base is cataloged.
+  { id: "claude-sonnet-5", contextWindow: CONTEXT_1M, supportsReasoningEffort: true },
   { id: "claude-4-sonnet", contextWindow: CONTEXT_200K },
   { id: "claude-4-sonnet-1m", contextWindow: CONTEXT_1M },
   { id: "claude-4.5-haiku", contextWindow: CONTEXT_200K },
   { id: "claude-4.5-sonnet", contextWindow: CONTEXT_200K },
   { id: "claude-4.5-opus", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
-  { id: "claude-4.6-opus", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
-  { id: "claude-4.6-sonnet", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
-  { id: "claude-opus-4-7", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
-  // Opus Fast families: live GetUsableModels (260822) lists ONLY effort-suffixed wire ids
-  // ({base-without-fast}-{effort}-fast; the bare id returns not_found), so every entry
-  // carries a tier picker. Live-verified: claude-opus-4-8-high-fast completed a turn.
-  // Tiers per the 260822 dump (devlog 260822_senpi_cursor_transfer/300).
-  { id: "claude-opus-4-7-fast", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
-  { id: "claude-opus-4-8-fast", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
-  { id: "claude-opus-4-8", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
-  // claude-opus-5 (bare) removed from the seed: GetUsableModels lists it but every Run returns
-  // not_found (quarantined via CURSOR_KNOWN_UNCALLABLE_MODEL_IDS; -fast/-thinking families stay).
-  { id: "claude-opus-5-fast", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
-  { id: "claude-fable-5", contextWindow: CONTEXT_200K, supportsReasoningEffort: true },
+  { id: "claude-4.6-opus", contextWindow: CONTEXT_1M, supportsReasoningEffort: true },
+  { id: "claude-4.6-sonnet", contextWindow: CONTEXT_1M, supportsReasoningEffort: true },
+  { id: "claude-opus-4-7", contextWindow: CONTEXT_1M, supportsReasoningEffort: true },
+  { id: "claude-opus-4-8", contextWindow: CONTEXT_1M, supportsReasoningEffort: true },
+  // claude-opus-5: regular variant is quarantined (not_found on every Run) but
+  // the umbrella row routes the THINKING variant, which is live — so the base
+  // row returns to the seed under the umbrella (resolver never sends the
+  // quarantined regular wire id for the bare slug).
+  { id: "claude-opus-5", contextWindow: CONTEXT_1M, supportsReasoningEffort: true },
+  { id: "claude-fable-5", contextWindow: CONTEXT_1M, supportsReasoningEffort: true },
 
   { id: "composer-1", contextWindow: CONTEXT_200K },
   { id: "composer-2.5", contextWindow: CONTEXT_200K },
@@ -301,17 +307,6 @@ export const CURSOR_STATIC_MODELS: readonly CursorModelInfo[] = normalizeCursorM
   { id: "gemini-3.6-flash", contextWindow: CONTEXT_GEMINI, supportsReasoningEffort: true },
   { id: "gemini-3.7-flash", contextWindow: CONTEXT_GEMINI, supportsReasoningEffort: true },
 
-  // Explicit-thinking variants (260825 live roster). Exposed as first-class ids the same way the
-  // Opus Fast families were in 831810c13: `isCursorModelAvailableForAccount` matches a base id
-  // against `{base}`, `{base}-{effort}` and the family's wire form, and none of those ever
-  // matched a `-thinking` id, so every one of these was invisible in the routed catalog.
-  // Suffix ORDER differs per family; `cursorWireModelIdWithEffort` owns that mapping.
-  ...CURSOR_THINKING_MODEL_IDS.map(id => ({
-    id,
-    contextWindow: CONTEXT_200K,
-    supportsReasoningEffort: cursorModelHasEffortTiers(id),
-  })),
-
   { id: "gpt-5-codex", contextWindow: CONTEXT_272K },
   { id: "gpt-5-fast", contextWindow: CONTEXT_272K },
   { id: "gpt-5-mini", contextWindow: CONTEXT_272K },
@@ -344,17 +339,15 @@ export const CURSOR_STATIC_MODELS: readonly CursorModelInfo[] = normalizeCursorM
   { id: "kimi-k2.7-code", contextWindow: CONTEXT_262K },
   // kimi-k3: cursor.com/docs/models/kimi-k3; account-verified via GetUsableModels (2026-07-28) —
   // ships only as effort-suffixed kimi-k3-{low,high,max}, so the tier picker is exposed.
-  { id: "kimi-k3", contextWindow: CONTEXT_262K, supportsReasoningEffort: true },
-  // kimi-k3-1m: synthetic ultra/Max-Mode picker variant (CURSOR_ULTRA_1M_MODEL_IDS) — wire sends
-  // kimi-k3-<effort> with maxMode=true; 1M context user-verified live on the Ultra plan
-  // (devlog 260826_cursor_responses_gap/025). inferCursorContextWindow maps "1m" ids to 1M.
-  { id: "kimi-k3-1m", contextWindow: CONTEXT_1M, supportsReasoningEffort: true },
+  // kimi-k3 folds the old synthetic kimi-k3-1m row into the umbrella: the base
+  // is maxModeVerified (user-verified 1M on the Ultra plan, devlog 260826/025),
+  // so the ultra effort rung arms Max Mode on the wire and the separate picker
+  // row is gone. cursor/kimi-k3-1m stays routable as an alias.
+  { id: "kimi-k3", contextWindow: CONTEXT_1M, supportsReasoningEffort: true },
 
   { id: "grok-4.5", contextWindow: 500_000, supportsReasoningEffort: true },
-  { id: "grok-4.5-fast", contextWindow: 500_000, supportsReasoningEffort: true },
   // 260813 preemptive: grok-4.6 seeded ahead of Cursor's lineup update (mirrors grok-4.5).
   { id: "grok-4.6", contextWindow: 500_000, supportsReasoningEffort: true },
-  { id: "grok-4.6-fast", contextWindow: 500_000, supportsReasoningEffort: true },
 ]);
 
 export function cursorModelIds(models: readonly CursorModelInfo[] = CURSOR_STATIC_MODELS): string[] {